Abstract
Underwater imaging is significant but the images are always subject to degradation, which varies in different underwater environments. Factors such as light scattering, absorption, and environmental noise can affect the quality of underwater images, leading to issues such as color shift, low contrast, and low definition. Polarization information benefits image recovery, and existing learning-based polarimetric imaging methods ignore multiple water types (viewed as domains) and domain generalization. In this paper, we collect, to the best of our knowledge, the richest polarization color image dataset with different water types and present a specially designed neural network UPD-Net firstly employing the domain-adversarial learning strategy to recover the degraded color images. The designed water-type classifier and domain-adversarial learning strategy enable the multi-encoder to output domain-independent features, the decoder outputs clear images consistent with the ground truths with the help of the discriminator and generative-adversarial learning strategy, and there is another decoder responsible for outputting DoLP image. Comparison experiments demonstrate that our method is state-of-the-art in terms of visual effect and value metrics and that it has a strong recovery ability in the source and unseen domains, including in water with high turbidity. The proposed approach has significant potential for underwater imaging and recognition applications in varied underwater environments.