Breaking the water dilemma: Transmission-guided bilevel adaptive learning for underwater imagery

Sihan Xie, Peiming Li, Jiaxin Gao, Ziyu Yue, Xin Fan, Risheng Liu

Abstract: Simultaneously enhancing the visual effects and resolution of underwater images poses a challenging task as it involves two types of image enhancement tasks, underwater image enhancement and image super-resolution. In spite of the emergence of various deep learning models, almost all existing methods are tailored to one specific enhancement task, rendering them unsuitable for super-resolution in underwater scenes, which consequently resulting in color distortion, unpleasant artifacts and missing high-frequency details. To address this challenge, we propose a multi-level degradation removal enhancer that utilizes underwater transmission prior to improve the quality of underwater images, dubbed as SimUESR. Specifically, the proposed SimUESR is designed to be guided by multiple sets of transmission-inspired guidance, which are cascaded with multi-stage degradation removal modules via a feature modulation operation. Through this, the underwater prior is used as modulation information to modulate contrast and color deviation, gradually embedded through the transmission-guided modules at the feature level. Then the enhanced features are incorporated into a multi-level degradation removal module to generate lossless image content. To release the burden of manually designing loss, we introduce a novel bilevel adaptive learning strategy that combines finite-difference approximation to automatically search for the desired loss, effectively improving visual perception performance. The experimental results demonstrate the remarkable superiority of the proposed method for underwater enhancement and super-resolution tasks, achieving improvements of 0.57 dB and 2.87 dB in PSNR on the UFO-120 and EUVP datasets, respectively. The code is available at https://github.com/lpm1001/SimUESR.