Abstract
Drop brackets (DB) and swivel clevises (SC) are critical components of railway catenary systems, playing a key role in maintaining cantilever stability. The condition of these components significantly impacts the safe operation of the catenary, necessitating periodic inspections to detect defects. This task is typically performed by onboard cameras using computer vision. However, traditional image processing methods often focus on shallow features, making it difficult to handle the interference from complex structures of components. While deep learning methods have strong capabilities in capturing semantic features, the lack of crack samples makes reliable crack identification challenging. Therefore, a joint approach for crack detection based on structural perception is proposed. The approach integrates three main components: object structure perception, stick structure perception, and crack defect detection. A multistream catenary components segmentation network (MCSnet) is employed to extract structural features of the DB and SC. Subsequently, an adaptive stick perception method (ASPM) is applied to identify potential crack candidates based on stick structure. The combined structural features enable effective detection of crack defects. Experimental results validate the effectiveness of the proposed approach.