基于深度学习的SAR图像目标在地表环境中的检测与伪装效果评估技术  

Target Detection and Camouflage Effect Evaluation in Ground Environment of SAR Image Based on Deep Learning

作　　者：刘青[1] 顾乃威 张学文 卢卫建 LIU Qing;GU Nai-wei;ZHANG Xue-wen;LU Wei-jian(Beijing Institute of Space Launch Technology,Beijing 100076,China)

机构地区：[1]北京航天发射技术研究所,北京100076

出　　处：《强度与环境》2022年第5期178-186,共9页Structure & Environment Engineering

基　　金：装备发展部预研领域基金。

摘　　要：本文研究了深度学习的端到端学习技术,采用了YOLO系列第三个版本YOLO-V3作为检测模型,根据真实雷达SAR图像-MSTAR数据集进行灰度融合构造目标背景图集,利用数据驱动的方式自主寻找适于区分不同类别的特征,对目标类型做出准确判断,实现检测模型在测试集平均精度为97.87%,最后利用独立同分布假设的识别性能下降理论来评价伪装性能,比较原型样本与伪装样本的检测概率的数值差异得出伪装效果的量化评价,实现SAR图像中的地面目标检测和伪装效果评估。This paper studies the end-to-end technology of deep learning and uses the third version of YOLO series YOLO-V3 as the detection model, and develops an algorithm. The real SAR image data set-MSTAR is used to construct the background target image by gray level fusion technology. Using data-driven mode, the most suitable features to distinguish different types are found to judge target types accurately and make the comment accuracy of YOLO detection model on the test set is 97.87%. Finally, the performance of camouflage is evaluated by using the theory of identification performance degradation under the assumption of independent identical distribution. The numerical difference of detection probability between normal samples and camouflage samples are compared, and the quantitative evaluation of camouflage effect are gotten. Ground target detection and camouflage effect evaluation in SAR images are realized.

关 键 词：SAR图像 YOLO-V3模型 地表环境 目标检测 伪装效果评估 

分 类 号：V557[航空宇航科学与技术—人机与环境工程]