A Two-Stage Insulator Defect Detection Network With Sequence Transduction

Insulators play a vital role in power systems, and the application of computer vision has effectively improved the detection efficiency of insulator defects. However, most existing methods do not fully utilize the spatial correlation between the insulator region and the defect region. This study proposes a two-stage defect detection network based on sequence transduction, which detects insulator defects by contrasting differences between defect patches and normal patches. In the first stage, an improved oriented region-based convolutional network (improved oriented RCNN) is used to accurately locate the insulator and preliminarily identify insulator defects. Next, the insulator region is segmented based on the detection results from the first stage to serve as input for the second stage. Insulator defects are then precisely located using the insulator defect detection (IDD) module, which employs a change-weight multihead self-attention (MHSA) structure for patch segmentation and comparison. Finally, the restricted non-maximum suppression (Restricted NMS) is utilized to integrate and analyze the defect detection results from both stages to derive the final recognition results. 2006 images from the power grid were utilized to train and test the model’s performance in NVIDIA A30. The experimental results clearly indicate that the proposed network achieves a significant 9.1% improvement in mean average precision (mAP) compared to the best-performing oriented object detection method. This validates the effectiveness of defect detection through patch comparison.