基于改进FPN的复杂场景下SAR图像船舶目标检测  

Ship target detection with SAR images in complex scenes based on improved feature pyramid network

作　　者：周慧[1] 李迎秋[1] 陈澎[2] 沈宇军 朱煜锋 ZHOU Hui;LI Ying-qiu;CHEN Peng;SHEN Yu-jun;ZHU Yu-feng(School of Computer and Software,Dalian Neusoft Information University,Dalian 116023,China;Navigation College,Dalian Maritime University,Dalian 116026,China;Hangzhou Hikvision Digital Technology Co.Ltd.,Hangzhou 310051,China)

机构地区：[1]大连东软信息学院计算机与软件学院,辽宁大连116023 [2]大连海事大学航海学院,辽宁大连116026 [3]杭州海康威视数字技术股份有限公司,杭州310051

出　　处：《大连海事大学学报》2022年第4期76-83,共8页Journal of Dalian Maritime University

基　　金：国家重点研发计划项目(2021YFC3320300);辽宁省普通本科高等学校校际合作重大科研项目。

摘　　要：针对合成孔径雷达(SAR)图像近岸船舶目标受背景杂波影响,导致SAR图像船舶目标检测率低和小尺度舰船目标漏检率高的问题,提出一种适用于复杂背景下SAR图像近岸舰船目标检测的改进FPN模型。该模型基于FPN目标检测算法,在特征提取网络中利用可变形卷积更加精确地确定目标采样点位置,以增强目标的特征提取能力,提高复杂背景下SAR图像舰船标的检测率;同时,采用通道注意力机制来捕获特征提取网络中不同通道图之间的特征依赖关系,降低漏检率。在公开的SAR图像舰船数据集上的测试结果表明,该模型在复杂场景下的检测精度为87.95%,相比原始FPN提升了8.46%,其中,针对小尺度舰船目标检测精度为95.14%,相比原始FPN检测精度提升了5.28%;对比Yolo5和mask RCNN,改进FPN模型平均检测精度分别提升了11.21%、2.98%。Aiming at the problem that the nearshore ship targets in synthetic aperture radar(SAR)images were affected by the background clutter,that resulted in lower detection rate of ship targets in SAR images and higher false alarm rate and missed detection rate of small-scale ship targets,an improved feature pyramid network(FPN)model for near-shore ship target detection in SAR images under complex backgrounds was proposed based on the FPN target detection algorithm.The deformable convolution was used in the feature extraction network to determine the target sampling point position more accurately to enhance the target feature extraction ability and improve the detection rate of ship targets in SAR images under complex backgrounds.At the same time,the channel attention mechanism was used to capture the feature dependencies relationship between different channel graphs in the feature extraction network and reduce the missed detection rate.Test experiments on the public SAR image ship dataset show that the detection accuracy of the model in complex scenes is 87.95%,which is 8.46%higher than the original FPN.There in,the detection accuracy for small-scale ship targets is 95.14%,which is 5.28%higher than the original FPN.Compared with Yolo5 and mask RCNN,the average detection accuracy of the improved FPN model increases by 11.21%and 2.98%respectively.

关 键 词：合成孔径雷达(SAR)图像 船舶目标检测 改进FPN模型 可变卷积 通道注意力 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]