The Impact of a Deep Learning Self-Adaptive Colour Restoration Pipeline for Deep Underwater Images in 3D Reconstruction

Abstract. Underwater photogrammetry is challenged by image degradation caused by water absorption and scattering, which negatively impacts feature detection and 3D reconstruction quality. This study aims to evaluate the effectiveness of a Feedforward Neural Network-based colour correction method designed to enhance underwater image quality, thereby improving feature matching and subsequent 3D reconstruction processes. The proposed approach leverages deep learning to correct colour distortions without relying on physical models of underwater light propagation. Evaluation was performed using established feature detection algorithms, such as SIFT and SURF, applied to multiple underwater datasets capturing diverse imaging conditions. The goal is to determine whether neural network-based colour correction can increase the number of valid feature correspondences, improve sparse and dense point cloud generation, and ultimately support more accurate and robust 3D reconstructions. By integrating this correction method within a photogrammetric workflow, the study investigates the potential benefits and limitations of data-driven colour enhancement in underwater environments. The findings are intended to inform future development of hybrid approaches that combine physical modelling with deep learning, aiming to optimize both visual clarity and geometric fidelity in underwater mapping and documentation. This work contributes to advancing underwater photogrammetry by addressing critical challenges related to image quality and reconstruction accuracy, with implications for archaeological surveys, marine research, and underwater infrastructure inspection.