智能系统学报2025，Vol.20Issue(5)：1123-1135,13.DOI:10.11992/tis.202410017
基于提案增强的解耦特征挖掘旋转检测器
Decoupled feature mining rotational detector based on proposal enhancement
摘要
Abstract
In remote sensing images,small and cluttered objects often appear intertwined,presenting considerable chal-lenges for object detection.These challenges are even further amplified in rotational object detection tasks.Aiming to address these challenges,this paper proposes a decoupled feature mining rotational detector based on proposal enhance-ment(PDMDet).First,a single-stage detector is employed to replace the region proposal network of the two-stage de-tector,generating high-quality proposals to reduce background redundancy.Second,self-attention is applied within the same feature dimensions and cross-attention across different dimensions,aiming to enhance intradimensional features and fuse interdimensional features,thereby improving the capability of the detector to identify objects of varying sizes.Finally,recognizing that classification and oriented bounding box regression tasks have different feature sensitivities,this paper proposes a decoupled feature refinement strategy that processes the two tasks separately.Experiment results demonstrate that PDMDet achieves single-scale mAP scores of 78.37％,72.35％,and 98.60％on DOTA-v1.0,DOTA-v1.5,and HRSC2016 datasets,respectively,outperforming existing algorithms in terms of detection accuracy and demonstrating strong competitiveness in complex rotational object detection scenarios.关键词
遥感图像/目标检测/旋转目标检测/小目标/高密度/倾斜边界框/跨尺度融合/两阶段检测器Key words
remote sensing image/object detection/rotated object detection/small object/high density/oriented bound-ing box/cross-scale fusion/two-stage detector分类
信息技术与安全科学引用本文复制引用
赵振博,付天怡,董红斌,张小平..基于提案增强的解耦特征挖掘旋转检测器[J].智能系统学报,2025,20(5):1123-1135,13.基金项目
国家自然科学基金项目(82374621). （82374621）
