Simultaneous Optimization of Abnormality Discriminator and
Illumination Conditions for Image Inspection of Textile Products
Yuma Nishikawa, Fumihiko Sakaue and Jun Sato
Nagoya Institute of Technology, Japan
Keywords:
Anomaly Inspection, Textile Product Inspection, Illumination Optimization.
Abstract:
In this study, we propose a method to simultaneously learn and optimize the illumination conditions suitable
for anomaly inspection and a neural network for anomaly inspection in textile product anomaly inspection.
In the inspection of abnormalities in industrial products such as textile products, it is necessary to optimize
the imaging environment including the lighting environment, but this process is mostly done manually by
trial and error. In this study, we show that highly accurate inspection of abnormalities can be achieved by
using a display whose light source position and brightness can be easily changed, and by presenting a proof
pattern suitable for abnormalities on the display. Furthermore, we show how to simultaneously optimize the
neural network and illumination conditions used for such anomaly inspection. We also show that the proposed
method can appropriately detect anomalies using images actually taken.
1 INTRODUCTION
The process of product inspection at industrial prod-
uct manufacturing operations is essential to assure the
quality of manufactured products. In this inspection
process, products are checked for any abnormalities,
such as scratches and irregularities in coloring, which
can occur in a wide range of products. Therefore, it is
difficult to automatically determine abnormalities us-
ing image processing technology, etc., even if they are
easily determined to be abnormal by humans. There-
fore, methods to automatically determine abnormali-
ties are still being researched and developed. These
methods can be classified into those that learn only
normal data and detect abnormalities[(Akcay et al.,
2018; Akc¸ay et al., 2019)], and those that collect both
abnormal and normal data and learn the differences
between them[(Bergmann et al., 2019; Bergman et al.,
2020)]. Since each of these methods can be used in
different situations, they are used for different pur-
poses.
It is well known that the inspection of abnormal-
ities in textile products such as fabrics is particularly
difficult. This is partly due to the fact that textile
products have various patterns and colors, and the
appearance of textile products varies greatly depend-
ing on these patterns and colors when inspecting for
abnormalities. In addition, textile products are com-
posed of fine threads, so that even a slight change in
Figure 1: Example of change in appearance of flaws on tex-
tile products.
the condition of the fibers or the surrounding condi-
tions can change the appearance of the textile product.
This means that the appearance of anomalies in tex-
tile products also changes significantly depending on
the conditions under which the textile product is im-
aged. Figure 1 is an example of an image of the same
flaw (abnormality) taken under different light source
environments, and it can be confirmed that the appear-
ance changes depending on the situation. Due to these
characteristics, human inspectors visually inspect by
changing the direction of observation and illumina-
tion conditions to create a situation in which it is easy
to detect anomalies.
Considering the above, it is necessary to optimize
the arrangement of cameras, illumination, etc. for ab-
normality inspection in accordance with the abnor-
mality to be detected. The importance of optimiz-
ing the imaging environment is well known in image
processing-based anomaly inspection, and when con-
structing an actual system, the optimal imaging con-
Nishikawa, Y., Sakaue, F. and Sato, J.
Simultaneous Optimization of Abnormality Discriminator and Illumination Conditions for Image Inspection of Textile Products.
DOI: 10.5220/0013322300003912
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 3: VISAPP, pages
755-760
ISBN: 978-989-758-728-3; ISSN: 2184-4321
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
755

ditions are studied by repeatedly capturing images of
actual products. Since this process requires a lot of
time and cost, an efficient method to perform it is
needed. In this study, we propose a method for au-
tomatically optimizing the imaging conditions.
For this purpose, we propose a method to perform
highly accurate anomaly inspection by using a dis-
play used for video presentation as a light source to
illuminate the object and optimizing the illumination
patterns displayed on the display. There are an al-
most infinite number of patterns that can be displayed
on a display used as a light source, and it is pos-
sible to create a very complex illumination environ-
ment. On the other hand, the number of variations
makes it impractical to optimize them through trial-
and-error. Therefore, the proposed method, which can
automatically optimize the lighting environment, has
a great advantage when constructing an abnormality
inspection system. In addition, the proposed method
trains a neural network used for anomaly detection
in conjunction with such lighting pattern optimiza-
tion. Combined with the derived patterns, the pro-
posed method can detect anomalies with extremely
high accuracy.
