A Fast and Low Cost Embedded Test Solution for

CMOS Image Sensors

Julia Lefèvre, Philippe Debaud, Patrick Girard, Arnaud Virazel

To cite this version:

Julia Lefèvre, Philippe Debaud, Patrick Girard, Arnaud Virazel. A Fast and Low Cost Embedded
Test Solution for CMOS Image Sensors. ITC 2021 - IEEE International Test Conference, Oct 2021,
Online, United States. pp.1-9, �10.1109/ITC50571.2021.00007�. �lirmm-03377562�

HAL Id: lirmm-03377562

https://hal-lirmm.ccsd.cnrs.fr/lirmm-03377562
Submitted on 14 Oct 2021

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 2021 IEEE International Test Conference (ITC)

A Fast and Low Cost Embedded Test Solution

for CMOS Image Sensors
J. Lefevre 1, 2 P. Debaud 1 P. Girard 2 A. Virazel 2

1 2
STMicroelectronics LIRMM
Imaging division, DFT University of Montpellier / CNRS
Grenoble, France Montpellier, France
[email protected], [email protected] <lastname>@lirmm.fr

Abstract—This paper presents a novel test solution directly purpose, test at the end of the manufacturing process is a
embedded inside CMOS Image Sensors (CIS) to sort out PASS crucial step to exhibit potential failures and check various
and FAIL dies during production test. The solution aims at functional, optical and electrical parameters.
reducing test time, which can represent up to 30% of the final
product cost. By simplifying the way optical tests are usually Various solutions for manufacturing test of CIS have been
applied with an ATE, the proposed Built-In Self-Test (BIST) proposed so far. A brief summary of these solutions is given
solution overcomes the drawbacks of long test time and huge in Section III.B. In the meantime, industrial practices for CIS
amount of test data storage. We experimented our solution by testing are mainly based on using external test approaches
considering that roughly half of the tests usually performed with based on the use of an ATE (Automatic Test Equipment).
an ATE can be embedded and applied using the proposed fast Unfortunately, these practices have a significant impact on
and low cost BIST engine. Results obtained on more than 2,400 the final cost of the product [4]. Depending on the targeted
sensors have shown that our solution reduces test time by about
application, the test cost for a complex IC (Integrated Circuit)
30% without impacting the defect coverage. The area cost of our
solution is about 1% of the digital part of the sensor, i.e., like a CIS may represent up to 30% of the final product cost,
approximately 0.25% of the total sensor area. The proposed which is now considered as prohibitive [5].
embedded CIS test solution outperforms existing solutions in In this paper, we present a fast and low cost BIST solution
terms of area overhead and test time saving, thus encouraging for CIS that aims at verifying the good functionality of the
its future implementation in an industrial production flow. optical part (pixel array) inside the sensor. The main objective
Keywords—CMOS image sensor, BIST, optical test, pixel array. is to reduce the significant CIS test time required by typical
I. INTRODUCTION external test procedures at a low cost. Our strategy consists
of using the proposed BIST solution to screen out local
CMOS Image Sensors (CISs) are widely used in many defects (singlets, couplets, clusters) and to use ATE-based
electronic devices for various industrial applications, such as tests to target global defects. To deal with output images
smartphones, autonomous vehicles, night vision systems, provided by a CIS, the BIST solution reuses the way pixels
medical imaging equipments, etc., as well as applications for are scanned in conventional image processing algorithms
the space conquest, such as cameras used by the Perseverance (convolution and median filtering) applied with an ATE. By
Mars rover in 2021. Although the Charge Coupled Device transforming two-dimensional computations into one-
(CDD) technology was initially dominant due to superior dimensional computations realized inside the sensor, test
sensitivity and picture quality, various improvements in CISs complexity and test data storage requirements are limited,
have led them to surpass CCD sensors in shipment volume and test time is saved.
since about 15 years. In particular, they can benefit from their
Performances of our solution have been evaluated owing
low fabrication cost, low power consumption, as well as the
to a software emulation of the developed BIST engine.
possibility of combining analog and digital functions [1, 2].
Validation of the solution has been done through experiments
As any other electrical sensor, a CIS translates physical performed on output images coming from 2,400 CMOS
information into electrical information [3]. Its architecture is image sensors, showing that test time can be reduced by
divided into two parts: the optical part contains an array of roughly 30% without impacting the defect coverage when
pixels (picture elements) that capture photons (light compared to a full ATE-based test solution. Moreover, the
information) and the electrical part converts this light silicon area estimated to implement the BIST engine is about
information into electrical signals that are transmitted outside 1% of the digital part of the sensor, i.e., approximately 0.25%
the sensor to display an image composed of pixel values. of the total area of a CIS.
Irrespective of the application field, verifying and ensuring The rest of this paper is organized as follows. Section II
that a CIS, especially the pixel array, will be able to operate presents the general architecture of a CIS and details the
correctly during in-field application is mandatory. To this

