MSAE–DETR: a multiscale adaptive-enhancement UAV small-object detection algorithm

Abstract Detecting small objects in unmanned aerial vehicle (UAV)–captured images presents substantial challenges because of insufficient target clarity, intricate environmental contexts, concentrated target arrangements, and restricted computational capabilities. To address these issues, this study introduces a multiscale adaptive enhancement detection transformer (MSAE–DETR), which is a detection framework incorporating three principal innovations. First, the multiscale adaptive feature extraction block utilizes dual-scale adaptive attention mechanisms and dual-domain fusion networks, enhancing small-object feature extraction via deformable convolutions and frequency-domain modeling. Second, the dual-histogram spatial reordering attention encoder introduces bidirectional histogram-sorting attention mechanisms that effectively capture fine-grained features while suppressing background interference. Third, the multiscale adaptive enhancement pyramid combines spatial information preservation mechanisms, cross-stage adaptive multikernel modules, and adaptive multikernel extractors to boost feature fidelity and multiscale fusion effectiveness. Comprehensive evaluations using the VisDrone2019 benchmark demonstrate that MSAE–DETR achieves competitive results with mAP 50 of 50.7% and mAP 50 :₉₅ of 31.1%, representing 3.4% and 2.5% improvements over the RT–DETR baseline, respectively, while reducing the number of parameters by 13.4% and maintaining real-time inference at 132.1 frames per second. Additional validation using TinyPerson and HIT–UAV benchmarks confirms the method’s consistent generalization performance across diverse small object detection applications.