基于特征与区域定位增强的遥感舰船目标检测  

Remote Sensing Ship Target Detection Based on Feature and Region Localization Enhancement

作　　者：宋志娜 李莎 杨建明 徐川 SONG Zhina;LI Sha;YANG Jianming;XU Chuan(School of Computer Science,Hubei University of Technology,Wuhan 430000,China;Service Support Department,Rocket Army Command Academy,Wuhan 430012,China)

机构地区：[1]湖北工业大学计算机学院,武汉430000 [2]火箭军指挥学院勤务保障系,武汉430012

出　　处：《计算机工程》2023年第8期257-264,共8页Computer Engineering

基　　金：湖北工业大学博士启动基金(BSQD2020056)。

摘　　要：高分辨率遥感图像在海上监视、海上搜救、海上运输等军用和民用领域的舰船检测方面有着广泛的应用。然而高分辨率光学遥感图像舰船目标检测通常存在背景复杂、目标方向任意、尺度多变等问题,导致检测精度不高。提出一种基于特征和区域定位增强的旋转检测算法RetinaNet-MPD。通过添加一个多尺度特征融合模块,充分融合不同尺度、不同层级的特征信息,以增强不同尺度特征图的特征表示能力。针对复杂背景下的舰船目标检测,提出极化双重注意力网络,通过在注意力网络后加入极化函数,充分提取目标的关键特征,同时抑制不相关信息,以有效区分目标和背景。此外,为更准确地定位舰船目标,在对正负样本进行训练时采用一种动态锚学习方法,从而动态选择目标区域内具有良好定位潜力的高质量锚,提高舰船目标检测精度。实验结果表明,RetinaNet-MPD算法在DOTA舰船和HRSC2016数据集上的检测精度分别为89.3%和85.8%,相比现有旋转目标检测算法的检测精度有所提升。The use of high-resolution remote sensing imagery for ship detection has a wide range of applications in military and civilian fields,such as maritime surveillance,search and rescue,and transportation.However,in high-resolution optical remote sensing images,complex environment as well as arbitrary directions and variable scales of ship targets lead to poor detection accuracy.To address these limitations,a rotation detection algorithm,known as the RetinaNet-MPD,is proposed based on feature and region localization enhancement.First,the RetinaNet-MPD adds a multi-scale feature fusion module,which entirely integrates feature information at different scales and levels,to enhance the feature representation ability of feature maps at different scales.Second,a Polarized Dual-Attention Network(PDANet)module is proposed for ship target detection in a complex environment.By adding a polarization function after the attention network,the key features of the target are entirely extracted,and irrelevant information is suppressed to effectively distinguish the target from its surrounding.In addition,a Dynamic Anchor Learning(DAL)method is adopted when training the positive and negative samples to dynamically select high-quality anchors with good localization potential in the target region and improve the accuracy and precision of ship target detection.The experimental results show that RetinaNet-MPD algorithm achieved detection accuracy of 89.3%and 85.8%on the DOTA-Ship and HRSC2016 data sets,respectively.Consequently,the average detection accuracy was improved effectively compared with other existing rotating-target detection models.

关 键 词：高分辨率遥感图像 舰船目标检测 多尺度特征融合 极化双重注意力网络 动态锚学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]