2 ANOMALY DISCRIMINATION
USING VISION
TRANSFORMER (ViT)
We first describe methods for detecting anomalies us-
ing image information. Two methods are possible for
image abnormality detection: one is to use only nor-
mal images and discriminate deviations from them
as abnormalities(Perera and Patel, 2019; Bergmann
et al., 2019), and the other is to collect abnormal and
normal images in advance and discriminate them by
2-class identification. While the former method is
easy to collect training data, it has the problem that
it is difficult to improve the accuracy of abnormality
discrimination compared to the method that directly
uses abnormality data. In this paper, assuming that
the anomaly images can be collected to some extent,
we proceed with the discussion with the main objec-
tive of anomaly detection by 2-class discrimination.
In traditional anomaly detection before deep
learning, classification of abnormal and normal im-
ages has been attempted based on the statistical prop-
erties of image features, such as principal component
analysis and linear discriminant analysis. In these
discriminant analyses, various types of information
have been obtained not only by directly using im-
age intensity information, but also by using infor-
Figure 2: Representative architecture of Vision Transfomer.
mation such as edges obtained by filtering. On the
other hand, in recent years, methods using neural net-
works represented by CNNs (Convolutional Neural
Networks) have been widely used in anomaly detec-
tion(Perera and Patel, 2019). Recently, neural net-
work structures that do not use convolution, called
Vision Transformer (ViT) in particular, have been
widely used,(Dosovitskiy et al., 2021). Transformer
architecture in neural networks is a method proposed
in the field of natural language processing, and the
structure that applies this to image information is the
Vision Transformer, which is widely used for image
identification.
The Vision Transformer divides the input image
into a set of small patches and processes these as to-
kens using the Transformer structure. In this case,
each token is input to a network structure called ViT
Encoder, as shown in Fig.2. The ViT Encoder uses
a structure called “multi-head attention” to represent
the relationship between tokens, enabling it to capture
global image features even from a finely segmented
patch set. The features obtained by the Encoder are
input to MLP (Multi-Layer Perceptron), which calcu-
lates the final identification result.
It is known that a large amount of training data is
required to achieve sufficient performance when con-
structing a discriminator using Vision Transformer,
compared to using a CNN. However, it is known that
this problem can be avoided by training (pre-training)
a neural network in advance using a large amount
of image data, and then learning the neural network
in transfer learning according to the task. This has
shown that the Vision Transformer can be applied
to a variety of tasks even when training data is lim-
ited(Dosovitskiy et al., 2021). In the discriminator
used in this study, we aim to achieve highly accurate
anomaly discrimination by transfer learning a neural
network that has been trained with a large amount of
data in advance, using the target data.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
756

Figure 3: Imaging environment with display as light source.
3 SIMULTANEOUS
OPTIMIZATION OF
ILLUMINATION CONDITIONS
AND ANOMALY DETECTION
ViT
3.1 Lighting Environment Using
Controllable Lighting
In this study, anomaly discrimination is performed us-
ing ViT as described in the previous section. How-
ever, in order to achieve stable and robust abnormal-
ity detection, it is necessary to acquire optimal im-
ages suitable for abnormality detection. As men-
tioned above, the appearance of a textile product,
which is composed of a collection of fine threads,
varies greatly depending on the lighting environment
in which it is photographed. Therefore, images taken
under inappropriate illumination conditions do not
sufficiently show the characteristics of the textile in
the image, making it difficult to perform proper ab-
normality detection. To avoid this problem, a method
of capturing images of objects under various illumi-
nation conditions in advance and using these images
for identification can be considered. However, this
method increases the time required to capture images
and increases the cost of equipment. Therefore, it is
necessary to prepare optimal illumination conditions
suitable for abnormality detection in order to perform
abnormality identification at the lowest possible cost
and in the shortest possible imaging time.
3.2 Determination of Illumination
Environment Using 1 × 1
Convolution
When a display is used as a light source, there is a
nearly infinite variation of possible illumination pat-
terns. Therefore, it is not realistic to optimize them
manually. In this study, we propose a method to
Figure 4: Example of a pattern on a display.
simultaneously optimize the illumination conditions
for anomaly detection and the neural network for
anomaly discrimination by utilizing the linearity of
the light source in the image.
In this method, a rectangle with a certain bright-
ness is displayed on the display as shown in Fig.3 and
is used as the light source. These rectangles divide the
display area and reproduce various lighting environ-
ments by changing the brightness of each rectangle as
shown in Fig.4. Let N be the number of light sources
(rectangles) displayed on the display, and let I
i
be the
image captured in an environment where only each
light source is turned on. In this case, the image I
captured by lighting multiple light sources simultane-
ously can be expressed by the following equation.
