Change detection in side-scan sonar imagery based on deep learning feature matching methods

This paper explores change detection in repeat-track side-scan sonar imagery through feature matching. It addresses insufficient matching accuracy and stability in low-contrast, noisy, and geometrically distorted side-scan sonar imagery typically collected from surface vehicles. The experiment included a comparison of classical, convolutional, and transformer-based feature matching methods (SIFT, DISK, SuperPoint, LoFTR, and LightGlue) on two real-world datasets, Atlantic and Baltic. The results were evaluated quantitatively and qualitatively. Quantitative evaluation used displacement, angular stability, and reprojection error metrics, as well as resource consumption metrics like execution time and memory usage. In addition, matching maps and change maps for pairs of images were generated and analyzed qualitatively. All methods produced interpretable change maps for the low-noise Baltic dataset, whereas the wave-affected Atlantic dataset with stripe- and speckle noise only occasionally produced consistent maps. The SuperPoint + LightGlue method ­demonstrated the highest ratio of inlier correspondences after RANSAC filtering (43.4% and 65.6%) and the lowest mean reprojection error (36.0 and 3.9 px), while LoFTR provided the densest coverage (up to 97%) consuming up to 15× more computational resources. These results confirm the advantage of transformer-based matching methods under challenging conditions due to their global receptive field. In contrast, CNN-based methods performed better in low-noise, well-aligned images. Overall, the findings indicate that deep feature matchers can improve the applicability and reliability of change detection in tasks such as humanitarian demining, autonomous underwater navigation, image mosaicking, and related applications