ECF-STPM: A Robust Crack Detection Method for Railway Catenary Components
Yujing Zhang, Junping Zhong, Zhigang Liu, Zhiwei Han
Abstract
Catenary components such as the Messenger wire base (MWB) and Rotary double ear (RDE) are important connection components in railway power supply systems. They may exhibit crack or break defects, which can be monitored using a recent advanced anomaly detection model called student-teacher feature pyramid matching (STPM), with advantages in accuracy and speed. However, the STPM has several problems with MWB and RDE crack detection. 1) Crack defects are not obvious in the low-light catenary images due to inadequate illumination, which may lead to difficulties in detecting dark cracks. 2) Target knowledge is not used during knowledge distilling, which weakens the generalization ability of the detection model. 3) Complex surface conditions of the components can cause detection errors, but it lacks a further correction step. This paper proposes a novel crack detection method called ECF (Enhanced Coarse-to-Fine)-STPM to address these issues. First, an image enhancement method, self-calibrated illumination learning (SCI), is introduced to improve image quality. It can make the features of dark cracks more obvious and easier to extract. Second, a novel coarse-to-fine model is constructed for crack recognition. In the coarse stage, a new teacher network that introduces the target component classification task rather than the original general object classification task is proposed. In the correction stage, a binary classification CNN based on data augmentation is built to correct the error labels of crack segments predicted by the coarse stage. In experiments, this paper evaluates the effect of each proposed module independently and the overall performance of the whole model on a real-life catenary component dataset. The test results show that the proposed ECF-STPM method effectively detects MWB and RDE cracks in catenary systems.