Download the full version and explore a variety of ebooks

or textbooks at https://ebookmass.com

A fiber-segmentation algorithm for composites

imaged using X-ray microtomography_ Development
and validation Peter J. Creveling & William W.
Whitacre & Michael W. Czabaj

_____ Tap the link below to start your download _____

https://ebookmass.com/product/a-fiber-segmentation-
algorithm-for-composites-imaged-using-x-ray-
microtomography_-development-and-validation-peter-j-
creveling-william-w-whitacre-michael-w-czabaj/

Find ebooks or textbooks at ebookmass.com today!

Here are some recommended products for you. Click the link to
download, or explore more at ebookmass.com

Elsevier Weekblad - Week 26 - 2022 Gebruiker

https://ebookmass.com/product/elsevier-weekblad-
week-26-2022-gebruiker/

X-Ray Fluorescence Spectroscopy for Laboratory

Applications Michael Haschke

https://ebookmass.com/product/x-ray-fluorescence-spectroscopy-for-
laboratory-applications-michael-haschke/

Handbook of Mineral Spectroscopy: X-Ray Photoelectron

Spectra: Volume 1: X-ray Photoelectron Spectra 1st Edition
J. Theo Kloprogge
https://ebookmass.com/product/handbook-of-mineral-spectroscopy-x-ray-
photoelectron-spectra-volume-1-x-ray-photoelectron-spectra-1st-
edition-j-theo-kloprogge/

First-Year Teaching For Dummies W. Michael Kelley

https://ebookmass.com/product/first-year-teaching-for-dummies-w-
michael-kelley/
Prohibition: A Concise History W J Rorabaugh

https://ebookmass.com/product/prohibition-a-concise-history-w-j-
rorabaugh/

Jock Seeks Geek: The Holidates Series Book #26 Jill

Brashear

https://ebookmass.com/product/jock-seeks-geek-the-holidates-series-
book-26-jill-brashear/

(eTextbook PDF) for Abnormal Psychology 2nd Edition by

William J. Ray

https://ebookmass.com/product/etextbook-pdf-for-abnormal-
psychology-2nd-edition-by-william-j-ray/

(eTextbook PDF) for Organizational Communication by

Michael W. Kramer

https://ebookmass.com/product/etextbook-pdf-for-organizational-
communication-by-michael-w-kramer/

Wool Fiber Reinforced Polymer Composites Sabu Thomas

https://ebookmass.com/product/wool-fiber-reinforced-polymer-
composites-sabu-thomas/
 Composites Part A 126 (2019) 105606

Contents lists available at ScienceDirect

Composites Part A
journal homepage: www.elsevier.com/locate/compositesa

A fiber-segmentation algorithm for composites imaged using X-ray T

microtomography: Development and validation
⁎
Peter J. Crevelinga, William W. Whitacreb, Michael W. Czabaja,
a
Department of Mechanical Engineering, University of Utah, Salt Lake City, UT 84112, United States
b
Draper Laboratory, Cambridge, MA 02140, United States

A R T I C LE I N FO A B S T R A C T

Keywords: This study proposes a new algorithm for three-dimensional segmentation of fibers in multi-directional fiber-
A. Polymer-matrix composites reinforced composites (FRCs) imaged using X-ray microtomography (µCT). The proposed algorithm uses syn-
B. Microstructures thetic fiber-templates and template matching to detect fibers in 2D. Detections are used to track individual fibers
C. Micro-mechanics in 3D using a Kalman-filter estimator. Validation of the algorithm was performed using synthetic image stacks,
D. CT analysis
which mimicked the quality and resolution of existing state-of-the-art X-ray systems. Segmentation of synthetic
images revealed that the proposed algorithm estimates the total number of fibers, the 3D positions of fiber cross-
sectional centroids, and fiber lengths with an error less than 7% when the image resolution is greater than 9
pixels per average fiber diameter. However, the segmentation algorithm was found to underestimate fiber cross-
sectional areas due to beam-hardening artifacts. This study concludes by demonstrating the versatility of the
algorithm on a multi-directional FRC imaged using X-ray µCT.

1. Introduction manufacturing, resulting microstructure, and performance of FRCs.

Additionally, segmentation of X-ray µCT images is needed to create
In recent years, high-resolution X-ray microtomography (X-ray µCT) high-fidelity 3D finite-element-based micromechanical models of as-
has emerged as a popular method for non-destructive three-dimensional manufactured FRCs. These models are expected to become an integral
(3D) imaging of fiber-reinforced composites (FRCs) at the constituent part of future multiscale “virtual testing” tools, which hold promise for
length scale. Modern X-ray µCT systems with sub-micron resolution reducing our reliance on pure empiricism in design and certification of
capabilities (e.g., Zeiss Versa-520) are providing an unprecedented in- FRC structures, and may enable development of revolutionary struc-
sight into microscale structures of individual fibers, manufacturing-in- tural-life prognosis tools like the Airframe Digital Twin [14].
duced imperfections, and damage [1–9]. Additionally, recent advances For most FRC systems imaged using X-ray µCT, image segmentation
in synchrotron-based high-speed X-ray µCT have enabled 4D (i.e., entails identification of pixels/voxels associated with individual fibers,
spatial and temporal) imaging of deformation and failure for a broad pre-existing flaws (e.g., voids), and damage (e.g., matrix cracks or fiber
range of composite systems, loading conditions, and environments breaks). If the contrast between individual features is adequately high,
[9–13]. Despite the increasing popularity of X-ray µCT, one of the segmentation is often performed using simple gray-scale thresholding
outstanding challenges associated with the effective use of this method algorithms, which are available in several open-source or commercially
is the large amount of image data generated. Depending on the desired available software packages such as ImageJ, Avizo, or Dragonfly.
quality of data (e.g., image-stack spacing or image bit-depth), a single However, if the contrast between features is low, segmentation based
X-ray µCT scan can produce an image stack that is multiple gigabytes in on gray-scale thresholding is inadequate. For FRC systems that exhibit
size, making it difficult to visualize, let alone analyze. low contrast between individual phases (e.g., carbon/epoxy or SiC/SiC
The sheer amount of data derived from X-ray µCT, coupled with composites), segmentation of individual fibers is particularly difficult
inherently complex microstructures of FRCs, has created a pressing due to: (1) small number of pixels representing fiber cross-sections; (2)
need to develop tools for automatic segmentation (i.e., classification) of dense packing of fibers and the resulting blurred boundaries between
physical features from 2D images or 3D image stacks. Accurate ex- adjacent fibers; and (3) varying size, orientation, and curvature of in-
traction of the reduced-order segmentation data from high-resolution X- dividual fibers.
ray µCT images is necessary to investigate the relationships among Given these inherent issues, high-fidelity 3D segmentation of

⁎
Corresponding author.
E-mail address: [email protected] (M.W. Czabaj).

https://doi.org/10.1016/j.compositesa.2019.105606
Received 12 June 2019; Received in revised form 20 August 2019; Accepted 21 August 2019
Available online 24 August 2019
1359-835X/ © 2019 Elsevier Ltd. All rights reserved.
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

individual fibers requires approaches that are more advanced, and ty- which made estimation of fiber diameters or local fiber orientations not
pically comprise of two distinct steps. The first step involves “detection” possible, limiting its application to unidirectional composites with
of individual fiber cross-sections in each X-ray µCT image (i.e., tomo- constant fiber radii. The new algorithm removes this limitation by using
gram) of the entire tomographic image stack. In general, this involves “synthetic” fiber-templates, which enables detection of fiber centroids,
identifying size and location of features that are either circular or el- diameters, and local fiber orientations relative to the scan-axis. The new
liptical, depending on the orientation of fibers relative to the X-ray µCT algorithm also includes procedures for fiber-track smoothing, track
scan axis. A non-comprehensive list of methods used for detection of stitching, and overlap removal, which minimizes errors associated with
fiber cross-sections includes: (1) Bayesian inferences [15], (2) template- the use of low-contrast and low-resolution X-ray µCT image data. The
matching (TM) [16,17], (3) trained fiber centroid characteristic dic- new fiber detection and tracking algorithm is evaluated using synthetic
tionaries [18,19], (4) Hough transforms [19–22], and (5) convolutional X-ray µCT image data that are developed to mimic the resolution and
neural networks [22]. All of the aforementioned methods can be used to quality that can be obtained with existing state-of-the-art high-resolu-
estimate centroids of individual fibers, but only a few are capable of tion systems (e.g., the Zeiss Ultra-810, Zeiss Versa-520, or Advanced
determining individual fiber diameters and local fiber orientations re- Light Source Beam Line 8.3.2). The benefit of using synthetic image
lative to the scan axis [19,20,22]. The second step in fiber segmentation data, as opposed to real high-resolution data from other imaging
involves the conversion of 2D fiber detections into continuous 3D fiber sources (e.g., see [20]), is that all information about fiber positions,
“tracks.” A Nearest Neighbor Search (NNS) is the most common method shapes, and sizes are known precisely, making validation more
of establishing 3D fiber tracks [18,19,23,24]. The NNS method is re- straightforward.
latively straightforward to implement and can be computationally ef- In what follows, this manuscript begins with a description of the
ficient; however, it can also result in non-smooth and non-continuous new fiber segmentation algorithm, including synthetic TM and Kalman-
tracks if the fiber-detection data contains spurious and/or missing in- filter fiber tracking for 3D segmentation. Next, the process for gen-
formation. To address this issue, one approach is to use a more com- erating synthetic X-ray µCT images are presented and then used to
putationally-intensive tracking algorithm based on a Kalman filter evaluate TM and fiber tracking over a range of image resolutions and
[16,20,22], which results in smooth and continuous tracks, even if fiber architectures. This manuscript concludes with the segmentation of
some of the fiber-detection data is spurious and/or missing. Other a real multi-directional X-ray µCT dataset.
methods include high orientation gradients [25], heavy ball method
[26], normalized correlation coefficient and Bayesian inferences [15], 2. Fiber segmentation algorithm
and 3D structuring elements within the Fiji plugin F3D [17].
Despite the wide range of available approaches for detecting and The fiber segmentation algorithm described herein has been de-
tracking cross-sections of individual fibers imaged using X-ray µCT, veloped specifically for continuous unidirectional and multi-directional
there is currently no consensus in the literature regarding which carbon/polymer composites imaged using X-ray µCT. The entire seg-
method is most suitable for the broadest range of possible FRC systems, mentation process, depicted using a block diagram in Fig. 1, is com-
scan resolutions, and quality of image data. Consequently, the applic- prised from three main components: (1) 2D TM to detect locations and
ability of existing fiber segmentation methods appears to be limited to radii of individual fiber cross-sections in individual tomograms; (2) 3D
data types for which they were developed. Additionally, in most cases, fiber tracking to estimate location and size of fiber segments using a
the accuracy of the existing segmentation methods have not been sys- Kalman filter; and (3) track smoothing, track selection, track stitching,
tematically validated, which makes selection and use of any one and overlap removal. Each component of the proposed algorithm is
method difficult. Of notable exception is the study performed by described in the subsequent sub-sections. Implementation of the entire
Emerson et al., where images of a glass-fiber reinforced polymer were algorithm was performed in Matlab®.
acquired at varying spatial resolutions using X-ray µCT, optical micro-
scopy, and scanning electron microscopy [19]. These images were used 2.1. 2D template-matching
to validate the accuracy of estimating unidirectional fiber diameters
and positions using a circular Hough transform and probabilistic fea- For each image in a tomographic stack, the TM algorithm is used to
ture labeling. The study by Emerson demonstrated that fiber cross- measure/detect centroids of fiber cross-sections, their radii, and mag-
sections could be detected with high precision, particularly for low- nitude of fiber tilt relative to the global scan axis and independently of
resolution image data. However, this study validated the algorithm the image resolution. This information is later passed on to the 3D fiber
with unidirectional fiber orientations, and its performance for detecting segmentation algorithm, which is described in Section 2.2. In this work,
and tracking multi-directional fibers in 3D remains to be seen. the TM algorithm assumes that only two phases (i.e., fibers and sur-
In view of the above discussion, this study aims to advance the rounding matrix) are present within a well-consolidated, void-free
current state-of-the-art in fiber segmentation from X-ray µCT images by composite. With that assumption, the TM algorithm generates synthetic
(1) developing a new 3D fiber segmentation algorithm which estimates, fiber templates based on two parts: (1) statistics on the distribution of
with high-accuracy, diameter and orientation of individual fibers in all pixel intensities within a tomogram and (2) fiber geometry (i.e., a
multi-directional FRC microstructures; and (2) proposing a new meth- potential range of fiber diameters and orientations of circular or ellip-
odology for evaluating fiber segmentation algorithms using synthetic X- tical fiber cross-sections). The first part determines the likelihood that a
ray µCT image data. The new segmentation algorithm developed herein given pixel intensity is associated with one of two Gaussian distribu-
is an extension to the existing approach proposed by Czabaj et al. that tions: one corresponding to fibers and one to the surrounding matrix, as
used TM to detect fiber cross-sections, and a Kalman filter to estimate seen in Fig. 2a. Expectation Maximization (EM) [27] is used to de-
(or segment) individual fibers in 3D [16]. The TM algorithm presented termine these distributions by fitting a Gaussian mixture model to the
in Ref. [16] was based on manual selection of “real” fiber-templates, histogram of pixel intensities. As a result, there are two outputs from

Fig. 1. Block diagram of the fiber segmentation

μCT images 3D digital fibers process.

template 2D detections 3D Kalman 3D fiber tracks smoothing, selection,

matching filtering stitching, overlap removal

2
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

5000

4000

changing orientation
3000
fibers
count

2000

1000
matrix

0
0 30 60 90 120 150 180 210 240

8-bit grayscale pixel intensity changing size

Fig. 2. Template-matching combines (a) global statistics on the distribution of pixel intensities and (b) synthetic templates of varying size and orientation to detect
fiber cross-sections.

