MixedTeacher: Knowledge Distillation for Fast Inference Textural
Anomaly Detection
Simon Thomine
1,2
, Hichem Snoussi
1
and Mahmoud Soua
2
1
University of Technology Troyes, Troyes, France
2
AQUILAE, Troyes, France
Keywords:
Anomaly Detection, Texture, Knowledge Distillation, Layer Selection, Unsupervised.
Abstract:
For a very long time, unsupervised learning for anomaly detection has been at the heart of image processing
research and a stepping stone for high performance industrial automation process. With the emergence of
CNN, several methods have been proposed such as Autoencoders, GAN, deep feature extraction, etc. In this
paper, we propose a new method based on the promising concept of knowledge distillation which consists
of training a network (the student) on normal samples while considering the output of a larger pretrained
network (the teacher). The main contributions of this paper are twofold: First, a reduced student architecture
with optimal layer selection is proposed, then a new Student-Teacher architecture with network bias reduction
combining two teachers is proposed in order to jointly enhance the performance of anomaly detection and its
localization accuracy. The proposed texture anomaly detector has an outstanding capability to detect defects
in any texture and a fast inference time compared to the SOTA methods.
1 INTRODUCTION
Anomaly detection in industry is a vast topic since
there is a lot of possible applications. For instance,
defect detection aims at identifying specific anomaly
classes and locations in industrial manufacturing
processes (K
¨
ahler et al., 2022). This detection is
crucial for ensuring the high quality of final products
(Minhas and Zelek, 2019). A common property
of defects is that their visual texture is inherently
different from the defect-free surface. The specificity
of textures is the pattern structure which, if known,
allows the detection and the extraction of anomalies.
However, the texture anomaly generally appears in a
small region in few samples, which makes it difficult
to build consistent normal and abnormal datasets
to be used in supervised learning methods. Hence,
unsupervised anomaly detection networks are very
suitable for industrial scenarios as they represent the
strong basis for building a detection model without
any annotated samples (Huang et al., 2022). Several
unsupervised anomaly detection methods have been
introduced for texture anomaly detection. These
methods could achieve high performance up to
99.6 AUROC. However, they suffer from complex
networks and high latency.
In another context, knowledge distillation has been
introduced with the purpose of reducing the network
size while increasing performance. Knowledge dis-
tillation aims to train a smaller network (student) to
imitate pretrained one or several larger ones (teach-
ers) on normal samples. As the teacher is pretrained,
it has the ability to generalize even if the sample con-
tains an anomaly, whereas the student won’t be able.
Hence, by comparing the extracted features between
the teacher and the student networks, an abnormal
sample could be detected. According to some stud-
ies (Iglesias and Zseby, 2015), using too many fea-
tures can significantly reduce the accuracy of anomaly
detection. Recently, a Student-Teacher Feature Pyra-
mid Matching (STPM) method has been proposed in
(Wang et al., 2021), where the first three network lay-
ers are used in order to focus on edges, colors and
shapes instead of context information. Even if using
layer selection technique is an interesting approach,
there is still a lack of explanation concerning the layer
choice and the relevance of the relative information.
Looking at the same layers for an object and for a
texture reduces the relevance of the extracted infor-
mation. For example, looking at context informa-
tion in a texture is pointless and for an object, pure
edge/color/texture information may not be the most
interesting information.
Thomine, S., Snoussi, H. and Soua, M.
MixedTeacher: Knowledge Distillation for Fast Inference Textural Anomaly Detection.
DOI: 10.5220/0011633100003417
In Proceedings of the 18th Inter national Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages
487-494
ISBN: 978-989-758-634-7; ISSN: 2184-4321
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
487

Another recurrent problem is the classifier bias. The
best current methods use a pretrained classifier net-
work on imageNet which is biased by the classes of
imageNet and can have an impact on the localization
and the detection of defects.
The main contributions of the paper are as follows:
• A new reduced student architecture for texture-
specific object category.
• In order to reduce the classification bias, we pro-
pose a new architecture combining two teachers
pretrained on imageNet but with different archi-
tectures (respectively ResNet-18 (He et al., 2015)
and EfficientNet-b0 (Tan and Le, 2020)) and a sin-
gle student network. This new mixed Teacher net-
work structure outperforms competitive state-of-
the-art methods both in inference time and SOTA
scores, on anomaly datasets such as MVTEC AD
textures and BTAD textures (Mishra et al., 2021).
