This work completes the parallel design of convolution neural network based on OpenCL, and verifies the feasibility and correctness of the designed algorithm on heterogeneous parallel platform. Based on the analysis of the basic structure of convolution neural network of handwritten numeral recognition, this paper summarizes the optimization method of convolution neural network training based on OpenCL, and proposes a single convolution process parallel, multi-convolution task parallel, multi-convolution data Parallel and batch processing and other openCL-based parallel optimization program, and complete the optimization program programming. The training, testing process and simulation of CNN algorithm are realized on Intel CPU, AMD GPU and NVIDIA GPU platform. The test results show that the parallel optimization scheme proposed in this paper is about 375 times faster than the serial execution method in the case of the same training accuracy rate.