1
conversion flow from the physical information to the output sequencer block which controls the system in conjunction
image. A brief summary of the state of the art in the field is with an external clock. When the pixel is addressed by the
also proposed. Section III provides a taxonomy of defects in sequencer, the output voltage of the pixel is converted into a
a CIS and shows how they can be detected by the proposed digital signal thanks to an Analog-to-Digital Converter
test solution. Section IV presents the novel BIST solution and (ADC). ADC settings, e.g. gain of the ramp used to sample
its hardware implementation. Section V summarizes the and to quantify the analog signal, allows to define the digital
results achieved and compared to optical tests usually value of the pixel encoded in a bit word.
performed with an ATE. Section VI concludes the paper.
II. BACKGROUND ON CIS TESTING
A. CMOS image sensor overview
Thanks to his optical and electrical blocks, a CIS is able to
transform the light information into electrical information
through the following conversion flow. The light arrives
directly on a micro lens located on top of the pixel so as to do
not miss any light rays and to focus them on the
photosensitive element. This element inside the pixel,
generally a photodiode, converts photons in electrons owing
to the photoelectric effect [6]. The general anatomy of a pixel
Figure 2: General architecture of a CIS
in the array of a CIS is shown in Fig. 1.
From the image containing the digital values of all pixels,
various operations are carried out by the Image Signal
Processor (ISP) such as white balancing to calibrate and
correct the color dominance according to the ambient light,
noise reduction to avoid “pepper and salt” aspect, sharpening
to ensure good contrast in the image, etc. [8]. The ISP also
performs an interpolation on the output RGB image between
pixels of the same type (Red, Green or Blue) to rebuild the
CIS image on the common three-channels RGB. This
principle, also called demosaicing, is illustrated in Fig. 3. The
goal of the demosaicing operation is to fill the white boxes as
shown in Fig. 3 with values computed from the same-colored
Figure 1: General anatomy of an active pixel in a CIS [6] neighbors. The output image of CIS is a superposition of the
three-channel images and can be displayed on a display
To deal with the color aspect, a color filter organized in a module thanks to the interface block (cf. Fig. 2).
Bayer pattern (Red, Green and Blue) is located on the pixel
array, bellow each micro lens (cf. red color filter as example
in Fig. 1), so that one pixel contains one color information.
Note that this color filter contains more green filtering
patterns than red or blue patterns to mimic the human eye
features with the dominance of green color.
A pixel can be characterized as passive or active. A passive
pixel contains only the photosensitive element and a single
transistor to select and read the information (electrons) from
the photosensitive element. An active pixel has a quite similar
architecture with the addition of a reset transistor and an
amplifier transistor (cf. Fig. 1). The reset transistor is used to
completely discharge the photodiode and hence be sure that
there are no parasitic electrons from the previous capture that Figure 3: From Bayer RGB to 3 channels RGB image by the ISP
could contaminate the pixel output [7]. The amplifier demosaicing operation
transistor allows to verify the electron-to-voltage The output image of a CIS does not reflect exactly the
amplification directly inside the pixel. Due to the addition of physical scene in front of the sensor but rather is an
transistors in an active pixel, a drawback is that the area approximation. There are sources of inaccuracy such as the
occupied by the photosensitive element inside the pixel is gap, called pitch, between pixels that creates a loss of
reduced. The advantage is that the signal degradation is lower information, the potential loss of light rays on the
since the parasitic capacitance from the direct connection photosensitive element (even if the micro lens catches the
between the photodiode and the column bus is reduced. majority of rays), or the light level definition. Note that, a
At the high level, row select bus and column bus transistors light level is the interpretation of the brightness. Naturally,
of each pixel are addressed by row and column decoders to the light level seen by a human eye is proportional to the
scan and to read out the whole pixel array row by row. These arithmetical value of the pixel, i.e., the lighter the pixel, the
blocks are depicted in Fig. 2. They are managed by a higher is its value.