EM: (a) the probability of each pixel within the image being a fiber with the following equation:
given the intensity, (P(F|Cij)), or (b) the probability of a pixel selected at
P (F Cij ) ⎞
random being from the fiber distribution (P(F)). A GMM was selected SLLR = ∑ log ⎛
⎜ ⎟

based on pixel intensity distributions observed in high-resolution X-ray ij ∈ RF ⎝ P (F ) ⎠ (1)

µCT images of void-free FRCs, where the distributions are typically bi-
where RF is used to denote the region of the template expected to be a
modal and contain two approximately normal distributed point sets.
fiber. The primary benefits of the synthetic template approach are as
Note that for lower resolution/contrast X-ray µCT images, other models
follows: (1) using multiple templates gives parametric information
(e.g., chi-squared or Weibull distributions) may be more appropriate.
about the fibers (including orientation and size), (2) use of multiple
Future work will focus on expanding the list of potential models to best
templates is straightforward where the score is the max over all of the
fit other distributions of pixel intensities. Here Cij denotes a pixel value
templates, and (3) a score of zero is a natural score threshold indicating
in the ith row and jth column, and F indicates a pixel being a fiber.
that the template is more likely to be correct than not.
In the second part, a priori knowledge of possible fiber geometries,
including cross-sectional shapes, orientations, and sizes, is used to
generate a set of synthetic templates similar to ones shown in Fig. 2b. 2.2. 3D fiber segmentation
The red portion of the template pertains to pixels being a fiber, and the
blue portion corresponds to either fibers or matrix, where each region is TM algorithm is applied to each tomogram within the image stack to
determined using the statistics collected on the distribution of all pixel generate a set of 3D fiber detections. The 3D fiber segmentation algo-
intensities within the image. The exact configuration of fiber positions rithm is used to estimate the number of fibers, their position, and their
is unknown, so the method used in this work assumes that the blue radii from the set of detected centroids, radii, and magnitude of fiber tilt
pixels are either fibers or matrix. This is done to detect events when red estimated during TM. Estimation of the fiber radius, compared to pre-
pixels from two or more neighboring fibers are touching, or if red pixels vious work [16], was added to the algorithm (1) to account for varia-
from a single fiber are completely surrounded by the matrix. The syn- bility in TM measurements of fiber cross-sections corresponding to the
thetic templates are generated by specifying the range of possible fiber same fiber, and (2) to enable overlap removal between neighboring
radii, the range of possible fiber orientations in 2D and 3D, the allow- fibers by adjusting their radii. More specifically, fiber tracking de-
able detection overlap, and the number of EM optimization steps to be termines an estimate of the state of each fiber for each tomogram, de-
performed. Fiber overlap removal consists of three steps. First, fiber ̂ . It is assumed here that
noted for the ith virtual fiber at image k as x i,k
cross-sections are approximated using the MATLAB function inpolygon. the fiber state, x, is composed of 2D position, p, 2D velocity, v, and
Second, the function quadprog is used to determine which cross-sections radius, r, for each fiber and image. In this work, the velocity of a fiber
are the most likely set of fibers without too much overlap. Third, cross- track is measured using the polar and azimuthal angles in spherical
sections that overlap are rotated, translated, and/or have their radius coordinates, where the initial velocity is based on the fiber tilt mea-
reduced until the remaining overlap is within the allowable threshold. sured from TM in the first image in the stack. In addition to the state
These parameters can be estimated based on visual inspection of the X- estimate, the segmentation process also generates an associated error
ray µCT images, or they can be made purposefully broad if such esti- covariance for each virtual fiber state for each image, denoted for the
mation is not possible. As shown in Fig. 2b, the overall template shape ith virtual fiber at the kth image as Pi,k. Generating a 3D reconstruction
is square with dimensions equal to the largest ellipse used. of the FRC microstructure from 2D detections in X-ray CT images
The procedure to detect individual fiber cross-sections is to move consists of five key steps: (1) tracking, (2) smoothing, (3) track selec-
each synthetic template across a tomogram while simultaneously tion, (4) track stitching, and (5) non-overlapping fiber constraint sa-
computing a Log-Likelihood Ratio (LLR) score for each template at each tisfaction. A detailed description of these algorithms is highlighted in
pixel location. Scores determine how well each template best explains the following sections.
that portion of the image using the overall global image statistics, and
the template with the highest LLR score is kept for each pixel location. 2.2.1. Tracking
Positions of fiber centroids are estimated by determining local maxima Details of tracking have been previously presented in [16] and are
of the LLR scores within the image. Computing the LLR score is done restated here for completeness. A fiber tracker makes a forward pass

3
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

tracks

tracks and assigned

detections updated tracks
extrapolate update
possible unassigned
detections gate and assignments detections new tracks
prediction assign initialize
score
tracks not assigned
detections
terminate

Fig. 3. Block diagram of the fiber tracking process.

between images within the stack and links neighboring 2D detections with the measurement matrix, H, for position, velocity, and radius de-
into a 3D track. This process is repeated for all detections to generate fined as
multiple 3D fiber tracks. The tracker uses a global nearest neighbor data
Hp = [ I2 02 01],
association scheme [28], to determine which 2D detections between
image cross-sections have close proximity, and performs estimation Hv = [ 02 I2 01], and
using a Kalman filter [29]. The key steps of the tracker are shown in the Hr = [ 0 0 0 0 I ], (6)
block diagram in Fig. 3 and described further throughout this section. where I2 denotes the 2 × 2 identity matrix, 02 denotes the 2 × 2 zero
As seen in Fig. 3, the first step of fiber tracking is to predict the matrix, 01 denotes the 2 × 1 zero matrix; and the measurement error
location of tracks using the 2D detections from TM for each image. For covariance, R, is assumed to be constant for all measurements and
the first image, k = 1, the detection centroid is used as the initial po- images.
sition, the initial velocity is assumed to be zero, and an initial track Next, for all possible detection to track pairings, three levels of
likelihood score is based on the ratio of true detections to false detec- gating are used to determine which pairing should be considered in the
tions, which is described in more detail later in this section. Ad- assignment step. First, there is a rectangular gate that compares each
ditionally, in the event a new detection appears within the field of view measurement dimension independently. For a potential pairing of the
(i.e., a fiber had a missing detection on a previous image) or enters the jth measurement zj,k+1 to track i on image k + 1, and for each dimen-
field of view at some later image, the same initial conditions are used. sion of the measurement, ξ, the following inequality must hold
Prediction is performed using the state estimate and covariance matrix
for each track from image k to image k + 1 using the standard Kalman (z i, k + 1 k (ξ ) − zj, k + 1 (ξ ))2 ⩽ 4Si, k + 1 (ξ , ξ ) (7)
filter propagation given as
where the parameter 4 is chosen such that the gate is a “2σ” window.

x i, k + 1 k = Φ
x i, k k This choice of a relatively tight window is to ensure that false detections
Pi, k + 1 k = ΦPi, k k ΦT + Q, are not allowed into the track. Although this increases the potential for
(2)
missed detections resulting in shorter tracks, this is not a problem as the
and a white noise constant velocity dynamic model and radius are used tracks can be stitched together and is discussed further in Section 2.2.4.
to model the change in fiber state from image to image, where Φ is the Potential pairings that pass the rectangular gate are then subjected to
state transition matrix, Q is the process noise covariance, P is the error an elliptical gate based on the squared Mahalanobis distance, η, given
covariance matrix, and x ̂ is the state estimate. For details on de- by
termining the process noise covariance and the state transition ma-
ηij, k + 1 = νijT, k + 1 Si−, k1+ 1 νij, k + 1 (8)
trices, the reader is directed to Section 4.2.2 in Ref [28]. The state
transition matrix is defined to be where the measurement residual, νij,k+1, between the ith track and the
0 0 0⎤ jth measurement, on image k + 1, is
⎡I τ
⎢0 I 0 τ 0⎥ νij, k + 1 = z i, k + 1 − zj, k + 1 (9)
Φ = ⎢0 0 I 0 0⎥
⎢0 0 0 I 0⎥ For a potential pairing to pass the elliptical gate, the squared
⎢0 0 0 0 I⎥ (3)
⎣ ⎦ Mahalanobis distance must be less than a threshold, ηthresh, that is set to
and the process noise covariance to be enclose 99% of the probability for a distribution of degree two and is
calculated using the Matlab® function chi2inv. The final gate is based on
3 2
⎡τ 0 τ 0 0⎤ the delta score, δ ij,k+1, which is a negative log-likelihood score and is
⎢3 3 2 ⎥ computed as
⎢0 τ 0 τ2
0⎥
⎢ 3 2 ⎥ 1
Q = q ⎢ τ2 δij, k + 1 = η + ln(2πβ Si, k + 1 )
0 τ2 0 0⎥ 2 ij, k + 1 (10)
⎢2 ⎥
⎢ τ2 ⎥ where β is a false detection density parameter. The false detection
⎢0 2 0 τ 2 0⎥
⎢0 0 0 density is determined by the number of false detections per unit of
⎣ 0 0⎥ ⎦ (4)
measurement space. The smaller the delta score, the more likely the
where I denotes the identity, q is the process noise scaling factor, and τ pairing is correct. Therefore, pairings which produce a negative delta
is the spacing between images, which is analogous to the time step in score are saved for possible assignment in the assignment step.
the typical description of the state transition matrix and process noise. The assignment step determines the most likely detection-to-track
The next step is to gate and score all possible pairings of detections pairings using a GNN association. The association makes an optimal
to tracks. This is performed by computing the predicted measurement assignment across all measurements of fiber tracks simultaneously from
z i,k+1, and the residual covariance matrix, Si,k+1, as the following: the list of possible pairings and their delta scores. Since the delta score
for each pairing is computed as a negative log-likelihood score, the
z i, k + 1 k = Hxi, k + 1 k
most likely set of detection to track pairings is given by the set with the
Si, k + 1 k = HPi, k + 1 k H T + R (5) lowest combined delta scores. In selecting the most likely set of

4
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

detection to track pairings, the assignment step must assign at most one true, if track i overlaps track j ⎫
Cij = ⎧
detection to a track and no detection to more than one track. Solving ⎨
⎩ false, otherwise ⎬
⎭ (13)
the assignment problem was performed using the Munkres algorithm
[30]. Note that the conflict matrix is originally defined to capture the
Each track that is assigned a detection in the assignment step is conflicts between all possible tracks. However, clustering is employed
updated with the standard Kalman filter update. Furthermore, each to reduce the problem. In the clustering step, the smallest possible
track has an associated cumulative likelihood score where the delta clusters are formed such that any two tracks that conflict are in the
score from gating and scoring is added to the tracks′ cumulative like- same cluster and therefore partitions the conflict matrix accordingly.
lihood score. The update equations for the ith track taking the jth de- The selection of tracks is then performed on each cluster one at a time,
tection at image k + 1 are the following: which reduces one large problem into multiple smaller problems.
Given the conflict matrix and the likelihood scores of each of the
K = Pi, k + 1 k H T Si−, k1+ 1, tracks in a cluster, solving for the most likely set of non-conflicting

x i, k + 1 k + 1 = 
x i, k + 1 k + Kνij , and tracks can be formulated formally as a binary integer programming
Pi, k + 1 k + 1 = Pi, k + 1 k − KHPi, k + 1 k . (11) problem [32], which is NP-complete [33]. Since this problem is
common, particularly in Multiple Hypothesis Tracking, multiple ap-
The “terminate” step stops a track from being extrapolated further, proximation algorithms have been developed [34–36]. However, these
which is done for tracks that have gone too long without an update (i.e., approximation algorithms are specifically designed to be efficient sol-
the end of a fiber track). For this work, any track missing more than one vers when the track selection problem is derived from multiple images,
update opportunity is terminated. This strict requirement is im- which is not the case here. For that reason, the implementation of these
plemented to maintain pure tracks, and is required due to the nature of algorithms did not seem appropriate or justified. Here a hybrid of two
the false alarms, which may appear repeatedly in the same place. As a approaches is used. When the number of tracks is small (< 15) the
result, short, pure tracks may be estimated, but these tracks can be selection problem is solved by brute force evaluation of all possible
stitched into long, continuous tracks later as described in Section 2.2.4. track combinations. When the number of tracks in a given cluster ex-
The last step is to initialize new tracks based on any detections that ceeds 15, then the problem is solved using the bintprog function in
are not assigned to an existing track. This is performed to handle de- Matlab®, which uses a linear programming based branch-and-bound
tections that were not assigned to a fiber track. Initial values for these algorithm.
tracks are assigned based on the initial conditions described previously
in the prediction step.
2.2.4. Track stitching
2.2.2. Smoothing As described in the tracking selection step, track deletion parameter
The tracker uses the detections only in a forward pass. Once com- settings favor track purity over track length resulting in pure but short
pleted, all detections are available and can be used to improve the tracks that do not necessarily span the entire length of the scanned
tracks for all images. For this work, a two-pass smoother was im- volume. These tracks must be linked into longer continuous tracks in a
plemented based on the algorithm presented in Ref. [31]. Since the process known as track stitching. Track stitching is widely used in
forward pass is performed by the tracker, only the backward pass is ground target tracking where potentially thousands of targets are being
made by the smoother. Given the state estimates and covariance for the tracked with frequent track drops due to line of sight occlusions or
ith track from its initial image, Ni,I, to its final image, Ni,F, the smoothed vehicles stopping. As such, several sophisticated track stitching ap-
estimates are given by recursive application of the following backwards proaches have been developed [37–39]. The track stitching problem
pass equations for k = Ni,F to Ni,I: faced here is comparatively simple, and therefore, a straightforward
track stitching algorithm is employed. The track stitching algorithm
P + = ΦPi, k − 1 k − 1 ΦT + Q, consists of five steps, which are described further as follows.
C = Pi, k − 1 k − 1 ΦT (P +)−1, Step 1: the first step is analogous to the gate and score step in the

x i, k − 1 Ni, F = x i, k − 1 k − 1 + C (
 x i, k Ni, F − Φ
x i, k − 1 k − 1), and tracker. The goal is to determine the possible track stitches and provide
a likelihood score for each. The search for possible stitches is con-
Pi, k − 1 Ni, F = Pi, k − 1 k − 1 + C (Pi, k Ni, F − P +) C T . (12) strained in four ways: (1) only tracks that, when stitched together, fit
within the boundary of the specimen are considered, (2) the first track
must end before the second track begins, (3) the number of images
2.2.3. Track selection between the end of the first track and the start of the second is limited
Track selection is used in two contexts within this work. First is in to a user-defined value to prevent incorrect stitches, and (4) when the
the selection among multiple possible track stitches, which is discussed first track is propagated between the end of the first track and the start
further in Section 2.2.4. The second is to determine which tracks to of the second, their potential pairing passes the elliptical gate described
keep in the case of significant overlap, which is discussed further here. previously.
TM may find fibers at the intersection of many neighboring real fibers Step 2: a conflict matrix is computed analogously to the conflict
even though there are no fibers there. If these false detections are found matrix discussed in the track selection step. Note that two stitches
consistently from image to image, the tracker will form them into re- conflict if they share the end of the first track or the beginning of the
latively long continuous tracks. To reconstruct the microstructure ac- second track.
curately, overlapping fibers must be discarded, which raises the ques- Step 3: the third step clusters the possible stitches in the same way as
tion of how to automatically determine which fibers are real and which tracks described in Section 2.2.3. The purpose is to reduce a single large
are not. In this paper, this question is formulated as a track selection problem into multiple smaller problems. Then the selection algorithm
problem. That is, tracks that overlap too much are considered to con- can solve each cluster separately.
flict and cannot both be real. In this work, too much overlap was de- Step 4: the selection algorithm described in Section 2.2.3 is used to
fined as 20% of the fiber diameter. However, a statistical measure based select the most likely set of non-conflicting stitches in each cluster.
on Mahalanobis distance would potentially be more appropriate and Step 5: the chosen stitches are used to develop a refiltered/smoothed
will be explored in future work. state estimate for the now longer track. Note that each of the stitches
To perform the track selection, the conflict matrix, C, is defined are pairwise, so multiple stitches can be used to form a single long
initially such that the ijth component is given by track.

