Zero-shot illumination adaption for improved real-time underwater visual perception

High quality of underwater images can provide great visual effect for improving numerous undersea investigations, such as underwater archaeological exploration and underwater topographic surveying, etc. Underwater images often suffer from distortions at various depths due to harsh conditions, leading to inconsistencies that compromise their reliability. Existing approaches to underwater image enhancement (UIE) often struggle with limited adaptability and may even fail when confronted with previously unseen real-world underwater conditions. Meanwhile, the inefficiency of model inference significantly constrains its feasibility for high-demand real-time underwater visual perception. To address these issues, we need to reduce the reliance on extremely limited labeled data and develop a lightweight scheme to adapt to diverse scenes. With this idea in mind, we present a joint CNN-Transformer framework termed AIDUIE that contains a zero-shot illumination disentanglement network (ZSIdNet) and a self-attention based dynamic camera response function (SADCRF) with long-range dependency modeling capability. We develop a multi-step iterative training mechanism for illumination disentanglement, which employs zero-shot learning at each step using pseudo labels generated by an image degradation module. Accordingly, we can adjust the disentangled illumination using SADCRF with parameter maps of adaptive pixel retention factor (PRF) to reconstruct the enhanced image. We substantiate the diverse advantages of our scheme over existing methods through a meticulous and comprehensive experimental evaluation process, illustrating its superiority in both quality and efficiency under unexplored and convoluted underwater circumstances.