基于FPGA的CNN分类器设计  

Design of CNN Classifier Based on FPGA

作　　者：方子卿 林瑞全[1] 孙小坚 FANG Zi-qing;LIN Rui-quan;SUN Xiao-jian(Institute of Electricaland Automation Engineering,Fuzhou University,Fuzhou 350108 China)

机构地区：[1]福州大学电气工程与自动化学院,福建福州350108

出　　处：《电气开关》2024年第1期64-68,共5页Electric Switchgear

摘　　要：传统CNN存在参数多,计算量大,部署在CPU与GPU上推理速度慢、功耗大的问题,为满足将卷积神经网络(Convolutional Neural Network,CNN)部署于嵌入式设备,实现实时图像采集与分类的需求,提出了一种基于FPGA平台的Mobilenet V2轻量级卷积神经网络分类器的设计方案。采用Cameralink相机采集图像,设计了裁剪、乒乓缓存和量化的图像预处理方式,实现连续的图像采集,CNN每层分别占用资源与计算结构,实现连续图片处理。设计了一种PW与DW的流水线结构,全连接层的稀疏化计算优化策略,减少了计算量和处理延迟。单张图片分类耗时1.25ms,能耗比为14.50GOP/s/W。In order to meet the requirements of deploying Convolutional Neural Network(CNN)on embedded devices to realize real-time image acquisition and classification,traditional CNN has many parameters,large computation,slow reasoning speed and large power consumption when deployed on CPU and GPU.This paper presents a design scheme of Mobilenet V2 lightweight convolutional neural network classifier based on FPGA platform.Using Cameralink camera to capture images,the image preprocessing methods of cutting,ping-pong cache and quantization are designed to achieve continuous image acquisition.Each layer of CNN occupies resources and computing structure respectively to achieve continuous image processing.A pipeline structure of PW and DW is designed,and the sparsity calculation optimization strategy of fully connected layer is designed to reduce the computation amount and processing delay.The classification time of single image is 1.25ms,and the energy consumption ratio is 14.50GOP/s/W.

关 键 词：FPGA CAMERALINK CNN 流水线结构 稀疏化 

分 类 号：TN402[电子电信—微电子学与固体电子学]