5
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

2.2.5. Non-overlapping fiber constraint satisfaction considerable number of fibers contained within the field of view, a
As defined previously, the track selection process removes all virtual smaller 205 × 205 × 80 µm sub-volume was extracted for this work. In
fibers that have significant overlap. This is normally caused by the addition to the high-resolution scan, four additional scans were per-
tracker developing tracks for fibers that do not truly exist. However, formed at progressively lower resolutions. In total, five sets of real
even when all of the virtual fibers correspond to real fibers, minor image data were acquired: one tomographic image stack to evaluate the
overlaps may still exist due to statistical errors and variations in fiber robustness of the fiber tracking algorithm, and four scans to form the
radii. One of the important requirements for the resulting digital re- range of image resolutions for synthetic data generation.
construction is that it contains no overlapping (or intersecting) virtual
fibers. This requirement can be formulated as a constraint between each 3.2. Generation of synthetic X-ray µCT data
pair of fibers where the overlap removal is treated as a constrained
estimation problem. The constraint is applied to the estimate as pro- Synthetic X-ray µCT data can be generated in two possible ways.
pagated by the Kalman filter defined by the work of Julier and LaViola The first method involves creating a phantom FRC microstructure,
[40], but is interpreted here as a constraint on the digital micro- performing a Radon transform to create corresponding sinograms, and
structure returned from the 3D segmentation routine. The solution to then reconstructing the sinograms using filtered back-projection. The
this problem is based on the solution of a maximum a posteriori (MAP) second method involves directly creating synthetic tomograms that
optimization and is specifically designed to remain tractable even as the mimic the resolution and quality of images obtained from existing
number of constraints/tracks grows. This work employs a numerical imaging systems. The first method accounts for artifact formation
approach based on a polygon intersection approximation to the ellipses. caused by filtered backprojection; however, it does not account for
A MAP estimate is sought from the unconstrained probability density noise and artifact formation created during image acquisition. Given
function (pdf) that satisfies the constraints. The unconstrained pdf for that the latter effect was thought to have a stronger influence on per-
the tracks derived from the previous Sections 2.2.1–2.2.4 define p(xi|Zi) formance of fiber segmentation algorithms, the second method was
as Gaussian for each image and each fiber independently. The MAP used in this study.
approach seeks the maximum of this distribution that does not violate Generation of synthetic tomograms was based on several key as-
the non-overlapping constraint. This results in a quadratic cost function sumptions. First, it was assumed that the distribution of noise in real X-
where the constraint is defined as the maximum overlap distance be- ray µCT data is Gaussian. In general, the actual distribution of noise is
tween polygon approximations. Mathematically, the maximum overlap difficult to determine, and is likely dependent on the material imaged
distance is written as the following equation: and/or the X-ray µCT system. Second, it was assumed that only two
constituents (fibers and the surrounding matrix) were present within
d = max ( min ∥pi − pj ∥ )
pi ∈ Poly1 pj ∈ Poly 2 (14) the synthetic microstructures. Features such as cracks or voids were
excluded from the synthetic data due to the difficulty in quantifying/
where d is the maximum overlap distance, and Poly1 and Poly2 are two representing their geometries systematically. Third, it was assumed that
circular or elliptical polygon approximations of fiber cross-sections fibers are perfectly cylindrical, and, when mimicking the fiber cross-
using points pi and pj. section profile from X-ray µCT image data, it was assumed that the pixel
intensity profile had circular symmetry (i.e., symmetry about the
2.2.6. Final track radius perimeter). These assumptions were used to construct representative 2D
After performing tracking, smoothing, track selection, track tomograms, which were subsequently converted into 3D image stacks
stitching, and non-overlapping fiber constraint satisfaction, the radius as described in the following two sub-sections.
for any given track may vary across its length. In this study, it is as-
sumed that fibers maintain a constant radius, thus the radius for each 3.2.1. Generation of 2D synthetic tomograms
track is determined using the minimum value from the list of estimated Generation of 2D synthetic tomograms was performed in five steps.
radii. The minimum radius is selected to avoid overlap between The first step involved creating a single 2D 232 × 209 µm 16-bit
neighboring tracks. grayscale image, with a resolution of ~0.42 µm/pixel, that contained
1000 hexagonally-packed fibers with randomly varying diameters and a
3. Generation of real and synthetic data fiber volume fraction of Vf = 0.81. The assignment of random fiber
diameters, which had a mean value of 7.1 µm and a standard deviation
In this study, real and synthetic X-ray µCT images are used to of ± 0.83 µm, was based on optical measurements obtained from an
evaluate the performance of the segmentation algorithm. Real X-ray AS4/3501–6 carbon/epoxy system [42]. In this image, the intensity of
µCT images were collected to provide baseline information needed to pixels corresponding to fibers was initially set to 0, while pixels cor-
generate synthetic images and were used to demonstrate segmentation responding to the matrix were set to a value of 35,727. In the second
of as-manufactured multi-directional microstructures. Synthetic images step, 38 fiber cross-sections were removed at random to simulate ma-
were used to validate the accuracy of TM and tracking algorithms. trix-rich regions, which resulted in 962 hexagonally-packed fibers with
Details of acquiring/generating these image data are described in the a fiber volume fraction of Vf = 0.79. In the third step, fiber centroids
following sub-sections. were randomly perturbed according to the algorithm developed by
Wongsto et al. [43]. In the fourth step, a mask (i.e., a profile that mimics
3.1. Acquisition of real X-ray µCT data the distribution of pixel intensities for carbon fibers imaged using X-ray
µCT) was applied to each fiber centroid. To generate this mask, pixel
Real X-ray μCT images were acquired on 1 mm × 1 mm section intensities across seven fiber cross-sections from real image data were
extracted from IM7/8552 carbon/epoxy laminate, which contained sampled using the open-source software Fiji [44], as shown in Fig. 4a.
plies oriented at +45°, −60°, and +60° with individual fibers having a The sampling of pixel intensities from real image data was necessary to
nominal diameter of 5.2 µm [41]. Imaging was performed using a Zeiss quantify the extent of beam hardening [45] that is observed on the
Xradia 520 Versa CT microscope over a 360° rotation using 4000 pro- periphery of individual fibers imaged with X-ray CT. The profiles of
jections, 60 kV voltage, 5 W power, 14 s exposure time per projection, pixel intensities for individual fibers were fitted with a sixth-order
and a 40× objective lens. The resulting voxel size was approximately polynomial function to obtain a general profile of pixel intensities
410 nm, and the total scan time was approximately 16 h. Reconstruc- across a typical fiber (see Fig. 4b). This polynomial function was used to
tion of attenuation data was performed using filtered back-projection generate a row of pixel intensities representing a generic fiber, which
producing 949 cross-sectional, gray-scale images. Due to the was subsequently revolved about the fiber centroid to produce the fiber

6
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

40k 538H × 2 W images, where each image in the stack had noise, blurring,
5

16-bit pixel intensity

1 and histogram matching applied separately.
6
35k Image stacks with off-axis fibers were produced by vertically
4 fiber #1 stretching the side-by-side H × W images within the 0°/0° stack to
2 30k fiber #2
fiber #3 create elliptical fiber cross-sections that represent +45° or −60° fiber
fiber #4
3 7 fiber #5
angles. For the +45°/−60° image, the newly stretched side by side
25k
fiber #6 images were of size 2 H × W and 2H × W, respectively. Next, images
fiber #7
6th degree polynomial fit on either side of the interface were translated vertically using a Matlab®
20k
0 0.25 0.5 0.75 1 function imtranslate such that each image in the stack was translated
position across fiber diameter
further than the previous. This was repeated for all image slices in the
stack to give fiber cross-sections the appearance of moving at a “fiber
velocity” corresponding to the desired orientation of a given ply. For
example, in the +45°/−60° dataset, left-side images were translated
vertically up, and right-side images were translated vertically down.
After translation, the resulting image stack resembled a parallelepiped,
which was subsequently cropped to a new width 2 W and height H to
produce a rectangular prism. Examples of stacks containing off-axis
fibers after cropping are shown in Fig. 6b and c.
The resulting synthetic and real image stacks are shown side-by-side
in Fig. 6e–h. In this figure, the 0°/0°, ± 45°, and +45°/-60° synthetic
image stacks have an average fiber diameter of 17 pixels and include
Fig. 4. Process of generating synthetic images based on a fiber mask generated 538, 537, and 499 images, respectively. Image stacks in Fig. 6e–g
from X-ray µCT data.
contain 623, 1441, and 1408 individual fibers, respectively. The pro-
cess for generating 3D tomographic stacks was repeated for three lower
mask, as shown in Fig. 4c. To account for varying fiber diameters, the resolutions synthetic images, 4.1, 8.8, and 14.4 pixels/davg, to produce
polynomial function was appropriately scaled for each unique dia- a total of 12 volumetric datasets.
meter. A unique mask was applied at each fiber centroid within the
synthetic microstructure using the convn function in Matlab®, as shown 4. Results and discussion
in Fig. 4d. In the fifth step, Gaussian noise, Gaussian blurring, and
histogram matching were applied to the synthetic image. This was done Evaluation of the 3D fiber segmentation algorithm was performed
to mimic the histogram of pixel intensities seen in real image data. Both by examining how: (1) the TM algorithm detects individual fiber cross-
noise and blurring were performed using Fiji, while histogram matching sections in 2D tomographic images, and (2) the fiber tracking algorithm
was performed using the function imhistmatch in Matlab®. The steps estimates positions of individual fibers in 3D based on the TM input.
mentioned previously describe the generation of a synthetic tomogram Both 2D detection and 3D estimation were performed on synthetic
at one resolution; however, these steps can be adapted for any resolu- images of four separate spatial resolutions: 4.1, 8.8, 14.4, and 17 pixels/
tion and number of fibers. davg.
The highest resolution synthetic 2D tomogram had a resolution of
0.417 µm/pixel, which corresponds to 17 pixels per average fiber dia- 4.1. Evaluation of the TM algorithm
meter (pixels/davg). To generate synthetic data at lower resolutions, the
“high-resolution” synthetic tomogram was down sampled using bilinear Evaluation of the TM algorithm was performed using synthetic 2D
interpolation in Fiji, which was followed by additional blurring and images described in Sections 3.2.1–3.2.2. For the highest resolution 2D
histogram matching. In this work, tomograms were generated at 4.1, image, TM, which required 82 unique templates, was completed in 3.3
8.8, 14.4, and 17 pixels/davg. These four resolutions correspond to an wall-clock seconds by running on 14 cores in parallel on a desktop
existing range produced by state-of-the-art X-ray µCT systems. Fig. 5 workstation with a 14 core Intel Xeon E5-2690v4, 2.6 GHz CPU with
shows the synthetic image generation applied to fiber locations ex- 256 GB of RAM. Fig. 7 shows a small subset of 0° fibers with TM de-
tracted from real X-ray μCT images, where positions and radii of fiber tections (red) and corresponding ground truth (yellow). From this
cross-sections were determined using results from the TM algorithm figure, it is evident that TM detects the location of fibers quite well for
applied to each real image. The top row shows original X-ray µCT all image resolutions. However, for all resolutions considered, the TM
images, the middle row shows the corresponding synthetic images, and algorithm underestimates the diameter of fibers, particularly at low-
the bottom row corresponds to histograms of pixel intensities extracted resolution (see Fig. 7a). To quantify this further, the TM algorithm was
from real and synthetic images. Qualitatively, the proposed method for evaluated based on the following three metrics: (1) the percent detec-
generating synthetic tomograms mimics the contrast, noise, and pixel tion of individual fibers, (2) the average error between true and mea-
distributions of real data quite well across the range of resolutions sured positions of fiber centroids, and (3) the average percent error
considered. between true and detected fiber cross-sectional areas.
Results from evaluating the first metric are shown in Fig. 8, where
3.2.2. Generation of 3D synthetic tomographic stacks the percent of detected fibers is plotted on the y-axis, the image re-
The high-resolution 2D synthetic image described in the previous solution is given on the x-axis, and the red, green, and blue bars cor-
section was used to generate three distinct synthetic 3D tomographic respond to fibers oriented at 60°, 45°, and 0°, respectively. As seen in
stacks, each representing two “plies” with 0°/0°, ± 45°, and +45°/ this figure, TM performs quite well detecting the total number of fiber
−60° fiber orientations, respectively. For the 0°/0° dataset, the 2D cross-sections, even at the lowest image resolution. TM does not detect
synthetic image was first cropped to a new width W = 125.7 µm and any spurious fibers, which is likely due to the dense packing of fibers in
height H = 125.7 µm to reduce the computational time associated with the synthetic images (i.e., spurious detections usually occur in large
3D fiber segmentation. Next, the ply interface between the two 0° plies resin-rich areas). Despite the very high rate of detection, a few fibers in
was created by placing two H × W cropped synthetic images together the 0° images were missed. The missed fibers were not detected because
side-by-side, resulting in a new image with a width 2 W and height H their diameters were smaller than the smallest pre-defined synthetic
(see Fig. 6a). The entire 0°/0° dataset was generated by using fiber template. The low range of potential fiber diameters was made

7
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

4.1 pixel/davg 8.8 pixel/davg 14.4 pixel/davg 17 pixel/davg

real data

synthetic
data

1600
real 375 1000
100 synthetic
300 1200
frequency

75 750
225
800
50 500
150

25 250 400
75

0 0 0 0
21k 25k 29k 33k 37k 19k 23k 27k 31k 26k 31k 36k 41k 46k 26k 30k 34k 38k 42k
16-bit pixel intensity
Fig. 5. Comparison between real X-ray CT (top) and synthetic (middle) at various resolutions. Bottom row compares histograms of real and synthetic images.

slightly larger than the smallest fibers to avoid detection of spurious the y-axis, the image resolution is plotted on the x-axis, and the color
fibers in the resin-rich areas. In other words, the detection of all pos- bars correspond to the fiber orientations. As evident for Fig. 9, the
sible fibers was sacrificed to generate high-quality true detections. An average error and standard deviation decrease with the increasing
alternative approach would be to select a range of templates that en- image resolution, and increase with the increasing fiber angle. The in-
compasses all fiber diameters, generate spurious detections, and allow crease of error with fiber angle may be related to the larger variability
the fiber tracker to filter these out during track selection and stitching. of pixel intensities within any given fiber cross-section. Fibers with
Results of evaluating the second metric are shown in Fig. 9, where higher angles have larger cross-sectional areas, and thus are more likely
the average error in detecting fiber centroids is computed by taking the to contain noisy pixels which are expected to decrease the detection
difference between true and measured positions of fiber centroids and accuracy. Regardless, for all three fiber orientations analyzed, the
normalizing these values by the corresponding true fiber diameter. In highest average error was 10.4% with a standard deviation of 6.2%,
this figure, the error and its standard deviation (error bars) is plotted on obtained from the 60° fibers and a resolution of 4.1 pixels/davg. This

Fig. 6. Images of synthetic (a–c) and real (d) 2D cross-sections and corresponding 3D volumes (e–h) used to evaluate the performance of the 3D fiber segmentation
algorithm.

8
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

18

TM position % error
15

12

0
4.1 8.8 14.4 17.0
pixel/davg
Fig. 9. Average error between true and TM-measured positions of fiber cen-
troids versus image resolution.

60

Fig. 7. TM results (red) and corresponding ground truth fibers (yellow) for four

cross-sectional
area % error
image resolutions: (a) 4.1, (b) 8.8, (c) 14.4, and (d) 17.0 pixels/davg. (For in- 45
terpretation of the references to color in this figure legend, the reader is referred
to the web version of this article.)

