High-speed railway catenary components detection using the cascaded convolutional neural networks

With the massive construction of the China high-speed railway, it is of a great significance to propose an automatic approach to inspect the defects of the catenary support devices. Based on the obtained high resolution images, the detection and extraction of the components on the catenary support devices are the vital steps prior to their defect report. Inspired by the existing object detection Faster R-CNN framework, a cascaded convolutional neural network (CNN) architecture is built to successively detect the various components and the tiny fasteners in the complex catenary support device structures. Meanwhile, some missing states of the fasteners on the cantilever joints are directly reported via our proposed architecture. Experiments on the Wuhan-Guangzhou high-speed railway dataset demonstrate a practical performance of the component detection with good adaptation and robustness in complex environments, feasible to accurately inspect the extremely tiny defects on the various catenary components.