A Nested Residual Encoder-Decoder Network for Overhead Contact System Fastener Anomaly Detection

An overhead contact system (OCS) is key to providing power high-speed railways. OCS detection an important measure ensure the safe operation of a railway. At present, anomaly mainly relies on manual analysis images regularly collected by 4C system, which very inefficient and can easily miss anomalies. Although some classification object methods based deep learning be used for detection, effective training networks difficult support due small number image samples. Considering that most faults are abnormal fasteners, we propose method normal images, called nested residual encoder-decoder network (NRE-Net). This consists two networks, where encoder shared part, structure added encoding decoding branches enhance feature expression ability. The experimental results show greatly improve accuracy CIFAR-10 dataset fastener dataset. Compared with previous state-of-the-art approaches, $F_{1}$ score proposed classes in has increased 10.8% 11.9%, respectively.