Damaged Wire Detection and Segmentation Method Based on Lightweight Channel Attention Mechanism and Transfer Learning

Detecting and segmenting damaged wires remains challenging due to varying lighting conditions and the constraints imposed by limited sample datasets, both of which severely degrade detection and segmentation accuracy. To address these challenges, this paper proposes a novel SECA (Squeeze-and-Excitation Channel Attention) mechanism integrated with YOLOv8n, aimed at enhancing key feature extraction under varying lighting conditions. The incorporation of transfer learning strategies further mitigates the challenges posed by insufficient sample datasets. We built an RGB dataset comprising 3,760 images representing diverse wire damage morphologies and expanded the lighting test sets to span six environments, with illuminance levels ranging from 50 to 200,000 lux, to comprehensively evaluate the model's robustness. Experimental results demonstrate that the proposed model markedly improves detection and segmentation precision as well as robustness across varying lighting conditions.