UMPAN: Multipriors Aggregation Network for Single Underwater Image Enhancement

Due to the lack of dewatered images as ground truth, existing deep underwater image enhancement methods usually resort to synthetic underwater images to alleviate the problem of insufficient training data. However, existing underwater image synthesis methods often struggle to generate high-quality underwater data, leading to learned models not always being effectively generalized to real-world scenes. Besides, we observe that much of the underwater domain-specific knowledge is almost completely discarded in existing deep models, which further limit their applications due to the fully training-data-dependent nature. Motivated by these two aspects, this paper focuses on improving the quality of underwater synthetic data and designing more effective deep architectures to advance this field. Firstly, a new hybrid underwater image synthesis model (HUIS), which fuses physical priors and data-driven cues, is proposed to create more realistic underwater training data. Specifically, a multi-stage attenuation underwater physical imaging model is proposed, which establishes a more comprehensive underwater light absorption and light scattering processes to better simulate underwater light attenuation phenomena. Meanwhile, an underwater image detail transfer model works as a supplement of the physical model to simulate other factors in a data-driven manner, thereby generating finer details of underwater images. Secondly, a novel underwater multiple priors aggregation network (UMPAN) is proposed, which introduces multiple underwater domain knowledge as adaptive auxiliary information to enhance the generalization of the deep model. To alleviate the effects of underwater degradation, a multiple priors-aware enhancement module is proposed at the feature extraction side, which performs physical inversion-constrained feature-level enhancement to learn more enhancementfriendly content features. At the image reconstruction side, underwater multiple priors are still utilized, and a multiple priors-guided modulation module is proposed, which can adaptively select and integrate intrinsic information from multiple underwater domain knowledge based on the importance of each prior information to generate richer reconstructed features for better enhancement. Extensive experiments prove the effectiveness and superiority of our HUIS and UMPAN.