The proposed MixedTeacher model uses a score
and anomaly localisation function based on each
complementary teacher features with a careful
feature selection.
The paper is organized as follows. In section 2,
we review the related work, especially on MVTEC
dataset and present the different approaches proposed
in literature. In section 3, we compare the results
of training with different architectures and different
layer selection schemes and introduce our proposed
texture-specific reduced student architecture. Section
4 is dedicated to describing a novel mixed Student-
Teacher network. In section 5, we compare our results
to the SOTA methods for both the reduced student ar-
chitecture and the MixedTeacher in terms of AUROC,
pixel-AUROC and inference time.
2 RELATED WORK
Anomaly detection is a problem that pops up in many
areas and is often very difficult to deal with. Indeed,
detecting the “abnormal” is a rather vague concept
and is difficult to define according to the use cases,
which makes research on this subject very specific.
For several years, the rise of deep learning has
never ceased to impress with high quality results and
interesting methods. Most of these methods are based
on an unsupervised representation approach to dis-
criminate outliers. Some specific work has been done
for fabrics defect detection such as the multi-scale
Convolutionnal denoising autoencoder (Mei et al.,
2018). For unsupervised anomaly detection in gen-
eral, we can also cite the GEE, a gradient based VAE
(Nguyen et al., 2019) or the Gaussian mixture model
VAE (Nguyen et al., 2019). Another common way
to detect anomaly is to use generative adversarial net-
works (Goodfellow et al., 2014). Ano-GAN (Schlegl
et al., 2019) was one of the first utilization of GAN for
anomaly detection but since then a lot of approaches
emerged such as G2D (Pourreza et al., 2021) and
OCR-GAN (Liang et al., 2022). Other interesting
approaches rely on pretrained models especially on
imageNet, using the feature extraction of pretrained
network to extract useful information about a given
sample. The idea is to extract features with a pre-
trained model and then train a normalizing flow model
on good samples, so that the model is ready to find
out if a given sample is an anomaly by looking at
the reconstruction error. An advantage of normaliz-
ing flow is the reversible aspect which is useful to lo-
cate the anomaly pixel-wise. Many techniques based
on this concept have been proposed such as differ-
Net (Rudolph et al., 2021a) and CS-FLOW (Rudolph
et al., 2021b) which consider multi-scale normalizing
flow and FastFlow (Yu et al., 2021) based on a 2D
normalizing flow.
Recently, the concept of knowledge distillation
has also been used for unsupervised anomaly detec-
tion. The student-teacher method consists of train-
ing a student teacher based on the output of a larger
teacher model which is pretrained on ImageNet. The
student network will learn to imitate the teacher on
good samples only. Then, when an abnormal sample
is tested, the teacher will be able to generalize and the
student won’t be, the difference between the output
of the teacher and the output of the student will allow
the detection of the anomaly. On the MVTEC dataset,
four methods have been implemented, STPM (Wang
et al., 2021) which trained the student on the 3 first
layers of ResNet-18, RSTPM (Yamada and Hotta,
2022) which is basically the same method but with
an attention layer, reverse distillation (Deng and Li,
2022) and CFA (Lee et al., 2022).
3 LAYER SELECTION AND
REDUCED STUDENT
In this section, after a comparative study of layer se-
lection methods for optimal texture anomaly detec-
tion, we present a new student architecture that both
increases performance and reduces the inference time.
3.1 Layer Selection
In deep neural networks, a common observation is
that deep layer features contain context information
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
488

and shallow layer features contain color, texture and
contour information. In a case of detection of defects
on the fabric or on a generic texture, the context infor-
mation is less important than the texture information,
therefore, we will turn to shallow layer features. As
reported in table 1, different combinations of shallow
layers have been tried in order to select the optimal ar-
chitecture with respect to detection performance eval-
uated by the AUC.
Table 1: Layers selection results.
Measures Layer 1 and 2 AUC Layer 2 and 3 AUC
Mean objects 0.876 0.910
Mean textures 0.990 0.971
3.2 Reduced Student
ResNet-18 architecture has been retained for the
teacher network. As texture specific anomaly detec-
tion is the main objective of this work, we propose to
add the ResNet-18 first layer after the first convolu-
tion to extract even more textural information. The
second objective was to alleviate the student architec-
ture to decrease inference time and possibly perfor-
mance. As ResNet-18 presents several residual blocks
with two identical convolutional layers, we first de-
cided to take only one layer for each block in our
student architecture. The classifier bias is another
known problem while dealing with pretrained classi-
fier and we tackled this problem by reducing features
size with an adaptive average pooling layer in each
Resnet residual block’s output as presented in figure
1.
