ASEG Extended Abstracts

ISSN: 2202-0586 (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/texg19

Automated Structure Detection and Analysis in

Televiewer Images

Daniel Wedge, Eun-Jung Holden, Mike Dentith & Nick Spadaccini

To cite this article: Daniel Wedge, Eun-Jung Holden, Mike Dentith & Nick Spadaccini (2015)
Automated Structure Detection and Analysis in Televiewer Images, ASEG Extended Abstracts,
2015:1, 1-4, DOI: 10.1071/ASEG2015ab152

To link to this article: https://doi.org/10.1071/ASEG2015ab152

Published online: 22 Mar 2019.

Submit your article to this journal

Article views: 175

View related articles

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=texg19
Automated Structure Detection and Analysis in Televiewer Images
Daniel Wedge 1, Eun-Jung Holden 1, Mike Dentith1 Nick Spadaccini2
1 2
Centre for Exploration Targeting School of Computer Science and Software Engineering
The University of Western Australia The University of Western Australia
35 Stirling Highway, Crawley WA 6009, Australia 35 Stirling Highway, Crawley WA 6009, Australia
[email protected]

Automated planar structure detection in borehole images has

been an active research field for some time. Several methods
SUMMARY (Torres, 1992; Hall et al., 1996; Saito et al., 1999; Wang et al.,
2007) are based on the Hough transform or the related Radon
Borehole televiewer data is an important source of data transform (van Ginkel et al., 2003). These methods detect
on structural and stratigraphic discontinuities in both the features in the borehole image, then for each feature pixel,
mining and petroleum industries. Manually picking possible sinusoid parameters are determined that fit that
features in downhole image logs is a labour-intensive and feature. By accumulating votes from each feature, parameter
hence expensive task and as such is a significant values with a large amount of support can be identified and
bottleneck in data processing. It is also a subjective sinusoids created with those parameters. The frequency, depth
process. and orientations of the equivalent planes can then be
determined. Assous et al. (2014) exploited odd symmetry
We present a new algorithm and workflow for present in sinusoids to accumulate votes for finding likely
automatically detecting and analysing planar structures in structure locations. A more direct approach was taken by Sun
downhole acoustic and optical televiewer images. First, and Pallottino (2003) who traced a structure across an entire
an image complexity measure highlights areas most borehole image, exploiting the connection between the left and
suitable for automated structure detection. Changes in the right sides of the image in applying a shortest path algorithm.
image complexity can be used to locate geological
boundaries. Second, structures are automatically detected, We present a system for automatically detecting planar
with each structure having an associated confidence structures in acoustic televiewer (ATV) and optical televiewer
level; users can apply a threshold to the confidence (OTV) images and a workflow providing feedback for
values to adjust the quality and quantity of the detected assisting in processing the detected structures and maximising
structures based on the image quality and geological the value of the automated results. We do this in two ways.
complexity. Third, structures that have been detected but Firstly, we provide an image complexity measure highlighting
that do not meet the structure confidence threshold can be regions of the hole most suited for automated structure
interactively assessed and if necessary selected. We also detection. This complexity measure can be used for
provide tools for rapidly picking sets of equivalent automatically detecting geological boundaries within the hole.
structures and reducing structures to a set of Second, each automatically detected structure is assigned a
representative picks. confidence value. Structures are detected by exploiting
vertical and horizontal sinusoidal symmetry in a novel
Key words: downhole imaging, televiewer, planar manner. We provide time-saving workflow tools including
structure detection, auto picking filtering automated picks by interactively controlling the
structure confidence values; rapidly detecting structures with a
specified orientation (e.g. sets of parallel structures); and
INTRODUCTION
reducing densely picked structures to sparse representative
picks separated by orientation and depth thresholds. Our
Downhole image logs are routinely collected from boreholes
system enables rapid processing of downhole images and
in the petroleum and minerals industries. An image log is an
removes the bottleneck in processing large quantities of data.
unwrapped image of the cylindrical borehole wall. A planar
structure cutting directly through the borehole is imaged as
one period of a sinusoid (Rider and Kennedy, 2011). IMAGE COMPLEXITY AND ZONE DETECTION
Structures visible in these logs are analysed in many
applications, including understanding stratigraphy, Image quality is important for automated analysis. There are
sedimentary facies analysis, modelling fluid flow, and for regions of a borehole that a human logger would choose to
stability analysis. With some companies collecting hundreds ignore when manually picking structures. These regions
of kilometres of image logs annually, processing this data should also be avoided by automated processing methods.
manually is time-consuming and hence expensive. It is also These regions may correspond to breakouts or geologically
subjective due to its complexity, e.g. sinusoids may be complex zones or simply areas of poor data quality. Our
difficult to fit to the image if structures are not quite planar, system computes an image complexity (or image quality)
only part of the structure is visible, or due to poor imaging measure using a sliding window over the downhole image.
conditions in the presence of mud or sediment.
For ATV images, planar structures with rough surfaces and
complex structures cause the acoustic signal to scatter rather

