UNDERWATER IMAGE ENCHANCEMENT BASED ON CONDINITIONAL DENOISING DIFFUSION PROBABILISTIC MODEL

Abstract

Underwater images often suffer from color distortion, low contrast, and noise due to light attenuation and scattering in water. Traditional enhancement methods such as CLAHE, DCP, MIP, RGHS, and ULAP improve specific image characteristics but generally lack adaptability across different underwater environments and lighting conditions. To overcome these limitations, DiffWater, a conditional denoising diffusion probabilistic model (DDPM)-based underwater image enhancement method, is proposed. The model incorporates a color compensation strategy in the RGB color space to guide the denoising process and restore degraded underwater images effectively. By leveraging conditional guidance and diffusion-based image generation, DiffWater achieves robust enhancement, improved visual quality, and stronger generalization across diverse underwater scenarios. Experimental results on multiple real underwater image datasets demonstrate that DiffWater outperforms existing methods in terms of enhancement quality, effectiveness, and robustness.