Image Registration Based on Correspondence of Seabed Structures Reflecting the Sensor Characteristics of Scanning Sonar

The accurate registration of scanning sonar images is essential for underwater navigation and mapping, especially in environments where navigation sensors such as doppler velocity log or digital compasses are unreliable. Conventional sonar image registration methods often rely on navigation data or brightness-based keypoints, which are highly sensitive to intensity attenuation, reflection nonuniformity, and acoustic shadows, resulting in unstable correspondence under real underwater conditions. To overcome these limitations, this study proposes a structure-based registration algorithm reflecting the sensor characteristics of scanning sonar. The proposed method extracts structurally consistent feature points using azimuthal peak-response sampling and selects stable anchor points through multiscale accumulation. An uncertainty-weighted triplet structural hash descriptor is then constructed to model geometric consistency while suppressing unreliable correspondences. Reliable inlier pairs are identified without relying on any navigation data using k-dimensional tree based nearest neighbor search and mutual-consistency filtering. the experimental results demonstrate that the proposed method achieves a mean registration error of 4.6 pixels in translation and 0.29 °in rotation. These results confirm that the proposed algorithm can robustly align sonar images even under severe intensity distortions and shadow regions, providing a foundation for structure-based underwater image registration that reflects sonar sensor characteristics.