Two-Stage Graph Neural Architecture Search for Visible-Infrared UAV Small Target Detection

Unmanned aerial vehicle (UAV) detection is critical for low-altitude airspace security, yet remains challenging due to small target size, limited texture, and severe visible-image degradation under nighttime or adverse weather conditions. Although infrared imaging alleviates illumination dependence, simply fusing visible and infrared features without explicit crossmodal alignment introduces modality interference and fails to exploit their complementarity. We propose a dual-stream detection framework built on YOLOv11s that addresses this via a twostage graph neural architecture search (NAS) fusion approach. In Phase 1, a cross-modal bipartite graph is constructed at each feature scale, where visible and infrared nodes interact exclusively across modalities via a 3-layer GNN whose operators are automatically selected from a compact 10-operator search space using DARTS-style differentiable optimization with shared architecture parameters, eliminating the need for manual fusion design. In Phase 2, global average pooling collapses spatial dimensions and a fully-connected graph over the three-scale channel descriptors enables cross-scale semantic reasoning, with a zero-initialized gated sigmoid attention modulating Phase 1 features for stable training. Experiments on the AntiUAV300 benchmark show that our method achieves mAP@ 50 of 0.991 and mAP@ 50: 95 of 0.639 with only 18.58 M parameters, outperforming state-of-the-art multimodal detectors including ICAFusion, RsDet, and SuperYOLO across all key metrics. These results demonstrate the effectiveness and robustness of the proposed approach for reliable UAV detection in complex low-altitude scenes, offering a promising solution for surveillance, security, and airspace management applications.