Diffusion Probabilistic Models for Underwater Image Super-Resolution

In recent years, single image super-resolution (SISR) has been extensively employed in the realm of underwater machine vision. However, the unique challenges posed by the underwater environment, including various types of noise, blurring effects, and insufficient illumination, have rendered the recovery of detailed information from underwater images a complex task for most existing methodologies. In this paper, we address and propose solutions to these challenges inherent in the application of super-resolution techniques in underwater machine vision. We introduce a novel underwater SISR diffusion probability model, termed as DiffUSR. This marks the first instance of utilizing a diffusion probability model in the domain of underwater SISR. Our innovative model enhances the data likelihood by employing a unique variant of variational constraints. Notably, DiffUSR is capable of providing diverse and realistic super-resolution (SR) predictions by progressively transforming Gaussian noise into SR images based on low-resolution (LR) inputs via a Markov Chain process. This approach represents a significant advancement in the field of underwater image super-resolution.