Method for identifying micro-defects in UAV panoramic stitched images of distribution line corridors based on YOLOv8

Distribution network line corridors are highly vulnerable to environmental degradation, leading to frequent micro-defects in components like insulators. However, conventional unmanned aerial vehicle (UAV) panoramic stitching often loses fine-scale features during multi-scale transmission. Moreover, the standard YOLOv8 model suffers from limited global context modeling and significant information decay during layer-by-layer propagation, making small-target identification extremely challenging. To address these limitations, this paper proposes an advanced micro-defect identification method for UAV panoramic stitched images based on an improved YOLOv8 framework. First, a standardized data acquisition scheme constrained by ground sampling distance is designed across three coordinated flight strips to ensure raw image quality. A deep hierarchical optical flow registration and gradient-weighted fusion algorithm is then developed to generate high-fidelity panoramic images, effectively suppressing artifacts while preserving high-gradient defect edges. Crucially, systematic enhancements are implemented in YOLOv8 across three dimensions: the backbone network incorporates a parallel convolutional attention fusion module to synergistically capture local textures and global contextual dependencies; the neck network introduces a low-stage gather-and-distribute mechanism to establish a lossless transmission channel for shallow positional details; and the detection head adopts an enhanced three-branch decoupled structure integrated with adaptive threshold focus loss and differentiated angle-aware intersection-over-union loss. Experimental results on a 6.8 km distribution corridor demonstrate that the proposed method achieves a mean average precision (mAP@0.5:0.95) of 63.2% and a small-object recall rate of 88.8%, outperforming the original YOLOv8 by 2.5 and 11.8%, respectively, with a parameter increase of only 1.2 million.