A Real-Time Siamese Network Based on Knowledge Distillation for Insulator Defect Detection of Overhead Contact Lines

As an essential component of the high-speed railway overhead contact lines (OCLs), the insulator supports OCLs while maintaining the insulation between OCLs and earth. Because of the lack of defect samples and the variety of defect types, achieving full automation of insulator defect detection using computer vision is, however, still challenging. To overcome these problems, this article proposes a real-time, unsupervised learning Siamese defect detection network (SDDN) based on knowledge distillation. It includes a teacher network (TN) and a student network (SN). Our method is mainly divided into two stages. In the first stage, insulators are quickly and accurately localized from OCL images. Then, insulators are sampled into small patches under the sliding window. These small patches are fed into the SDDN for defect detection in the second stage; furthermore, the defect scores of samples are determined by SDDN. If the time cost of ImageNet-1k pretraining for the TN can be afforded, we provide a faster version: Faster SDDN. During the training phase, whether it is SDDN or Faster SDDN, TN, however, only uses normal samples to distill the knowledge of the deep features to SN. The dissimilarity between the distilled features of SN and TN is applied to score the samples’ defect scores at the testing phase. The defect detection experiment using the insulator dataset of the Linzi-Qingzhou City north high-speed railway proves the effectiveness of our method.