In this study, a large display is used as illumina-
tion, as shown in Fig.3, and the aim is to realize highly
accurate anomaly discrimination by controlling it ap-
propriately. The display used here can emit light at
arbitrary locations and at arbitrary brightness. By dis-
playing various patterns on the display, a very com-
plex lighting environment can be created. Therefore,
by estimating and presenting display patterns suitable
for anomaly detection, it is expected to be possible
to create a situation in which anomaly detection can
be performed with higher accuracy than with general
lighting.
I =
N
∑
i=1
E
i
I
i
, (1)
where E
i
is the brightness of each light source and is
the ratio of the brightness when I
i
is taken. When I
i
is illuminated at the maximum brightness that can be
displayed when I
i
is taken, 0 ≤ E
i
≤ 1. Therefore,
the determination of the optimal lighting environment
can be said to be the determination of E
i
as shown in
the Eq.(1).
In determining E
i
, we consider the simultaneous
optimization of these parameters and the neural net-
work for anomaly detection. As shown in Eq. (1),
an image taken under optimal illumination conditions
can be expressed as a weighted sum of images taken
under a single light source. If N images are con-
sidered as N channel images, the estimation of the
optimal illumination image can be thought of as the
process of merging N channel images into a single
Simultaneous Optimization of Abnormality Discriminator and Illumination Conditions for Image Inspection of Textile Products
757

Figure 5: Proposed architecture.
channel image. Since the linear sum of images is syn-
onymous with the linear sum of each pixel in the im-
age, the linear sum of images can be expressed as the
convolution of 1 × 1 × N kernels(Ueda et al., 2024).
Therefore, the determination of optimal illumination
conditions can be redefined as the optimization of this
convolution kernel. Based on the above, we define the
network structure including the lighting optimization
as shown in Fig.5
In Fig.5, N images are input and M images are ob-
tained by the convolution of N images with M kinds
of 1 × 1 kernels. By inputting this to the ViT ar-
chitecture, it determines whether the input image is
a normal image or an abnormal image. This con-
volution process with 1 × 1 kernel is equivalent to
the process described in Eq(1). Therefore, the im-
age obtained by this convolution is equal to the image
taken under the illumination pattern based on the ker-
nel weights. Therefore, while N images taken under
each light source must be acquired when training the
network, only M images taken under the obtained op-
timal illumination can be used for inspection. Since
M can be any integer greater than or equal to 1, it can
be set according to the allowable imaging time and
required inspection accuracy to accommodate various
situations. This allows the neural network to learn and
the illumination conditions suitable for it at the same
time.
4 EXPERIMENTAL RESULTS
4.1 Environment
In this section, we present the results of an experi-
ment conducted to verify the effectiveness of the pro-
posed method in the previous section. In this exper-
iment, the proposed method was used to simultane-
ously train the illumination condition and the neu-
ral network, which was then used to detect anoma-
lies on the fabric product. For this purpose, we took
images of flaws on the fabric using a needle, and
examined whether it is possible to discriminate be-
tween flawless areas (normal areas) and flawed areas
(abnormal areas). Examples of normal and abnor-
mal images are shown in Fig.6 These images were
captured under the illumination of the display shown
in Fig.3. The input images are the 224×224 im-
(a) Normal image (b) Abnormal image
Figure 6: Examples of input image.
ages extracted from the captured images. The Vision
Transformer used to discriminate anomalies is a pre-
trained one using JFT-300M(Sun et al., 2017) and Im-
ageNet1k(Russakovsky et al., 2015). This model was
also trained as a 2-class discriminator with a 1×1 con-
volutional kernel that indicates the illumination con-
dition using the structure shown in Fig.5. Since ViT
is trained with RGB 3-channel images as input, after
reproducing the illumination condition using the 1×1
kernel, three 3×3 convolutional kernels were applied
to convert the image to a 3-channel image, which was
used as input to the discriminator.
On the display, a rectangle is displayed as shown
in Fig.3, which is used as the light source. Twenty-
five (5 × 5) are displayed on the device. The bright-
ness of each light source was determined by the pro-
posed method. The number of kernels to be trained
was set to 1 (one captured image) and 2 (two captured
images), and the performance of each case was exam-
ined. For training, 50 sets of normal images and 50
sets of abnormal images were used, and 1000 sets of
data (500 normal images and 500 abnormal images)
were used for evaluation. For comparison, additional
training of the discriminator was performed on im-
ages taken under a single illumination using the same
data, and the performance of the discriminator was
examined. For the comparison method, the lighting
condition with the best discrimination result among
25 input images was selected as the first image, and
for the second image, the condition with the best dis-
crimination result was selected when discrimination
was performed using two images, the lighting condi-
tion with the best discrimination result and another
image.
