Unsupervised Underwater Image Enhancement Combining Imaging Restoration and Prompt Learning

Underwater image enhancement (UIE) based on deep learning often requires a large amount of labeled data. This study proposes a simple but effective unsupervised UIE method from the perspectives of content and style improvement by incorporating the traditional imaging restoration model and prompt learning techniques. We initially train the fast language-image pre-training (FLIP) model using unpaired underwater images to learn image-perceptive prompts for distinguishing between high and low image qualities. Meanwhile, we perform underwater image restoration via the underwater dark channel prior's method to obtain pre-restored images. Next, a simple U-net network is employed to further enhance the pre-restored image, while utilizing the FLIP prompts as the loss to guide style and content and using a consistency loss to maintain texture structure. The experiment compared the proposed UIE method with nine other methods on three datasets, and evaluated them with the metrics of UCIQE, UIQM, and NUIQ. Experimental results show that the method performs better than traditional and SOTA unsupervised methods, improving image quality and robustness, and helping to solve the problem of scarce underwater data.