Cvpr21 unsupervised anomaly detection cutpaste:self supervised learning for anomaly detection and localization

Original address

CutPaste CVPR21 original text

Thesis reading methods

Three times thesis method

First time to know

This paper mainly proposes a new augmentation method “CutPaste”, Based on this method, a two-stage unsupervised anomaly detection method is proposed ,① Build a classifier to distinguish normal samples from abnormal samples , As an agent （proxy task） Learning feature representation ;② Extract features according to the trained network , Used to detect anomalies . In this method MVTec New performance on data sets SOTA( Test task 96.6AUC, Positioning tasks 96AUC).

CutPaste： Cut a block area randomly from the image , Then paste it randomly on an area of the image .

Know each other

The core technology

CutPaste
The figure above shows a normal sample 、 Abnormal samples and Cutout and CutPaste Equal augmented method . First of all, the author will Cutout Apply to the two-stage framework , This seems to be a classification problem that can be solved with a simple filter , But what is surprising is that it can learn some discriminant features , For anomaly detection . So the author makes the task more difficult , Change the scale and change the color (Scar), This further improves performance , See the experimental results below .

On this basis, the author further puts forward CutPaste：1. Cut a small rectangular area with a certain area ratio and length width ratio from the original drawing ;2. Select this image block and color jitter ;3. Paste it back to a certain position in the original drawing . Besides , It also produced scar variant , Although the two operations are very similar , But the image effect is very different , So we use 3 classification , take Cutpaste and Cutpaste-scar As two classes , The effect has been further improved .

Calculation of abnormal score
This paper adopts a simple density estimation method with parameters （ Nonparametric estimation requires a large number of samples , High computational complexity ）GDE（ Gaussian kernel density estimator ）, As shown below ：

Patch-Level
For precise positioning , Use Patch-level The way , Use patch Level image block as input . Before training, an image block is randomly cropped from the original image as input , Then the rest of the operation is the same as before . For each image block , After getting the abnormal score , Use Gaussian smoothing （Gaussian Smoothing） Pass the score to each pixel .

The specific experiment is ： Zoom the image to 256x256,patch The size is 64x64, When testing , In steps of 4 Intensive Forecasting , After getting an abnormal score , First, Gaussian blur is performed, and then up sampling is performed to the original image .

Finally, the whole framework is as follows ：

The theoretical analysis

This part is very important , Improve the height of the paper , It also brings me great inspiration

CutPaste The success of can be analyzed from two aspects , One is from abnormal exposure (outlier exposure) The angle of , Compared with using natural images as abnormal samples ,CutPaste Build directly during training , And the task is more difficult , This further urges the model to learn this irregularity .

On the other hand ,CutPaste It can be seen as a simulation of real anomalies , from t-SNE The picture shows , Tectonic CutPaste Although it does not overlap with the real abnormal samples , However, it widens the distance between normal samples and abnormal samples in the feature space .

experimental analysis

Here we mainly put Detection The experimental results of the task , For more information, please refer to the original

review

When I first read this article ,CVPR Not released yet , Later, when I investigated again , I found them hit . Judging from the whole article , The method is really simple and effective , And the intention is very high , The experiment is also well done .

However, there are also some areas worth improving , such as object Quasi random cutpaste It is possible to paste the background area directly onto the background ？ Will forcibly dividing them into different categories damage the learning of representation ？ Can this expansion form be well adapted to each task ？

In fact, the article also tells us the answer , about object Some classes of , The effect is really bad （ Even in fine-tuning The effect is better than pre-trained Worse ）

Open source code （ unofficial ）

The code is not officially open source , An unofficial code is attached here first , Thank you for the selfless dedication of the author .

https://github.com/Runinho/pytorch-cutpaste