Self-supervised Railway Pantograph Image Component Retrieval with Geometry Prior

The pantographs are important infrastructures in the railway traction power supply, therefore, whose serving status are frequently monitored, in order to detect any faults or anomaly as early as possible. To visually understand the inspection images, the pantograph pixels should be grouped with markings to indicate specific components. In this paper, a novel unsupervised image component retrieval method is proposed for pantograph visual inspection. To fully utilized the prior knowledge of the interested artificial objects, predefined 3D models are used to estimate latent geometric pose parameters, so as to assist the retrieval of the specified component. Particular deep Q-network based reinforcement learning is designed and trained with the help of an environmental simulator to interactively search optima in high-dimensional parameter space with a global envision. Experiments on the synthesis and real datasets proved the effectiveness of the proposed method in pantograph monitoring.