InsDef: Few-Shot Learning-Based Insulator Defect Detection Algorithm With a Dual-Guide Attention Mechanism and Multiple Label Consistency Constraints

Abstract

Due to inefficiency and insecurity of manual insulator inspection, research on intelligent insulator inspection has received widespread attention. Previous deep learning-based defect detection methods rely excessively on Big Data training. However, in reality, defective insulator samples are too scarce to meet the training requirements of traditional neural networks. To overcome this problem, this article proposes a few-shot learning based two-stage insulator defect detection algorithm (InsDef). InsDef consists of two stages: the insulator extraction stage and the defect recognition stage. The insulator extraction stage uses YOLOv5 to locate and extract insulators in aerial image. Then, in the defect recognition stage, we use the proposed few-shot learning based defect recognition network (FSDef) to diagnose the health of insulators. FSDef employs a dual-guide attention mechanism (DGA) and multiple label consistency constraints (MLC), which enhances the network's ability to detect unobvious defects and improve network's generalization. The experimental results show that InsDef achieved a precision of 77.83%, outperforming other advanced methods. Additionally, the model size of 22.62 mb and a run time of 47.01 ms demonstrate the practical application value of InsDef.