Insulator Defect Detection and Segmentation Algorithm Based on Deformation Convolution

In this study, an algorithm named Mask R-DCN was examined and able to detect insulator self-explosions on overhead lines to achieve automatic detection and segmentation of insulator self-explosion defects in images, thus improving the efficiency and accuracy of power inspections. First, data augmentation techniques were used to expand the insulator dataset. Then, we optimized and improved the Mask Scoring R-CNN model using deformable convolution and deformable pooling layers to construct a feature extraction network; this process enhanced the network’s capability to extract features from the insulator defect images. Throughout the training, we implemented a cosine annealing learning rate schedule to optimize the model’s training effectiveness. The test results revealed that our model achieved an 86.2% recognition rate for detecting self-explosion defects in insulators, which was 3.27%higher than the Mask Scoring R-CNN model. Also, the accuracy of defect area segmentation reached 96.2%, surpassing the Mask Scoring R-CNN model by 3.4%.