2
B. State of the art in CIS testing to measure resistances. The other solution uses a vector
CIS testing is done at various stages of a sensor life, from network analyzer equipment to drive the RF impedance used
characterization of the die on the wafer to online test during to detect partial defective lines. Despite its efficiency, this
sensor lifetime. The work presented in this paper in intended solution needs additionnal equipment (line delay detector,
to be used at a production test level in order to sort out PASS delay controler, pulse detector, etc.) to detect defective lines.
and FAIL sensor dies. Test of a CIS comprises electrical and Note that all these BIST solutions in [14-18] aimed at
optical tests, all usually performed on an ATE, in order to test reducing optical test time, a common objective with the BIST
all the elements inside the sensor. When one test fails (optical solution proposed in this paper.
or electrical), the sensor is declared as defective and is
withdrawn from the production line. III. DEFECTIVE PIXEL DEFINITION AND DETERMINATION
Electrical tests can be functional, structural, or parametric A defect is an unwanted particularity of a device that can
depending on the part of the circuit (analog, digital or mix) create a failure. It may come from particle deposition (e.g.,
which is under test [9]. Electrical tests include continuity test, bridging between two metal lines, bad contact definition) or
leakage control, IDDQ test, conventional stuck-at fault tests, process deviation (e.g., diffusion depth issue, misalignment
as well as test for memories, ADC, PLL, etc. [10]. of masks) during manufacturing.
Optical tests are performed on the output images coming Regarding CIS, defects can often be seen after reading of
from a CIS and use image processing algorithms like defect the pixel array, when the output images are available. Figure
tracking, dark current capture, signal imbalance verification, 4 shows some examples of image defects. Note that A*A is a
gain verification, etc. [11]. Results of these optical tests allow Kernel containing A² pixels, with A being a positive odd
to check if output images are correct according to predefined integer. A Kernel is used to stake out an area in an image that
image quality criteria [12]. During optical testing, the CIS allows to classify a defective pixel with respect to its
under test is put in two main states (dark or light) to avoid neighbors. The choice of A depends on the sensor resolution
error sources such as non-uniformity and presence of a scene and on the customer specifications.
or colours. Moreover, there is not only “light on” and “light Several categories of defects have been defined [6]:
off” extremes cases but a panel of settings variations • A singlet is a defective pixel with no immediate
(different converter gains, integration time, etc.) [13]. Note defective neighbors and located at the center of a Kernel
that it is possible to detect a physical defect on the pixel array of A*A pixels, as shown in Fig. 4;
of a CIS in dark and in light images or independently in dark • A couplet is a defective pixel with only one immediate
or in light images. defective neighbor in a Kernel of A*A pixels. One of the
Optical tests are performed thanks to image processing defective pixels is at the center of the Kernel;
algorithms applied on the output images of the CIS under test. • A cluster is a group of more than two defective pixels in
Usually, it is necessary to wait for several images to be a Kernel of A*A pixels. A defective row or a defective
captured entirely and stored by the ATE before computing column (continuous, partial, dotted, etc.) are two
metrics and determine if the sensor images pass the optical particular cases of clusters.
tests or not. This mean that test time and data storage
dedicated to test may become prohibitive, e.g. test time can
go from 10 seconds to 1 minute [14].
In the literature, some solutions have been proposed to
embed test facilities inside a CIS and hence save test time.
Authors in [14-16] and in [17-18] propose to reduce the
amount of optical tests and proceed with a fully electrical test
set so as to avoid precise and restrictive settings needed in Figure 4: Examples of possible image defects [18]
optical tests. Authors in [14], [15] and [16] propose a solution
The classification illustrated in Fig. 4 is based on the
based on the generation of a pulse (voltage stimuli) which is
defective pixel definition and each of these categories can
sent on the anode of the photodiode inside the pixel to
reflect a physical defect. In order to determine whether or not
simulate the light illumination. Authors indicate that the
a pixel is defective, its light level and arithmetic value need
implementation of this method is very dependent on the pixel
to be considered. In a fault-free case, the light level of the
architecture. This structural-based test is done on the pixels
pixel needs to be in accordance with the light level of its
of the array independently of each other, without any
neighbors so as to ensure uniformity. The arithmetic value of
management of clustering pixels. Experimental results
the pixel is used to evaluate its lightness level with respect to
demonstrate a relatively good correlation in dark conditions
a predefined range of values. This range depends on the
with respect to tests performed with an ATE, but the
sensor features. For example, if the pixel value is encoded on
comparison is unreliable in the light conditions due to the use
8 bits, the range of pixel values will be [0; 255].
of a light source not stable enough.
There exist three different types of defective pixels: hot,
The work presented in [17] and [18] proposes a BIST
dead and weak. A hot pixel is a pixel with the highest
framework to detect defective lines used to read out or
arithmetic value and a dead pixel is a pixel with the lowest
addressing pixels, thanks to the extraction of the impedance
value [19][20]. A hot pixel can be seen as a pixel stuck at ‘1’
of lines. Authors use two solutions for the detection of
and a dead pixel can be seen as a pixel stuck at ‘0’. A weak
continuous and partial defective lines. One solution deals
pixel is a pixel with a lower arithmetic value when compared
with continuous defective lines by reusing the ADC circuitry
to a threshold computed with the values of its neighbors [21].

3
 To detect a defective pixel, a local average has to be IV. NEW BUILT-IN SELF-TEST SOLUTION
determined to check if the pixel value is in accordance to the
A. Principle of the proposed test approach
values of its neighbors. Several local averages are calculated
by using different mask sizes. A mask is composed of a group CIS testing consists of electrical and optical testing phases.
of several neighboring pixels located on the same row. For A CIS can pass electrical tests and fail optical tests, or vice
example, a local average can be computed with a 1x5 size versa. In both cases, the CIS is declared as FAIL if at least
mask (i.e., a 1D array of five pixel values) and is computed one test reports a failure. However, electrical tests are
by excluding the pixel at the center, i.e. the pixel under test. significantly faster than optical tests, thus motivating the need
A defective pixel has a pixel value which is outside a range to minimize the duration of optical tests for CIS.
defined by its higher and lower bounds, called “thresholds” In order to save test time of a CIS, a solution consists in
in the sequel. To implement the proposed BIST solution, using an embedded test solution (BIST) to screen out local
these thresholds need to be unified, i.e., they need to be defects (singlets, couplets, clusters) and to use ATE-based
calculated in the same way by using constant parameters, tests to target global defects, such as pixel values uniformity
named amin and bmin for the lower threshold, and amax and bmax along the array or noises like Fixed Pattern Noise (FPN) [1],
for the higher threshold. As can be seen in Equations (1), which represents the variation of pixel values in the array.
these thresholds are represented by affine functions in which BIST allows to test each pixel on-the-fly without the need to
constants are defined depending on whether the CIS operates store a full image before testing as is the case with ATE-based
in a dark or in a light environment (cf. Fig. 5). testing. Moreover, BIST is performed at-speed so that test
time can be greatly reduced compared with ATE-based test.

