A Defect Detection Method for Aerial Insulators Based on an Improved YOLOv8n Model

This paper proposes a defect detection method for insulators in transmission lines, addressing the numerous challenges they face in complex outdoor environments, including contamination, fine cracks, and missing components. The method is based on an improved YOLOv8n model, which replaces the standard convolutional layers after the first two layers with Adaptive Downsampling layers (ADown) to enhance feature extraction capabilities through multiple convolution and pooling operations. Additionally, Dynamic Snake Convolution (DySnakeConv) is introduced into the C2f module before the Spatial Pyramid Pooling - Fast (SPPF) layer, incorporating dynamic adjustment properties to more accurately adapt to and capture the complex shapes and detailed features of targets. To fully utilize the interactivity of the Path Aggregation Network (PAN), a Content-Guided Attention Fusion mechanism layer (CGAFusion) is added before the detection head, which introduces channel, spatial, and pixel-level attention information interactions to help the model better understand and handle the non-uniform fog distribution in images, thereby strengthening attention to key defect areas. Experimental results show that compared to the original YOLOv8n model, the proposed method improves detection accuracy by 5.3%, with a 0.9% increase in mAP at 0.5 IoU. Additionally, the model demonstrates enhanced convergence speed and generalization capabilities, with a parameter size of only 3M, showcasing excellent lightweight characteristics.