Shape-Matching Model Optimization Using Discrete-point Sampling and Feature Salience

The component classification and potential fault region locating in the full-automatic inspection system of a freight train require a computer vision method with the ability of classifying quickly and locating precisely, addressing anti-nonlinear deformations, and being able to perform extensible learning. Inspired by these requirements, this paper specifically optimizes the three elements of a shape-matching model, including the scene map, the shape template, and the matching. Our method uses a discretepoint sampling map (DPSM) as an intermediate representation, to enhance the stability of the scene maps, uses the criterion function based on feature salience to select a better shape-template group, and matches hand-sketches with regions in DPSMs to reduce the difficulty of the matching calculation. Based on our optimized shapematching model, we set up a new procedure for component classifications and potential fault region locating in the fullautomatic inspection system for freight trains, which has been applied successfully on more than 10 parts of freight train cars in the railway for more than 2 years. The results of anti-noise testing in laboratory and daily operation at several inspecting stations show that our method has a strong ability to survive with nonlinear deformations, and has a good extensibility to be used with different parts, which meet application demands for the full-automatic inspection system.