Figure 5: Example of light and dark images of a CIS [19]

Figure 6: From a 2D Kernel to a 1D mask on a pixel under test
In a dark environment, the distribution of pixel values for
Of course, the area overhead required by the BIST
a given mask should be uniform in a fault-free case. So, we
hardware limits the amount and complexity of possible
just need to add a fix number to the local average (bmin and
computations, so that embedding conventional image
bmax in Equations (1)) to compute the thresholds and
processing algorithms used for optical test in a BIST engine
determine whether the pixel is defective or not. In a light
can be too costly. Indeed, convolution and filtering methods
environment, a lens is used to diffuse while focusing the light
used in image processing algorithms are generally based on
on the sensor, thus provoking a shaded aspect (cf. example of
two-dimensional data that are available when an entire image
a light image presented in Fig. 5) on the output image [22].
has been stored. However, when only a flow of pixels coming
Due to the difference of light repartition, a fixed value cannot
from a row (i.e. a mask) of the pixel array is available, using
be used to compute the local average. Indeed, a pixel at the
such methods becomes impossible. To alleviate this issue, we
center of the image has a high arithmetic value due to the light
simply reuse the way pixels are scanned in 2D convolution
concentration. Therefore, it could be falsely declared as
and 2D median filtering methods to deal with 1D masks of
defective if the computation would be done in the same way
pixels (cf. Fig. 6). With this modification, we avoid the need
as for pixels located at the borders of the pixel array, which
of additional hardware to store data during the reading of the
have a lower arithmetic value due to the darkness. To
array, and we limit the complexity of operations that are
compute the thresholds (“Thr”), an alternative is to use a
specific to 2D data management, thus reducing test time.
percentage of the local average (amin and amax) based on the
values of the neighbors of the pixel: B. BIST workflow
Thr =a × local +b To test the pixel array, the BIST engine needs to access
Thr =a × local +b
(1) each arithmetic pixel value inside the array. The array is
streamed row by row at the pixel rate thanks to the sequencer
block presented in Fig. 2. When a pixel is dealt with the BIST
Equations in (1) are defined in a generic manner to allow module, first, the pixel type (Red, Green or Blue) is
their use irrespective of the CIS under test, considering that determined from the pixel coordinates. Then, the arithmetic
the threshold values may change from one CIS to another value of the pixel is compared with the two thresholds
with the modification of the pixel architecture, the gain of the computed from a local average (cf. Equations (1)). If the pixel
ADC, the number of pixels in the array, etc. This generic value is outside the defined range, the pixel is defective and
formulation also allows the computation of thresholds values its data (value and coordinates) are stored in a memory for the
irrespective of the detection environment (light or dark). Note next phase. Otherwise, these data are not saved.
that the constant parameter values in (1) are chosen
empirically based on the acceptable rate of defective pixels Once all the pixels of the array have been evaluated by the
in CIS output images. BIST, the data of each defective pixels stored inside a

4
memory is read by a software program which is able to count This allows to do parallel computations on several
singlets, couplets and determine whether there is a cluster or accumulators to pipeline the detection on several pixel types.
not into the pixel array of the CIS under test. If a pixel is located near the borders of the array, it is
C. General BIST architecture mandatory to compute a padding value to detect if the pixel
is defective or not. This suggests a coordinate management
In our work, CIS die sorting is done by resorting to two
inside the BIST module, that can be done with counters
types of resources: the first one is a BIST infrastructure
operating at the pixel rate and returning row and column
embedded in the CIS for test and defective pixel detection,
relative positions in the array.
and the second one is an external memory for test data
storage. This distribution of tasks has been decided to achieve Then, the pixel is processed by the Defective pixel & line
the best tradeoff between test time efficiency and CIS detection block. More precisely, the value of the pixel is
hardware overhead. An external CPU (Central Processing compared to the thresholds (Thrmin & Thrmax) to determine if
Unit) is used for defect classification. the pixel is correct or defective. The pixel data, i.e. the value
of the pixel, its coordinates and its type, are stored in the
One important point is the location of the BIST module
memory of the sensor in case it is declared as defective.
inside the CIS architecture. In order to avoid pre-processing
on pixel values by the ISP, that could potentially correct and The information “the pixel is defective” in the form of a
hide potential defects, pixel values must come directly from boolean will launch the writing of the pixel data inside the
the pixel array without going through intermediate hardware external memory. Note that, to avoid test hardware overhead,
modules like correction modules, filtering block, scaler the external memory is a memory already existing in the CIS
block, etc. [8]. Fig. 7 shows the integration and interfacing of but not used during the test phase and hence available to store
the BIST module inside a CIS. pixel data of defective pixels. The coordinates and values of
a defective pixel are encoded by the Data formatting block
before storage in the memory.

Figure 7: Insertion of the BIST module inside a CIS

Figure 9: Singlet or couplet according to the size of the Kernel
The proposed BIST solution has been implemented in
When all the pixels in the array have been evaluated and
Verilog language. The general BIST architecture is depicted
have passed through the BIST engine, the memory fulfilled
in Fig. 8 where each digital block is dedicated to one function.
by all the defective pixels data is read by a python program
embedded on the CPU of the system to classify the defective
pixels. By using information about the Kernel size, this
program is able to compare pixel coordinates to determine
whether a defective pixel can be linked with neighboring
pixels and placed in a given category. Let us consider the
example in Fig. 9. Here, the program will count two singlets
in the left image if the size of the Kernel is defined by 3*3
whereas it will count one couplet in the right image if the
Kernel size is 5*5. The number of singlets, couplets and the
presence of clusters are information that can make the die
Figure 8: General BIST architecture passing from PASS to FAIL. If the number of singlets or
All the blocks detailed in this architecture are couplets in the dark or light images is too high considering a
synchronized by a Finite State Machine (BIST FSM) which is given limit, the CIS is tagged as FAIL. Note that the presence
used to manage the BIST process by launching appropriate of only one cluster in dark or light conditions is a FAIL factor.
signals and collecting required information. Communications
V. VALIDATION OF THE BIST SOLUTION
between the various blocks is done owing to wires (1-bit
signal carriers) or bus (word of several bits carrier). In order to validate the proposed BIST solution, we
In order to test all pixels in a CIS array, each pixel is performed experiments by using a dataset composed of
selected based on its type (red, green or blue) since a pixel of images taken by CISs in dark and light environments. The
one type can only be tagged as defective with respect to pixels output of the BIST engine is a PASS/FAIL information
of the same type. The block Pixel data mgt guides the pixel related to the CIS under test as well as the number of singlets,
in dedicated accumulators located inside the Defective pixel couplets and cluster associated to each image. The goal of our
& line detection block to store pixel values of the same type. experiments is to demonstrate the efficiency of the BIST
solution in reproducing part of optical tests for CIS as they

