Multiscale Feature Optimization for Accurate Small Object Detection in Remote Sensing Imagery

Detecting small, overlapping objects in high-resolution remote sensing imagery is crucial for applications such as smart city monitoring and disaster response. However, challenges such as severe feature confusion and spatial misalignment hinder accurate localization. This paper introduces Multiscale SOG-DETR, a systematic redesign of the RT-DETR framework tailored for remote sensing small-object detection. We propose a lightweight Multiscale Overlapping-Object Decoupling Network (MOODNet) to significantly reduce feature entanglement in overlapping regions. Additionally, our specialized fusion neck, comprising the Residual Spatial-Alignment Progressive Fusion Module (SAPFM), E-CGAFusion, and WTConv2d modules, enhances multiscale semantic focus and preserves high-frequency details cost-effectively. Experimental results on the RSOD, VisDrone2019, and NWPU VHR-10 datasets demonstrate that Multiscale SOG-DETR achieves superior detection accuracy with significantly fewer parameters compared to the baseline RT-DETR model, increasing AP_ _IoU=50 by 3.1%, 3.0%, and 5.2%, and AP _IoU=50:95 by 5.1%, 2.1%, and 8.5%, respectively. These findings position Multiscale SOG-DETR as a promising solution for efficient and accurate small-object detection in remote sensing applications.The source code is publicly available at https://github.com/AaronWang-code/Multiscale-SOG-DETR.