30
102

101
15
fiber % detection

100

99 0
98 4.1 8.8 14.4 17

97
pixel/davg
Fig. 10. Average error between true and TM-measured fiber cross-sectional
96 areas versus image resolution.
95
TM
synthetic image profile
0 dactual
4.1 8.8 14.4 17.0
pixel intensity

ddetected
pixel/davg
Fig. 8. Percent detection of individual fibers by the TM algorithm versus image
resolution.
beam
hardening
error corresponds to the detected positions being off by approximately
0.42 pixels, which is quite small relative to the average fiber diameter. Fig. 11. The difference in distribution of pixel intensities across real/synthetic
Results from evaluating the third metric are shown in Fig. 10, where fibers (dashed line) and that assumed by the synthetic templates in the TM
the average percent error between true and measured fiber cross-sec- algorithm (red line). (For interpretation of the references to color in this figure
tional areas is plotted on the y-axis, and the image resolution is plotted legend, the reader is referred to the web version of this article.)
on the x-axis. Examining Figs. 7 and 10, it is evident that the TM al-
gorithm had difficulty detecting the true area of fiber cross-sections for
synthetic template (solid red line) assumes there is a sharp change in
all image resolutions. At the lowest resolution, the error was 53.2% on
pixel intensities when transitioning between a fiber and the sur-
average for all three fiber orientations, and as the resolution was in-
rounding matrix. However, fibers imaged using X-ray µCT exhibit beam
creased, the error remained relatively constant at approximately 33%.
hardening, which is manifested by a gradual drop in pixel intensities
For each fiber angle, the standard deviation decreased from 8.4% at 4.1
around the perimeter of individual fibers, as shown by the dashed black
pixels/davg to 1.9% at 17 pixels/davg.
line in Fig. 11. The low-intensity pixels caused by beam hardening are
One explanation for the larger error in measuring fiber area may be
excluded after EM is performed; and as a result, TM underestimated the
the simplistic model used to generate synthetic templates (shown in
true area of fiber cross-sections. Two possible remedies for this issue
Fig. 2b). The limitation of this model is illustrated in Fig. 11, where the
include: (1) modeling the synthetic templates to model beam hardening

9
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

artifacts, or (2) artificially increasing the estimated diameter/area of - -

detected fibers based on a priori knowledge of fiber sizes. The latter
TM ± Tracker ±
option will only work for fiber types that exhibit a constant fiber dia-
meter.
Based on the results presented in Figs. 8–10, it can be concluded
that the proposed TM algorithm is quite accurate when fibers contain,
15

fiber position % error

on average, at least 9 pixels across the diameter. At this resolution,
error in detecting the number and position of fiber cross-sections is less
than 5% regardless of fiber orientation. Although this error increases
12
for lower resolution image data, TM remains accurate in detecting the
number and position of fiber cross-sections, even when individual fibers
cannot be discerned visually. A surprising outcome of validating the TM 9
algorithm is the persistent error in measuring fiber cross-sectional area.
Correction of this issue will be the subject of future work.
6
4.2. Evaluation of the 3D fiber tracking algorithm

Evaluation of the fiber tracking algorithm was performed using

3
synthetic volumes described in Section 3.2.2. To this end, for each
synthetic volume, TM was applied to all images within the stack, the 0
resulting detections were handed to the fiber tracking algorithm, and 4.1 8.8 14.4 17.0
subsequently processed as outlined in the block diagram shown in
Fig. 3. The greatest computational cost was associated with tracking the pixel/davg
highest resolution +45°/−60° dataset, which took approximately 1 h
of wall-clock time on the same desktop workstation described in Section Fig. 12. The average percent error between true and estimated fiber centroid
4.1. The lowest computation cost was 47.5 min and corresponded to the positions versus image resolution.
0°/0° dataset at the lowest resolution, which does not include the time
taken to evaluate the resulting fiber tracks. When executing the fiber smaller than the smallest pre-defined synthetic fiber template. The
tracking algorithm, the greatest computational cost is devoted to the degradation in the quality of TM detections near the boundaries results
gating and scoring steps. For a few datasets analyzed, fiber tracking in decreased quality of fiber tracks, which affects the estimation of fiber
produced short fiber-tracks near the periphery of the synthetic volumes, area, length, and orientation. To lessen this effect, it is recommended
which resulted from clipped measurements during TM. These tracks, that fiber segmentation be performed on slightly larger image stack
caused by missing data, could not be stitched into longer continuous volumes (e.g., volumes that are at least one major fiber diameter larger
tracks, and thus were removed from evaluation of the fiber tracking than needed on all sides) and track data at the boundaries be discarded.
algorithm described in the subsequent discussion. Results from evaluating the second metric are shown in Table 1,
The 3D fiber tracking algorithm was evaluated based on the fol- which gives the mean, µ, and standard deviation, σ, of the percent error
lowing five metrics: (1) the average percent error between true and in the estimated track lengths for each dataset. Overall, the 3D esti-
estimated fiber centroid positions, (2) the average percent error in es- mation algorithm determines the true length of fibers well. As the
timated track length, (3) the average percent error between true and image resolution increases, the mean and standard deviation remains
estimated fiber track cross-sectional areas, (4) the average percent error roughly constant for each dataset. For the +45°/−60° and +45° da-
in estimated fiber track orientation, and (5) the percentage of resulting tasets, the mean and standard deviation values are similar regardless of
tracks that correspond to true fibers. the image resolutions, while for the 0°/0° datasets the error is nearly
Results from evaluating the first metric are shown in Fig. 12. In this zero. As mentioned previously, for the off-axis fibers, the error in the
figure, the average percent error in detected fiber centroid positions estimated track length is related to the degradation of the fiber-detec-
after 3D estimation for the +45°/−60°, ± 45°, and 0°/0° datasets is tion data near the image boundaries. Specifically, as the fiber tracks
plotted on the y-axis using red, green, and solid blue bars, respectively. begin to diverge near the boundary, they are terminated once the next
As in Fig. 9, the position error was normalized by the true fiber dia- estimated position fails to meet the gating and scoring criterion (see
meter, and the standard deviation is represented with error bars. In Section 2.2.1).
addition, for each dataset, Fig. 12 shows the average error in detected Results from evaluating the third metric are shown in Fig. 13, where
fiber centroid positions after TM but prior to tracking, which is re- the average percent error between true and measured fiber cross-sec-
presented using hatched bars. As seen in Fig. 12, on average, the error tional areas for the +45°/−60°, ± 45°, and 0°/0° datasets is plotted on
and standard deviation decrease with increasing image resolution for the y-axis using red, green, and solid blue bars, respectively. Standard
all three datasets. deviation is represented with error bars, and the image resolution is
When comparing results of TM to 3D estimation for the +45°/ given on the x-axis. Examining Fig. 13, there are three separate trends
−60°, ± 45° datasets, the error and standard deviation increases after worth noting. First, for the 8.8, 14.4, and 17 pixels/davg resolutions,
3D fiber estimation was performed, while for the 0°/0° datasets the increase in fiber angle decreases the error, but increases the standard
error and standard deviation decreases or remains constant. The in- deviation. Second, the error in cross-sectional area for the 0°/0° datasets
crease in the position error after 3D estimation is related to degradation decreases monotonically with the increasing resolution. Third, the error
of track quality near the boundaries of the image volume. Specifically, in cross-sectional area for the ± 45° and +45°/−60°datasets decreases
for the off-axis datasets, as individual fibers approach the image between the two lowest resolutions, but remains essentially constant
boundary, their cross-sections are progressively clipped. For such cases, thereafter. For all three data sets, the standard deviation always de-
the TM algorithm attempts to fit smaller templates within the clipped creases with increasing image resolution.
cross-sections, which results in detected fiber centroids that follow the The trends observed in Fig. 13 for all three datasets are directly
volume boundary (rather than the true fiber centroid). This process is related to how the final fiber radius is determined for each track. As
repeated for all fibers exiting the field of view until their fiber cross- described in Section 2.2.6, the fiber radius is determined based on the
sections are fully clipped, or the detected diameters of clipped fibers are

10
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

Table 1
The mean and standard deviation of the percent error in estimated track length.
Resolution

4.1 pixel/davg 8.8 pixel/davg 14.4 pixel/davg 17 pixel/davg

Dataset µ [%] σ [%] µ [%] σ [%] µ [%] σ [%] µ [%] σ [%]

+45˚/−60˚ 6.2 4.5 7.6 4.8 6.6 3.9 5.9 3.8

± 45˚ 7.4 4.6 7.5 5.1 6.7 4.1 6.2 4.1
0˚/0˚ 0.1 1.3 0.0 0.0 0.2 1.7 0.0 0.0

90 which gives µ and σ of the percent error in the estimated track or-
- ± ientations for each dataset. Track angle is calculated using the or-
75 ientation of a best fit straight line through the coordinates of the cor-
responding fiber track. As seen in this table, the 3D estimation
cross-sectional
area % error

algorithm determines true fiber orientation quite well. The greatest

60 error occurred at the second lowest resolution for the +45°/−60° da-
taset, where the fiber orientation error has a mean of 1.04% and a
45 standard deviation of 4.2%. This orientation error corresponds to fibers
being off by 1.5° on average. Similar to the other metrics, the slight
error in estimating fiber orientations for off-axis datasets is related to
30
the degradation of fiber detections near the boundaries of the tomo-
graphic image stacks.
15 Results from evaluating the fifth metric are shown in Fig. 14, where
the percentage of resulting tracks that correspond to true fibers for the
0 +45°/−60°, ± 45°, and 0°/0° datasets is plotted on the y-axis using
4.1 8.8 14.4 17 red, green, and solid blue bars, respectively. The percentage is calcu-
lated based on the number of resulting tracks that correspond to a true
pixel/d avg fiber track, independent of their length, divided by the total number of
true fiber tracks. As seen in Fig. 14, 3D fiber estimation performs quite
Fig. 13. The average percent error between true and estimated fiber cross- well at determining the total number of existing fibers within the syn-
sectional area versus image resolution. thetic image stacks for all datasets analyzed, particularly for the 0°/0°
datasets. For each dataset and independent of resolution, 3D fiber es-
smallest estimated radius along the length of a given track, which is timation determined approximately 96.8% of the existing fiber tracks. A
done to ensure no overlap between adjacent fibers. For the 0°/0° da- majority of the missing tracks for off-axis datasets were quite short, and
tasets, the fiber estimation was noisy at low image resolutions, which were removed during the track deletion step (see Section 2.2.1). In
translated into large errors in the fiber cross-sectional area. As the re- general, the removal of tracks occurred at locations where fiber cross-
solution was increased, the radius estimation was much smoother and sections were clipped during TM, and where tracks spanned across 10
closer to the nominal fiber diameters measured by TM. For datasets images or less.
with off-axis fibers, on average, estimation of fiber radius is a lot less In summary, the 3D fiber tracking algorithm performs well at esti-
noisy since an increased cross-sectional area provides more pixels from mating position, length, orientation and number of existing fibers when
which to determine the boundary, resulting in a better overall estima- fibers angles are small relative to the X-ray µCT scan axis. Estimation of
tion. However, as mentioned before, for off-axis fibers estimation of fiber cross-sectional area is the only metric that improves with in-
fiber radii was affected by degradation of TM data near the image creasing fiber angle. Considering the relative magnitude of error for all
boundaries, which in turn resulted in overall higher scatter in the es- five metrics, 3D estimation performs best when the fiber cross-sections
timated fiber area. Based on the results concluded in Fig. 13, the use of contained at least 9 pixels across any given fiber diameter. At this re-
the smallest estimated radius along any given track can be considered solution, detection error for the position of fiber centroids is on average
overly conservative. Future work will address this issue by defining the ~5% for all fiber orientations analyzed, estimated track length error is
fiber radius based on a mean or median value from the list of estimated on average less than 5%, fiber orientations are estimated on average
fiber radii for a given track and allow for the MAP solution (see Section within 1% of their true values, and more than 95% of the existing fibers
2.2.5) to remove the overlap between neighboring fiber tracks, if such are detected. Comparing the results obtained from evaluation TM and
overlap exists. fiber tracking, 3D estimation is quite sensitive to the quality of detec-
Results from evaluating the fourth metric are shown in Table 2, tions, thus to produce high-quality fiber tracks, high-quality detections

Table 2
The mean and standard deviation of the percent error in orientation of fiber tracks.
Resolution

4.1 pixel/davg 8.8 pixel/davg 14.4 pixel/davg 17 pixel/davg

Dataset µ [%] σ [%] µ [%] σ [%] µ [%] σ [%] µ [%] σ [%]

± 45˚ 0.79 2.08 1.04 4.24 0.05 0.14 0.06 0.30

+45˚/−60˚ 0.53 1.59 0.17 0.83 0.03 0.08 0.04 0.33
0˚/0˚ 0.01 0.01 0.01 0.12 0.01 0.19 0.00 0.00

11
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

101
corresponding to true tracks
to synthetic templates not accounting for artifacts caused by beam
- ± hardening around the perimeter of individual fibers. Despite this issue,
100 comparing Fig. 15a and c, the fiber segmentation algorithm provided a
99 very good approximation of the imaged microstructure. Further ana-
% of fiber tracks

lysis of the segmented, reduced-order data shown in Fig. 15 was beyond

98 the scope of this work, and will be the subject of future efforts.
97
5. Conclusion
96
This study aims to advance the current state-of-the-art in fiber
95 segmentation from X-ray µCT images by (1) developing a new algo-
94 rithm for 3D segmentation of individual fibers in multi-directional FRC
microstructures, and (2) proposing a methodology for evaluating fiber
93 segmentation algorithms using synthetic X-ray µCT image data. The
new fiber segmentation algorithm uses synthetic fiber-templates and
0
template matching to detect, in 2D, centroids, diameters, and orienta-
4.1 8.8 14.4 17.0
tions of individual fibers. The 2D detections are used to track individual
pixel/davg fibers in 3D using a Kalman-filter estimator. To obtain high-quality fiber
tracks and minimize errors associated with spurious or sparse detection
Fig. 14. The percentage of resulting tracks that correspond to true fibers for 3D data, the segmentation algorithm incorporates fiber-track smoothing,
estimation versus image resolution. stitching, and overlap removal. The new segmentation algorithm was
evaluated using synthetic X-ray µCT image data that were generated to
from TM are needed. mimic the scan quality and resolution achieved by existing state-of-the-
art high-resolution X-ray µCT systems. Synthetic images of unidirec-
tional and multi-directional microstructures were used to assess seg-
4.3. 3D fiber segmentation of real X-ray µCT image data mentation quality by evaluating errors associated with detecting cen-
troids and areas of fiber cross-sections, as well as orientations, lengths,
The initial evaluation of the fiber segmentation algorithm was fol- and total number of fibers in a tomographic image stack. Finally, the
lowed by segmentation of fibers in the IM7/8552 carbon/epoxy com- effectiveness of the fiber segmentation algorithm was demonstrated by
posite system imaged as described in Section 3.1. Results of this seg- analyzing real X-ray µCT data containing three distinct nominal fiber
mentation are presented in Fig. 15, where Fig. 15a shows the 3D orientations.
rendition of the actual X-ray CT image stack; Fig. 15b shows the re- The key findings resulting from this study include:
sulting CAD model of the segmented fibers color-coded according to the
fiber orientation angle, α: and Fig. 15c shows the pdf of α. In Fig. 15b-c, • Synthetic X-ray µCT images can be tailored to a variety of image
α is defined as the angle between a given fiber and the x-axis. As de- resolutions and fiber orientations and are a good analog for real
scribed in [16], the direction of each fiber was approximated as a tomographic images.
straight line that intersects the fiber endpoints, and which is projected • The proposed fiber segmentation algorithm is quite accurate in es-
onto the x–y plane. As evident from Fig. 15b and 15c, the majority of timating the number, 3D position, and length of fibers in synthetic
fibers are aligned with the +45°, −60°, and +60° directions, which is image stacks, especially when the image resolution is greater than 9
consistent with the laminate ply layup used. For each distinct ply, in- pixels per average fiber diameter.
dividual fibers deviate to some extent from the nominal fiber directions, • For all image resolutions considered, the segmentation algorithm
which is evident by the texture of colors seen in Fig. 15b and the exhibits a large and systematic error in estimating the true cross-
standard deviation of each major peak in Fig. 15c. In addition to pro- sectional area of individual fibers. This error is related to simplifi-
viding information related to the location and orientation of individual cations used to generate synthetic templates, which do not account
fibers, the fiber segmentation algorithm was used to determine the for beam-hardening artifacts occurring on the perimeter of in-
average area of fiber cross-sections. For the data set shown in Fig. 15, dividual fibers imaged using X-ray CT.
the average fiber area was 17.52 ± 2.27 µm2, which is approximately • After segmentation and analysis of both synthetic and real X-ray CT
17.5% smaller than the nominal area of the IM7 fibers reported by the image data, a requisite improvement to the proposed algorithm was
manufacturer [41]. As described in Section 4.1, this difference is related identified. Specifically, the synthetic fiber-templates used in

0.20
PFD of misorientation

°
0.15

° 0.10
°
z 0.05

°
y 0.00
x -90 -60 -30 0 30 60 90
(deg.)
a) b) c)
Fig. 15. Segmenting of real X-ray CT data: (a) 3D rendition of tomographic reconstruction, (b) CAD rendition of segmented fibers, and (c) pdf of fiber orientation, α.