5
are usually applied with an ATE without the drawbacks of presence of a cluster is represented by a Boolean indicator (0
long-time duration and huge amount of data storage. or 1) and not by an integer value. So, in this case, we count
the number of common images with a cluster reported by the
A. Experimental setup
BIST and by the ATE-based tests.
A dataset composed of output images coming from more
To quantify the efficiency of the proposed BIST solution,
than 2,400 CMOS image sensors originating from the same
we have computed the mean of the relative error for all output
packet were used in our experiments. Among these 2,400
images owing to Equations (3) shown below:
CISs, a part of them were identified as FAIL by former ATE-
∑A∈"H;JKL$ 8828@ABCDEF " $
;< =' _ =
based optical tests. Images coming from these CISs were
selected so as to get a representative sample of various defect ! M
∑A∈"H;JKL$ 8828NOPQDEF " $
;< =' _ =
categories, i.e., singlets, couplets, defective columns, (3)
defective rows, clusters of various sizes, etc. For each CIS, 1234 ! M
we collected the same number of dark and light images so
that the dataset was split into two equivalent sets of dark and To determine the threshold values as defined in Section III,
light images as shown in Fig. 10. we need to calculate the local average for each pixel under
test and define values of constant parameters. The local
average is calculated by considering a mask size and the
values of pixels inside the mask. Several mask sizes (1x3, 1x5
and 1x7) have been considered based on experience. The
values of constant parameters used to determine higher and
lower thresholds were empirically chosen as shown in (4).

=RST_U/VW = 1; YRST_U/VW = +120

= = =RST_U/VW ; Y =.
Constants for dark img : \]^__`abc
_U/VW _U/VW d

=RST_f+gh- = 1 + 0.15; YRST_f+gh- = 0

(4)
Constants for light img : e
Figure 10: Dark & Light image distribution of the dataset [6]
= _f+gh- = 1 . 0.15; YR _f+gh- = YRST_f+gh-
The subset of images coming from the FAIL CISs was
organized into several categories with respect to the defect
The two sets of constants given in (4) are chosen to detect
type. For each image, the numbers of singlets and couplets
defective pixels with values outside the range [local_avg-60;
are known, as well as the presence of cluster, thanks to the
local_avg+120] for dark images (“img”) and outside the
output data from former optical tests performed on ATE.
range [local_avg*(100%-15%); local_avg*(100%+15%)] for
Many other information coming from electrical tests formerly
light images. These values are very dependent on the type of
performed on these CISs are also available for each image.
CIS and have been chosen empirically.
For this reason, there is a need to analyze all data related to a
given image to sort out relevant information and to be sure TABLE I. SINGLET AND COUPLET MEAN ERROR (IN PIXEL) IN DARK
AND LIGHT IMAGES DEPENDING ON THE MASK SIZE
that the FAIL classification of a given CIS came from an
optical failure and not from an electrical failure. To this end, Mask size 1x3 1x5 1x7 3x3
these information are stored in a csv file to be further used in
a postprocessing stage and observe correlation between Singlet 3.70 4.40 -66.20 4.80
Dark
results from the BIST solution and those from the ATE-based Couplet -0.05 -0.08 -0.26 -0.08
tests.
Singlet -1.90 -2.10 30.00 -1.80
B. Results and correlation Light
Couplet 0.05 0.00 0.06 0.01
In order to evaluate our BIST solution, we define some
error metrics to quantify the relative error between results Table I reports the mean error calculated for singlets and
obtained with the ATE-based solution, considered as a couplets with respect to a given mask size. Three differents
reference, and those produced by the emulated BIST solution. 1D mask sizes (1x3, 1x5 and 1x7) have been experimented.
For each defect category, an error metric has been defined by The selection of these mask sizes has been done so as to limit
Equations (2) as follows: storage of pixel values used in the local average computation.

! "#$ = %#&'( )*+,- "#$ . %#&'( )/-0 "#$

The 3x3 mask size, which is based on a 2D computation

"#$ = 5 67( )*+,- "#$ . %#&'( )/-0 "#$

performed by the ATE-based test, is used to compare results
1234 !
7 % &5 8 "#$ = 95(6%) *+,- "#$ == 5(6%) /-0 "#$:
(2) with a 1D computation. We can observe that the singlet mean
13 ! error for dark and light images are acceptable for 1x3 and 1x5
mask sizes. Indeed, a mean error of 4.4 pixels with a 1x5
where “i” is an integer in [0; N-1] associated to a given image, mask size for dark images represents a deviation of 0.0003%
N being the total number of images in the dataset. with respect to the total number of pixels in the CIS array,
For all experiments, errors were computed for dark and which is a negligible error. Conversely, we can observe that
light images in the dataset. A negative error indicates that the the 1x7 mask size is not approppriate for the detection of
BIST solution has identified a lower number of defects than singlets in dark or light images. Indeed, 66.2 defective pixels
the ATE-based solution. Note that applying such type of error are missing with the BIST solution for the dark images and
computation for the cluster category is unfeasible as the 30 additional pixels have been found for the light images,
which is not acceptable. The detection of couplets in dark or

