UNET [basic network]

background

Introduce ： Classic model of semantic segmentation network

advantage ： You can get context information and location information at the same time , It is very conducive to the extraction of surface information , It performs well in segmentation that requires shallow information .

Model structure

The first part （ Trunk feature extraction ）： Get five preliminary feature layers . Convolution + Maximum pool composition , Remove the fifth largest pool layer and the following part , Other structures and VGG-16 identical .

The second part （ Strengthen feature extraction ）： Sample the last feature layer , Stack with the penultimate feature layer （ Channels add up ）, After two convolutions, the feature layer after feature fusion is obtained , Then repeat the above steps with different preliminary feature layers to obtain the final feature layer three times .

The third part （ forecast ）： adopt 1*1 Convolution adjusts the number of channels to the number of classes that need to be predicted , Use the final feature layer to classify each pixel .

Reference resources

Han Pi's semantic segmentation and remaking 6——Pytorch Build your own Unet Semantic segmentation platform _Bubbliiiing The blog of -CSDN Blog