Physics-driven prior learning-based deep unrolling for underwater image enhancement

We propose a physics-driven prior learning-based algorithm unrolling approach for underwater image enhancement that leverages the advantages of both model- and learning-based approaches while overcoming their limitations. Model-based algorithms are theoretically robust because of prior knowledge of the underlying physics but may degrade image quality due to modeling inaccuracies. On the other hand, learning-based algorithms exhibit better adaptivity but inferior interpretability due to their black-box models and neglect of domain knowledge. In this work, we first formulate underwater image enhancement as a joint optimization problem with physics-based underwater-related priors and two learnable regularizers to compensate for modeling inaccuracies. Then, we solve the problem by reformulating it as a set of subproblems, which are then solved iteratively. Finally, we unroll the iterative algorithm into a deep neural network comprising a series of blocks, in which the optimization variables and regularizers are updated using closed-form solutions and learned deep neural networks, respectively. Experimental results on several datasets demonstrate that the proposed algorithm outperforms state-of-the-art underwater image enhancement algorithms on both quantitative and qualitative comparisons. The source code and pretrained models will be available at https://github.com/thithuypham/BLUE-Net.