基于MPSoC的遥感图像目标检测算法硬件加速研究  

Accelerator of Remote Sensing Image Object Detection Based on MPSoC

作　　者：李强[1,2] 武文波[2] 何明一[1] LI Qiang;WU Wenbo;HE Mingyi(Northwestern Polytechnical University,Xi’an 710072,China;Beijing Institute of Space Mechanics&Electricity,Beijing 100094,China)

机构地区：[1]西北工业大学,西安710072 [2]北京空间机电研究所,北京100094

出　　处：《航天返回与遥感》2022年第1期58-68,共11页Spacecraft Recovery & Remote Sensing

基　　金：科工局民用航天项目(D040101)。

摘　　要：遥感图像目标实时检测是遥感应用领域的关键技术问题之一。深度神经网络遥感图像目标检测准确率高,但此类网络通常结构复杂、参数多、计算量大,对计算资源和存储的需求较高,设计轻量化软硬件系统实现星载边缘端部署较为困难。针对上述问题,文章提出了一种基于多处理器片上系统(MPSoC)现场可编程门阵列(FPGA)的遥感图像目标检测算法硬件加速方案。首先研究了适合星载边缘端部署的目标检测算法;然后设计了深度卷积神经网络并行加速计算结构和引擎,采用有限精度运算实现网络参数,使其数字量减少了75%,显著降低了计算和存储开销;最后基于MPSoC FPGA处理器实现了飞机目标检测的原型演示验证系统。实验结果表明,文章提出的遥感图像目标检测系统方案的目标检测精度可达92%以上;与基于嵌入式CPU、CPU、GPU的方案相比,单帧图像推理时间从100s、1000ms、100ms缩短至10ms级,可以满足遥感图像目标检测实时处理要求,具有一定的工程应用价值。Real-time object detection from remote sensing image is a hot topic in the field of remote sensing applications.Deep neural network has advantage of high accuracy in target detection,however it is difficult to design a lightweight hardware and software system to realize the on-board edge deployment because of the complexity of deep neural network’s structure,large amount of computation and high demand for computing resources and storage.In order to solve these problems,a lightweight remote sensing object detection system is proposed based on Multi-Processor System on Chip(MPSoC)Field-programmable Gate Array(FPGA)processor.Firstly,the lightweight object detection algorithm suitable for the deployment of satellite edge devices is studied.Then,a parallel accelerator for deep convolutional neural network and network’s digital parameter compression are studied to further reduce the amount of digital parameters by 75%and the calculation and storage overhead.Finally,a prototype verification system based on MPSoC FPGA processor is designed to realize aircraft object detection.The experimental results show that the aircraft target detection accuracy can reach more than 92%,and the reasoning time for single frame image can be reduced to 10ms level,from 100s(embedded CPU based),1000ms(CPU based)and 100ms(GPU based)levels,which can meet the real-time processing requirement of remote sensing target detection.

关 键 词：目标检测 多处理器片上系统 现场可编程门阵列 深度卷积神经网络 嵌入式 硬件 加速 遥感应用 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]