ASEG-PESA 2015 – Perth, Australia 1

than reflect, producing dark areas in the image. We move a determined similarly. We also apply a confidence multiplier in
sliding window over the image and compute a measure based order to emphasise some characteristics of structures as
on the mean acoustic reflectance value of the window observed by the human visual system, including contrast,
compared to the mean and standard deviation of the entire structure thickness and sinusoidal symmetry. Adjacent
image. Higher mean reflectance values are indicative of structures sharing similar dip/azimuth angles and depths are
regions of low complexity; lower reflectance values merged into a single structure.
correspond to more complex regions.

We use a different complexity function for OTV images as

they record optical rather than physical properties, and so
complexity is represented differently. First the coloured OTV
image is converted to greyscale and then the local contrast,
orientation and type of image features are obtained using a
monogenic filter (Felsberg and Sommer, 2001). The
complexity measure incorporates: the proportions of feature
types (e.g. lines or steps) and orientations within the sliding
window, and the horizontal distribution of features around the
borehole within the sliding window, where an even
distribution is desired as it indicates likely structures spanning
the entire image/borehole.

The image complexity values are displayed as a colour map (a) (b) (c) (d)
next to the downhole image. Cool colours, e.g. blue and green, Figure 1. Detecting image features in an ATV image (a)
indicate regions of low image complexity most suitable for using phase symmetry to find small features (b) and large
automated processing, and hot colours, e.g. yellow and red, features (c); brighter pixel values indicate stronger feature
highlight the most complex regions. Zone boundaries coincide responses. The line segments fitted after thresholding and
with large step changes in the image complexity. thinning the feature outputs are shown in (d), where the
thickness of each line segment corresponds to the width of
AUTOMATED STRUCTURE DETECTION the source features in (b) and (c).

Sinusoids are identified by fitting short line segments to image

features. Line-like features are detected using phase symmetry
(Kovesi, 1997) which is a contrast-invariant feature detection
method. Phase symmetry is run at first to detect small-scale
features and then again to detect larger features. Sinusoids are
only fitted to features of the same scale, improving the
robustness of our algorithm. Hysteresis thresholding is applied (a) (b)
to the phase symmetry output and then the thickness of each
Figure 2. Fitting a sinusoid from line segments exhibiting
detected feature in the thresholded image is obtained from the (a) reflective symmetry about the green line and (b) out of
distance transform (Danielsson, 1980), before reducing the
phase symmetry, the green lines mark points 180° apart.
thresholded features to one pixel thickness (Lam et al., 1992).
Next we fit line segments to these features, using a minimum
line segment length of 5 pixels to discard potential false
positive features. These steps are shown in Fig. 1.

A novel pairing algorithm exploits the symmetry present in

sinusoids to find seed pairs of line segments. Two cases of
sinusoidal symmetry are exploited, shown in Fig. 2: i)
reflective symmetry about a local extremum, where a first line
segment’s gradient has the same magnitude but opposite sign
of a second line segment, and the segments have a similar
vertical location within the image; and (ii) out of phase
symmetry, where one line segment’s gradient is the negative
of the other, but the segments are separated by half the image
width (i.e. they are located on opposite sides of the borehole)
and may have different vertical positions. We solve for the
sinusoid parameters in a least squares manner (Rust, 2002).

