An improved subpixel-level registration method for image-based fault diagnosis of train bodies using SURF features

Abstract

Using computer-vision-based fault-diagnosis technologies, anomalies in train bodies can be automatically identified as part of daily maintenance. Nevertheless, since the train speed changes constantly and nonlinearly during the process of image acquisition, the obtained images can be severely distorted. As such, it is necessary to first register the collected images to correct the distortion, so as to accurately extract the key parts of the train body for fault detection. This paper proposes a new accurate registration method for train images based on speeded-up robust features, which improves on the traditional segment registration (SR) method by adding procedures for feature selection, piecewise linearization, and interpolated registration. The novelty of the proposed method is that up to subpixel-level registration accuracy can be achieved. Meanwhile, the robustness of the algorithm is ensured by its strong ability to screen out non-matching pairs when selecting features for registration. To demonstrate the performance of the proposed scheme, experimental comparison and verification are conducted, which show that, compared with the traditional SR method, the registration accuracy is remarkably improved in terms of mean square error, normalization cross-correlation, and structural similarity index.