Dual Adversarial Contrastive Learning for Underwater Image Enhancement

Underwater images are highly distorted, which makes high-level computer vision tasks difficult. Existing underwater image enhancement algorithms mainly focus on restoring the appearance of images. As a result, enhanced images may not be useful for high-level computer vision tasks. The lack of label images also makes most supervised learning networks impractical. In this paper, a dual adversarial contrastive learning enhancement network is proposed, which is both visually friendly and task-oriented. A circular network is proposed to achieve self-supervised learning between unpaired images. We also introduce a contrastive prior between the enhanced and degraded results in feature space to ensure the good visual appearance of the enhanced results. Furthermore, the high-level detection task is also used to constrain the enhanced results. The experiments were carried out on a popular underwater dataset, the enhanced images of the proposed method showed better visual quality and improve tracking performance as well.