Given a training dataset of images without
anomaly D = [I
1
,I
2
,...,I
n
], our goal is to extract the
information of the L bottom layers. For an image
I
k
∈ R
w∗h∗c
where w is the width, h the height and
c the number of channels, the teacher outputs fea-
tures F
l
t
(I
k
) ∈ R
w
l
∗h
l
∗c
l
and F
l
s
(I
k
) ∈ R
w
l
/2∗h
l
/2∗c
l
/2
with l > 1 and F
l
s
(I
k
) ∈ R
w
l
∗h
l
∗c
l
if l = 1. For the loss
function, we took the l2 distance of normalized fea-
ture vectors like in the STPM original paper (Wang
et al., 2021) while using an adaptive average pooling
on teacher features where l > 1 and just sum all fea-
ture maps of all layers to obtain our loss with the same
ratio for all layers (Eq.1).
F
l>1
t
(I
k
) = AAP(F
l>1
Resnet18
(I
k
)) (1)
where AAP refers to the Adaptive Average Pooling.
Pixel loss is defined in the following Eq.2:
loss
l
(I
k
)
i j
=
1
2
∥norm(F
l
t
(I
k
)
i j
) − norm(F
l
s
(I
k
)
i j
)∥
(2)
and for the layer l, the loss is defined as:
loss
l
(I
k
) =
1
w
l
h
l
w
l
∑
i=1
h
l
∑
j=1
loss
l
resNet
(I
k
)
i j
(3)
and finally for the total loss is written as:
loss(I
k
) =
l
∑
loss
l
(I
k
) (4)
Performance and inference speed are later re-
ported in section 5 with comparison with SOTA net-
works on anomaly detection.
Figure 1: Reduced student architecture with AP for adaptive average pooling.
MixedTeacher: Knowledge Distillation for Fast Inference Textural Anomaly Detection
489

4 MIXED TEACHER
In this section, we introduce our new student teacher
network structure that combines two teachers with the
purpose of reducing the classifier bias, taking bene-
fits from the two networks and exploiting the different
layers in an optimal way.
4.1 Observation and Main Ideas
While testing our new student reduced architecture on
the MVTEC AD textures, we obtained good results,
but some noise still degrade results in terms of default
localisation on specific images or texture-specific nor-
mal variation. Different teacher network architec-
tures have been tested to conclude that ResNet-18
remains the best in terms of average precision and
speed. However, interesting behaviors have been ob-
served on the noise localisation for each architecture.
In fact, every classifier had the capacity to locate the
anomaly, but with output noise and anomaly detection
mistakes.
The combination of two pretrained classifier net-
works has therefore been proposed with the purpose
of interpolating their defect localisation to cancel
noise and false detection/segmentation.
EfficientNet-b0 has been proposed as the second
teacher when considering its performance in terms of
precision and speed. For this network, it has been
observed that for the bottom layers, one has good lo-
calisation but with a huge noise and with top layers, a
coarse defect localisation but with minimal noise has
been obtained, as illustrated in figure 2.
Figure 2: Difference between top layers and bottom layers
for EfficientNet-b0 architecture.
4.2 Method Description
The learning architecture is composed of two teach-
ers: the ResNet-18 as main teacher and EfficientNet-
b0 as a localisation confirmation teacher. For the
ResNet-18 part, the reduced student proposed in
section 3 is used in order to ensure a good infer-
ence speed and precision on texture samples. For
EfficientNet-b0 student, we used one convolution for
each efficientnet block without pooling because we
used deepest layers. In the student architecture, there
is no communication between the networks except for
the loss function as shown in figure 3.
