Pytorch 20 realizes corrosion expansion based on pytorch

In some image processing projects , Sometimes it is necessary to carry out corrosion expansion and various value processing on the image , introduce opencv After processing , Data cannot be processed quickly , Value range interception 、 Use mask、where Operation etc. . therefore , have access to torch Achieve corrosion expansion , So as to avoid the modification of data structure , At the same time, it ensures that all operations can be realized in one model .（ Of course ,opencv It is also very convenient to intercept the value range in , For details, please refer to section 7 in the following link c++ On opencv Common array operations _a486259 The blog of -CSDN Blog ）

however , be based on torch Achieved corrosion expansion , There are certain assumptions ： All structural elements are square operators . After implementation, it can act on C++ Under the libtorch in , however ksize The size of can only be fixed , Unable to change input . In practical use , Whether the image is a binary graph , It won't affect the result . Only the element with the largest value in the corresponding region of the operator will be selected for replacement .

Core code ：

#ksize Must be odd , Otherwise, the size of the pool will change . Besides ,stride=1, padding=int((ksize-1)/ 2) It's a must 
ksize=21
max_pool = torch.nn.MaxPool2d(kernel_size=ksize, stride=1, padding=int((ksize-1)/ 2))  # Adjustable kernel_size

Expansion operation

Inflation