Zero-reference single underwater image enhancement

Underwater images play an essential role in acquiring and analyzing underwater information. Autonomous Underwater Vehicles (AUVs) highly rely on the quality of the captured underwater images, in order to carry out several activities. Due to the poor lighting conditions and the limited capacity of the optical imaging device, captured underwater images usually contain severe color distortions and contrast reduction. To this end, most existing deep learning-based underwater image enhancement methods synthesize the pseudo ground-truth, or employ the in-air clear images as references to train the models. However, the synthesized or selected reference images are generally unsatisfying due to the lack of diversity and applicability. This paper presents a novel underwater image enhancement approach based on training an end-to-end underwater image enhancement network, without using any reference image. A novel encoder-decoder network structure and a set of non-reference loss functions are designed to measure the enhancement quality. The subjective and objective evaluations show that the proposed algorithm outperforms the state-of-the-art approaches.