Research on a convolution operation method based on domain transformation in deep learning

In deep learning, convolution operation is an important part of convolutional neural network. Various transformations have been used to deal with convolution problems, but complex number operation and the inability to carry out convolution in the DCT domain have been introduced. In this paper, the relationship between DCT transform of output matrix and DCT transform of input matrix is found. At the same time, in view of the reality of partitioning operation in the case of large matrix, the method of overlapping of blocks is introduced. Finally, experiments on LeNet show that the proposed method can achieve the accuracy of the traditional method and can be used in practical applications.