Improved Dense 3D Reconstruction for Underwater Infrastructure Inspection Using Physics-Based Image Restoration

Vision-based inspection using underwater robots is increasingly employed for the structural health monitoring (SHM) of marine infrastructure. However, underwater optical images are degraded by wavelength-dependent attenuation and backscatter, limiting the quality of three-dimensional (3D) reconstruction. This paper presents a revised physics-based image restoration and reconstruction pipeline for underwater pile inspection. Building on the Sea-Thru framework, fundamental mathematical inconsistencies in a widely used open-source implementation are identified and corrected, and implementation-level optimisations are introduced. The revised pipeline achieves an average processing speed improvement of 89% while producing more physically consistent colour restoration. Integration with a stereo-based reconstruction workflow demonstrates significant improvements in surface coherence and texture fidelity on real underwater inspection data. The results show that careful implementation of physics-based restoration methods can materially improve vision-based underwater SHM workflows.