UVE-LUT: Learnable Attenuation-Aware Lookup Table for Underwater Video Enhancement

Deep learning techniques are increasingly being employed for underwater video enhancement (UVE), with the key objectives not only encompassing color cast correction and image quality improvement but also maintaining temporal consistency. Most existing underwater image enhancement (UIE) methods, when directly applied to video sequences, often introduce inter-frame flickering artifacts, thereby impairing the comprehension of video content. Although 3D convolutional neural network-based approaches can preserve temporal coherence, they are characterized by substantial parameter counts and computational complexity, leading to inefficient model performance. To address these limitations, we propose a learnable Look-Up-Table (LUT) based on attenuation-aware and wavelet prior for UVE, abbreviated UVE-LUT, which effectively maintains inter-frame consistency in color and brightness distribution by leveraging look-up table (LUT) technology. The proposed method incorporates a learnable Attenuation-Aware LUT (AA-LUT) module that performs low-latency, low-complexity adaptive enhancement on the low-frequency components of frame sequences in the wavelet domain. Additionally, we design a high-frequency detail enhancement module to conduct three-dimensional detail enhancement on the high-frequency components. Extensive subjective and objective evaluations on underwater video and image datasets demonstrate that the proposed UVE-LUT achieves a favorable balance between image quality and temporal consistency. Our source code will be available at Github.