UHD Underwater Image Enhancement via Frequency-Spatial Domain Aware Network

Yiwen Wei, Zhuoran Zheng, Xiuyi Jia; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 299-314

Currently, carrying ultra high definition (UHD) imaging equipment to record rich environmental conditions in deep water has become a hot issue in underwater exploration. However, due to the poor light transmission in deep water spaces and the large number of impurity particles, UHD underwater imaging is often plagued by low contrast and blur. To overcome these challenges, we propose an efficient two-path model (UHD-SFNet) that recovers the color and the texture of an underwater blurred image in the frequency and the spatial domains. Specifically, the method consists of two branches: in the first branch, we use a bilateral enhancement pipeline that extracts the frequency domain information of a degraded image to reconstruct clear textures. In the pipeline, we embed 1D convolutional layers in the MLP-based framework to capture the local characteristics of the token sequence. In the second branch, we develop U-RSGNet to capture the color features of the image after Gaussian blurring to generate a feature map rich in color information. Finally, the extracted texture features are fused with the color features to produce a clear underwater image. In addition, to construct paired high-quality underwater image enhancement dataset, we propose UHD-CycleGAN with the help of domain adaptation to produce more realistic UHD synthetic images. Experimental results show that our algorithm outperforms existing methods significantly in underwater image enhancement on a single GPU with 24G RAM. Codes are available at https://github.com/wyw0112/UHD-SFNet.