For the training loss function, we used basically
the same loss function as the one for the reduced
teacher and we add an α factor to smooth the layer ac-
tivation difference from the two teacher networks. As
feature difference in efficientNet was about 10 times
bigger than in ResNet-18, α has been set to 0.1.
loss
l=5,6
e f f Net
(I
k
)
i j
=
1
2
∥norm(F
l
t
(I
k
)
i j
) − norm(F
l
s
(I
k
)
i j
)∥
(5)
and
loss
l=5,6
e f f Net
(I
k
) =
1
w
l
h
l
w
l
∑
i=1
h
l
∑
j=1
loss
l
e f f Net
(I
k
)
i j
(6)
and for Resnet-18 part :
loss
l=1,2,3
resNet
(I
k
) =
1
w
l
h
l
w
l
∑
i=1
h
l
∑
j=1
loss
l
resNet
(I
k
)
i j
(7)
with loss
l
resNet
(I
k
)
i j
defined as in section 3. For the
total loss with the α factor :
loss
tot
(I
k
) =
3
∑
l=1
loss
l
resNet
(I
k
) + α
6
∑
l=5
loss
l
e f f Net
(I
k
)
(8)
As in every knowledge distillation method, the
loss only impacts the student.
4.3 Anomaly Score and Localisation
In the test phase (inference), we want an anomaly map
M of the original image size where every pixel at po-
sition (i, j) has an anomaly score M
i j
. With a test im-
age I and F
l
tResnet
, F
l
tE f f Net
the two teachers features
of lth layer and F
l
sResnet
, F
l
sE f f Net
their corresponding
lth layer student features, we perform an upsample on
the difference between the corresponding layers. The
coarse localisation output of the efficientNet layers is
obtained by summing each layer’s anomaly map.
The anomaly map is obtained in the same way for
the resnet part. Respectively :
A
mapE f f net
=
6
∑
l=5
U psample(F
l
tE f f Net
− F
l
sE f f Net
)
(9)
and :
A
mapResnet
=
3
∑
l=1
U psample(F
l
tResnet
− F
l
sResnet
) (10)
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
490

Figure 3: MixedTeacher architecture.
We then multiply the resnet anomaly map by the
normalization of the effnet anomaly map multiplied
by its mathematical extent. With A
mapE f f net
, the
anomaly map of efficientNet layers and A
mapResnet
the
anomaly map of resnet layers, the final anomaly map
is then defined as :
M = A
mapResnet
∗ (max(A
mapE f f net
)−
min(A
mapE f f net
))A
mapE f f net
(11)
The anomaly score is defined as :
score =
w
∑
i=1
h
∑
j=1
M
i, j
(12)
with w and h are respectively the width and height of
the anomaly map.
5 EXPERIMENTS
5.1 Datasets
We experiment our methods on the textures from the
MVTEC AD (Bergmann et al., 2019) dataset which
consists of 15 categories : 5 textures and 10 objects
with a total of more than 5000 high resolution im-
ages. This dataset is used for unsupervised anomaly
detection therefore it contains only anomaly free im-
ages for the training. For the test part, it shows a
good variety of defects with ground truth masks for
anomaly localisation. We also used the texture of the
BTAD (Mishra et al., 2021) dataset which is an un-
supervised anomaly dataset with three different cate-
gories including one texture figure 4.
Figure 4: Overview of textures from MVTEC AD and
BTAD dataset, samples with anomaly and ground truth.
These images are only used for testing and unseen during
the training.
The performance is evaluated with AUROC met-
ric image-level and pixel-level to compare our results
with other methods.
MixedTeacher: Knowledge Distillation for Fast Inference Textural Anomaly Detection
491

5.2 Implementation and Training
Metrics
Training and inference were done on an rtx 2080ti.
To test the student reduced, we used the features
of the first three blocks and the layer before the first
block of ResNet-18. The Resnet network was pre-
trained on imageNet. We used stochastic gradient de-
scent with a learning rate of 0.4 for 100 epochs with
a batch size of 16. To test the MixedTeacher, we used
the output features of the first two blocks and the layer
before the first block of ResNet-18 and the output fea-
tures of block 5 and 6 of EfficientNet-b0. We used
stochastic gradient descent with a learning rate of 0.4
for 200 epochs with a batch size of 16. Both networks
are pretrained on imageNet. We resized all the im-
ages to a size of 256x256 keeping 80% for training
and 20% for validation. We kept the checkpoint with
the lowest validation loss.
5.3 Reduced Student
5.3.1 Performance Results
In this paragraph, we will compare reduced student
AUROC results to SOTA methods. In 2, we present
AUROC performance results of CFA (Lee et al.,
2022), PatchCore (Roth et al., 2021), FastFlow (Yu
et al., 2021), STPM (Wang et al., 2021), CutPaste (Li
et al., 2021) and our reduced student on MVTEC AD
textures.
