Abstract: 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.

Key words: Remote sensing images, Oriented object detection, Dual-label assignment, Inconsistency issues, Sampling position refinement