Research on Defect Detection Algorithm Based on Improved YOLOX-S

Traditional object detection algorithms face issues such as excessive redundant windows, high computational load, and slow speed, especially in complex backgrounds or situations where the target shares similar colors or shapes with the background. These challenges often result in suboptimal detection accuracy and higher false positive and false negative rates. To address these problems, we propose an insulator defect detection algorithm based on an improved YOLOX-S. The paper introduces the single-stage YOLOX-S algorithm and compares it with the Faster R-CNN algorithm in detecting defects in glass and ceramic insulators. Additionally, the CAM attention mechanism is incorporated into the YOLOX-S algorithm, and the VariFocalLoss loss function is used to explore the effects of these two improvements on the detection accuracy of insulator defects. As a result, the defect detection accuracy improved from 86.3% to 96.2%, and the mAP value increased from 66.4% to 75.1%. For detecting discharge marks on ceramic insulators, the accuracy improved from 85.3% to 94.2%, and the mAP value rose from 59.1 % to 73.2%. These improvements significantly enhanced detection accuracy with minimal impact on computation speed.