TSF-DETR: Two-Stage Frequency-Domain Enhancement DETR for Small Object Detection in Remote Sensing

Small object detection in remote sensing imagery is inherently constrained by several persistent factors within the conventional spatial domain. On one hand, the small proportion of tiny targets in the entire image results in sparse spatial feature representations that are easily submerged by complex background noise. On the other hand, effectively capturing long-range dependencies in dense-object scenarios remains difficult, which limits the model's capacity for global context modeling. Collectively, these factors represent significant performance bottlenecks for existing spatial-based object detection algorithms. To remedy these shortcomings, TSF-DETR, a novel detection framework anchored by a Two-Stage Frequency-Domain Enhancement strategy, is proposed in this study. Specifically, a Remote Sensing Frequency Enhancement Module (RSEM) is integrated within the backbone as the first stage to implement a hierarchical differentiated mechanism for frequency-domain feature refinement. In the subsequent neck stage, a Multi-Scale Frequency Domain (MSFD) module is leveraged to execute multi-scale and multi-orientational Gabor filtering, establishing a robust “Global FFT Enhancement – Local Gabor Optimization” synergy to suppress background noise. Furthermore, to fortify the model's global perception, a tailored Linear Multi-Head Self-Attention (LMSA) mechanism is incorporated. Distinguishable from standard linear attention designs, this modified LMSA enhances the model's ability to capture long-range dependencies and aggregate global semantic context, thereby improving detection precision while maintaining high computational efficiency. Extensive evaluations on four benchmark datasets—VisDrone2019, DIOR, RSOD and USOD—demonstrate that TSF-DETR outperforms the RT-DETR baseline and several state-of-the-art methods, validating its efficacy in strengthening small object representations and enhancing detection accuracy across diverse remote sensing scenarios.