6
light images performs very well irrespective of the mask size respectively 0.0008% and 0.0005% of the total number of
(mean errors are around 0), thus demonstrating the efficiency pixels in the CIS array.
of the BIST solution in detecting double defective pixels.
From the results reported in Table I, we can conclude that the
detection of singlets and couplets performs very well for 1x3
and 1x5 mask sizes.
Table II reports the number of images containing a cluster
with respect to a given mask size. The 1x3 mask allows to
detect a higher number of bad images with the BIST
compared to the ATE-based solution (92 against 69) in a light
environement. However, as only 63 images over the 92 are in
common with the ATE-based solution, we can conclude that
Figure 11: Distribution of singlet errors in dark and light images
using a 1x3 mask size is not appropriate.
TABLE II. NUMBER OF IMAGES WITH A CLUSTER DEPENDING ON THE We can also observe in Fig. 11 that the distribution of
MASK SIZE singlet error in dark images is shifted to the positive values
and the distribution in light images is shifted to the negative
Mask size 1x3 1x5 1x7 3x3 values. This means that BIST testing has detected a bit more
Found by ATE-based test 33 singlets than the based-ATE tests in dark images and a bit less
in light images. A solution to even more increase the
Dark Common images 32 33 31 33 effectiveness of our BIST solution would consist in adjusting
Found by BIST solution 34 36 33 35 the constant values used in Equations (1) in order to shift the
two distributions in Fig. 11 towards ‘0’.
Found by ATE 69
With a mask size fixed at 1x5 thanks to information given
Light Common images 63 65 60 64 in Tables I and II, we can now adjust the constant values used
Found by BIST solution 92 65 67 68
to define the thresholds. Note that by changing constant amax
or bmax, constants amin and bmin will change too due to the
Conversely, results regarding cluster detection obtained dependency between the high and low thresholds of the
with a 1x5 mask size perfectly correlates the results found by range. Table III gives the mean error for singlet and couplet
the ATE-based solution for dark images (33/33 images in detection on dark images when the constant bmax_DARK varies.
common). Moreover, this 1x5 mask size allows to detect As can be observed, the choice of bmax_DARK = +120 is
three additional images containing a cluster with the BIST validated in the dark as it is the only one which has the lowest
solution (36/33). After analysis, these images were confirmed singlet error number. Note that, if the BIST engine detects
to be bad images actually containing a cluster (defective row). more singlets with +110, it is because the higher threshold is
This is an interesting feature of our BIST solution which was set too low and the lower threshold is too high, resulting in a
able to detect the cluster on dark and light images, whereas very restricted range of good pixel values. The same
the ATE found the cluster only on light images. comment can be done with bmax_DARK = +140. The BIST
engine misses a lot of defective pixels as the thresholds create
Regarding light images, the use of a 1x5 mask prevented
a large range where all pixel values have been tagged as not
the BIST solution to identify four bad images containing a
defective. For the couplet mean error rate, we can see that the
cluster (65 versus 69). However, a careful analysis revealed
BIST classifies correctly double defective pixels even if the
that these four images contained shaded rows that were found
constant bmax_DARK varies.
by the BIST solution on the dark images. So, finally the BIST
do not miss any defect on any sensor, thus demonstrating the TABLE III. MEAN ERROR RATE FOR SINGLET AND COUPLET
DETECTION IN DARK IMAGES WITH VARIATION OF BMAX_DARK
effectiveness of our solution.
Finally, for the sensor type considered in our experiments, bmax_DARK 110 120 130 140
we have chosen the 1x5 mask size, that offers the best
Singlet 37.30 4.40 -21.90 -44.30
tradeoff between lowest singlet and couplet error rate and Dark
good cluster detection. Couplet 0.05 -0.08 -0.15 -0.22

Figure 11 depicts the singlet error distribution for dark and TABLE IV. NUMBER OF DARK IMAGES WITH A CLUSTER DEPENDING
light images given in the form of a histogram. Each bar of the ON BMAX_DARK
histogram refers to the number of images having the same
singlet errors. Ideally, the two singlet error distributions bmax_DARK 110 120 130 140
should be placed around the “0” meaning that there is no error Found by ATE-based test 33
between ATE-based and BIST solutions. The standard
deviation (σ in Fig. 11) is used to show that the dispersion of Dark Common images 33 33 32 31
values around the mean is low, with 1.6 pixels in light images Found by BIST solution 38 36 34 35
and 2.3 pixels in dark images. From the standard deviation
definition, it means that 99.7% of the data is included inside Table IV gives the number of detected dark images with a
[mean-3σ; mean+3σ] [23]. For dark and light images, we can cluster when bmax_DARK varies. The same comments as done
see that the “mean±3σ” extrema values are +11.2 pixels for Table II can be raised since, for example, the BIST
(dark) and -6.8 pixels (light). These values represent solution with bmax_DARK = +110 allows to detect five more

