Dual-Branch Underwater Image Enhancement Utilizing Bidirectional Supervision Loss, Contrast Loss, and Style Transfer Loss

Due to the complexity of underwater environments, underwater images often exhibit multiple forms of degradation. Despite recent notable advancements in underwater image restoration, the scarcity of paired datasets poses significant challenges for future progress in this field. This paper introduces a dual-branch underwater image enhancement network that includes an image restoration branch and an illumination estimation branch, trained using both paired and unpaired datasets to mitigate the scarcity of paired data. The network is jointly trained through supervised and unsupervised methods: in the supervised phase, it uses paired datasets guided by contrast loss, perceptual loss, and bidirectional supervision loss to learn the mapping from underwater to clear images, in the unsupervised phase, it employs unpaired datasets with bidirectional supervision loss and style transfer loss to learn the style of natural images and enhance network generalizability. Experimental results demonstrate that our proposed method performs comparably or better than other methods across multiple datasets, providing valuable insights and techniques for underwater image restoration.