Underwater image enhancement using an edge-preserving filtering Retinex algorithm

Abstract

We develop a novel edge-preserving filtering retinex algorithm for single underwater image enhancement, in which gradient domain guided image filtering (GGF) priors of reflection and illumination are embedded into a retinex-based variational framework for promoting image structures and reducing artifacts or noise. We transform an underwater image enhancement issue into a two-phase objective function. We first employ a general retinex-based method to generate guidance reflection and illumination, and then we use GGF to fuse fine structures of guidance reflection and illumination into ideal reflection and illumination. Meanwhile, the l2 norm is efficiently imposed on GGF priors which measure gradient errors between latent and ideal estimations of reflection and illumination. Then we derive an efficient optimization scheme to address the proposed model, which is fast implemented on pixel-wise operations and requires no prior knowledge about imaging conditions. Final experiments demonstrate the effectiveness of the proposed method in structures promotion, artifacts or noise suppression, naturalness and color preservation. Compared with several leading methods, the proposed method yields better subjective results and objective assessments. Furthermore, the utility of our method is extended for enhancing sandstorm and low illumination images.