计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (1): 301-309.DOI: 10.3778/j.issn.1002-8331.2209-0287

• 工程与应用 • 上一篇    下一篇

大幅宽SAR图像嵌入式舰船实时检测系统设计

陆天宇,徐湛,崔红元,龚昊,王琤   

  1. 1.北京信息科技大学 信息与通信工程学院,北京 100101
    2.中国科学院大学 计算机科学与技术学院,北京 100089
    3.北京雷鹰科技有限公司,北京 100080
  • 出版日期:2024-01-01 发布日期:2024-01-01

Design of Embedded Real-Time Large Size SAR Image Ship Detection System

LU Tianyu, XU Zhan, CUI Hongyuan, GONG Hao, WANG Cheng   

  1. 1.School of Information and Communication Engineering, Beijing Information Science and Technology University, Beijing 100101, China
    2.School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing 100089, China
    3.Beijing Raying Technology Co., Ltd., Beijing 100080, China
  • Online:2024-01-01 Published:2024-01-01

摘要: 针对星载或机载高分辨率合成孔径雷达(synthetic aperture radar,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3U VPX FPGA+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10 km×10 km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

Abstract: In applications of real-time ship detection in large size synthetic aperture radar (SAR) images after real-time imaging of spaceborne or airborne high-resolution SAR, it is difficult for traditional FPGA+DSP embedded system to realize both of the SAR imaging process and the artificial intelligence-based ship detection in real-time for large size SAR images. In this paper, a large size SAR images oriented real-time ship detection system on 3U VPX FPGA+GPU is proposed and a YOLOv5s based ship detection model is proposed as well, which applies L2-norm sparsity penalty scaling factor control method for lightweight. The average detection accuracy of the proposed lightweight ship detection model is 0.968, where the number of the parameters of the model is reduced by 47.39%, and the computational cost is reduced by 18.67%. The lightweight ship detection model is applied to the embedded ship real-time detection system for large size SAR images. For the typical large size images application scenario of 10 km×10 km, the embedded real-time detection system is designed and implemented by utilizing multi-threading and GPU-based many-core parallel programming. Performance evaluation are conducted on a public SAR image data set. Experimental results verify that the proposed system is able to meet the requirements of the real-time ship detection for large size SAR images under different resolutions.

Key words: synthetic aperture radar (SAR) , YOLOv5s, lightweight, graphics processing unit (GPU) , real-time ship detection