ECCV 2022 | can be promoted without fine adjustment! Registration based anomaly detection framework for small samples

author |  Wang Yanfeng 、 Zhang Ya

Company |  Shanghai Jiaotong University 、 Shanghai Artificial Intelligence Laboratory

source  |  Almost Human

In recent years , Anomaly detection in industrial defect detection 、 Medical diagnosis , Automatic driving and other fields have a wide range of applications .“ abnormal ” Usually defined as “ normal ” Opposites , That is, all samples that do not meet the normal specifications . Generally speaking , Compared to normal , The types of abnormal events are inexhaustible , And very rare , Difficult to collect , Therefore, it is impossible to collect detailed abnormal samples for training . therefore , Recent research on anomaly detection focuses on unsupervised learning , That is, use only normal samples , By using a single category （one-class） classification , Image reconstruction （reconstruction）, Or other self supervised learning tasks to model normal samples , after , Detect anomalies by identifying samples that are different from the model distribution .

Most existing anomaly detection methods focus on training a special model for each anomaly detection task . However , In real scenarios such as defect detection , Considering hundreds of industrial products to be processed , It is not cost-effective to collect a large number of training sets for each product . Regarding this , Shanghai Jiaotong University MediaBrain The team and the intelligent medical team of Shanghai Artificial Intelligence Laboratory have proposed a small sample anomaly detection framework based on registration , By learning the common model shared among multiple anomaly detection tasks , No need to adjust model parameters , It can be extended to new anomaly detection tasks . at present , The study has been ECCV2022 Received as Oral The paper , The complete training code and model have been open source .

Paper title ：

Registration based Few-Shot Anomaly Detection

Thesis link ：

https://arxiv.org/abs/2207.07361

Code link ：

https://github.com/MediaBrain-SJTU/RegAD

Method introduction

In this work , The training of a general model for anomaly detection with small samples is inspired by how humans detect anomalies . in fact , When trying to detect anomalies in the image , People usually compare the test sample with a sample that has been determined to be normal , To find out the difference , The part with difference can be regarded as abnormal .

In order to realize this process similar to human comparison , The author of this paper adopts the registration technology . The author of this paper holds that , For registration networks , Just know how to compare two extremely similar images , The actual semantics of images are no longer important , Therefore, the model is more suitable for new tasks of anomaly detection that have never been seen . Registration is especially suitable for anomaly detection with few samples , Because registration can be very convenient for cross category promotion , The model can be quickly applied to new anomaly detection tasks without parameter tuning .

The above figure outlines the framework of small sample anomaly detection based on registration . And conventional anomaly detection methods （one-model-per-category） Different , This work （one-model-all-category） First, a general anomaly detection model based on registration is trained by using multi category data . Normal images from different categories are used together in the joint training model , Randomly select two images from the same category as training pairs . At testing time , A support set consisting of several normal samples is provided for the target category and each test sample . Given the support set , A statistical based distribution estimator is used to estimate the normal distribution of the registration characteristics of the target category . Test samples that exceed the statistical normal distribution are considered abnormal .

This work uses a simple registration network , At the same time, it refers to Siamese [1], STN [2] and FYD [3]. To be specific , With twin neural network （Siamese Network） For the framework , Insert spatial transformation network （STN） Realize feature registration . For better robustness , The author uses the registration loss of feature level , Instead of pixel by pixel registration like typical registration methods , This can be regarded as a relaxed version of pixel level registration .

experimental result

In comparison with other anomaly detection methods with small samples ,RegAD Whether in testing performance 、 It is applicable to the adaptive time of new category data , Compared with the benchmark method TDG [4] and DiffNet [5] Have significant advantages . This is because other methods require multiple rounds of iterative updating of the model for new category data .

in addition ,RegAD Compared with the version without joint training of multi category feature registration （RegAD-L）, Performance has also been significantly improved , It shows that the training of general anomaly detection model based on registration is very effective . In this paper, anomaly detection data set MVTec [6] and MPDD [7] Experiment on . For more experimental results and ablation experiments, please refer to the original paper .

Besides , The author also shows the results of visualization of anomaly location . You can see , Joint training can make the anomaly location of the model more accurate .

T-SNE The visualization of also shows , Registration based training can make normal image features of the same category more compact , Thus, it is conducive to the detection of abnormal data .

summary

This work mainly explores a challenging but practical setting of anomaly detection ：1） Train a single model for all anomaly detection tasks （ It can be promoted without fine adjustment ）;2） Only a few new category images are available （ Few samples ）;3） Only normal samples are used for training （ Unsupervised ）.

Trying to explore this setting is an important step towards the actual large-scale industrial application of anomaly detection . To learn category independent models , This paper proposes a solution based on comparison , This is very different from the popular methods based on reconstruction or single classification . The specific registration model is based on the existing registration scheme , Full reference to the existing outstanding work [1,2,3], Without parameter adjustment , Impressive detection results have been achieved on the new anomaly detection data .

reference

[1] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. CVPR. 2021.

[2] Max Jaderberg et. al. Spatial transformer networks. NeurIPS. 2015.

[3] Ye Zheng et. al. Focus your distribution: Coarse-to-fine non-contrastive learning for anomaly detection and localization. arXiv:2110.04538. 2021.

[4] Shelly Sheynin et. al. A hierarchical transformation-discriminating generative model for few shot anomaly detection. ICCV. 2021.

[5] Marco Rudolph et.al. Same same but differnet: Semi-supervised defect detection with normalizing flows. WACV. 2021.

[6] Paul Bergmann et. al. MVTec AD--A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. CVPR. 2019.

[7] Stepan Jezek et. al. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. ICUMT. 2021.

Read more

# cast draft   through Avenue #

  Let your words be seen by more people  

How to make more high-quality content reach the reader group in a shorter path , How about reducing the cost of finding quality content for readers ？ The answer is ： People you don't know .

There are always people you don't know , Know what you want to know .PaperWeekly Maybe it could be a bridge , Push different backgrounds 、 Scholars and academic inspiration in different directions collide with each other , There are more possibilities . 

PaperWeekly Encourage university laboratories or individuals to , Share all kinds of quality content on our platform , It can be Interpretation of the latest paper , It can also be Analysis of academic hot spots 、 Scientific research experience or Competition experience explanation etc. . We have only one purpose , Let knowledge really flow .

  The basic requirements of the manuscript ：

• The article is really personal Original works , Not published in public channels , For example, articles published or to be published on other platforms , Please clearly mark  

• It is suggested that  markdown  Format writing , The pictures are sent as attachments , The picture should be clear , No copyright issues

• PaperWeekly Respect the right of authorship , And will be adopted for each original first manuscript , Provide Competitive remuneration in the industry , Specifically, according to the amount of reading and the quality of the article, the ladder system is used for settlement

  Contribution channel ：

• Send email ：[email protected] 

• Please note your immediate contact information （ WeChat ）, So that we can contact the author as soon as we choose the manuscript

• You can also directly add Xiaobian wechat （pwbot02） Quick contribution , remarks ： full name - contribute

△ Long press add PaperWeekly Small make up

Now? , stay 「 You know 」 We can also be found

Go to Zhihu home page and search 「PaperWeekly」

Click on 「 Focus on 」 Subscribe to our column

·