Hybrid CNN-Transformer Network for Two-Stage Underwater Image Enhancement with Contrastive Learning

Conventional and deep learning methods have demonstrated notable success in the field of underwater image enhancement. However, the majority of previous methods have only employed a single type of enhancement, which is not sufficiently adaptable to complex underwater environments. Therefore, we design a two-stage underwater image enhancement method that combines traditional methods and deep learning methods, thereby combining the advantages of both to make our method more adaptable to complex underwater environments. In the first stage, we infer the optimal outputs from five traditional enhancement methods using NIQE, which are then employed as inputs for the second stage. After that, the difficulty of following enhancement is reduced. In the second stage, we design a deep underwater image enhancement network with Transformer embedded in CNN and combine it with contrastive learning to improve the enhancement performance at the feature level while obtaining better feature representation. Experiments demonstrate that the enhancement performance of our proposed method on the underwater image enhancement dataset is more in alignment with human vision compared to other methods.