Nips18 (AD) - unsupervised anomaly detection using geometric transformations using geometric augmentation

Original address

Original text and additional materials

Thesis reading methods

Three times thesis method

First time to know

But from the summary and introduction , In this paper, the idea It's simple and novel ( After all 18 Conference articles in ), The core is to perform various geometric augmentations on the input image to get different versions . Then construct a multi classification model , Let the model predict the expanded categories （ Like rotation 、 Flip or pan, and so on ）.

By performing such a task , The author observed that the model can learn some about anomaly detection （ Image semantic level ） Useful features . At testing time , The test image is also augmented and predicted , According to the model softmax Probability vector to calculate the abnormal score , So as to carry out anomaly detection .

Know each other

The core technology

In fact, the previous section has basically summarized , For a training set containing only normal samples S, Different geometric changes are used for each training sample , Each geometric change represents a certain kind , It should be noted that it also includes identity transformation （ That is to do nothing , The kind that outputs the original picture ）.

Train a multi classification model , The goal is to predict each input image （ It has been expanded by some geometry ） Geometric transformation category of , The loss function is the common cross entropy loss .（ In the process , Any image classification trick Both can be used. ）.

Abnormal score
About abnormal scores , The original Xiaoxiao Sasa describes a large paragraph , There is no lack of a lot of statistical reasoning , But in the end, a simple version was adopted , It is shown in the following formula ：

among ,y(Tj(x)) That is, the input sample adopts j After geometric transformation, it is sent into the trained classification model , And then you get softmax vector , Take the second one j dimension ( This is also the model prediction Tj(x) by j The probability of a class ).k Two transformations (0~k-1), Get the corresponding softmax probability , Make a weighted average , The final value is taken as the anomaly score .

It's not difficult to understand. , In this way, the score of normal samples is generally higher than that of abnormal samples , Abnormal samples may also score well in some form of amplification , but k There are always some changes that stretch , So the score is different from the normal sample .

This is just an intuitive understanding , The specific original text gives reasoning .

experiment

The classification model adopts Wide ResNet, Experiments were carried out on different data sets . See the experimental results in the figure below in detail , It can be seen that the benchmark level has been greatly improved .

among CIFAR-100 Is aimed at 20 Experiments done by subclasses , It shows the effect of the model on multi class distribution anomaly detection .
For multi class distribution , Two classification heads are constructed during training , One is the classification of categories , The other is the classification of geometric transformation ; However, only geometrically varied softmax Calculate the exception score .

There are also some interesting experiments , You can read the original . Including the use of gradient ascent Carry out visual reconstruction to verify the effectiveness of the model , And with 8 and 3 The example of shows the detection effect under different augmentation .

Geometric transformation

This paper is mainly based on Flip horizontal + translation + rotate constructed 72 There are different types of geometric augmentation , The original text did not elaborate , I found the corresponding content from the supporting materials ：

among b Indicates whether to use horizontal flipping ,sw,sh Indicates whether translation is applied in both directions （ translation 25%, Front and back directions ）,k Indicates the angle of rotation （ In steps of 90°）.

So we end up with 2x3x3x4=72 Two geometric transformation methods .

About why geometric transformation is used , The author really shows that some non geometric transformations were used in the early stage , The effect is not good , So only geometric transformation is preserved . Doubt is non geometric transformation （ Gaussian smoothing 、 Sharpen, etc ） It makes the image lose effective features .

review

Published in Nips2018, At that time, self supervision was not on fire , In fact, it is also a self-monitoring task , By constructing a proxy task, Learn useful features for downstream tasks . Look at the full text , It's really two words , Simple 、 It works , A simple idea, Go with a sota Result .

The article is very interesting , The author also mentioned some improvements , I agree with one thing very much . It is currently used for all categories 72 Species augmentation , Is there a way to learn appropriate augmentation for a given data set , There is also the application of this idea to other tasks .

Open source code

Official tf Code ：
https://github.com/izikgo/AnomalyDetectionTransformations