大幅宽SAR图像嵌入式舰船实时检测系统设计.pdfVIP

⦾工程与应用⦾

大幅宽SAR图像嵌入式舰船实时检测系统设计

摘要：针对星载或机载高分辨率合成孔径雷达（syntheticapertureradar，SAR）实时成像后的大幅宽SAR图像舰船

实时检测的应用需求，传统的基于FPGA+DSP的嵌入式系统很难同时实现SAR成像处理和基于人工智能技术的大

幅宽SAR图像舰船实时检测，为此设计了一种基于3UVPXFPGA+GPU架构的大幅宽SAR图像嵌入式舰船实时

检测系统；提出了一种基于YOLOv5s的舰船检测模型，采用基于L2-范数稀疏性惩罚的缩放因子控制法进行轻量

化，轻量化舰船检测模型的参数量减小了47.39%，计算量减少了18.67%，平均检测精度为0.968；将轻量化舰船检测

模型应用于大幅宽SAR图像嵌入式舰船实时检测系统，并针对典型的10km×10km的大幅宽图像应用场景，设计开

发基于多线程技术和基于GPU的众核并行计算技术的大幅宽SAR图像嵌入式实时检测系统软件；通过公开的SAR

数据集进行功能验证和性能评估，该系统能够满足不同分辨率的大幅宽SAR图像舰船实时检测需求。

关键词：合成孔径雷达（SAR）；YOLOv5s；轻量化；图形处理器（GPU）；实时舰船检测

DesignofEmbeddedReal-TimeLargeSizeSARImageShipDetectionSystem

Abstract:Inapplicationsofreal-timeshipdetectioninlargesizesyntheticapertureradar(SAR)imagesafterreal-time

imagingofspaceborneorairbornehigh-resolutionSAR,itisdifficultfortraditionalFPGA+DSPembeddedsystemto

realizebothoftheSARimagingprocessandtheartificialintelligence-basedshipdetectioninreal-timeforlargesizeSAR

images.Inthispaper,alargesizeSARimagesorientedreal-timeshipdetectionsystemon3UVPXFPGA+GPUispro-

posedandaYOLOv5sbasedshipdetectionmodelisproposedaswell,whichappliesL2-normsparsitypenaltyscaling

factorcontrolmethodforlightweight.Theaveragedetectionaccuracyoftheproposedlightweightshipdetectionmodelis

0.968,wherethenumberoftheparametersofthemodelisreducedby47.39%,andthecomputationalcostisreducedby

18.67%.Thelightweightshipdetectionmodelisappliedtotheembeddedshipreal-timedetectionsystemforlargesize

SARimages.Forthetypicallargesizeimagesapplicationscenarioof10km×10km,theembeddedreal-timedetection

systemisdesignedandimplementedbyutilizingmulti-threadingandGPU-basedmany-coreparallelprogramming.Perfor-

manceevaluationareconductedonapublicSARimagedataset.Experimentalresultsverifythattheproposedsystemis