DQSA-DETR: A Density-Guided Adaptive DETR Framework for Tiny Object Detection in Remote Sensing Images

Abstract

Tiny object detection in remote sensing images remains highly challenging due to extremely small object sizes, complex backgrounds, weak feature representation, and substantial variations in object density across images. Although DETR-like methods have significantly improved detection performance by introducing the Transformer architecture, their performance remains suboptimal for this task. This is primarily because they struggle to effectively enhance weak tiny-object features under background interference, adapt the query budget to varying object densities, and alleviate the misalignment between classification confidence and localization quality for tiny objects. To address these issues, we propose DQSA-DETR, a density-guided adaptive DETR framework for remote sensing tiny object detection. Specifically, DQSA-DETR integrates a Density-Guided Feature Enhancement (DGFE) module after the encoder to enhance tiny-object feature representation and suppress background interference, particularly in locally dense regions; incorporates a Density-Driven Adaptive Query (DDAQ) mechanism before the decoder to dynamically allocate the query budget according to image-level object density while reducing decoder computation; and introduces a Quality-Aware Classification Loss (QCL) during training to improve the consistency between classification confidence and localization quality for tiny objects. Experimental results on AI-TOD-V2, VisDrone-2019, DOTA-v1.0, and RS-STOD show that DQSA-DETR achieves AP scores of 30.5%, 35.9%, 48.2%, and 34.7%, respectively, outperforming strong baselines and several state-of-the-art methods. Compared with DQ-DETR, DQSA-DETR achieves a higher AP while reducing GFLOPs by 53.6%, demonstrating a favorable accuracy–computation trade-off under large-query detection settings. The code for DQSA-DETR is available at https://github.com/gfnanxi/DQSA-DETR