An Image-Based Deep Learning Approach with Improved DETR for Power Line Insulator Defect Detection

Abstract

Insulators are basic parts of high-voltage transmission, and detecting faults of insulators is a critical task. Most state-of-the-art methods contain two or more stages, including insulator detection and defect locating. Some also involve hand-designed components to improve the performance due to the complicated and misleading background of the wild. To automatically detect faults in UAV-captured insulator images, this paper presents a method that introduces DETR into insulator defect detection. With the self-attention mechanism in Transformer, the model can naturally exploit its advantage in focusing on the target area. However, training DETR requires large data sets and long training schedules to establish spatial relations in sparse locations, which makes it generally not feasible to train in small data sets. To explore the possibility of training a well-performing model with a data set that minimizes the cost of collecting insulator images, transfer learning techniques were applied to this process. To compensate for the disadvantage of DETR in detecting small objects at more precise scales, an improved loss was transplanted to this model. The results show that our proposed method can detect defects directly from UAV images without the need to locate the insulator first, while providing competitive performance with a lower cost of collecting training samples.