Underwater Image Enhancement Based on Accelerated Conditional Diffusion Probabilistic Model

ABSTRACT Underwater images often suffer from significant color distortion and blurred features due to optical loss and dispersion. This degradation can hinder tasks such as underwater object detection. To address this issue, this study proposes an underwater image enhancement (UIE) model based on an accelerated conditional diffusion probabilistic model (UW‐DDPM). This model is a rapid denoising diffusion probabilistic model designed specifically for UIE. The UW‐DDPM directly establishes a diffusion generation relationship between degraded and reference images based on the conditional diffusion probabilistic model (CDDPM) redesigning an implicit accuracy diffusion model for direct image translation, which not only improves the quality of image enhancement but also addresses the slow sampling speed issue of the CDDPM. Simultaneously, speed‐UIE was designed for processing training on conditional images, which is a lightweight model network. Specifically, we combined a pre‐trained diffusion model with a lightweight UIE algorithm, using speed‐UIE to guide conditional generation. The diffusion prior mitigates the drawbacks of poor‐quality synthetic images, whereas the lightweight model addresses the issue of the diffusion model lacking high‐quality prior conditions, resulting in higher‐quality images. Ablation experiments demonstrate that enhancing the conditional images before inputting them improves the visual quality of the output images. Extensive experiments on publicly available UIE datasets have verified that the UW‐DDPM outperforms existing traditional and deep learning‐based methods in terms of full‐reference, no‐reference image quality assessment metrics, and generation speed. The UW‐DDPM and other state‐of‐the‐art (SOTA) methods are used to compare the image enhancement experiments of underwater robots in the field. The UW‐DDPM still demonstrated excellent robustness in practical applications.