7
images than the ATE-based solution, but nevertheless detects C. Discussion
the same 33 bad images found than the ATE-based solution. With a correct setting of constant values, the proposed
We have done the same comparison for light images by BIST solution is able to appropriately sort the dark or light
varying constant amax_LIGHT. Table V shows the mean error images into good or bad images. Nevertheless, PASS and
rate for singlets and couplets. As for dark images settings, the FAIL sensors classification from dark and light images
choice of 15% has been found to be the most appropriate due require further analysis. More precisely, if a dark or a light
to the lowest singlet error rate and considering that the image is labelled as bad by the BIST solution, the sensor will
couplet error rate is very good irrespective of the percent be declared as FAIL. A sensor will be declared as PASS if
value. both dark and light images are identified as good. To be
TABLE V. MEAN ERROR RATE FOR SINGLET AND COUPLET “good”, an image must not contain any cluster and we need
DETECTION IN LIGHT IMAGES WITH VARIATION OF AMAX_LIGHT to verify that the number of singlets and couplets are under
admitted values defined according to a given application.
amax_LIGHT 10% 15% 20%
Considering the whole dataset of 2,400 sensors, the BIST
Singlet 16.00 -2.10 -4.20 solution has a 99.95% correlation with the PASS/FAIL
Light classification provided by the ATE-based solution. Only one
Couplet 0.64 0.00 -0.01
CIS has been misclassified as FAIL by the BIST engine.
As for dark images, the number of images with cluster was Results obtained with our BIST engine shows that, from
computed and results are given in Table VI. These results quite simple computations used to detect defective images, it
show that 15% is indeed the most appropriate constant value. allows to sort reliable PASS and FAIL information for a CIS
TABLE VI. NUMBER OF LIGHT IMAGES WITH A CLUSTER DEPENDING under test. The optical part of a CIS is tested directly inside
ON AMAX_LIGHT the sensor, without any interface that can add sources of
errors (potential contact resistance, pad continuity, etc.).
amax_LIGHT 10% 15% 20% Regarding the coverage of optical algorithms usually
Found by ATE 69 performed on ATE, we estimate that more than 50% of these
algorithms can be embedded inside the CIS under test, thus
Light Common images 67 65 59 leading to a reduction of approximately 30% of the test time
Found by BIST solution 143 65 62 for each sensor. This estimation is based on a deep analysis
of optical tests, demonstrating that 50% of them usually target
From all the above experiments and results, we can local defects and hence can be embedded. The area taken by
conclude that correctly setting values of constants amax, bmax the BIST architecture has been estimated in terms of
and amin, bmin is crucial, as well as correctly selecting the mask additionnal logic gates. It represents roughly only 1 % of the
size. A precise setting is mandatory to avoid image digital part of a sensor.
missclassification due to a false number of singlets or As BIST solutions presented in Section II.B, the proposed
couplets. Regarding couplet detection, the proposed BIST BIST engine does not embed all optical tests usually
performs always very well irrespective of the mask size or the performed on an ATE. Compared to existing BIST
constant values. approches, the advantages of the proposed BIST solution are
that i) it represents less than a few percents of the total CIS
area, far below what can be achieved with some of the
existing solutions, and ii) all potential local defects in the
pixel array can be detected, while only a part of them are
covered by existing BIST solutions.
VI. CONCLUSION AND DISCUSSION
In this paper, we have presented a novel BIST solution for
CIS testing. It is based on the detection of defects such as
singlets, couplets and cluster in dark or light output images
from a CIS under test. These detection information are used
Figure 12: Singlet mean error for dark images with respect to to define whether images are good or bad and, finally, if the
bmax_DARK constant variation CIS is PASS or FAIL. A software emulation of the BIST
engine has been done to validate our solution. Experiments
To be even more precise in the analysis of the bmax_DARK
carried out on images coming from 2,400 CIS have shown
parameter, linear approximations have been done to find the
that our proposed solution adequately classify CIS into PASS
optimal constant value. Figure 12 reports the result of the
and FAIL categories in 99.95% of cases. In addition to be
linear approximations and shows that bmax_DARK should be
useful for testing CIS, this type of embedded test solution is
precisely set to +122 to get a perfect correlation between
also valuable for diagnosis purpose, since the coordinates of
number of singlets found by the BIST solution and those
defective pixels stored inside the memory are easily
found by the ATE-based solution.
reachable.
As for bmax_DARK, linear approximations have also been Despite its effectiveness, it is still possible to improve our
done to find the optimal value for amax_LIGHT. The result led to solution by considering the following aspects:
14.4%.

