CDDU: Convolutional Dictionary Deep Unfolding Network for Underwater Image Enhancement

Underwater images often suffer from low contrast, color distortion, and blurred details due to the absorption and scattering of light by the underwater medium, thereby increasing the difficulty of underwater tasks. Underwater image enhancement, as a key technology, has garnered significant attention in recent years, particularly in optimizing efficiency for maritime operations and advancing automatic machine learning in the shipping industry. Despite notable advancements in using deep learning algorithms to improve underwater image quality, many methods often rely on black-box models, lacking interpretability. In this article, we propose an innovative approach named CDDU, a convolutional dictionary deep unfolding network for underwater image enhancement. Specifically, we model in both color and brightness spaces to extract unique and shared features. The decomposed features undergo implicit regularization, and their modeling features are solved using a gradient descent algorithm. Simultaneously, we introduce a learnable spatial joint residual aggregation block (SJRAB) to replace the proximal operator and, finally, utilize a feature aggregation block (FAB) to merge features, reducing the loss of spatial features. Extensive experimental studies were conducted on three benchmark underwater datasets, demonstrating the superior performance of our proposed algorithm. Our model and code can be found at https://github.com/makabala/CDDU