Coarse-to-Fine Non-Rigid Registration for Side-Scan Sonar Mosaicking

TL;DR

The paper tackles non-linear, spatially varying distortions in large-scale side-sCan sonar mosaicking by introducing a coarse-to-fine registration framework that fuses global structure via Thin Plate Splines with locally adaptive refinement. A superpixel-guided patch strategy enables context-aware local registration, where an unsupervised SynthMorph model provides dense deformation fields, subsequently fused with the global TPS to form a smooth, coherent mosaic. Extensive experiments on challenging sonar data demonstrate superior accuracy, structural consistency, and deformation stability compared with rigid, non-rigid, and learning-based baselines, while also highlighting the value of global-to-local fusion and artifact suppression. The work offers a practical, unsupervised local refinement paradigm tailored to low-texture sonar imagery, advancing large-scale seabed mapping through robust, scalable non-rigid registration.

Abstract

Side-scan sonar mosaicking plays a crucial role in large-scale seabed mapping but is challenged by complex non-linear, spatially varying distortions due to diverse sonar acquisition conditions. Existing rigid or affine registration methods fail to model such complex deformations, whereas traditional non-rigid techniques tend to overfit and lack robustness in sparse-texture sonar data. To address these challenges, we propose a coarse-to-fine hierarchical non-rigid registration framework tailored for large-scale side-scan sonar images. Our method begins with a global Thin Plate Spline initialization from sparse correspondences, followed by superpixel-guided segmentation that partitions the image into structurally consistent patches preserving terrain integrity. Each patch is then refined by a pretrained SynthMorph network in an unsupervised manner, enabling dense and flexible alignment without task-specific training. Finally, a fusion strategy integrates both global and local deformations into a smooth, unified deformation field. Extensive quantitative and visual evaluations demonstrate that our approach significantly outperforms state-of-the-art rigid, classical non-rigid, and learning-based methods in accuracy, structural consistency, and deformation smoothness on the challenging sonar dataset.

Figures (8)

• Figure 1: Coarse-to-fine hierarchical registration framework for side-scan sonar images. Global TPS warping compensates large-scale distortions, followed by superpixel-based local partitioning. Each local patch undergoes artifact suppression preprocessing and unsupervised deep registration to estimate dense deformation fields. Local deformations are fused into a global transformation, combining TPS-based global modeling and deep learning-based local refinement for accurate non-rigid alignment.
• Figure 2: Initial global non-rigid alignment and local patch generation pipeline. Radiometric correction mitigates range-dependent intensity inhomogeneity through column-wise amplitude scaling and local contrast enhancement. Sparse control points in overlapping regions guide TPS estimation to produce a coarse global deformation field, aligning the moving image to the fixed image. Subsequently, superpixel segmentation on the fixed image generates structurally consistent regions, which are enclosed by bounding boxes to define regular local patches. Corresponding patch pairs are extracted from the fixed and globally warped moving images, forming the basis for localized non-rigid refinement.
• Figure 3: Patch-wise non-rigid registration workflow. Each superpixel-derived patch pair undergoes preprocessing to suppress stripe artifacts via non-local means filtering and enhance structural details using a Laplacian sharpening filter. The preprocessed patches, resized to fixed dimensions, are input to an unsupervised SynthMorph network that predicts dense local displacement fields. These displacement fields warp the moved patches to align with fixed patches. The resulting local deformation fields capture fine-scale geometric variations essential for precise alignment.
• Figure 4: Visualization of the global registration process. Top row: Left — original and preprocessed fixed/moving images in grayscale and pseudo-color formats (where higher intensities appear yellow-green, lower intensities appear blue-green). Right — manually selected control point pairs and the resulting initial TPS alignment on the moving image. Bottom row: Left — separation of overlapping and non-overlapping regions after initial TPS registration. Overlapping results and deformation fields are shown first, followed by superpixel segmentation and bounding box partitioning on both the fixed and warped images. Middle — refined global registration result (overlapping region only) and its deformation field. Right — extrapolated non-overlapping results from TPS and final mosaic output.
• Figure 5: Local visualization of the patch-wise non-rigid registration process. This figure presents local registration results across 5 selected patch regions (rows). Each row includes the following columns: (1) Fixed patch (reference); (2) Preprocessed fixed patch; (3) Warped patch (extracted from globally warped image via bounding box); (4) Preprocessed warped patch; (5) Locally registered patch using SynthMorph;(6) Corresponding local displacement field (visualized as a dense displacement grid).