DFSO-DETR: dual-domain Fourier fusion DETR for small object detection in remote sensing images

Small object detection in remote sensing imagery remains challenging because many targets occupy only a few pixels, appear in dense or arbitrarily oriented layouts, and are embedded in cluttered, low-contrast, or haze-degraded backgrounds. Although DETR-style detectors offer end-to-end prediction and long-range contextual reasoning, their decoding accuracy is still constrained by the quality of the encoded multi-scale features. Once weak boundaries and fine spatial details are suppressed by the backbone or neck, query-based decoding cannot reliably recover them. This article proposes DFSO-DETR, a Dual-Domain Fourier Fusion DETR for remote sensing small object detection, built upon the efficient RT-DETR framework. Instead of enlarging the detection head, DFSO-DETR strengthens feature extraction and fusion through two coordinated components. The Directional Fourier-Enhanced Network (DFENet) combines Pinwheel-Shaped Convolution with an Edge Fourier Block to capture directional local structures and fuse Scharr-enhanced spatial responses with learnable Fourier-domain representations. The Dual-domain Small Object Pyramid (DSOP) uses space-to-depth transformation and dual-domain attention to retain low-level details while aggregating semantic context. Experiments on VisDrone2019, HazyDet, and AI-TOD-v2 report 51.1%, 78.9%, and 59.9% mAP50, respectively, with approximately 20.6 M parameters. Ablation studies show that the directional, edge-frequency, and detail-preserving pyramid components contribute complementary gains. These results indicate that spatial-frequency feature refinement is an effective and practical strategy for end-to-end UAV and aerial small object detection in dense, hazy, and tiny-object scenarios.