Defect Recognition of Insulators on Catenary via Multi-Oriented Detection and Deep Metric Learning

Nowadays in electrified railway industry, 4C running vehicle is widely used to record images of catenary for defect recognition of insulators. This paper proposes a vision-based solution that applies deep convolutional neural network to recognize defect of insulators. Two main stages are included in the proposed method. First, a framework called Angle-enhanced Faster RCNN is proposed as a multi-oriented detector to localize insulators on catenary. Then, a discriminator is proposed to decide whether an insulator is defective using deep metric learning. Compared to conventional form of detection, the detector precisely takes out insulators into horizontal position so that defect discrimination with similarity measurement can be applied. For the discriminator, a novel feature extractor based on convolutional neural network is designed and triplet loss is applied as the loss function, which is more robust than classification method when facing a lack of defective samples. Experiments indicate that this solution achieves a good performance of accuracy.