摘要
近年来,深度学习方法在合成孔径雷达(SAR)图像目标检测中得到了广泛的应用。船舶出现在近海、港口、岛礁、远洋等各种场景中,同时海洋环境复杂多变,使得船舶目标检测很难排除混乱背景的干扰。对于大纵横比、任意方向、密集分布的目标,精确定位变得更加复杂。本文基于深度学习的方法提出用于SAR图像目标检测的改进RetinaNet模型,使用深度残差网络自主获取图像特征,利用基于圆形平滑标签(Circular Smooth Label,CSL)的旋转框检测方法实现精准定位,在分类与定位网络中加入了注意力机制以增强网络特征提取能力。经SSDD数据集实验验证,本文方法目标检测精度达到88.63%,比传统RetinaNet模型提高了8.74%,表现出了良好的检测效果。
In recent years,deep learning method has been widely used in target detection in synthetic aperture radar(SAR)images.Ships appear in various scenes such as nearshore,port,island and reef,ocean.The complex and changeable marine environment also makes it difficult for ship detection to eliminate the interference of chaotic background.For targets with large aspect ratio,arbitrary direction and dense distribution,accurate positioning becomes more difficult.In this paper,an improved RetinaNet model for target detection in SAR images was proposed based on deep learning method.The depth residual network was used to obtain image features independently.The rotate anchor based on circular smooth label(CSL)was used to achieve accurate positioning.The attention mechanism was added to the classification and detection network to enhance the network feature extraction ability.Experimental results on SSDD dataset showed that the detection accuracy of the proposed method reached 88.63%,which was 8.74%higher than that of the conventional RetinaNet model,showing a good detection performance.
作者
岳冰莹
陈亮
师皓
盛青青
YUE Bingying;CHEN Liang;SHI Hao;SHENG Qingqing(Radar Research Lab,School of Information and Electronics,Beijing Institute of Technology,Beijing 100081,China;Beijing Institute of Technology Chongqing Innovation Center,Chongqing 401120,China;Beijing Key Laboratory of Embedded Real-time Information Processing Technology,Beijing 100081,China)
出处
《信号处理》
CSCD
北大核心
2022年第1期128-136,共9页
Journal of Signal Processing
基金
国家重大科研仪器研制项目(部门推荐)(31727901)
国家自然科学基金重大研究计划集成项目(91738302)。
