Seeing through scattering media via unfolded sparse-dark-channel model

When seeing through scattering media, scattering effects of the particles always degrade the quality of the acquired images. Therefore, it is crucial to restore the ground-truth image from its low-quality observation, which is often called image dehazing. Although deep learning solvers have boosted the area of image dehazing recently, challenges such as the loss of interpretability and data dependency also remain. In this paper, we propose a dehazing network that combines imaging models with data-driven learning techniques using algorithm unfolding. In the first stage, a regularization dehazing model based on a sparse-dark-channel prior is developed, which integrates dark and bright channel priors for transmission map estimation. And results in a more accurate parameter estimation in complex scenes. Then, an unfolded network is designed to optimize the model, where these regularization parameters are learned from the training data rather than heuristically determined. Numerical experiments show the unfolded network with only 6 layers efficiently restores image details, achieving improvements in both visual quality and measure metrics compared to these recent methods. Our work offers a new perspective for integrating physical informed models and data-driven learning methods in scattering imaging.