Transmission Map and Background Light Guidedenhancement of Unpaired Underwater Image

Leveraging paired low- and high-quality images to enhance underwater lowquality image is a common way. Unfortunately, the complexity of underwaterenvironment blocks the acquirement of paired images. Unpaired underwater image enhancement using adversarial networks has received attentionconsequently. However, existing unpaired methods are mainly driven byvisual-quality and lack consideration about the characteristics of adversarialnetworks, which can easily lead to over-enhancement or under-enhancementof generated images. To this end, this paper proposes an unpaired underwater image enhancement algorithm based on the physical maps guidance.Specifically, inspired by underwater physical models, we conclude that thetransmission map and background light can guide not only enhancementbut also degradation for underwater images. Therefore, by estimating boththe transmission map and background light in both processes respectively,we correspondingly design UIIE-Subnet for constructing high-quality imagesand UDE-Subnet for obtaining low-quality images. The UDE-Subnet canfurther benefit the UIIE-Subnet training in adversarial networks. In addition, we derive the physical map consistency assumption, thereby proposingthe transmission map loss and background light loss to constrain the networktraining. Experiments on three public datasets (e.g. EUVP, UIEB, UFO)show that the proposed method produces more natural images by training with unpaired data, which outperforms most paired and unpaired enhancement methods.