Next, supporting segments for each sinusoid are determined,

i.e. line segments that are approximately tangential to the (a) (b) (c) (d)
sinusoid. The sinusoid’s structure confidence is computed Figure 3. Our interface displays the downhole image (a),
from the percentage of the image width where there are with detected structures overlaid (b). Structure confidence
supporting segments. When computing the structure values are shown in (c) as filled triangles for displayed
confidence, if the upper half of the sinusoid has at least double structures, or dotted outlines for structures not meeting
the coverage of the lower half, we mark the sinusoid as a half- the selected structure confidence threshold. The image
sinusoid and use the coverage percentage of the upper half complexity map (d) shows in this example that the bottom
only as the structure confidence. Lower-half-sinusoids are of this image has more complex features than the top.
Fig. 3 shows a screenshot of our interface. The user can select rest as a dotted line; full sinusoids are drawn as continuous
a structure confidence threshold by judging the confidence of solid lines. Open structures are filled with diagonal lines.
structures shown by the triangles and the presence or absence
of corresponding image features. As the threshold is changed, Fig. 4 shows results for three structure confidence thresholds
the sinusoids are updated in the interface in real-time, thus for a processed ATV image. As expected, at higher structure
adding or removing many structures simultaneously. confidence thresholds a larger proportion of the structures
Increasing the threshold will display fewer and higher- correspond to visible geological features; lower thresholds
confidence structures, decreasing the threshold displays more increase the number of false positives returned. Here, half-
and low-confidence structures. The threshold can be set for sinusoids are fitted to actual half-structures e.g. at 31.2m. In
either a zone within the image, the entire borehole, or a addition, some full structures only had half-sinusoids fitted as
selected depth range only. We recommend setting one they are only approximately sinusoidal in shape, e.g. below
structure confidence threshold per zone, since the image and 30.2m. Since there are fewer image features tangential to these
geological complexity are typically fairly constant within a sinusoids, these structures have lower structure confidence
zone, but can vary between zones. values as demonstrated by them not meeting the higher
confidence thresholds. Sinusoids with high confidence were
After applying a threshold, there may be some structures that fitted in complex regions; of note is a scenario at 29.4m where
were detected but did not satisfy the structure confidence sinusoids are fitted correctly to two structures dipping in the
threshold. An individual sinusoid can be confirmed as an opposite direction. For this image, a structure confidence
actual structure by first hovering the mouse over either the threshold of 40% gives best results, though of course
structure in the image or the point of the structure confidence individual structures with lower structure confidence values
triangle outline, which will temporarily display the sinusoid on can be added with a double click.
the downhole image. Then, by double clicking on the same
point, the user can confirm that this is an actual structure. In Fig. 5, results for an OTV image are shown. Subtle
variations in shape where imaged structures are not quite
A structure confidence threshold can also be applied to sinusoidal are handled by fitting half-sinusoids, e.g. at 9.2m
structures with a range of orientations selected on a stereonet and 9.9m. Some open structures are also fitted, e.g. at 9.5m.
display, e.g. a higher threshold can be applied only to bedding For this OTV image, a structure confidence threshold of 60%
orientations, keeping only bedding structures with high gives the best results. This contrasts with the 40% threshold
confidence while preserving other structures. The orientations used for the ATV image because of differences in the image
can be combined with a selected depth range, allowing modality, imaging conditions and geological complexity.
specific tuning of the results.
Finally, the result of rapidly fitting a set of parallel structures
ANALYSIS TOOLS is illustrated in Fig. 6; the maximum permitted difference in
orientation is defined by the user. The more steeply-dipping
A set of parallel structures can be identified by selecting a structure below and tangential to the highlighted sinusoid was
single structure with the desired orientation. This structure’s outside of the dip-difference threshold and was not detected.
shape is fitted to single line segments detected in the image.
Here, sinusoids are fitted using position and gradient
constraints rather than using sinusoidal symmetry. Supporting
line segments are determined, the structure confidence
calculated and the user can then apply a structure confidence
threshold as before.

Representative picks, i.e. a subset of detected structures with

each structure separated by some orientation or depth, can be
selected automatically. We cluster structures based on
orientation and depth attributes and the structure with the
highest confidence in each cluster is selected. This way the
representative picks are selected objectively. Where multiple
teams using the same dataset have different objectives, e.g.
structural geologists may be interested only in changes in the
general orientations of structures, whereas geotechnical
engineers analysing rock stability may require every structure
to be picked, selecting representative picks reduces processing
and ensures that a consistent dataset is used by both teams.

Our system also provides functionality for editing the detected

structures, e.g. altering the dip, azimuth, depth or aperture.
Users can also manually pick or delete structures.

