Insulator Self-explosion Defect Detection Based on Image Semantic Segmentation

Insulators, a special insulation control, play a role in overhead transmission lines. Insulators exposed to the natural environment for a long time are prone to self-explosion, resulting in line failure and interruption of power transmission. Therefore, insulator self-explosion defect detection is important. To detect insulator self-explosion defect in complex background images, we decompose the target task into two steps: image segmentation and object location and finally realize the automatic detection of insulator self-explosion defect on our segmentation images. For the segmentation task, we propose an insulator string segmentation model based on affinity and contour boundary modelling, which improves the segmentation performance. For the detection task, Yolov4 is used to locate insulator self-explosion defects based on our segmentation images. The experiments show that the AP50 on our segmentation images is 3.6% higher than that on the original images, and the F1-score on our segmentation images is 0.18 higher than that on the original images.