Table 2: Image-AUROC comparison on MVTEC AD : Re-
duced Student.
Category CutPaste CFA PatchCore STPM FastFlow Ours
carpet 100 97.3 98.7 95.4 99.4 100
tile 100 99.4 98.7 94.9 100 98.7
grid 96.2 99.2 98.2 98.2 100 99.7
wood 99.1 99.7 99.2 96.1 99.2 99.6
leather 95.4 100 100 98.9 99.9 99.7
Mean 98.1 99.1 99.0 96.7 99.7 99.5
For FastFlow, we choosed to take the results from
Anomalib as we were not able to reproduce their pa-
per results (99.9 AUROC in paper). As seen in ta-
ble 2, our reduced student is better than CFA for
texture anomaly detection, which is the best actual
knowledge distillation unsupervised anomaly detec-
tion method and is close to the SOTA results. We
manage to gain 2.8 points against classic STPM with
a network reduction and a wise layer selection aiming
for texture specific anomaly detection.
5.3.2 Inference Time Results
In table 3, we compare the reduced student inference
time to other SOTA methods. The main purpose of
reduced student was to propose a high processing
speed to manage real time for several high resolu-
tion images. To get inference time results, we employ
Anomalib. All the additional results come from this
library to make sure the tests were carried out under
the same conditions.
Table 3: Inference time results.
Category PatchCore FastFlow STPM Ours
FPS 5.8 21.8 83.2 108.1
Latency (ms) 172 45.9 12 9.2
The presented results are based on Anomalib in-
ference time. In a self made code, we were able to
obtain a 10x better inference time for STPM and re-
duced student. The most important thing to consider
is that the STPM is by far the fastest anomaly detec-
tor and reduced student reduced its inference time by
30%.
5.4 MixedTeacher
5.4.1 Performance Results
Unlike the reduced student, the MixedTeacher main
purpose is performance and not inference time. In ta-
ble 4 we compared AUROC of several SOTA methods
in texture anomaly detection.
Table 4: Image-AUROC comparison on MVTEC AD :
MixedTeacher.
Category CutPaste CFA PatchCore FastFlow ReducedStudent Ours
carpet 100 97.3 98.7 99.4 100 99.8
tile 100 99.4 98.7 100 98.7 100
grid 96.2 99.2 98.2 100 99.7 99.7
wood 99.1 99.7 99.2 99.2 99.6 99.6
leather 95.4 100 100 99.9 99.7 100
Mean 98.1 99.1 99.0 99.7 99.5 99.8
Our method is the new state of the art texture
anomaly detection on the MVTEC AD dataset.
5.4.2 Anomaly Localisation
Even though anomaly localisation was not our main
purpose, our approach uses EfficientNet-b0 with the
objective of making the location more precise. To this
end, we present in table 5 and table 6, our anomaly
location results on textures from MVTEC AD dataset
and BTAD respectively and we compare these results
to the SOTA methods.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
492

Table 5: Pixel-AUROC comparison on MVTEC AD :
MixedTeacher.
Category CutPaste PatchCore FastFlow Ours
carpet 98.3 98.9 99.1 99.0
tile 90.5 95.6 96.6 95.9
grid 97.5 98.7 99.2 97.5
wood 95.5 95 94.1 94.9
leather 99.5 99.3 99.6 99.4
Mean 96.2 97.5 97.7 97.3
Table 6: Image-AUROC comparison on BTAD:
MixedTeacher.
Category FastFlow Ours
1 (wood from btad) 96.0 97.0
5.4.3 Inference Time Results
In terms of inference speed, our MixedTeacher is 3x
slower than the reduced student since it used two
teacher networks and a more complex student archi-
tecture.
6 CONCLUSION
In this paper, we proposed two methods for effi-
cient unsupervised anomaly detection using the prin-
ciple of knowledge distillation applied to unsuper-
vised anomaly training. Both methods offer differ-
ent benefits. The reduced student proposes a high
speed texture anomaly detector with an AUROC per-
formance close to the state of the art, this method is
to be used in situations where inference time is the
most important priority (mobile device, low computa-
tional power, cost efficiency). The MixedTeacher pro-
pose the highest actual performance on anomaly de-
tection with a localisation close to the state of the art
on the MVTEC AD textures with still a fast inference.