RESULTS

We present structure detection results for ATV and OTV

images, and for detecting a set of parallel structures from a (a) (b) 20% (c) 40% (d) 60% (e)
single structure. Half-sinusoids are drawn with one half-period Figure 4. ATV results. (a) Borehole depth. Structures are
as a solid line (corresponding to the detected features) and the shown for structure confidence thresholds of (b) 20% (c)
40% (d) 60%. (e) Image complexity.
 Further functionality is provided for automatically detecting: a
set of structures with a particular orientation based on picking
one structure and specifying orientation tolerances; for
reducing the set of detected structures to a representative
subset; and assigning a structure type to a subset of structures.
These tools empower geoscientitsts to maximise the value of
the automated results and enable faster processing of
downhole images.

ACKNOWLEDGMENTS

This research was funded by Rio Tinto Iron Ore from 2010 to
2013. We thank Advanced Logic Technology for allowing us
to publish the datasets in this paper.

REFERENCES

Assous, S., Elkington, P., Clark, S., and Whetton, J., 2014,
Automated detection of planar geologic features in borehole
images: Geophysics, 79(1).

Danielsson, P.-E., 1980, Euclidean distance mapping:

Computer Graphics and Image Processing, 14, 227-248.

Felsberg, M. and Sommer, G., 2001, The monogenic signal:

(a) (b) 30% (c) 60% (d) 90% (e) IEEE Transactions on Signal Processing, 49(12), 3136-3144.
Figure 5. OTV results. (a) Borehole depth. Structures are
shown for structure confidence thresholds of (b) 30%, Hall, M.G.J., Ponzi, M. and Maletti, G., 1996, Automatic
(c) 60% and (d) 90%. (e) Image complexity map. extraction and characterisation of geological features and
textures from borehole images and core photographs: PSWLA
37th Annual Logging Symposium.

Kovesi, P., 1997, Symmetry and asymmetry from local phase:

Proceedings of AI’97 Tenth Australian Joint Conference on
Artificial Intelligence, 185-190.

Lam, L., Lee, S.-W. and Wuen C.Y., 1992, Thinning

methodologies - a comprehensive survey: IEEE Transactions
on Pattern Analysis and Machine Intelligence, 14(9), 869-885.

Rider, M. and Kennedy, M., 2011, The Geological

Interpretation of Well Logs 3rd ed.: Rider-French Consulting.

(a) (b) (c) Rust, B.W., 2002, Fitting nature's basic functions Part III:
Exponentials, sinusoids, and nonlinear least squares:
Figure 6. Detecting sets. (a) Structures detected using our
pairing algorithm. (b) Further structures detected by Computing in Science and Engineering, 4(4), 72-77.
fitting a set to the highlighted sinusoid in (a) with dip and
Saito, N., Bennett, N.N., and Burridge, R., 1999, Method of
azimuth variations of up to 15°. (c) Line segments used for
determining dips and azimuths of fractures from borehole
fitting. Different colours are used to identify segments
images: U.S. Patent 5,960,371.
joined end-to-end; segments’ thicknesses reflect the
feature size in the televiewer image.
Sun, C. and Pallottino, S., 2003, Circular shortest path in
images: Pattern Recognition, 36(3), 709-719.
CONCLUSIONS
Torres, D.O., 1992, Method for determining dip and strike
We have developed software for assisting geoscientiststs to angles in borehole ultrasonic scanning tool data, U.S. Patent
rapidly and objectively process downhole images. An image 5,162,994.
complexity map highlights regions most suitable for
automated structure detection; this map can be used to van Ginkel, M., Kraaijveld, M., van Vliet, L.J., Reding, E.,
automatically determine zone boundaries. Our structure Verbeek, P., and Lammers, H., 2003, Robust curve detection
detection algorithm detects planar structures imaged as using a Radon transform in orientation space: Lecture Notes in
sinusoids or half-sinusoids in ATV and OTV images. Each Computer Science: Proceedings of the 13th Scandinavian
detected structure is assigned a structure confidence based on Conference on Image Analysis, 2479, 125-132.
the image evidence for the structure. A structure confidence
threshold can be used to interactively select a subset of Wang, Y., Etchecopar, A., Onda, H., and Mathieu, G., 2007,
detected structures. Workflow tools are provided to rapidly Method of determining planar events from borehole or core
pick further structures by double-clicking on the image. images: U.S. Patent 7,236,887.