Fastflow [abnormal detection: normalizing flow]

Preface

The blog is from 2021.11CVPR A paper on ,papers with code The website is organized in MVTec The data set is ranked 2（ As of the time of posting ）.

Source code （ unofficial ）：https://github.com/gathierry/FastFlow

background

Methods to obtain global and local features ： Distribution 、 Multiscale 、 The sliding window . Typical distributions are center of an n-sphere/Gaussian/KNN(k nearest neighbors)

thought ： Obtain normal image features through pre training network , Then use statistical methods to model the distribution , Test by distribution .Normalizing Flow It is a kind of statistical method . The training process is continuous fitting to achieve the maximum probability of the occurrence of this pile of normal image features .

FastFlow It can be used as an insertion module , Use different feature extractors （vision transformer（ The last layer of features has a strong ability to capture the relationship between local and global ）or resnet）.

Model principle

Feeling is Glow Modification of the model , The coupling layer changes the convolution block （ Alternating convolution kernels are designed to balance accuracy and speed ）, Sorting layer use RealNVP Random ordering of models , Is the road to Jane .

experiment

Comparative experiments ： Compare the execution speed with different models , Parameter quantity , Image detection 、 Positioning accuracy ; Compare with different models on different data sets .

Ablation Experiment ： Choose different convolution kernel combinations and different feature extractors for comparison .

Add

The paper mentioned CutPaste（ Density estimation , Multi dimensional Gaussian distribution calculates mean variance , Using clustering algorithm ）,vision transformer（ Be able to learn global and local information , Two typical methods Deit、CaiT）

Reference resources ：

FrEIA：https://blog.csdn.net/qq_41804812/article/details/124477478