12
P.J. Creveling, et al. Composites Part A 126 (2019) 105606

template matching algorithm must be modified to include an addi- doi.org/10.1155/2011/154798.

tional phase of pixel-intensities, which for FPRs imaged using X-ray [15] Sencu RM, Yang Z, Wang YC, Withers PJ, Rau C, Parson A, et al. Generation of
micro-scale finite element models from synchrotron x-ray CT images for multi-
CT, corresponds to artifacts caused by beam hardening. This mod- directional carbon fibre reinforced composites. Compos Part A Appl Sci Manuf
ification is expected to dramatically decrease the error associated 2016;91:85–95. https://doi.org/10.1016/j.compositesa.2016.09.010.
with estimating fiber cross-sectional areas, which was determined to [16] Czabaj MW, Riccio ML, Whitacre WW. Numerical reconstruction of graphite/epoxy
composite microstructure based on sub-micron resolution x-ray computed tomo-
be equal to or greater than 33% for the resolutions considered.
•
graphy. Compos Sci Technol 2014;105:174–82. https://doi.org/10.1016/j.
Based on segmentation and analysis of synthetic X-ray CT images, it compscitech.2014.10.017.
was observed that to obtain accurate estimates of fiber orientation [17] Ushizima DM, Bale HA, Bethel EW, Ercius P, Helms BA, Krishnan H, et al. IDEAL:
images across domains, experiments algorithms and learning. JOM
and length, the overall volume of interest should be increased by at 2016;68:2963–72. https://doi.org/10.1007/s11837-016-2098-4.
least one fiber diameter in all directions. This additional volume can [18] Emerson MJ, Jespersen KM, Dahl AB, Conradsen K, Mikkelsen LP. Individual fibre
be used during template matching and fiber tracking but should be segmentation from 3D X-ray computed tomography for characterising the fibre
orientation in unidirectional composite materials. Compos Part A Appl Sci Manuf
discarded as “lower-quality” in the later stages of analysis (e.g.,
2017;97:83–92. https://doi.org/10.1016/j.compositesa.2016.12.028.
during the generation of finite element models) [19] Emerson MJ, Dahl VA, Conradsen K, Mikkelsen LP, Dahl AB. Statistical validation of
individual fibre segmentation from tomograms and microscopy. Compos Sci
6. Data availability Technol 2018;160:208–15. https://doi.org/10.1016/j.compscitech.2018.03.027.
[20] Bricker S, Simmons JP, Przybyla C, Hardie R. Anomaly detection of microstructural
defects in continuous fiber reinforced composites. Int Soc Opt Photonics
Access to the synthetic image data used in this work has been made 2015:94010A. https://doi.org/10.1117/12.2079679.
available using the following web address: https://hdl.handle.net/ [21] Sanei SHR, Barsotti EJ, Leonhardt D, Fertig RS. Characterization, synthetic gen-
eration, and statistical equivalence of composite microstructures. J Compos Mater
11256/988. 2017;51:1817–29. https://doi.org/10.1177/0021998316662133.
[22] Yu H, Guo D, Yan Z, Liu W, Simmons J, Przybyla CP, et al. Unsupervised learning
Declaration of Competing Interest for large-scale fiber detection and tracking in microscopic material images. arXiv
Prepr. arXiv1805.10256; 2018.
[23] Eberhardt CN, Clarke AR. Automated reconstruction of curvilinear fibres from 3D
The authors declared that there is no conflict of interest. datasets acquired by x-ray microtomography. J Microsc 2002;206:41–53. https://
doi.org/10.1046/j.1365-2818.2002.01009.x.
[24] Requena G, Fiedler G, Seiser B, Degischer P, Di Michiel M, Buslaps T. 3D-
Acknowledgments Quantification of the distribution of continuous fibres in unidirectionally reinforced
composites. Compos Part A Appl Sci Manuf 2009;40:152–63. https://doi.org/10.
This work was supported by the NASA Langley Research Center 1016/j.compositesa.2008.10.014.
[25] Viguié J, Latil P, Orgéas L, Dumont PJJ, Rolland du Roscoat S, Bloch JF, et al.
Internal Research and Development Program and the Center Innovation
Finding fibres and their contacts within 3D images of disordered fibrous media.
Fund sponsored by the NASA Space Technology Mission Directorate. Compos Sci Technol 2013;89:202–10. https://doi.org/10.1016/j.compscitech.
2013.09.023.
References [26] Martín-Herrero J, Germain C. Microstructure reconstruction of fibrous C/C com-
posites from X-ray microtomography. Carbon N Y 2007;45:1242–53. https://doi.
org/10.1016/j.carbon.2007.01.021.
[1] Kalfon-Cohen E, Kopp R, Furtado C, Ni X, Arteiro A, Borstnar G, et al. Synergetic [27] Moon TK. The expectation-maximization algorithm. IEEE Sig Proc Mag
effects of thin plies and aligned carbon nanotube interlaminar reinforcement in 1996:47–60.
composite laminates. Compos Sci Technol 2018;166:160–8. https://doi.org/10. [28] Blackman SS, Popoli RF. Design and Analysis of Modern Tracking Systems. Artech
1016/j.compscitech.2018.01.007. House; 1999.
[2] Maire E, Withers PJ. Quantitative x-ray tomography. Int Mater Rev 2014;59:1–43. [29] Kalman RE. A new approach to linear filtering and prediction problems. J Basic Eng
https://doi.org/10.1179/1743280413Y.0000000023. 1960;82(1):35–45. https://doi.org/10.1115/1.3662552.
[3] Wright P, Fu X, Sinclair I, Spearing SM. Ultra high resolution computed tomography [30] Munkres J. Algorithm for the assignment and transportation problems. J Soc Ind
of damage in notched carbon fiber—epoxy composites. J Compos Mater Appl Math 1957;5:32–8. https://doi.org/10.1137/010500.
2008;42:1993–2002. https://doi.org/10.1177/0021998308092211. [31] Bar-Shalom Y, Li Y, Kirubarajan XR. Estimation with Application to Tracking and
[4] Garcea SC, Wang Y, Withers PJ. X-ray computed tomography of polymer compo- Navigation. John Wiley and Sons; 2001.
sites. Compos Sci Technol 2018;156:305–19. https://doi.org/10.1016/j. [32] Morefield CL. Application of 0–1 integer programming to multitarget tracking
compscitech.2017.10.023. problems. IEEE Trans Automat Contr 1977;22:302–12. https://doi.org/10.1109/
[5] Scott AE, Mavrogordato M, Wright P, Sinclair I, Spearing SM. In situ fibre fracture TAC.1977.1101500.
measurement in carbon-epoxy laminates using high resolution computed tomo- [33] Karp RM. Reducibility among combinatorial problems. Springer; 1972. p. 85–103.
graphy. Compos Sci Technol 2011;71:1471–7. https://doi.org/10.1016/j. [34] Poore AB, Rijavec N, Liggins M, Vannicola VC. Data association problems posed as
compscitech.2011.06.004. multidimensional assignment problems: problem formulation. Opt Eng Phot Aerosp
[6] Garcea SC, Sinclair I, Spearing SM, Withers PJ. Mapping fibre failure in situ in Sens 1993;1954:552–64. https://doi.org/10.1117/12.157786.
carbon fibre reinforced polymers by fast synchrotron x-ray computed tomography. [35] Deb S, Yeddanapudi M, Pattipaii K, Bar-Shalom Y. A generalized S-D assignment
Compos Sci Technol 2017;149:81–9. https://doi.org/10.1016/j.compscitech.2017. algorithm for multisensor-multitarget state estimation. IEEE Trans Aerosp Electron
06.006. Syst 1997;33:523–38. https://doi.org/10.1109/7.575891.
[7] Clay SB, Knoth PM. Experimental results of quasi-static testing for calibration and [36] Pattipati KR, Deb S, Bar-Shalom Y, Washburn RB. A new relaxation algorithm and
validation of composite progressive damage analysis methods. J Compos Mater passive sensor data association. IEEE Trans Automat Contr 1992;37:198–213.
2017;51:1333–53. https://doi.org/10.1177/0021998316658539. https://doi.org/10.1109/9.121621.
[8] Clay SB, Knoth PM. Experimental results of fatigue testing for calibration and va- [37] Castañón G, Finn L. Multi-target tracklet stitching through network flows. 2011
lidation of composite progressive damage analysis methods. J Compos Mater Aerosp. Conf., IEEE; 2011. https://10.1109/AERO.2011.5747436.
2017;51:2083–100. https://doi.org/10.1177/0021998316670132. [38] Mori S, Chong C. Performance analysis of graph-based track stitching. 16th Int.
[9] Bale HA, Haboub A, Macdowell AA, Nasiatka JR, Parkinson DY, Cox BN, et al. Real- Conf. Inf. Fusion, IEEE 2013.
time quantitative imaging of failure events in materials under load at temperatures [39] van der Merwe LJ, de Villiers JP. Track-stitching using graphical models and
above 1600°C. Nat Mater 2013;12:40–6. https://doi.org/10.1038/nmat3497. message passIng. Int. Conf. Inf. Fusion, 2013, IEEE; 2013.
[10] Larson NM, Zok FW. In-situ 3D visualization of composite microstructure during [40] Julier SJ, LaViola JJ. On Kalman filtering with nonlinear equality constraints.
polymer-to-ceramic conversion. Acta Mater 2018;144:579–89. https://doi.org/10. Signal Process IEEE Trans 2007;55:2774–84. https://doi.org/10.1109/TSP.2007.
1016/j.actamat.2017.10.054. 893949.
[11] Larson NM, Cuellar C, Zok FW. X-ray computed tomography of microstructure [41] Hexcel. HexTow ® IM7 Carbon Fiber Data Sheet, 2018; 1–2. http://www.hexcel.
evolution during matrix impregnation and curing in unidirectional fiber beds. com.
Compos Part A Appl Sci Manuf 2019;117:243–59. https://doi.org/10.1016/j. [42] Hexcel. HexTow ® AS4 Carbon Fiber Data. Sheet, 2018; 1–2 . http://www.hexcel.
compositesa.2018.11.021. com.
[12] Ferguson JC, Panerai F, Lachaud J, Mansour NN. Theoretical study on the micro- [43] Wongsto A, Li S. Micromechanical FE analysis of UD fibre-reinforced composites
scale oxidation of resin-infused carbon ablators. Carbon N Y 2017;121:552–62. with fibres distributed at random over the transverse cross-section. Compos Part A
https://doi.org/10.1016/j.carbon.2017.06.013. Appl Sci Manuf 2005. https://doi.org/10.1016/j.compositesa.2005.01.010.
[13] Ferguson JC, Panerai F, Borner A, Mansour NN. PuMA: the porous microstructure [44] Schindelin J, Arganda-Carreras I, Frise E, Kaynig V, Longair M, Pietzsch T, et al. Fiji:
analysis software. Soft X 2018;7:81–7. https://doi.org/10.1016/j.softx.2018.03. an open source platform for biological image analysis. Nat Meth 2012;9:676–82.
001. https://doi.org/10.1038/nmeth.2019.Fiji.
[14] Tuegel EJ, Ingraffea AR, Eason TG, Spottswood SM. Reengineering aircraft struc- [45] Barrett JF, Keat N. Artifacts in CT: recognition and avoidance. RadioGraphics
tural life prediction using a digital twin. Int J Aerosp Eng 2011;2011:1–14. https:// 2004;24:1679–91. https://doi.org/10.1148/rg.246045065.

13
Other documents randomly have
different content
Padini. He elected that the supper should take place in his own room
at the hotel. You can guess why."

Jack began to see matters more clearly now. The task allotted to
himself was plain and simple. He would have preferred something
more in the way of adventure; but, after all, somebody must do the
ordinary work. He managed to see Anstruther in the afternoon, and
intimated to him that he was dining in Panton Square that night.
Anstruther replied that he was glad to hear it; possibly, Jack
thought, because there would be an ear-witness to prove the music
in the study.

It was nearly eight o'clock when Jack strolled into the drawing-room
of Panton Square, and found Claire alone there. He deemed it
prudent not to tell her too much of what had taken place the last
few hours; indeed, he was more concerned to hear the latest
information about Serena.

"I have not seen much of her to-day," Claire said. "I do not know
what to make of her at all. Last night late she came into my
bedroom, and we had a long talk about her boy. It is a very strange
thing, Jack, that only this morning a man arrived to see my
guardian--a man who seemed to be annoyed at Mr. Anstruther's
refusal to pay him a sum of money. I happened to overhear a few
words as they parted. The stranger declared that if he did not have
something definite by Saturday, 'he would send the kid back.' I
should have thought nothing of this unless I had heard Serena's
story last night, but, taken in conjunction with what she said, I
shouldn't wonder if the man in question had not the custody of the
poor woman's child."

"This is interesting," Jack said. "Did you take any particular note of
the man's appearance?"

Claire replied that she had not failed to do so. But she had not
followed him, though her suspicions were aroused. Jack debated the
thing in his mind for a moment before he spoke again.

