Special target detection in UAV remote sensing imagery is challenged by composite multi-type degradation, which collectively erodes target structure across every stage of a detection pipeline. Existing methods address individual degradation types in isolation and do not generalize to the composite conditions encountered in real deployment. We propose the Riemannian Hyper-Graph Detection Transformer (RHG-DETR), a degradation-robust end-to-end framework composed of the Dynamic Receptive-field Hyper-graph Attention Network (DRHANet), the Bi-directional Weighted Adaptive Fusion Network (BWAFN), and the Adaptive Sparse Multi-scale Encoder with Dynamic Normalization (ASMED). DRHANet introduces anisotropic dynamic depthwise separable convolutions to align receptive fields with local structural orientations and Riemannian hyper-graph fusion to aggregate multi-scale features on a manifold, preserving inter-scale angular relations that Euclidean fusion destroys under degradation. BWAFN employs a bi-directional weighted pyramid in which each fusion node learns per-scale contribution weights, correcting cross-scale semantic misalignment that fixed-weight single-pass aggregation cannot recover. ASMED combines saliency-conditioned sparse window attention to suppress background dilution, a spatially gated feed-forward branch to retain pre-attention spatial geometry, and a bounded dynamic normalizer to stabilize activations under extreme illumination and electromagnetic interference. On a self-constructed UAV special-target dataset spanning seven physics-based degradation types, RHG-DETR achieves 78.5% mAP50, a 3.7% absolute gain over RT-DETR at 34.4% lower GFLOPs and 28.8% fewer parameters at 84.2 FPS, outperforming restoration-then-detect pipelines in both accuracy and latency. Consistent improvements on VisDrone2019 and BDD100K confirm cross-domain generalization.
Wang et al. (Wed,) studied this question.