Aqua3DGS: Enhanced 3D Gaussian Splatting for Robust Underwater Scene Reconstruction

Underwater 3D reconstruction has emerged as a critical tool for applications in marine science, archaeology, and industrial inspection. However, the unique challenges posed by underwater environments, such as turbidity, color attenuation, and light scattering, often degrade image quality and hinder accurate 3D modeling. These challenges necessitate innovative approaches that can overcome the inherent limitations of underwater imaging systems and provide robust reconstructions.To address these issues, this paper proposes a novel framework that integrates 3D Gaussian Splatting with underwater-specific enhancements to achieve high-precision, fast, and robust 3D reconstruction. The proposed system begins with a tailored pre-processing module for underwater imagery, including advanced color correction and noise reduction. A Transformer-based enhancement network is introduced, incorporating a PAdaIN module, prior and posterior distribution estimation, consensus prediction, and a fine-grained detail enhancer. These modules collectively improve the visual quality of underwater images before 3D modeling. This pipeline includes a noise suppression module during point cloud generation to mitigate spatial uncertainty and ensure cleaner reconstructions. Experimental results demonstrate significant improvements in both visual quality and reconstruction accuracy.