"We know perfectly well," he said, "that Anstruther is terribly

pressed for ready money. He is certain not to send that check, and it
is equally certain that the man will call again for the cash on
Saturday morning. It will be an easy matter to get Bates to lend me
a plain clothes man and follow the fellow wherever he goes. But you
must understand----"

What more Jack would have said was prevented by the entrance of
Anstruther, closely followed by the announcement of dinner. It was
not a gay meal, for the host was moody and depressed. He talked
brilliantly at times, then lapsed into a reverie, and appeared not to
hear when spoken to. Claire rose presently with a sigh of relief, glad
to get away from the gloom of the dining-room and its depressing
atmosphere. Anstruther smoked half a cigarette, and then threw the
end down impatiently.

"I must really get you to excuse me," he said. "But my head is so
bad that I can hardly hold it up. I am afraid that even my music will
fail to soothe me to-night."

Jack murmured something in the way of polite sympathy. He was

glad of the opportunity to be able to escape to the drawing-room,
where he sat for a long time discussing the situation with Claire. It
was pleasant and soothing to sit there with his arm about her and
her head lovingly upon his shoulder; but, happy as they were, they
could not altogether shake off the feeling of impending evil. All this
time the music of the violin floated mournfully from the study.
Eleven o'clock struck, and still the melody went on. Claire roused
herself a little presently, and a look of pleased interest crossed her
pretty face.

"What a delightful little composition," she said . "I have never heard
that before. I am quite sure that is original."
"Listen very carefully," Jack said. "I want you to impress that piece
of music on your mind."

The piece was finished at length, and then repeated once more. As
the last strains died away, Claire rose from her comfortable seat and
crossed over to the piano. Very quietly, yet quite correctly, she went
through the whole composition.

"I am glad it has so impressed you," Jack said. "You will, perhaps, be
surprised to hear that Anstruther has never heard that piece of
music in his life, and that it was composed by Padini, who has never
played it to anybody till last night, when he performed it for Rigby's
benefit. Not only this, but he gave Dick Rigby the original manuscript
to get published for him. I know this is only a small matter, but these
small matters will make a mountain of evidence against Anstruther
when the time comes."

"It is very extraordinary," Claire murmured, "to think that that music
should sound so charming and natural, when we know that all the
time the player is a mile or two away. You are sure that my guardian
is not in his study, Jack?"

Jack was sure enough on that point. It was a few moments later
that Serena came quietly into the room with a request that Mr.
Masefield would go to the telephone, as some one desired to speak
to him on pressing business. Jack rose with alacrity.

"I shall soon be able to prove to you that Anstruther is a long way
off, or I am very much mistaken," he said. "Very well, Serena, I will
come down at once."

The voice at the other end of the telephone inquired cautiously if

that were Mr. Masefield. Jack replied that it was, but even then the
questioner did not appear to be satisfied.

"I think I recognize your voice," he said, "but one has to be very
careful in sending messages to Panton Square. How goes the music?
Anything original to-night?"

"One piece," Jack smiled. "I know what you mean, and I don't mind
making you a small bet that you are Inspector Bates."

The voice at the other end of the telephone chuckled.

"You have got it quite right, Mr. Masefield," he said. "I am Bates sure
enough. And you needn't worry about going down-stairs to see
whether or not Anstruther is playing at Paganini, because he isn't on
the premises at all."

"Where are you speaking from?" Jack asked.

Bates replied that he was speaking from a public call office in the
neighborhood of Mansion House Station. All he wanted to do was to
make sure that Jack was still in Panton Square, and now that his
mind was easy on this score, he could devote himself to the serious
business of the evening. Anstruther had just been shadowed outside
the Mansion House Station, where he was apparently waiting for the
substitute so kindly provided for him by "Simple Charlie."

The message ceased here, and the connection was cut off. Jack
would have been just a little surprised if he had seen the
transmogrified Bates who had been speaking to him over the line.
The inspector crossed the road and disappeared into the shadow.
Anstruther stood there, glancing impatiently up and down the road
as if waiting for somebody that was late. A figure slouched up to
him, and a hoarse voice whispered in his ear:

"Party of the name of Maggs," he said in his gin and fog voice. "Pal
of 'Simple Charlie.' Old Charlie couldn't get away to-night, so he sent
me instead. Don't you be disappointed, guv' nor; you will find me
just as clever with them bits of steel as Charles himself. Bit of
burglary, ain't it?"

Anstruther nodded curtly.

"We had better walk along," he said. "I suppose your friend
explained to you that this little job will put twenty pounds in your
pocket? It is a mere matter of opening a safe. The getting into the
premises is perfectly simple, because I have come provided with the
keys. You know the City and Provincial Bank?"

The other man grinned, and remarked that banks generally were a
bit above his form. Anstruther smiled as he reflected that he had the
keys of the bank premises proper in his pocket, so that there would
be no great difficulty in getting into the counting house, and from
there to Carrington's private office. As to the night watchmen--that
was another matter altogether. In the face of recent happenings,
they would be more alert than they had been in the past; but, at the
same time, their attention would be bestowed more upon the cellars
than the office.

The road was entirely deserted now, as Anstruther crossed the

street and gently turned the key in the outer door. A moment later,
and the pair were in Carrington's private office. They could afford to
turn the lights up, for the iron shutters outside made a perfect
screen. In one corner of the room stood the safe upon which the
man who called himself Maggs was intended to operate. Anstruther
pointed at it impatiently.

"Get to work at once," he said. "There is something inside that I

must take away to-night."

"A fine set of Cellini plate, I presume?" Maggs said, in an entirely

different voice. "No, you don't, Mr. Anstruther. If you put your hand
in your hip pocket, I'll blow your brains out. I have the advantage of
you here, and I am going to keep it."

"Who the deuce are you?" Anstruther stammered. His hands had
fallen to his side, and his face was pale and ghastly. "Who are you?"

The so-called burglar snatched away his wig and ragged beard, and
with a handkerchief changed the aspect of his face.
"I am Inspector Bates," he said. "Very much at your service."

CHAPTER XLV.

CAUGHT!

Bates had laid his plans very carefully and very well indeed. In many
respects Rigby had got the best of the detective, but this was as
much due to circumstances as anything else. Still, when it came to
the technical side of the case, Rigby was no match for the inspector.
It was nearly nine o'clock before Bates called at Carrington's rooms
and asked to see the latter. There was no occasion yet for Bates to
assume the very effective disguise with which he was to trick
Anstruther. There would be plenty of time for that. Carrington was
just finishing his dinner--so his man said. He was not very well, and
did not care to see anybody. But Bates put the man aside in his own
easy way, and walked into the dining-room without the trouble of
announcing himself.

That Carrington was suffering from some mental and physical

excitement was perfectly plain. His face was ghastly pale, his eyes
were bloodshot, and there was a twitching of his lips which told a
plain tale to an experienced officer like Bates. Carrington scowled,
and demanded the meaning of this unwarrantable intrusion.

"I don't think you will find it unwarrantable when you have heard me
to the finish," Bates said. "Nor will it pay you to take this tone with
me. I am an inspector from Scotland Yard, and unless you answer
my questions freely, I shall have to put them in a more disagreeable
form."

Carrington changed his note altogether. His face became still more
pallid. He motioned Bates to a chair. He would have found it hard to
have spoken just then. Bates waited a moment to give the other
time to recover. Carrington at length found words to ask Bates what
his business was with him.

"It is with regard to your affair at the bank," the inspector explained.
"You may not be aware of the fact, but the case has been placed in
my hands by my superiors."

"Oh, you are alluding to the burglary," Carrington said.

"We will call it a burglary for the present," Bates replied, with a
significance that there was no mistaking. "I have gone into the
matter carefully, and I have come to the conclusion that there was
no burglary at all."

Carrington jumped to his feet with a well-simulated air of

indignation. He advanced towards Bates threateningly.

"You insolent scoundrel!" he cried. "What do you mean? Do you

know you are dealing with a gentleman and man of honor?"

"Softly, softly," Bates replied. "I think we had better understand one
another. I have in my possession at the present moment a warrant
for your arrest for fraud and embezzlement, relating to certain
jewels and other valuables deposited in your keeping by various
clients. It is in my power to execute that warrant at once. The case
is much too serious a one for bail, and it is for you to say whether
you will remain for the present in your comfortable quarters, or pass,
at any rate, the next two months in jail."

Carrington made no further show of fight. He collapsed into his chair,

and wiped his wet forehead distractedly.
"You don't mean that," he groaned. "There must be some terrible
mistake here. Why, all the evidences pointed to an ingenious and
daring burglary. The night watchmen were drugged, as you know,
and the thieves employed dynamite to blow up the safes. No one
regrets the loss of all those valuables more than I do, but even
banks are not secure against the modern burglar. Those safes were
crammed full of valuables, as I could easily prove."

"They were," Bates corrected. "But I am in a position to prove a few

things, too. You would give a great deal, I suppose, to know where
those valuables are?"

Carrington replied to the effect that he would give half his fortune
for the desired information. Bates smiled.

"You need not worry about it," he said. "I have a list in my pocket of
the big pawnbrokers in London where most of the goods were
pledged. In three cases the pawnbrokers in question are in a
position to swear to the identity of the man who handled the jewels.
You would not, of course, mind meeting these people?"

But Carrington had no reply. He looked so helplessly at Bates that

the latter could not but feel sorry for him. "I am afraid the game is
up, sir," he said. "My investigations of this case prove most
conclusively that you are at the bottom of the whole thing. We know
perfectly well that recent speculations of yours have brought about a
financial crisis in your bank. In your desperate need, you realized the
securities which certain clients had left in your hands. It was only
when Lady Barmouth called for her gems that the situation became
acute. But that will form the basis of another charge."

"But that was all a mistake," Carrington gurgled eagerly. "I sent Lady
Barmouth her gems, but they proved to be those belonging to
somebody else. I assure you that was quite an error."

Bates shrugged his shoulders impatiently. He was getting annoyed

with this, man, who refused to follow his lead. "We know all about
that ingenious fraud," he said. "We are quite aware of that clever
business of the paste gems, for which you gave £200 at Clerkenwell.
You paid for that rubbish with Bank of England notes marked with
the stamp of your establishment. It was a very happy idea of yours
and Anstruther's."

Carrington groaned feebly; he began to fear the very worst.

"You seem to know everything," he said. "Perhaps you can tell me

the story of the burglary?"

"I am coming to that presently," Bates said coolly. "Now you were at
your wits' ends to know what to do. You knew perfectly well that
many of your clients would require their jewels for Lady Barmouth's
dance. They were not forthcoming, for the simple reason that they
had been pledged elsewhere. You had not the necessary cunning to
devise some scheme to shift the blame from your shoulders, so you
called in your friend Anstruther. It was he who hit upon the idea of
the burglary. It was you who placed temptation in the way of the
night watchmen through the medium of a couple of bottles of
drugged port wine. After that the rest was easy. You had only to
enter the bank with your own keys----"

"Stop a moment," Carrington cried eagerly. "You seem to forget that

even I cannot enter the vaults of the bank without duplicate keys in
the possession of various cashiers."

"Now, listen to me," Bates said impressively. "This discussion is

absolutely irregular. It is my plain duty to arrest you at once and
convey you to Bow Street. But if you help me, I may be in the
position later on to do you a service. We know precisely how
Anstruther used the dynamite; we know precisely what happened in
the vaults, and how most of the few valuables that remained were
conveyed to your own private safe. More than that, we are perfectly
well aware what fee Anstruther demanded for his trouble. Need I go
into the matter of that service of Cellini plate?"
Carrington threw up his hands with a gesture of despair. He was
crushed and beaten to the ground by the tremendous weight of
evidence with which Bates was overwhelming him.

"It is no use fighting any longer," he said. "I confess to everything. I

shall plead guilty, and afford you every information in my power. Do
you want me to come along with you now?"

On the whole, Bates rather thought not. He had effected his

purpose, and sooner or later Carrington would have to become his
prisoner. He knew that the latter would speak freely enough, like the
craven coward that he was; but there was Anstruther to be thought
of. Bates rose to leave.

"You can remain where you are for the present," he said. "But if you
will take my advice, you will make no attempt to escape--you are too
carefully watched for that; and now, good-night."

Bates went off in the direction of the City feeling that the last hour
had not been wasted. On the strength of recent information, he
would have felt justified in arresting Anstruther also. But he had a
wholesome admiration for that individual, and the more evidence
secured against him the better. Therefore it was that Bates was
about to carry out the latter part of the programme, in which he was
to play the part of substitute for "Simple Charlie." The programme
had been easily arranged. There had been no difficulty in persuading
the burglar to write the desired letter to Anstruther, and Bates had
made up his mind from the first that the mythical Maggs should be
none other than himself. From first to last the thing worked
admirably. Anstruther was utterly deceived by the detective's
admirable disguise, which he had assumed after leaving Carrington,
and had fallen headlong into the trap.

Therefore it was that the two men stood facing one another in
Carrington's office. Anstruther white and furious, Bates coolly
contemptuous, with a revolver in his hand.
"What have you to say for yourself?" Bates asked. "Have you any
reason to show why I should not take you straight to Bow Street on
the charge of burglary?"

Anstruther was fighting hard to regain possession of himself. Bates

could not but admire the marvelous courage of the man.
Anstruther's laugh had something quite genuine about it.

"We are making a great fuss over a little thing," he said. "I came
here because Mr. Carrington was not well enough to accompany me.
There are certain things of mine in my friend's private safe here, and
unfortunately he has lost the key. It was imperative that I should
have my property to-night, and that will, perhaps, explain my
presence here. Does that satisfy you?"

"I should be easily satisfied if it did," Bates said coolly. "I should like
to know, for instance, why you require the assistance of a
professional burglar. I know perfectly well that you called in the
assistance of 'Simple Charlie,' but I was in a position to force that
individual's hand--hence my appearance in his place."

"Really, Mr. Bates," Anstruther smiled. "I had expected better things
from you. You are perfectly well aware of the fact that I am
acquainted with half the thieves in London. It was no use asking any
safe-maker in London to try to pick that lock, because it happens to
be a French make. In such awkward circumstances as this it is no
new thing to call in a cracksman when things are wanted in a hurry."

"I am afraid that won't do," Bates said. "You had plenty of time to
call in legitimate assistance, whereas so recently as last night you
visited 'Simple Charlie' and left a note for him."

Anstruther smiled politely. He was perfectly cool and collected now--

a match for any detective in the force.

"We can settle the matter in two minutes," he said. "All you have to
do is to call in one of your men from outside and send a note to
Carrington, who will reply to the effect that I am here with his full
knowledge and consent."

"Can't do it," Bates said curtly. "I have no man to send. As a matter
of fact, I am alone in this business."

Anstruther bent down his head to conceal a smile. There was

something devilish in the cunning ferocity of his eyes. He had
discovered an important fact, and Bates did not seem to understand
for the moment what he had given away. He felt quite sure that he
had matters in his own hands now. He strolled slowly round the
table, and proceeded to examine carefully the lock of the safe.

"Do you really think you could open this?" he asked. "If you could I
should have no difficulty in proving to you----" Anstruther broke off
suddenly; his left foot shot out dexterously, and Bates came half
stumbling on his knees. Like lightning Anstruther grabbed for the
revolver. He had Bates's wrist in a grip of steel, forcing his hand back
till the fingers were bound to relax their grip on the weapon. A
moment later the revolver was kicked away, and the two men were
struggling desperately on the floor.

There was no mistaking the look on Anstruther's face. He was going

to murder Bates if he could. It would never do for any living soul to
know that he was here to-night. Once Bates's mouth was silenced
forever, he could hurry back to Panton Square, and there prove such
an alibi as would hold good in any legal court in the world. All these
things passed through that wily brain as his hands clutched closer at
Bates's throat.

