A multi-stage underwater image aesthetic enhancement algorithm based on a generative adversarial network

Existing underwater image enhancement algorithms rely on paired datasets, which enhance underwater images by learning the mapping relationship between low-quality and high-quality data. However, currently, high-quality data (which are called real data) are artificially selected by the dataset builders from the results of previous algorithms, and there are no real paired data in the true sense. In this paper, we used CycleGAN for underwater image enhancement, which is unsupervised learning. We designed the aesthetic loss and style consistency loss to constrain the generated image to make it more consistent with perception by human eyes and to improve the contrast. We used a two-stage generative network structure to compensate for the loss of information during the enhancement process and enhanced the details. We verified the superiority of our algorithm in the subjective and aesthetic aspects through a large number of comparative and ablation experiments as well as subjective and objective analyses.