Mspba [anomaly detection: representation_based]

Preface

2022.6CVPR A paper on , The effect ranks No 9（papers with code）

background

SVDD do baseline

Model principle

thought ： Train the network with all normal data sets （ The four losses used can reduce the distance difference between different pixels SVDD/Cos/Kmean, Direction difference SSL）, Then use the distance between all the characteristic maps output by all the normal maps through the network and the test map to detect and locate .

The process ： Extract 64*64/32*32/16*16 The block , Send in separately Encoder64/Encoder32/Encoder16（ In the process of the training VGG16 The Internet ） The extracted features , After four losses .

Four kinds of losses

SVDD Loss

effect ： Gather the same similar semantic information .

operation ： Random selection patch_i, Then randomly select those with pixel spacing patch_j,64/32/16 The pixel intervals of are 4/2/1.patch_k From away patch_i Far apart patch Randomly choose .

Cos Loss

effect ： Make two more distant patch The semantic similarity is small .

operation ： And SVDD identical .

Kmean Loss

operation ： Random selection patch_r.

SSL Loss

operation ： Random selection patch_m, Then choose randomly 64/32/16 Pixel spaced patch_n. The direction chosen is 30 Degrees are spaced 12 Angles .

test

classification - testing ： Use a sliding window to block the diagram , Then calculate the average value of pixels in the block , The maximum value of all block scores is taken as the detection score . This method can weaken the influence of outlier pixel values on the results .

Division - location ：64/32/16 Of patch Block by step 16/4/4 selection , Compare with all feature maps formed by all training sets to obtain the minimum value of the distance at each pixel position in each block , Then average the overlapping values at the same pixel position , Then put different sizes patch The value graphs formed by blocks are added , Get the score map .

experiment

Different patch Comparison of block size selection ; The impact of different losses on the network effect .