SMWG-DETR: DETR Enhanced by Fourier Spectral Modulation and Wavelet-Guided Fusion for Tiny Object Detection

Tiny object detection is a crucial task in the intelligent interpretation of remote sensing imagery, with significant applications in transportation, public security, and emergency management. However, the performance of existing detectors in remote sensing scenarios is still constrained by the extremely small object sizes and the presence of complex background clutter. In this article, we propose DETR enhanced by Fourier spectral modulation and wavelet-guided fusion (SMWG-DETR), which addresses the issues of spectral distribution bias during feature extraction as well as feature misalignment and detailed feature loss during feature fusion. First, Fourier spectral modulation is employed to suppress redundant frequency components in single-scale feature maps while preserving critical ones, thereby reducing spurious responses caused by cluttered backgrounds. Second, in the feature fusion stage, we apply the discrete wavelet transform (DWT) to lower level feature maps, where the resulting low-frequency and high-frequency sub-bands are used to guide higher level feature map upsampling and detailed feature refinement (DFR), thus leveraging the complementary information across multiscale features. Finally, a dynamic denoising query selection (DDQS) strategy is introduced to discard potentially misleading queries in the contrastive denoising (CDN) process, providing more accurate supervision during training. In experiments conducted on the AI-TOD and AI-TODv2 datasets, SMWG-DETR achieves average precision (AP) scores of 32.1% and 30.5%, respectively, achieving state-of-the-art (SOTA) performance.