A Double Knowledge Distillation Framework for Insulator Defect Detection

Insulator defect detection is essential for maintaining reliable power delivery systems. Recently, insulator image detection has emerged as a promising alternative to traditional manual inspections. However, existing datasets primarily focus on object detection, and current methods often suffer from limitations in accuracy and computational efficiency. To address these challenges, this article develops a specialized insulator defect segmentation dataset and a novel double knowledge distillation framework aimed at achieving superior detection performance with a lightweight model. The proposed method begins with fine-tuning a large-scale teacher model. A compact student model is then trained through the first knowledge distillation process, utilizing a distinctive contrastive learning strategy. Unlike existing methods that directly compare segmentation masks, we contrast the no-prompt output of the teacher model with the output mask generated by the student model, enhancing the granularity of knowledge transfer, with the underlying theory established and proven. In the second phase of distillation, generative adversarial networks facilitate deeper knowledge distillation, where the student model generates synthetic labels and the teacher model provides ground-truth labels, iteratively refining the student’s capabilities. Experimental results show that the proposed framework significantly reduces model size while maintaining high performance in insulator defect segmentation.