It was touch and go with the latter. The only thing he could do was
to fight for his breath, and husband his strength for a final effort
later on. He looked straight into the gleaming eyeballs of his
assailant now, but he could not see the faintest suggestion of pity
there. The world began to dance before his eyes; a thousand stars
seemed to be bursting from the dark sky; then came along the
corridor the echo of fast-approaching footsteps.

"Curse it," Anstruther muttered. "Another moment, and I should

have been safe. Take that, you hound."

With one final blow he jumped to his feet, and, sprinting across the
office floor, darted into the shadow of the night.

CHAPTER XLVI.

THE MUSIC STOPS.

Bates was sitting up in bed nursing an aching head, and plotting out
schemes whereby he could best retrieve the disaster of the previous
night. It was fortunate for the inspector that one of Carrington's
night watchmen should have heard something of the disturbance on
the previous night, and come hotfoot to his assistance. There was no
great damage done beyond a bruised face and a general shock to
the system. Bates felt all the better for a good night's rest, and was
quite ready now to carry on the campaign against his powerful foe.
It was some time in the afternoon before Jack Masefield put in an
appearance at Bates's lodgings, having been summoned there by a
special messenger. Jack smiled as he noticed Bates's somewhat
dilapidated condition.

"What's the matter?" he asked. "You do not seem to have been as

successful as you might--I mean over last night's business. Was the
thing a failure, or were you satisfied?"
Bates explained that up to now the battle was a drawn one. He had
a feeling that Jack would be able to help him, and that was why he
asked him to call this afternoon.

"I am not in the least dissatisfied with my last night's work," he

explained. "In the first place, we have Carrington absolutely at our
mercy. I let him know what we have discovered, and he will do
anything for us that we desire. After that, I played the part of the
mythical Maggs, and in due course disclosed myself to Mr.
Anstruther. Perhaps I was a little too confident; anyway, I gave him a
chance to murder me, and he responded to the opportunity with
absolute enthusiasm. But for the opportune arrival of the night
watchman, Scotland Yard would have lost one of its most
distinguished ornaments. It was a very near thing, I assure you."

"But what could he possibly gain by that?" Jack asked.

"Well, you see, I had let him know that I was quite alone in the
business," said Bates. "At the same time, he was not aware that my
information was so complete. If he could murder me and get safe
home without being detected, he was in a position to prove an
absolute alibi. Of course, I did not dream that I was running any risk
of my life--but that is not the point. You will remember my
suggesting to you yesterday the advisability of you dining in Panton
Square last night. I suppose that was all right?"

Jack replied that he had followed Bates's instructions out implicitly.

He had done all he could in that way.

"Very well, then. You see what I am driving at. I take it for granted
that Anstruther's mysterious musical friend was much in evidence
last night. I have no doubt that Miss Helmsley and yourself listened
with rapt attention to the music in the study."

"We had every opportunity of doing so," Jack said.

"That is precisely what I expected. Anstruther must have left the
house a little after ten o'clock, and I don't see how it was possible
for him to return much before half-past twelve. I suppose you didn't
happen to see him when he came in?"

"Indeed I did," Jack said. "It was quite half-past twelve when I was
leaving the house. The music was still in progress, but when I
slipped out of the front door, Anstruther was rapidly approaching the
house running across the lawn. He seemed very much annoyed and
put out when he saw me, and muttered something to the effect that
he had heard somebody trying the front door. I understood him to
say that he had not been out all the evening, but that was all
nonsense. I could see by his boots that he had been walking some
considerable distance. Of course, you see what the dodge is: he
does not leave the house by the door, but by the French window
leading from the study to the garden. This window he leaves
unfastened, so that he can get back at any time without a soul being
any the wiser. Of course, there was always a chance of somebody
finding the window unlatched, but that is a small matter."

"Is the window always left open?" Bates asked thoughtfully.

Jack replied that he thought so. Bates smiled with the air of a man
who is perfectly well satisfied.

"I am going to get up presently," he said. "After I have had a bath

and some tea, I shall be quite fit for duty again. I want you to find
some pretext for calling at Anstruther's just after dinner, because I
may need your assistance."

"What are you going to do?" Jack asked eagerly.

"Well, in the first place I am going to arrest Mr. Anstruther," Bates

replied. "In the second instance, I have another little scheme, which
we need not discuss now. I want you to go as far as Mr. Rigby's
chambers and get him to keep an eye on Padini, and see that last
night's programme is repeated, if possible. This is rather an
important thing. I think I can trust Mr. Rigby to manage it."

Jack went off obediently enough, and subsequently ran Rigby to

earth at the offices of the Planet. The latter seemed delighted at the
turn which affairs were taking. He began to see now that he would
be able to carry out for his paper the series of sensational articles
required by the proprietor. "We shall have a splendid scoop," he said.
"Indeed, one might almost make a three-volume novel out of it. I
am only too sorry that I can't be at Anstruther's to-night and witness
the arrest. I shall leave you to supply all the graphic details. I can
easily manage the Padini business this evening by writing to the
fellow that I have a check to pay over and shall call at his rooms late
to-night. I am sure to find him there. He is very hard up, and the
money is certain to fetch him."

"There are other things connected with this business," Jack said,
"which puzzle me. For instance, there is that affair of the mysterious
Mr. Ferris, whose acquaintance I made at the Great Metropolitan
Hotel. I am quite sure, also, that Seymour has some deep design on
hand. You may be absolutely certain that that business of the crystal
ball played off on Anstruther at Lady Barmouth's dance the other
night was not mere flummery."

Rigby was of the same opinion. He was anxious to know if anything

had been yet done in the matter of Carrington's private safe and the
service of Cellini plate which Anstruther had coolly appropriated for
himself. But on this point Jack had no information to offer. He did
not doubt that the whole thing would be explained in a few hours
now. He killed the day as best he could, and after dinner turned his
steps in the direction of Panton Square. Mr. Anstruther and Miss
Helmsley had practically finished, Serena explained, but they had not
yet left the dining-room. Anstruther raised his brows significantly as
Jack entered the dining-room, but his manner was polite and cordial
enough as he invited the visitor to a seat and a glass of claret. He
did not look in the least perturbed or put out; on the contrary, Jack
had seldom seen him so easy and self-possessed. His neuralgia was
quite gone. He had charmed it away as usual, he said with the
soothing aid of music.

"How is it you never bring your violin up to the drawing-room?"

Claire asked. "We hardly ever have any duets together."

"After next week," Anstruther promised. "Really, I am a great deal

more busy than I appear to be, and I feel it quite easy to play and
think at the same time."

Jack glanced across the table significantly at Claire, and she seemed
to divine what he was thinking about.

"I thought I knew most of your music," she said , "but there was
one little item last night that took my fancy immensely. I feel quite
sure that you composed it yourself."

Anstruther disclaimed any such gift. Fond as he was of his violin, it

had never occurred to him to try his hand at original composition.

"All the same, I really must get it," Claire persisted. "I am sorry that
you do not recall the piece at all. If you will come into the drawing-
room with me, and can spare me a few minutes, I will strum the
piece over to you. It so fascinated me that I committed it to
memory. Do come along for a moment."

Anstruther laughed, as Jack thought, rather uneasily. He tried

skilfully enough to divert the conversation into another channel, but
Claire's enthusiasm refused to be baffled. Anstruther's face darkened
for a moment, and there was a look in his eyes that boded ill to
somebody. He rose and walked across towards the door, and up the
stairs in the direction of the drawing-room..

"Very well, if you must," he said. "I can give you ten minutes. I dare
say it is some silly trifle that I have heard somewhere without
recognizing its source."
Claire seated herself at the piano, and played the little piece off with
both brilliancy and feeling. As a matter of fact, she had been
practicing it several times during the afternoon until she had it
absolutely correct. The slow, mournful chords died away at length,
and then Claire turned to her guardian with a smile.

"That is it," she said . "That is the little piece that so fascinated me
last night. Surely you can tell me the name of it and where it came
from?"

The question was apparently simple enough, but Anstruther

appeared to be absolutely incapable of answering it.

"Do you mean to say you could forget a thing like that?" Claire
protested. "It seems to me impossible."

"Perhaps it made less impression upon me than it did you,"

Anstruther muttered. "I haven't the slightest recollection of playing it
myself. In fact----"

Anstruther broke off in absolute confusion. The incident, trivial as it

seemed, had upset him altogether. He was about to betray himself
by saying that he had never heard the piece before, and that it had
no place amongst his music; but he pulled himself up just in time.
He bitterly blamed Padini's carelessness. It was no part of the
programme for his double to give him anything but pieces of music
with which he was absolutely familiar. What he might have said and
done was frustrated by the appearance of Serena, who announced
that a gentleman down-stairs desired to see Mr. Anstruther.

Jack felt his pulses beating a little faster, for he had had no reason to
inquire who the stranger was. Serena's eyes were demure and
downcast as usual as she replied to Anstruther's question that the
gentleman down-stairs was none other than Inspector Bates, of
Scotland Yard. Only just for an instant did Anstruther falter and turn
pale, then he was absolutely himself again. He almost wished now
that he had not waited so long. He had his ingenious alibi, it was
true, but even that might fail. There were so many meshes in the
nets of Scotland Yard. In a calm, even voice he ordered Serena to
show the stranger up-stairs. Bates came at length, a little pallid and
bruised, but otherwise little worse for his last night's adventure.

"And what might be your business with me, inspector?" Anstruther

asked. "It is some time since I had the pleasure of meeting you. Will
you please take a seat?"

"I do not see the necessity," Bates responded. "As my business is

private, perhaps you will be good enough to follow me to your study.
I will speak if you like, but----"

"You may say anything you please," Anstruther said defiantly.

"Then I arrest you on a warrant, charging you with attempted

burglary last night," Bates said pithily. "You were on the premises
belonging to the City and Provincial Bank with a felonious intent of
breaking into a safe between the hours of eleven and half-past
twelve. Need I say any more?"

"Amazing," Anstruther laughed. "Fortunately I have my witnesses at

hand to prove that I was not off these premises during the hours
you mentioned. As a matter of fact, I was in my study playing my
violin all the time."

"Sounds ingenious," Bates muttered, "but in these days of clever

mechanical contrivances--by the way, is not some one playing the
violin down-stairs now?"

Despite his command of himself, a furious curse broke from

Anstruther's lips. For even as Bates spoke, there came sounds of
liquid melody from the study. Not only was this so, but, furthermore,
the piece in question was precisely the same as the one that Claire
had just been playing over to her guardian. The girl rose to her feet,
and looked across at Jack significantly. Bates smiled in the manner
of one who has solved a great problem.
"Really, a most remarkable coincidence," he said. "I am afraid this
rather spoils the simple beauty of your alibi, Mr. Anstruther; unless,
perhaps, you have some friend who entertains your household at
such times as business calls you elsewhere. But let us go down-
stairs and see for ourselves."

"No, no," Anstruther cried furiously. "You shall not do it. You shall
not interfere. I'll kill you first."

"Come along," Bates responded. "Come with me and witness the

solving of the mysterious problem."

CHAPTER XLVII.

"A WOMAN SCORNED."

It was plainly evident that Bates believed in his ability to solve the
problem. Anstruther had quite thrown the mask off by this time, and
stood glaring vindictively at the inspector. It was absolutely
maddening to a man of his ability to be caught in a sorry trap like
this. One of the strongest points in Anstruther's schemes was the
fact that hitherto he had always been on the side of the police. He
had been regarded as one of them, so to speak, so that many of his
ingenious plots had been guided solely by the action of the
authorities. It had never once occurred to him that he might have
been an object of suspicion at Scotland Yard.

"You might just as well take it quietly," Bates said. "We know the
whole thing from start to finish. It will go a great deal easier with
you if you give us all the information that lies in your power and
save us trouble."

"That is the usual course, I believe," Anstruther sneered. "But you

have a different man to deal with in me. I am quite at a loss to
understand what you are doing here at all."

Bates shrugged his shoulders, and walked in the direction of the

door. He had no difficulty in seeing that Anstruther had made up his
mind to see this thing through to the bitter end. Therefore, it was
quite useless to try and get him to see matters in a reasonable light.
Anstruther stood there, white, silent, and furious, whilst all the time
the amazing music was going on in the study.

Mysterious as the whole thing appeared to be, there was almost an

element of farce in it. Here was the very man who relied upon his
devotion to his violin to save him in the hour of danger, actually
listening, so to speak, to his own performance. He had little doubt
what Bates meant to do, for the latter was already half-way down
the stairs on his way to the study. With a sudden impulse Anstruther
followed. He passed Bates with a rapid stride, and, standing with his
back to the study door, defied the inspector to enter.

"You do not seem to understand," Bates said. "The warrant I have

for your arrest gives me the right of searching the whole house. If
you persist in this absurd conduct, I shall have to call my men in and
remove you by force."

The two men faced one another, both angry and excited, and ready
to fly at one another's throats. And yet the whole time their ears
were filled with the beautiful melody of the music, as it floated from
the room behind.

"What are we going to do?" Claire asked. She was standing with
Jack at the top of the staircase. "Is it not time that we declared
ourselves?"
Jack whispered to Claire to remain where she was a moment, and
slipped out of the house into the garden unperceived. It had
suddenly occurred to him that perhaps the window leading from the
study to the garden was unfastened. He recollected that this was the
means by which Anstruther left and returned to the house. It would
have been imprudent on the latter's part to use the front door, and
there was not much risk in leaving the study window unlatched.

It was just as Jack had expected. The long French window gave to
his touch, and a moment later he was in the room. As it happened
on the previous occasion, he could see not the faintest trace of any
mechanism by means of which the melody was conveyed from the
Great Metropolitan Hotel to Panton Square. And yet the whole room
was flooded with it; rising and falling in triumphant strains, as if
mocking the intellect of a man who had brought this wonderful
result about. But there was no time to speculate on that, no time for
close investigation. On the other side of the door the voices of
Anstruther and Bates were rising to a still more angry pitch, and
Claire's tones of expostulation came to Jack's ears. As he crossed the
room he could see that the key was in the door. He flung it open,
and Anstruther came staggering backward into the room, closely
followed by the detective.

"You can see that the game is up," the latter said coolly. "Why not
make a clean breast of it? I shall find out how this is done, if I have
to pull down the house to do it."

Anstruther smiled in a scornful kind of way, and flung himself

doggedly into a seat. He bade Bates do his worst, and prophesied
that the police would suffer for this indignity. But Bates was not
listening. He was pacing rapidly round the room with his ear to the
wall, as if scenting out some clue to the mystery. A moment later,
and there came into the room the form of Serena.

One glance at her sufficed to show that she was not the Serena
whom Jack had known so long. The demure, downcast eyes were no
longer seeking the floor as of old; there was no shrinking and
timidity on the part of the woman now. She was changed almost
beyond recognition. She walked with a firm, elastic tread, her
shoulders were thrown back, and her head uplifted fearlessly. From
under his heavy brows Anstruther glanced at her suspiciously.

"Go away," he commanded hoarsely. "How dare you force yourself in

here like this! Go, woman."

But the tones of command had evidently lost their power. There was
no shrinking on Serena's part. She advanced into the middle of the
room as if the place belonged to her.

