卷积神经网络的FPGA实现及优化  

Implementation and optimization of convolutional neural networks on FPGA

作　　者：王开宇 生梦林 韩睿 李伯轩 刘晨阳 申人升 WANG Kai-yu;SHENG Meng-lin;HAN Rui;LI Bo-xuan;LIU Chen-yang;SHEN Ren-sheng(Faculty of Electronic Information and Electrical Engineering,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]大连理工大学电子信息与电气工程学部,辽宁大连116024

出　　处：《实验室科学》2018年第4期79-84,共6页Laboratory Science

基　　金：中央高校基本科研业务费专项资金(项目编号:No.DUT16QY32);辽宁省自然基金(项目编号:No.201602166)

摘　　要：卷积神经网络是神经网络的一个分支,通过卷积神经网络可以完成对图像的卷积处理。然而在传统的CPU上,由于并行性不强,会导致计算速度很慢; FPGA由于其并行的特点,逐渐被用到卷积神经网络的图像处理领域。通过设计一套完整的基于FPGA的图像卷积处理方案,利用串口实现上位机与FPGA通信,实现了实时的图像卷积处理,与前人相比,在充分发挥FPGA的并行性以提升运算速度的同时,减小了带宽和资源占用,具有一定实用价值。Convolution neural network is a branch of the neural network. Image processing can be completed through the convolution neural network. But in the traditional CPU, because the parallel is not strong, the computing speed will be slow. Because of its parallel characteristics, FPGA is gradually used in the convolution of the neural network image processing. The design of a complete image convolution processing scheme is completed based on FPGA. It uses serial port to realize the communication between PC and FPGA. So real time image convolution processing is realized. Compared with the pre- decessors, it gives lull play to the parallelism of FPGA to speed up the operation, and reduce the bandwidth and resource occupation, It has some practical value.

关 键 词：卷积神经网络 FPGA 图像处理 串口通信 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]