8
 • Our BIST solution currently focuses on local defects [6] R. Fei. “Alternative Solution to Improve the Production Test of Optical
(singlets, couplets and clusters) and not yet on global Sensors in CMOS Technology”, PhD dissertation, Grenoble Alpes
University, 2015.
defects such as shading (difference of values along the
[7] I.L. Fujimori, “CMOS Passive Pixel Imager Design Techniques“, PhD
array, change of uniformity) or noises like temporal dissertation, Massachusetts Institute of Technology , 2002, 50544970-
noise (non-desired variations of pixel values between MIT.pdf.
different captures). Even if the probability to have a [8] M. Tusch, G. Cooper, and T. Atherton, “Combining an ISP and Vision
local defect is higher than the probability to have a Processor to Implement Computer Vision”, Embedded Vision
Alliance, 2019, https://www.edge-ai-vision.com/2019/01/combining-
global defect (except for some noises such as FPN and an-isp-and-vision-processor-to-implement-computer-vision/.
dark current, which are very common in CISs), [9] G. Prenat, “Design of an Analog and Mixed-Signal BIST Architecture
complementing our solution to cover global defects is in Deep Submicron Technology,” PhD dissertation, Institut National
of real interest. To this end, simple modifications can be Polytechnique de Grenoble, 2005.
implemented to add functionalities to the BIST solution [10] A. Bounceur, “CAT Platform for Mixed-Signal Circuit Testing”, PhD
and hence cover these global defects. dissertation, Institut National Polytechnique de Grenoble, 2007.
• Bus latency to communicate between the external [11] J. Chen, “CIS Testing Technology”, Jetek Technology corporation,
2018. https://fr.slideserve.com/amos/cis-testing-technology
memory and the BIST module can be extensive and the
[12] M. Pedersen, J.Y. Hardeberg, “Full-Reference Image Quality Metrics:
size of the memory can be inadequate compared with Classification and Evaluation”, Foundations and Trends® in Computer
what we need to store (too large or too small pixel data). Graphics and Vision, vol. 7, no. 1, pp. 1-80., March 2012.
A possible solution would consist in adding a small [13] M.B. Kaplinski, G. Waligorski, V. Berezin, and E.R. Fossum, “Test
memory inside the CIS, together with the BIST Methodologies for Digital CMOS Camera-on-a-Chip Image Sensors”,
in Proc. IEEE Workshop on CCDs and Advanced Image Sensors, pp.
hardware. This could be done at the cost of additional 239-241, 1999.
area overhead. [14] L. Lizarraga, S. Mir, and G. Sicard, “Experimental Validation of a
• Location information about defects were not used in our BIST Technique for CMOS Active Pixel Sensors,” in Proc. IEEE VLSI
work to compute the correlation between optical tests Test Symposium, pp. 189 – 194, 2009.
performed on ATE and with the BIST solution. Indeed, [15] L. Lizarraga, S. Mir, and G. Sicard, “Evaluation of a BIST Technique
for CMOS Imagers,” in Proc. IEEE Asian Test Symposium, 2007.
the correlation relies only on the number of defects in
[16] L. Lizarraga, S.Mir, G. Sicard, and A. Bounceur, “Study of a BIST
each category (singlet, couplet and cluster). However, it Technique for CMOS Active Pixel Sensors,” in Proc. IFIP
may happen that two different defects are detected International Conference on Very Large Scale Integration, pp. 326-
respectively by the ATE-based test and the BIST 331, 2006.
solution, and hence are considered in the correlation [17] R. Fei, J. Moreau, and S. Mir, “BIST of Interconnection Lines in the
computation. Adding a comparison between Pixel Matrix of CMOS Imagers,” in Proc. IEEE International
Workshop on Advances in Sensors and Interfaces, pp. 174 –177, 2013.
coordinates of defective pixels detected by the BIST and
[18] R. Fei, J. Moreau, S. Mir, A. Marcellin, C. Mandier, E. Huss, G.
those detected by the ATE could provide even better Palmigiani, P. Vitrou, and T. Droniou, “Horizontal-FPN Fault
correlation information. Coverage Improvement in Production Test of CMOS Imagers,” in
Proc. IEEE VLSI Test Symposium, pp. 1-6, 2015.
REFERENCES [19] G.H. Chapman, J. Leung, I. Koren, and Z. Koren, “Tradeoffs in Imager
[1] A. Dupret, M. Tchagaspanian, A. Verdant, L. Alacoque, and A. Design with Respect to Pixel Defect Rates”, in Proc. IEEE
Peizerat, “Smart Imagers of the Future”, in Proc. IEEE Design International Symposium on Defect and Fault Tolerance in VLSI
Automation & Test in Europe, 2011. Systems, 2010.
[2] M. Bigas, E. Cabruja, J. Forest, and J. Salvi, “Review of CMOS Image [20] C.-Y. Cho, T.-M. Chen, W.-S. Wang, and C.-N. Liu, “Real-Time Photo
Sensors,” Microelectronics Journal, vol. 37, no. 5, pp. 433-451, May Sensor Dead Pixel Detection for Embedded Devices,” in Proc.
2006. International Conference on Digital Image Computing Techniques and
Applications, pp. 164 - 169, 2011.
[3] B. Luo, F. Yang, and L. Yang, “Key Technologies and Research
Development of CMOS Image Sensors”, in Proc. IITA International [21] “Basic Principles of Image Sensors“, Online courses “Optique
Conference on Geoscience and Remote Sensing, 2010. Ingénieur”, Maine University, http://www.optique-
ingenieur.org/en/courses/OPI_ang_M05_C06/co/Contenu_18.html
[4] Quality Testing Services, “Breakdown on Costs of Testing vs. Cost of
Not Testing,” May 1, 2019. https://qualitytest.net/product-testing-vs- [22] Advantest Corporation, “CMOS Image Sensor Test Solution”, 2017.
not-testing-costs/ http://www.koreatest.or.kr/sub02/2017data/%EB%B0%95%EA%B8
%B0%EC%88%98.pdf
[5] A. Grochowski, D. Bhattacharya, T.R. Viswanathan, and K. Laker,
“Integrated Circuit Testing for Quality Assurance in Manufacturing: [23] M. Perekupka, “Standard Normal Distribution: Definition &
History, Current Status, and Future Trends”, in Proc. IEEE Example,” Study.com, March 9, 2015.
Transactions on circuits and systems-II: analog and digital signal http://study.com/academy/lesson/standard-normal-distribution-
processing, vol. 44, no.8, 1997. definition-example.html