Two-stage Hybrid Insulator Defect Inspection from Aerial Images

To guarantee the stable operation of the electric system, regular inspection of the electrical components, such as electrical transmission lines and insulators, is necessary. With the process of deep learning and computer vision, automatable detection and inspection of insulators becomes possible. However, the defect detection of the insulators is non-trivial due to complicated backgrounds, various occlusions, light conditions, and motion blur. To handle these problems, we propose a two-stage framework to achieve the defect detection of insulators. The proposed framework consists of the insulator detection network and the defect identification network. First, we detect the position and size of the insulator based on modified YOLOv5. Then, the full image is cropped and fed to the defect identification network. To ease the negative impact of low luminance and image blur, we adopt gamma correction and image sharpening on cropped insulator images. This formulation will improve the accuracy of the defect identification network greatly. Quantitative and qualitative results on our collected dataset demonstrate the effectiveness of our proposed model.