Insulator is one of the most common and widely used electrical components in transmission lines, which plays a critical role in electrical insulation and mechanical support, ensuring that the current flows along the specified path and reducing electromagnetic interference with the surrounding environment. Since insulators are installed outdoors, they are exposed to wind, sunlight, rain, ice, frost and other bad weather for a long time, and their surface defects are inevitable. If the insulator appears self-explosion or drop string, which will cause leakage due to the loss of insulation, leading to electric shock accidents, thus resulting in huge economic losses. Relying on computer vision and deep learning technology, insulator defect detection from massive UAV aerial images has become an urgent problem for power operation and maintenance. However, the backgrounds of aerial images from overhead transmission line corridors are complex. Under different lighting conditions, shooting angles, shooting distances, etc., the scale of insulators in aerial images varies greatly, and the insulator strings are prone to occlusion, the defect area of the insulator is much smaller than the insulator itself. Therefore, there are numerous difficulties in detecting insulator defects in practical applications.

In recent years, compared with the traditional object detection methods, deep learning methods can quickly and accurately identify insulators and their defects from power inspection images. There is still a lack of comprehensive review of the latest progress in insulator defect detection in existing literature, without introducing object detection algorithms such as anchor free algorithm, YOLOv7, Transformer, and knowledge extraction techniques. In view of this, this article summarizes and analyzes a large number of visual methods for insulator defects detection, systematically reviews deep learning methods for insulator defect detection in drone aerial images, aiming to select appropriate detection methods for specific insulator defects and provide valuable reference for researchers engaged in transmission lines fault diagnosis.

This paper reviews the research progress of deep learning methods for insulator defect detection in UAV aerial images. Firstly, the current research status of transmission lines inspection based on deep learning is briefly reviewed. Then, the insulator defect detection methods based on deep learning are explained, mainly from the target detection models, lightweight network models, cascade detection models and other methods are summarized, which is conducive to the comparison between different deep learning methods and more helpful for power inspection personnel to select appropriate deep vision detection methods for fault diagnosis of specific electrical component. And the target detection models based on two-stage algorithms, one-stage algorithms and anchor-free algorithms are elucidated. The lightweight network models based on model pruning, knowledge distillation, low-rank decomposition, network quantization and the target detection model based on Transformer are summarized. Next, the self-built and public datasets for insulator defect detection are introduced. Due to the lack of training samples and unified dataset for insulator defect detection, scholars mostly conduct defect detection research through self-built datasets in different detection scenarios. Finally, the challenges faced by insulator defect detection methods based on deep learning are elucidated, including insufficient defect samples, low defect detection accuracy, difficulty in detecting small target defects, and feature extraction, etc. Based on existing deep learning techniques and recent research ideas, several important research directions in the future are pointed out, including expanding insulator defect samples, establishing unified performance evaluation indicators, small and zero sample learning, new defect detection frameworks, multi-level detection of small defects, deep fusion of multiple learning technologies, cloud-edge-end collaborative fusion, and improving network model stability and real-time performance.