Research on Accelerating Computation Model with FPGA Oriented YOLO Detection Networks
PEI Song-wen1,2,3,WANG Xian-rong1
1(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)2(State Key Laboratory of Computer Architecture,Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China) 3(Shanghai Key Laboratory of Data Science,Fudan University,Shanghai 200433,China)
Abstract:FPGA(Field Programmable Gate Array) with its characteristics of high parallelism and customizable can solve some problems of object detection networks,which are complexity of model structure,intensive computation,high storage cost and so on. This paper researches and implements the accelerating calculation model of YOLO (You Only Look Once) series neural networks with FPGA verification platform. Firstly,we perform dynamic fixed points quantification to reduce data storage and transmission delay. Secondly,for two typical types of convolution layers with high computational overhead in YOLO model,the strategies of pipeline,loop unroll and module fusion are adopted to realize the fast convolution processing engine based on Winograd and GEMM respectively,so as to improve the computational efficiency. Experimental results show that a performance of 64.9 GOP/s is achieved on the PYNQ-Z1 verification platform,which is 2.15 times higher than the performance based on sliding window convolution computation.
裴颂文,汪显荣. YOLO检测网络的FPGA加速计算模型的研究[J]. 小型微型计算机系统, 2022, 43(8): 1681-1686.
PEI Song-wen,WANG Xian-rong. Research on Accelerating Computation Model with FPGA Oriented YOLO Detection Networks. Journal of Chinese Computer Systems, 2022, 43(8): 1681-1686.