The evaluation value was set to be rejected if it
was less than a threshold value, and by changing this
threshold value, the rate at which an abnormal image
was judged as a normal image (false acceptance rate:
FAR) and the rate at which a normal image was er-
roneously rejected (false rejection rate: FRR) were
obtained, and each method was evaluated.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
758

Figure 7: ROC curves for proposed and conventional ViT.
(a) Optimal
illumination
(b) Best single
illumination
(c) Worst single
illumination
Figure 8: Images taken by the proposed method (a), images
taken with the best single illumination source (b), and im-
ages taken with the worst single illumination source (c).
4.2 Results
First, the ROC curves of the proposed method and the
method using a single illuminated image are shown
in Fig.7. This figure shows that all methods have
high discrimination performance, which confirms the
effectiveness of ViT in discriminating anomalies.
Among the methods, the proposed method shows the
best discrimination performance when two input im-
ages are used, which confirms its effectiveness. Ex-
amples of images actually taken under the optimal il-
lumination, images taken in the environment with the
best discrimination performance, and images taken in
the environment with the worst discrimination perfor-
mance are shown in Fig.8. These images show that in
the image taken under the optimal illumination, only
the abnormal area is emphasized, making it easier to
discriminate abnormalities. On the other hand, in the
worst-case environment, it is difficult to see the ab-
normal area, and this is thought to make it difficult to
discriminate abnormalities.
Next, the false rejection rate (FRR) when the false
acceptance rate (FAR) is set to 0 is shown in Tab.1.
This corresponds to over-detection, in which a normal
part is judged as abnormal in the inspection process.
Since missing anomalies are not acceptable in product
inspection, it is particularly important from the stand-
point of practicality how low the over-detection can
be suppressed when the over-detection is set to zero.
The table shows that the proposed method can sup-
press the false rejection rate to a low level even when
Table 1: False rejection rate (FRR) with false acceptance
rate set to 0[%].
FRR[%]
Proposed (1 image) 0.6
Proposed (2 images) 0.4
Single illumination (1 image) 10.9
Single illumination (2 images) 2.7
Figure 9: Change in false rejection rate with learning.
the number of input images is small. This indicates
that the proposed method can be used to construct a
very practical anomaly detection system.
In addition, how the FAR changed with learning
is shown in Fig.9. The results show that the proposed
method with illumination optimization not only keeps
the FAR low, but also always keeps the FRR low re-
gardless of the learning process. This indicates that
the proposed method is capable of stable learning.
These results confirm that the method proposed in this
paper can construct a stable and robust anomaly de-
tection system.
5 CONCLUSION
In this study, we proposed a method to improve the
accuracy of textile product abnormality inspection
by simultaneously optimizing the illumination condi-
tions and the abnormality detection neural network.
REFERENCES
Akcay, S., Atapour-Abarghouei, A., and Breckon, T. P.
(2018). Ganomaly: Semi-supervised anomaly detec-
tion via adversarial training. In Asian conference on
computer vision, pages 622–637. Springer.
Akc¸ay, S., Atapour-Abarghouei, A., and Breckon, T. P.
(2019). Skip-ganomaly: Skip connected and adversar-
ially trained encoder-decoder anomaly detection. In
2019 International Joint Conference on Neural Net-
works (IJCNN), pages 1–8. IEEE.
Bergman, L., Cohen, N., and Hoshen, Y. (2020). Deep
nearest neighbor anomaly detection. arXiv preprint
arXiv:2002.10445.
Simultaneous Optimization of Abnormality Discriminator and Illumination Conditions for Image Inspection of Textile Products
759

Bergmann, P., L
¨
owe, S., Fauser, M., Sattlegger, D., and Ste-
ger, C. (2019). Improving unsupervised defect seg-
mentation by applying structural similarity to autoen-
coders.
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn,
D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer,
M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby,
N. (2021). An image is worth 16x16 words: Trans-
formers for image recognition at scale.
Perera, P. and Patel, V. M. (2019). Learning deep features
for one-class classification. IEEE Transactions on Im-
age Processing, 28(11):5450–5463.
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,
S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,
Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).
ImageNet Large Scale Visual Recognition Challenge.
International Journal of Computer Vision (IJCV),
115(3):211–252.
Sun, C., Shrivastava, A., Singh, S., and Gupta, A. (2017).
Revisiting unreasonable effectiveness of data in deep
learning era. In Proc. ICCV2017.
Ueda, T., Kawahara, R., and Okabe, T. (2024). Learn-
ing projection patterns for direct-global separation. In
Proc. the 19th International Conference on Computer
Vision Theory and Applications (VISAPP2024), pages
599–606.
VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications
760