Real-time target recognition method for multi-UAV cooperative intelligent disaster assessment of distribution network lines under image enhancement

After extreme meteorological events such as typhoons, ice disasters, and floods, distribution network lines are often in complex visual environments intertwined with low illumination, dense fog and ice coating, high-humidity reflection, and motion blur. This leads to global contrast degradation in UAV inspection images. A single UAV is prone to single-view occlusion and distortion, insufficient feature alignment and small-target semantic expression, and poor recognition ability for tiny targets. To address these issues, a method integrating image enhancement and real-time multi-UAV cooperative target recognition is proposed. A 3D motion model with individual safety, formation configuration, and inter-UAV collision avoidance constraints is constructed. Combined with Euler discretization, recursive update of UAV states is realized to ensure full-area coverage and safe cooperative acquisition of post-disaster distribution network lines by the formation. A multi-scale feature enhancement module and spatial-channel attention mechanism are embedded into a cycle-consistent generative adversarial network, combined with power structure prior constraints, to achieve brightness restoration, noise suppression, and edge fidelity of low-quality images. A projective transformation model is used to complete spatial alignment of multi-view images. A position-attribute joint association criterion is constructed based on Jaccard intersection-over-union and Hellinger distance to realize accurate matching and feature fusion of multi-UAV image targets. The YOLOv11 network is improved, and a cross-enhanced feature fusion module is designed, which includes an edge-aware adaptive aggregation unit and dual-domain selection mechanism to register features, fully express small-target semantics, and improve the recognition ability for tiny and occluded targets. Experimental results demonstrate that the proposed method achieves a minimum recognizable size of 5 mm for tiny targets such as insulators and fittings, with a maximum average single-target positioning deviation of ≤ 2.0 m. The enhanced images attain a peak signal-to-noise ratio (PSNR) in the range of 33.8–37.5 dB, providing high-quality inputs for reliable detection. It exhibits excellent refined recognition performance and environmental robustness in various typical disaster scenarios.