Railway Track Image Stitching Based on Detail Assistance

Railway track images play an essential role in daily track inspection. However, the limited perspective presented by a single image inherently restricts its applicability. To address this issue, image stitching is applied to obtain images with a large view and enhance the detection efficiency. Most of the traditional image stitching methods utilize purely hand-crafted features for feature matching, while track textures are highly similar, which greatly increases the difficulty of feature extraction in the stitching tasks. In this work, we construct a detail-assisted method for railway track image stitching, which incorporates wavelet transform (WT) and an unsupervised convolutional neural network (CNN). More specifically, considering that railway track images have salient geometric characteristics of edges, WT is employed to extract frequency components of the data; and then CNN is employed to adaptively extract features and accordingly stitch railway track images. Experimental results on our railway image stitching dataset (RISD) verify the excellent performance of the proposed method.