"No, no," she cried in tones as clear and ringing as Anstruther's own.
"Your power has gone forever. For three long patient years I have
waited for this moment. God only knows what my life has been, and
what a hell your cruelty has created for me. But the cord is broken
now. Only to-night I have learned the truth. I have been your good
and faithful servant; I have stooped to do your hateful work; I have
been the ally of criminals--of your creature Redgrave, amongst
others; and all because I thought you held my life in the hollow of
your hand."

"Tell them the story of your boy," Anstruther sneered.

"I will tell them the truth," Serena cried. "You said you could hang
me if you liked. You pretended that in my delirium I had taken the
life of my darling child. You were shielding a murderess, as I
thought. But it was a black and cruel lie. Give me back my wasted
years, you coward; give me back my sleepless nights and dreary
days. But, thank God, that time has passed. My boy is alive--alive!
He is safe in the house at present!"

Anstruther started as if some loathsome insect had stung him, then

dropped sullenly back in his seat again. Bates turned to Serena and
called her attention to the music.
"You seem to be in a communicative mood to-night," he said. "You
need not fear any one for the future--Redgrave, or anybody else. I
understand this last scoundrel is safe in the hands of the New York
police, who were wanting him badly. Perhaps you can tell us the
meaning of this extraordinary concert we are listening to. If you will
be so good----"

Serena made no reply in words, but crossed to the side of the room
opposite the door, and tugged at a volume which was the centre of a
set of some classical dictionary. The volume came away quite easily
in her hand, bringing other dummy books with it; and then the
interested spectators saw that the books in question were no more
than painted gauze. In the orifice disclosed by the stripping away of
the sham, there appeared to be something that resembled a mouth
of a great silver trumpet. This was partly plugged with a set of
sensitive metal plates, which were evidently intended to act as a
diaphragm for the record of musical expression.

"There you have the whole thing in a nutshell," Serena said,

speaking quite naturally and quietly. "It is very ingenious, and yet, at
the same time, it is not entirely original. It is an adaptation of the
theatre-phone, in connection with a somewhat modified form of
telephone. The recording instrument is situated in my husband's in
the Great Metropolitan Hotel, and he has only to start his
performance there, and the music sounds here quite as distinctly as
if he were actually playing in this apartment. It seems exceedingly
simple, now that you know how it is done."

It did seem simple, indeed, after listening to Serena's explanation.

Bates turned to Anstruther, and asked him if he had anything to say;
but the latter shook his head doggedly. He felt quite sure that the
game was up, though he had no intention whatever of giving himself
away. And yet, despite his danger, he was still the connoisseur
enjoying the beautiful music made by Padini's violin. But to Claire,
who had crept into the room unobserved, the whole thing was
horrible and unnatural. Such lovely music as Padini was playing now
was but a sorry accompaniment to all this vulgar crime and intrigue.
The girl shuddered, and placed her hands over her ears as if to shut
out the liquid melody.

"Oh, I wish it would stop," she said . "I do wish it would stop."

As if in answer to this prayer, the long, wailing notes died away, and
the music fainted into nothingness. At the same time, Bates
approached the mouth of the trumpet, and blew shrilly on his police
whistle. There was a pause just for an instant, and then, to Jack's
surprise, came the voice of Rigby clear and distinct.

"Is that you, Inspector Bates?" he asked. "We have just finished at
this end. I am afraid there will be no more music to-night, as two of
your detectives have most inhospitably insisted upon breaking up
our concert, and escorting Signor Padini to Shannon Street police
station. Shall I come round there, or will you come round here? Do
you get my voice quite clearly?"

Bates replied grimly that he did. There was no occasion whatever to

trouble Rigby any further to-night. Then the inspector turned to
Anstruther, and tapped him on the shoulder.

"I think there is no reason to carry this farce any farther," he said.
"You will be good enough to consider yourself my prisoner. Would
you like to walk to Bow Street, or shall I call a cab?"

Anstruther intimated that it was all the same to him. He knew

perfectly well now that the whole thing was exploded. There was
something bitter in the reflection that he had been found out at last
and laid by the heels over so paltry a business as the bogus burglary
at the City and Provincial Bank.

"I think I'll walk," he said. "No, you need not call any of your men,
and you need have no fear of personal violence."
"All right," Bates said. "Though I am still suffering from the shaking
up you gave me last night. Come along."

"I must apologize for all this trouble," Anstruther said, turning to
Claire, and speaking in quite his natural manner. "I must leave you
to manage as best you can for the present. I dare say you will be
able to manage with Serena."

He turned curtly on his heel, and walked to the door. Of Jack he took
no notice whatever. A moment later the front door closed sullenly,
and Anstruther was gone.

"The house smells all the sweeter for his absence," Jack said. "My
dearest girl, you can see now what a narrow escape you have had. I
only hope, for your sake, that the fellow has not been tampering
with your fortune. You must not stay here after to-morrow. The
place will be simply besieged by newspaper reporters and
interviewers. I must find some house for you----"

"You need not trouble about that, Mr. Masefield," Serena said.
"There is one house where both of us will be welcomed with open
arms. Need I say that I am alluding to Lady Barmouth's?"

Jack gave a sigh of relief; for the moment he had quite forgotten
Lady Barmouth. At any rate, for to-night Claire and Serena could
stay where they were, and they could go to Lady Barmouth's in the
morning. Then Jack remembered all that Serena had gone through,
and warmly congratulated her upon the recovery of her boy. "It
means all the world to me," Serena cried. "It fell out exactly as Miss
Helmsley said it would. When that man called to see Mr. Anstruther
again, I told him who I was, and he took me to my child at once.
The stranger had been very kind to the lad. He knew nothing of the
rascality and villainy behind it all, and he was only too glad to see
mother and son united."

"And Padini?" Jack suggested. "You must not forget----"

"I want to forget everything about him," Serena cried. "I shall be
glad, really glad, to know that that man is outside the power of
doing mischief for the next three years. Do not ask me anything
else--do not ask me, for instance, why I was playing the deaf-mute
that night at Carrington's rooms. I don't know. I was a mere slave
and tool in Anstruther's hands, and had to do exactly as he told me.
It was only by the merest accident that I discovered how the trick of
the music was done, and that I should have had to have kept to
myself if my dear boy had not been so marvelously restored to me.
Perhaps at some future time, I may be disposed to tell you more.
For the present, all I want to do is to sleep. I am longing for that
one night's sweet repose which has been so cruelly denied to me the
last few years."

Jack said no more. He left the house presently with the intention of
seeing Rigby at once, and then of calling on Lady Barmouth the first
thing in the morning, and making such arrangements as would
conduce to the comfort of Claire and Serena.

CHAPTER XLVIII.

THE PROOF OF THE CAMERA.

Society generally had plenty to talk about in the way of scandal next
morning, when it became known that Spencer Anstruther had been
arrested in connection with the burglary of the City and Provincial
Bank. The only paper giving anything like the account of the arrest,
naturally, was the Planet,which paper vaguely hinted at further
disclosures in the early future. Jack read the account over the
breakfast table, and smiled as he recognized the hand of Rigby in all
this. He would see Rigby presently, and ascertain exactly what had
taken place last night at the Great Metropolitan Hotel. First of all, he
had to see Lady Barmouth, who had already heard something of the
news. She listened with vivid interest to all that Jack had to say,
then announced her intention of going to Panton Square at once.

"I shall bring my sister and Claire here," she said . "They shall stay
as long as they please. As to my sister and her boy, I shall be
delighted to have them. I presume there will be some sort of
proceedings against Anstruther this morning?"

To the great disappointment of the public, when Anstruther came to

be charged at Bow Street the evidence was purely formal. The
prisoner had elected not to be represented by a lawyer, and, with a
view of expediting the proceedings, had formally pleaded guilty to
the charge, and asked to be committed to the Central Criminal
Court, which took place a week from now. "Clever chap that," Bates
said, as he and Rigby, together with Jack, turned into Covent
Garden. "Pretty cool, too. He wants to save time, of course, and get
the thing over before we can complete our chain of evidence. But I
fancy that by the end of a week we shall be able to produce all the
witnesses we want."

"I expect so," Rigby said. "By the way, don't forget about that
service of plate. Seymour says it ought to be conveyed to Scotland
Yard and the photographs taken at once. I have a letter from
Seymour in my pocket in which he asks me to go round and see Sir
Frederick Ormond, induce that gentleman to take the sealed crystal
ball to your headquarters, and to see that the seal is not broken,
except in the presence of one of your leading officials. Then you can
get both sets of photographs done at once."

Bates had his hands full for the next few hours. Then, towards four
o'clock, he made his way to Carrington's flat. Under plea of
indisposition, the latter had not been out for a day or so; but, as a
matter of fact, Bates had given him a pretty broad hint to keep clear
of the bank premises, and to consider himself more or less as a
prisoner on parole. Carrington's knees knocked together, and his
face turned deadly pale as Bates came into the room.

"So you have come again," he stammered. "I hope, perhaps, that--
don't say I am your prisoner."

"I am afraid that's what it comes to," Bates said. "We can't let you
off altogether, you know. But you help us, and give us all the
information in your power, and I'll do my best to get you let off as
lightly as possible. It makes all the difference between two years'
imprisonment and seven years' penal servitude."

"Am I to come with you now?" Carrington managed to stammer out.

"Is there no such thing as bail?"

Bates shook his head. Carrington would have to pass the night, and
doubtless a good many succeeding nights, in the police cells; but,
first of all, they were going as far as the bank. Bates explained that
there was no reason, for the present, why Carrington should stand
confessed as a prisoner. The bank officials need know nothing
whatever about it. What Carrington had to do now was to hand over
the service of Cellini plate at present locked up in his private safe.
The detective gave his promise that the plate in question should be
restored to its proper owner in due course, though he refused to
gratify Carrington's curiosity as to why he had specially selected this
particular art treasure.

An hour later the Cellini plate was safe in Bow Street, together with
the crystal globe; and before the week was out both articles had
undergone some mysterious process of photography, not altogether
unconnected with sheets of glass. Meanwhile, Anstruther was
preparing his defense as best he could, and Carrington had been
twice remanded on a charge of fraudulently dealing with the
property of his clients. The two cases excited the greatest interest,
and on the following Monday morning the Central Criminal Court was
packed with society people eager to hear the charges against
Spencer Anstruther.

Anstruther stood there, quite calm and collected, with just the touch
of a cynical smile on his lips. He looked round the court as if in
search of acquaintances, but no one responded. Many people whom
he knew quite well affected to look over his head. But cool and
deliberate as he was, Anstruther had all his work cut out to keep his
feelings in control when the barrister who represented the Crown
proceeded to call witnesses. The name of Seymour resounded down
the corridor, and a tall man with his face muffled up and a slouch hat
on his head stepped into the box. He bowed gravely to the judge,
and apologized for wearing his hat. A moment later his hat and coat
slipped away, and he turned his face half defiantly to the light. There
was an instant's breathless pause, then a veritable shout of
astonishment, as the Nostalgo of the posters stood face to face with
those whose curiosity had been so deeply touched during the past
four months.

"My name is Seymour," he said quietly, as if quite unconscious of the

tremendous sensation his appearance had excited. "I have known
the prisoner for some years. Before I unfortunately made his
acquaintance, I was not the human wreck you see now, but a man
like my fellows. But I need not go into that. What I propose to do
now is to tell the story of the burglary at the City and Provincial
Bank.

"Previous to my visit to Mexico, I occupied with Mr. Carrington the

rooms which are now his. I have in my pocket a latch-key which
opens the front door. It matters little now why I wanted to make a
search of Mr. Carrington's rooms, but I did make that search, and I
was hidden in the conservatory behind the smoking-room with Mr.
John Masefield on the night that the prisoner and Carrington
planned the sham burglary at the bank. The whole scheme was
revealed to us, and I shall be prepared to tell the jury presently what
steps I took to see the so-called burglary carried out. It is sufficient
for the present to say that it was carried out, and that I witnessed
the whole proceedings in the company of Mr. Masefield and a
journalist on the staff of the Planet,Mr. Rigby by name.

"I should like, at this point, to call the attention of the jury to what
we saw when the bank strong room was forced. So far as valuables
are concerned, the safe was practically empty, save for a service of
Cellini silver plate. Other witnesses beside myself will tell you that
the prisoner claimed that plate as a reward for the ingenious way in
which he had plotted to preserve Carrington's reputation. When I
heard this, a sudden inspiration came to me. With a piece of greasy
rag I hastily smeared the surface of the set of plate. I will come to
my reason for doing that presently. When the whole affair had been
finished, the prisoner was half minded to take the service of plate
back with him at once to his house in Panton Square. But Carrington
dissuaded him from this on the grounds of prudence. Therefore the
prisoner carried the plate up-stairs and deposited it in Carrington's
private safe. There it remained for a day or two, pending some way
of conveying it to Panton Square.

"But in the meanwhile something happened which aroused the

prisoner's suspicions. He made up his mind that he would himself
remove the plate from Carrington's safe by means of another
burglary. Carrington refused to have anything to do with this, but the
prisoner got his own way by the simple expedient of stealing
Carrington's keys. The prisoner is more or less intimately acquainted
with some of the cleverest thieves and housebreakers in London.
There was no time to call in an honest expert to open Carrington's
safe, but the prisoner was equal to the occasion. He called upon a
well-known housebreaker who passes by the name of 'Simple
Charlie.' I know this, because for some time I have been watching
the man in the dock. I have my own reasons for keeping quiet and
living in an out-of-the-way place, and I have a set of rooms fitted up
in what is more or less a common lodging house.
"By good fortune the man known as 'Simple Charlie' had rooms in
the same block of buildings. When the prisoner called upon him the
housebreaker was out, so that a note was left for him. This note I
managed to get hold of and read. Together with a friend of mine
named Ferris, we laid a little plot for 'Simple Charlie.' We compelled
him to find a substitute who would operate upon the safe, and that
substitute was no other than Inspector Bates, as doubtless he will
tell you later on."

It must be clearly understood that Seymour did not stand in the box
and reel off his evidence in the glib way of one who is making a
speech for the prosecution. On the contrary, the fascinating evidence
he gave was in reply to questions asked by the representative of the
Crown, occasionally supplemented by a query or two from the judge.
All this time Anstruther stood in the dock, his face knitted in an ugly
frown. Despite his easy air, his confidence was fast deserting him.
Any other man would have been crushed and broken by the deadly
weight of a testimony like that of Seymour's. In his heart of hearts
Anstruther was sick and frightened. Never for a moment had he
dreamed of anything like this. Seymour stood before him without a
trace of expression on his scarred, repulsive face. And yet every
word he uttered was as another month on the long sentence he was
already anticipating.

Anstruther came out of a dream presently, and realized with a start

that Seymour's deadly revelations were still going on. A moment
later, and the Crown Counsel suggested that Seymour should stand
down for a moment, and that Bates should take his place. The
detective came into the box alert and smiling. He told how he had
impersonated the mythical Maggs, and how he had accompanied
Anstruther to the City and Provincial Bank.

"At this point I should like to ask you a few questions," said Counsel
for the Crown. "I understand that you have become possessed of
the service of silver plate to which the last witness has already
alluded. He spoke just now of some device of his whereby the
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookmasss.com