Research on Underwater Structured Light Three-Dimensional Reconstruction Based on Deep Learning

Abstract

Underwater structured light three-dimensional (3D) imaging is essential for applications in underwater exploration and inspection. However, it suffers from stripe brightness attenuation, contrast reduction, and random noise introduction due to absorption, scattering, and refraction across media interfaces, leading to degraded stripe decoding and reconstruction accuracy. This paper presents a diffusion model-based approach for underwater structured light 3D reconstruction. Without altering the core framework of structured light imaging, the method employs the diffusion model to learn the degradation process of underwater stripes and effectively restore them, thereby enhancing 3D reconstruction quality. The approach integrates an encoder-decoder denoising network in the reverse diffusion stage, taking noisy images, time-step embeddings, and turbidity conditions as input to progressively predict and eliminate noise, ensuring incremental detail recovery and stability. Time-step embeddings are incorporated to adaptively adjust the denoising strategy at varying degradation stages, maintaining stripe structure consistency. Experiments conducted on an underwater structured light platform, with turbidity levels controlled by milk injection, demonstrate the method's performance using multi-frequency phase-shifted fringes and geometric targets. Results show that, under the highest turbidity, the PSNR reaches 42.42 dB, SSIM is 0.97, and the 3D reconstruction RMSE remains below 0.1 mm, validating the method’s robustness and high-precision reconstruction capability in complex aquatic environments.