摘要
为实现实时图像识别的端侧部署,设计并实现了一种基于国产FPGA与自主设计类ASIC架构的嵌入式系统。软件层面,提出了一种轻量级神经网络NexusEdgeNet,以仅0.184 MB参数量,对39类农田病害图像的识别准确率达到94.22%。硬件层面,创新性地设计了一款完全采用Verilog HDL描述的类ASIC加速器,采用分布式存储,不依赖外存储器,支持任意形状卷积、池化及全连接等算子。通过近存并行计算、流水线、滑动卷积窗口及双缓冲存储等优化策略,该神经网络加速器在中科亿海微EP6HL130 FPGA上实现了399 f/s的高推理帧率,大幅降低了逻辑资源使用量,计算资源利用率高达85%。系统集成图像采集、处理与显示链路,支持视频流的实时处理与识别,在保持高精度的同时,具备优异的实时性与资源效率,为国产FPGA在边缘计算中的低成本应用提供了有价值的实践方案。
This paper presents an embedded system based on a domestically produced FPGA and a self-designed ASIC-like architecture for real-time edge deployment.On the software side,a lightweight neural network named NexusEdgeNet is proposed.It achieves 94.22%accuracy on 39 farmland disease categories with only 0.184 MB of parameters.On the hardware side,an ASIC-like accelerator fully described in Verilog is designed.It adopts a distributed on-chip memory structure,eliminating external memory access,and supports arbitrary-shaped convolution,pooling,and fully connected operations.Several optimization techniques are applied,including nearmemory parallel computing,pipelining,sliding convolution windows,and double buffering.The accelerator reaches 399 f/s inference speed on the EP6HL130 FPGA,with 85%resource utilization and significantly reduced logic consumption.The system integrates image acquisition,processing,and display,supporting real-time video stream recognition.It maintains high accuracy while achieving excellent real-time performance and resource efficiency.This work provides a practical,low-cost solution for edge computing applications based on domestic FPGAs.
作者
陈冠夫
兰小磊
陈镇城
张艺豪
陈林亮
李赛
CHEN Guanfu;LAN Xiaolei;CHEN Zhencheng;ZHANG Yihao;CHEN Linliang;LI Sai(School of Intergrated Circult Science and Engineering,Beihang University,Beijing 100191,China)
出处
《集成电路与嵌入式系统》
2026年第2期43-52,共10页
Integrated Circuits and Embedded Systems
