Speed-Up DDPM for Real-Time Underwater Image Enhancement

Underwater images often suffer from serious color bias and blurred features because of the effect of the water bodies on the light. To enhance underwater images, we present SU-DDPM, a method of real-time underwater image enhancement (UIE) based on a denoising diffusion probabilistic model (DDPM). SU-DDPM outperforms other baseline and generative adversarial network models in underwater image enhancement, thus establishing a new state-of-the-art baseline. SU-DDPM processes images more rapidly than the diffusion model, which makes it competitive with other deep learning-based methods. We demonstrate that if conditional DDPM is used directly for the UIE task, the processing speed is slow, and the enhanced images are of poor quality and show color bias. The quality of the enhanced image is improved by combining the degraded image with the reference image in the diffusion stage to create a fusion–DDPM model. The specificity of the UIE task allows us to accelerate the inference process by changing the initial sampling distribution and reducing the number of iterations in the denoising stage of the model. We evaluate SU-DDPM on the UIE task using challenging real underwater image datasets and a synthetic image dataset and compare it to state-of-the-art models. SU-DDPM ensures increased enhancement quality, and enhancement processing speed is comparable to the speed of real-time enhancement models.