Image processing method and related device

A computer apparatus comprising a processor, a memory, a communication interface and one or more programs stored in the memory and comprising instructions performed by the processor to implement steps of the image processing method, the image processing method comprising:

• grouping input image data of a dilation convolution to obtain Dh×Dw grouped image data;
• wherein Dh is a dilation rate of a convolution kernel corresponding to the dilation convolution on a height dimension thereof, Dw is a dilation rate of a convolution kernel corresponding to the dilation convolution on a width dimension thereof, and both Dh and Dw are positive integers;
• wherein the step of grouping input image data of a dilation convolution to obtain Dh×Dw grouped image data, comprises:
• filling the input image data to obtain first image data, according to configuration parameters of the dilation convolution;
• determining a size of each of the grouped image data according to a size of the first image data;
• performing dilation processing on the first image data to obtain second image data, according Dh, Dw and the size of each of the grouped image data, wherein the step of performing dilation processing on the first image data to obtain second image data, according to Dh, Dw and the size of each of the grouped image data, comprises:
• determining a size of the second image data according to Dh, Dw and the size of each of the grouped image data;
• and
• performing the dilation processing that padding Ph rows of zeros after an (H′−1)-th row of the first image data, and performing the dilation processing that padding Pw columns of zeros after a (W′−1)-th column of the first image data, to obtain the second image data;
• wherein the size of the first image data is H′×W′, the size of the second image data is H″×W″, Ph=H″−H′, and Pw=W″−W′;
• and
• grouping the second image data to obtain the Dh×Dw grouped image data;
• performing a convolution calculation on the Dh×Dw grouped image data, respectively, by a first convolution kernel, to obtain Dh×Dw grouped convolution calculation results;
• wherein the first convolution kernel is a convolution kernel of a convolution before the dilation convolution is dilated;
• and
• obtaining a dilation convolution calculation result of the input image data according to the Dh×Dw grouped convolution calculation results.