Design of Efficient CNN Accelerator Based on Zynq Platform

Abstract

With the rapid development of deep learning, convolution neural network is widely used in image recognition and object detection. However, CNN with extensive calculation is not suitable for the mobile terminals. In this paper, lightweight convolution neural network MobileNet is modified as backbone network for small object detection based on the Zynq platform. A method of data bypass is proposed which reserves shallow feature map information to the last convolutional layer, with which the detection accuracy increases by 8%. Additionally, 16 x 4 Wallace tree addition tree is designed instead of the original two-operand addition tree generated by Vivado HLS. The usage of hardware resources is reduced by 26%. The detection accuracy of the entire acceleration system is 73.7% and the FPS is 28. Based on the methods proposed, the accelerator achieve both high detection speed and accuracy, which are better than other MobileNet hardware acceleration works.