MUIR: Mamba for Underwater Image Rendering
Liyuan Chen, Weijia Li, Qingxia Yang, Lihan Tong, Erkang Chen, Bin Huang, Ruoxue Li · 2024
Underwater image rendering techniques are designed to generate images with a distinct underwater style from arbitrary clean images. These techniques are widely used in underwater image generation, virtual games, and camera filters. Recently, a novel image processing architecture based on the state space model (SSM), named VMamba, has been proposed, which effectively reduces computational complexity to linear while preserving a global receptive field. This paper explores the potential of VMamba for underwater image rendering, focusing on computational efficiency and model performance. To this end, we incorporate the VSS Block, the core module of VMamba, into the underwater image rendering network and propose MUIR. Specifically, we use real underwater images to provide the light field style and clean images to supply the image content with depth information. Additionally, we design the VMUNet architecture, which utilizes the VSS Block for feature fusion after each layer of down sampling. To the best of our knowledge, this is the first underwater image rendering model based on the Mamba architecture. Experiments demonstrate that our method, with a lightweight architecture of 5.06GMac FLOPs and 7.16M Params, achieves visual effects and quantitative metrics comparable to or surpassing previous state-of-the-art methods.