Improved YOLOv7 Models with Attention Mechanism for Inspection of Broken Glass Insulator on Power Line UAV Images

Glass insulators play a critical role in ensuring the proper operation of high-voltage transmission lines. Intelligent inspection of power lines by drone, therefore, requires robust, fast and accurate detection models. But since the defects of broken glass insulators are extremely small and the backgrounds of aerial images are complex, coupled with a relatively small dataset, defect detection becomes a challenging task for automatic transmission line inspection. Thus, to overcome these limitations, we build a strong Vietnamese-Moroccan database of broken glass insulators with proper augmentation techniques, which we partly share publicly, and establish a comparative study based on improved versions of YOLOv7, the new state-of-the-art real-time object recognition model. In this paper, 3 versions of YOLOv7 enhancements are proposed by introducing 3 different attention blocks (NAM, GAM and SE) along with a CSP Bottleneck into the YOLOv7 header to better detect anomalies. The improved and baseline YOLOv7 models were trained, tested and compared on our dataset. The experimental results show that the mean Average Precision mAP:50 of the improved models is at best 5% higher than the baseline model and 11% lighter. Thus, validating the improved YOLOv7 models by achieving real-time performance for broken glass insulator detection in complex aerial images. The open-source dataset can be accessed via this link: https://github.com/phd-benel/BGI .