This method is to be used in situations where perfor-
mance is the priority and the computational power is
big enough (monitoring server etc ...)
REFERENCES
Bergmann, P., Fauser, M., Sattlegger, D., and Steger, C.
MVTec AD — a comprehensive real-world dataset for
unsupervised anomaly detection. In 2019 IEEE/CVF
Conference on Computer Vision and Pattern Recogni-
tion (CVPR), pages 9584–9592. IEEE.
Deng, H. and Li, X. Anomaly detection via reverse distilla-
tion from one-class embedding.
Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B.,
Warde-Farley, D., Ozair, S., Courville, A., and Ben-
gio, Y. Generative adversarial networks.
He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learn-
ing for image recognition.
Huang, J., Li, C., Lin, Y., Lian, S., and Innovation, A. Unsu-
pervised industrial anomaly detection via pattern gen-
erative and contrastive networks.
Iglesias, F. and Zseby, T. Analysis of network traffic fea-
tures for anomaly detection. 101(1):59–84.
K
¨
ahler, F., Schmedemann, O., and Sch
¨
uppstuhl, T.
Anomaly detection for industrial surface inspection:
application in maintenance of aircraft components.
107:246–251.
Lee, S., Lee, S., and Song, B. C. CFA: Coupled-
hypersphere-based feature adaptation for target-
oriented anomaly localization.
Li, C.-L., Sohn, K., Yoon, J., and Pfister, T. CutPaste: Self-
supervised learning for anomaly detection and local-
ization.
Liang, Y., Zhang, J., Zhao, S., Wu, R., Liu, Y., and Pan, S.
Omni-frequency channel-selection representations for
unsupervised anomaly detection.
Mei, S., Wang, Y., and Wen, G. Automatic fabric defect
detection with a multi-scale convolutional denoising
autoencoder network model. 18(4):1064.
Minhas, M. S. and Zelek, J. AnoNet: Weakly supervised
anomaly detection in textured surfaces.
Mishra, P., Verk, R., Fornasier, D., Piciarelli, C., and
Foresti, G. L. VT-ADL: A vision transformer net-
work for image anomaly detection and localization.
In 2021 IEEE 30th International Symposium on In-
dustrial Electronics (ISIE), pages 01–06.
Nguyen, Q. P., Lim, K. W., Divakaran, D. M., Low, K. H.,
and Chan, M. C. GEE: A gradient-based explainable
variational autoencoder for network anomaly detec-
tion.
Pourreza, M., Mohammadi, B., Khaki, M., Bouindour, S.,
Snoussi, H., and Sabokrou, M. G2d: Generate to de-
tect anomaly. In 2021 IEEE Winter Conference on Ap-
plications of Computer Vision (WACV), pages 2002–
2011. IEEE. event-place: Waikoloa, HI, USA.
Roth, K., Pemula, L., Zepeda, J., Sch
¨
olkopf, B., Brox,
T., and Gehler, P. Towards total recall in industrial
anomaly detection.
Rudolph, M., Wandt, B., and Rosenhahn, B. Same same
but DifferNet: Semi-supervised defect detection with
normalizing flows. In 2021 IEEE Winter Conference
on Applications of Computer Vision (WACV), pages
1906–1915. IEEE. event-place: Waikoloa, HI, USA.
Rudolph, M., Wehrbein, T., Rosenhahn, B., and Wandt, B.
Fully convolutional cross-scale-flows for image-based
defect detection.
Schlegl, T., Seeb
¨
ock, P., Waldstein, S. M., Langs, G.,
and Schmidt-Erfurth, U. f-AnoGAN: Fast unsuper-
vised anomaly detection with generative adversarial
networks. 54:30–44.
Tan, M. and Le, Q. V. EfficientNet: Rethinking model scal-
ing for convolutional neural networks.
Wang, G., Han, S., Ding, E., and Huang, D. Student-teacher
feature pyramid matching for anomaly detection.
MixedTeacher: Knowledge Distillation for Fast Inference Textural Anomaly Detection
493

Yamada, S. and Hotta, K. Reconstruction student with at-
tention for student-teacher pyramid matching.
Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., and
Wu, L. FastFlow: Unsupervised anomaly detection
and localization via 2d normalizing flows.
VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications
494