Visible-Light Insulator Defect Detection Algorithm Based on Improved YOLOv4-Tiny

Abstract

Insulator self-explosion defect detection plays a crucial role in the stable operation of power systems. In recent years, with the development of technology, deep learning-based object detection models have been widely used in power inspection work. To address the problems of redundant structure, large size, and slow detection speed of traditional object detection models, a lightweight insulator self-explosion defect detection algorithm based on YOLOv4-Tiny (You Only Look Once version 4-Tiny) is proposed. Firstly, the model training strategy of transfer learning is adopted to improve the model's generalization ability to the insulator dataset. Secondly, the SE attention module is introduced in the feature pyramid network to strengthen the model's feature extraction and fusion capability. Finally, a small target detection layer is added to enhance the model's ability to recognize small targets. Experimental results show that the average precision of the improved algorithm in this paper is 91.45%, which is 8.46% higher than the model using transfer learning. Additionally, the model size is 22.9MB, indicating that the model can achieve high detection accuracy for insulator and defect recognition, and the model's size is smaller than that of general object detection models, making it suitable for deployment on front-end devices to achieve real-time detection.