Improved YOLOv8 Algorithm for Detection of Small and Complex Morphological Defects in Transmission Lines

In recent years, power grid companies have used artificial intelligence image recognition technology to conduct inspections of transmission lines. Although some results have been achieved, defects of small sizes and complex shapes are still easily missed or misdetected. In response to this problem, the DRM-YOLOv8 algorithm is proposed: in the feature extraction part, a C2f_DCNv3_DLKA module that combines DLKA (Deformable Large Kernel Attention) and DCNv3 (Deformable Convolution Network v3) is designed to improve the identification and extraction capabilities of important features; a heavy-parameterized cross-over algorithm is proposed The stage feature fusion local network optimizes and fuses high-level semantic features and low-level spatial features; finally, a multi-head adaptive spatial feature fusion detection head is designed. By adding an additional small target detection layer and improving cross-scale feature fusion, the detection capabilities of multiple small targets with significant size differences are improved. Experimental results show that on a data set containing twenty types of real transmission line defects, compared with the original YOLOv8n model, the overall average detection accuracy ([email protected]) reached 84.1%, an increase of 5.2%, for small size and complex shape defects The detection accuracy is increased by 8% and 2% respectively, which is significantly better than the existing algorithm, showing excellent robustness and generalization, and effectively improving the detection ability of multiple types of defects in transmission lines in complex scenarios.