基于双重标签分配的遥感有向目标检测方法  

Remote Sensing Orineted Object Detection Method Based on Dual-label Assignment

作　　者：董燕[1,2] 魏铭宏 高广帅 刘洲峰 李春雷[1] DONG Yan;WEI Minghong;GAO Guangshuai;LIU Zhoufeng;LI Chunlei(School of Information and Communication Engineering,Zhongyuan University of Technology,Zhengzhou 450007,China;School of Automation Engineering,University of Electronic Science and Technology,Chengdu 610000,China)

机构地区：[1]中原工学院信息与通信工程学院,郑州450007 [2]电子科技大学自动化工程学院,成都610000

出　　处：《计算机科学》2024年第S02期486-494,共9页Computer Science

基　　金：国家自然科学基金(62072489);中原科技创新领军人才项目(234200510009);河南省科技攻关项目(232102211002,232102211030)

摘　　要：由于遥感图像目标具有任意方向、大纵横比和密集排列等多样性分布特点,预设的锚框难以精准匹配所有真实目标,导致对大纵横比和密集排列的有向目标检测精度不高。为了解决上述问题,提出了一种基于双重标签分配的遥感有向目标检测方法。首先,提出双重标签分配策略为目标分配最大及次优交并比的候选框;其次,通过排斥损失(AP-Loss)和吸引损失(UP-Loss)约束相邻目标的候选框,以提高目标正确匹配概率;然后,为了提取适应于分类和回归分支的鲁棒特征,设计了一个特征增强模块(FEM),该模块基于偏振函数构造自适应特征,能够有效增强分类和回归任务所需的特征表达能力;最后,设计了一个定位指导分类(LGC)模块,该模块通过定位任务指导分类任务的采样位置,进行定位细化,以获取分类任务的关键特征,从而缓解分类与定位之间的不一致问题。在3个公开的遥感有向目标检测数据集DOTA,HRSC-2016和DIOR-R上进行了大量的实验,实验结果证明了所提方法的有效性,且优于现有主流方法。Due to the inherently diverse distribution characteristics of remote sensing image objects,such as their arbitrary orientation,large aspect ratio,and densely arranged,the utilization of preset anchor boxes makes it difficult to accurately match all real objects.This limitation results in low detection accuracy,particularly for oriented objects with large aspect ratios and densely arranged.To address this issue,an oriented object detection method in remote sensing images based on dual label assignment is proposed.Firstly,a dual-label assignment strategy is proposed to assign candidate boxes with maximum and suboptimal intersection union ratios to the real object.Then,the candidate boxes of adjacent objects are constrained by repulsion loss(AP Loss)and attraction loss(UP Loss)to improve the probability of correct object matching.In addition,to extract robust features suitable for classification and regression branches,a Feature Enhancement Module(FEM)is designed.This module constructs adaptive features based on polarization functions,which can effectively enhance the feature expression ability required for classification and regression tasks.Finally,a localization-guided classification(LGC)module is designed,which guides the sampling position of the classification task through localization tasks,performs localization refinement,and obtains key features of the classification task,thereby alleviating the inconsistency between classification and localization.A large number of experiments were conducted on three publicly available oriented object detection in remote sensing datasets,namely DOTA,HRSC-2016,and DIOR-R.The experimental results demonstrated the effectiveness of the proposed method and its accuracy(mAP)is better than existing mainstream methods.

关 键 词：遥感图像 有向目标检测 双重标签分配 不一致问题 采样位置细化 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]