Multi-scale diffusion model for underwater image restoration and enhancement

Abstract

Background: Underwater environments face challenges with image degradation due to light absorption and scattering, resulting in blurring, reduced contrast, and color distortion. This significantly impacts underwater exploration and environmental monitoring, necessitating advanced algorithms for effective enhancement. Objectives: The study aims to develop an innovative underwater image enhancement algorithm that integrates physical models with deep learning to improve visual quality and surpass existing methods in performance metrics. Methods: The proposed method employs a multi-level design combining physical insights from the diffusion model with deep learning. It utilizes an encoder-decoder pipeline for image decomposition, adaptive color adjustment through a color encoder network, and pixel optimization using an inverse denoising diffusion model to achieve enhanced visual quality. Results and conclusion: Experimental results demonstrate significant improvements, with an average increase of 6.61 in PSNR, 0.15 in SSIM, and 0.87 in UIQM. The integration of physical models and deep learning proves effective for underwater image enhancement. This research advances the field by providing a robust framework for addressing imaging challenges in underwater environments.