基于改进Efficient-Det的Video SAR动目标检测方法  

Video SAR Moving Target Detection Method Based on Improved Efficient-Det

作　　者：毛倩倩 詹梦洋 李银伟 MAO Qianqian;ZHAN Mengyang;LI Yinwei(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《现代雷达》2024年第10期58-64,共7页Modern Radar

基　　金：国家自然科学基金资助项目(61988102);上海市自然科学基金资助项目(21ZR1444300)。

摘　　要：动目标阴影检测是近年来视频合成孔径雷达(Video SAR)领域中的一项具有挑战性的任务。动目标阴影存在着尺寸易变、深浅不一的特点,使得视频SAR动目标检测更加困难。文中提出了一种基于改进神经网络Efficient-Det的视频SAR动目标检测方法。首先,为获得动目标阴影的更多空间位置信息,通过融入坐标注意力机制重建了Efficient-Det网络的主干部分;其次,为进一步减少计算量,对颈部网络剪裁并将深度可分离卷积替代头部网络中原本的卷积。在网络训练过程中,采用非极大抑制法来筛选预测框,损失函数采用回归损失和交叉熵损失的组合,其中的交叉熵损失是用来抑制正负样本极端不平均带来的负面作用。通过实验验证,文中方法可有效提升视频SAR动目标阴影检测的准确性与效率。Moving target shadow detection is a challenging task in the field of video synthetic aperture radar(Video SAR)in recent years.The moving target shadows are characterized by variable sizes and different depths,which make the detection of moving targets in video SAR more difficult.This paper proposes a method for video SAR moving target detection based on an improved neural network,Efficient-Det.Firstly,to obtain more spatial location information of the moving target shadows,the backbone of the Efficient-Det network is reconstructed by incorporating a coordinate attention mechanism.Secondly,to further reduce the computational cost,the neck network is pruned,and depthwise separable convolutions replace the original convolutions in the head network.During the training process,non-maximum suppression is used to select prediction boxes,and the loss function is a combination of regression loss and cross-entropy loss,where the cross-entropy loss is used to mitigate the negative effects of extremely imbalanced positive and negative samples.Experimental validation shows that this method effectively improves the accuracy and efficiency of video SAR moving target shadow detection.

关 键 词：视频SAR 动目标阴影 目标检测 深度学习 

分 类 号：TN957.52[电子电信—信号与信息处理]