Research on Power Insulator Defect Detection Based on Improved YOLOv8

To address the challenges of small insulator sizes, subtle defect features, and severe background interference in power transmission scenarios, traditional YOLOv8 models suffer from high missed detection rates for small targets and insufficient feature extraction capabilities, struggling to balance detection accuracy and real-time performance. This paper proposes an improved framework based on YOLOv8, integrating an attention mechanism and multi-scale feature fusion (AMF). The framework employs a channel attention module to filter critical feature channels, utilizes spatial attention to localize defect regions, and combines multi-scale fusion technology to enhance adaptability to targets of varying sizes. Experiments on a self-built power equipment dataset (including insulators and common defect samples) demonstrate that the traditional YOLOv8 achieves only 67.8% detection accuracy for tiny defects, while the AMF framework improves this to 76.3%, representing an 8.5% enhancement in accuracy while maintaining lightweight characteristics. This provides a feasible solution for real-time insulator defect detection in complex scenarios.