An underwater image enhancement model combining physical priors and residual network

Abstract Deep learning‐based methods have achieved notable performance for underwater image enhancement. However, previous studies are mostly focused on pursuing high similarity between the original image and the target, which incurs performance drop when the models are used for real‐world images. A new framework for underwater image enhancement is proposed to improve the generalization performance of enhancement. First, the coordinate attention module is integrated into the backbone network, which serves as a pre‐trained model, to strengthen the feature extraction capability of the network. Second, the backbone is finetuned by physical prior knowledge and real‐world images, in an unsupervised manner, to realize generalization from artificial images to real‐world images. Furthermore, a model protection mechanism is designed to guarantee the successful execution of the training. The experimental results indicate that the proposed method provides a powerful pre‐trained backbone network and the finetuning strategy can further solve the color distortion and improve the image sharpening, especially in the harsh real environment. Compared with relevant methods, the UCIQE and NIQE are, respectively, 0.525 and 4.149, with a 0.009–0.095 increase in UCIQE and a 0.256–1.032 decrease in NIQE compared to other methods.