CapNet: Learning Insulator Self-Blast From Bounding Box

Neural network detectors have been widely used to detect the insulators’ self-blast defect from aerial images. However, detector performance can be significantly limited by insufficient defect samples. To this end, this article proposes a novel detection method that can promote defect feature learning with normal samples and their bounding box annotation, so that detectors can be improved without any additional manual burden. First, an efficient contour-based segmentation network named CapNet is proposed to extract insulator caps from box annotation. Then, the memory mechanism and polar alignment loss (PAL) are designed to promote training convergence and segmentation quality, respectively. Finally, a lightweight algorithm named MirrorFill is proposed to erase randomly selected caps, thus multiple defect samples can be generated from a single normal insulator. To verify the effectiveness of the proposed method, 21 492 images with 46 387 normal and 745 self-blast insulators are collected for dataset establishment. The experimental results show that MirrorFill can effectively improve the detector performance of multiple architectures. Also, the proposed method is capable of unsupervised (Uns.) anomaly detection (AD), which can achieve 37% defect sample F1-score even training with defect-agnostic annotation.