Input High-Dimensional Expansion Convolution: Convolution Optimization for Spatially Varying Convolution

In this paper, in order to improve the execution speed of complex image processing functions in convolutional neural networks, we propose an optimization algorithm for convolution. This algorithm is aimed at optimizing the special convolution calculation of complex image processing functions in image processing, in which the weights of the kernel change with the position of the convolution kernel My algorithm mainly expands the image and variable convolution kernel to higher dimensions to reduce the number of cycles through vectorization operations, and optimizes the method of image expansion to higher dimensions. The experimental results show that my algorithm fully utilizes the parallel computing power of the CPU, which is more than 20 times faster than the direct method.