Resumen
Insulators are widely used in various aspects of the power system and play a crucial role in ensuring the safety and stability of power transmission. Insulator detection is an important measure to guarantee the safety and stability of the transmission system, and accurate localization of insulators is a prerequisite for detection. In this paper, we propose an improved method based on the YOLOv5s model to address the issues of slow localization speed and low accuracy in insulator detection in power systems. In our approach, we first re-cluster the insulator image samples using the k-means algorithm to obtain different sizes of anchor box parameters. Then, we add the non-local attention module (NAM) to the feature extraction module of the YOLOv5s algorithm. The NAM improves the attention mechanism using the weights? contribution factors and scaling factors. Finally, we recursively replace the ordinary convolution module in the neck network of the YOLOv5 model with the gated normalized convolution (gnConv). Through these improvements, the feature extraction capability of the network is enhanced, and the detection performance of YOLOv5s is improved, resulting in increased accuracy and speed in insulator defect localization. In this paper, we conducted training and evaluation on a publicly available dataset of insulator defects. Experimental results show that the proposed improved YOLOv5s model achieves a 1% improvement in localization accuracy compared to YOLOv5. The proposed method balances accuracy and speed, meeting the requirements of online insulator localization in power system inspection.