D2M-DETR: Differential Decoupling and Multiscale Dynamic Modeling for Remote Sensing Small Object Detection

Abstract

In recent years, small object detection in remote sensing images has been widely applied in many fields, yet it still faces several serious challenges. On one hand, small objects inherently have weak features and are easily disturbed by complex backgrounds. On the other hand, small objects in remote sensing images are typically densely distributed and often accompanied by occlusion, which easily leads to missed detections and false positives. To tackle these challenges, we introduce an end-to-end detection network named $\text {D}^{2}$ M-DETR, which integrates differential decoupling and multiscale dynamic modeling. The overall network comprises the differential decoupled salient features (DDSFs) strategy and the multiscale dynamic spatial context modeling (MDSCM) module. The DDSF strategy employs a dual-branch backbone for global and local feature extraction. It then applies differential decoupling to separate and enhance unique and basic features, effectively suppressing background interference and improving the discriminability of small objects. The MDSCM module jointly constructs multiscale dynamic local focus and multidirectional global spatial modeling, and enhances interactions among multiscale contextual information. This enables explicit modeling of structural priors and semantic relationships of small objects in densely occluded scenes. Our $\text {D}^{2}$ M-DETR achieves excellent detection performance on multiple remote sensing datasets, validating the efficacy of the proposed algorithm. The source code is available at https://github.com/Xidian-AIGroup190726/DDMDETR