Toward Accurate and Efficient Burn Marks Inspection for MAV Using Active Learning

Insulator burn mark detection is a critical component of Micro Air Vehicle (MAV) inspection, which has evolved into the mainstream of high-voltage power line inspection. This paper addressed the challenge of burn mark inspection on power line defects in real-time, given a low volume of the collected dataset and under less human supervision. We introduced two methods for the intelligent real-time inspection of burn marks based on passive and active approaches. The proposed Models were based on YOLOv5 and achieved a frame rate of 50 FPS. The test results indicated that the proposed active deep learning approach could effectively perform insulator inspection and accurately identify several types of burn mark defects. Our proposed approach can achieve a high accuracy of mAP 0.989% while reducing the annotation overhead to less than 57% and reducing the computational cost. Moreover, the proposed active inspection model actively seeks feedback from both human annotators and other sources to tune the burn mark detection. As a result, the model can continue to learn from its errors and enhance its precision over time. In addition, it can deliver better generalization and interpretability by actively seeking diverse and representative samples during the training process.