UIE-DDPM: Underwater Image Enhancement Based on the Integration of Physical Model and Conditional Denoising Diffusion Probabilistic Model

Underwater image enhancement (UIE) is crucial for underwater perception tasks but is challenged by complex physical degradations, including backscatter, wavelength-dependent attenuation, scattering, turbidity, and color cast. These factors severely reduce visibility and color fidelity in underwater scenes. This article presents a UIE method called UIE-DDPM, which is based on the conditional denoising diffusion probabilistic model and underwater physical model. UIE-DDPM innovatively enhances underwater images by integrating the Jaffe-McGlamery underwater physical model with the diffusion process and employing optical compensation as a conditional controller to regulate each iteration step with greater specificity. The UIE-DDPM consists of the variational autoencoder–optical compensation prediction network (VCP) and the underwater image semantic enhancement network (UISE). The VCP establishes the analytical relationship between optical compensation levels, performs Gaussian sampling and parameter renormalization, and estimates the distribution of optical compensation losses in the latent space. The UISE effectively enhances the essential semantics of degraded images by utilizing reference color-enhanced images as auxiliary information, thereby improving the understanding of details, contours, and contrasts. Meanwhile, the UIE-DDPM outperforms other baseline methods and generative adversarial network models in UIE, establishing a new state-of-the-art benchmark. In real-world application scenarios, the maximum underwater image quality metric increased by 1.336, and the maximum underwater color image quality evaluation increased by 0.158, demonstrating the effectiveness of the proposed method in enhancing underwater image quality under turbid conditions.