Abstract: In order to improve the processing speed of the convolutional neural network, we use the convolution method of zero-grad approximate treatment (grad convolution) to reduce the computation amount and improve the reuse rate of the data. The grad calculation of the data is performed in terms of the convolution kernel, and a flexible gradient threshold calculation strategy for different levels of different networks is adopted to rationally reuse the convolution results of adjacent windows. The key grad processing module and convolution calculation part are implemented on Field-Programmable Gate Array(FPGA), combined with pulsation array to improve resource utilization, and the data flow suitable for gradient convolution is designed for the problem of load imbalance. In the target detection experiment based on YOLOv3 model and Pascal VOC dataset, the software side reduced the computation by about 23.2%, and the combined hardware acceleration ratio was about 17.8%.