Mask-guided region-aware underwater image enhancement Mamba network

Underwater images are often affected by region-aware inconsistent degradation. Specifically, foreground objects, background regions, and boundaries of the image exhibit inconsistent and conflicting degradation states, which affects the processing performance of underwater image enhancement (UIE) networks. Existing UIE networks enhance the underwater image as a whole, ignoring the degradation differences between regions. This can lead to the destruction of foreground details, insufficient background enhancement, and the introduction of artifacts or distortions at boundaries. To address this issue, we proposed MRU-Mamba, a mask-guided region-aware underwater image enhancement network based on Mamba. The MRU-Mamba network uses a semantic mask to segment the foreground, background, and boundary regions of an image, and applies Region-Oriented Selective Scan (ROSS) strategies to different regions. ROSS utilizes Region-aware Spatial Selective Scanning (RSSS) and Region-aware Channel-oriented Selective Scanning (RCSS) to integrate long-range dependencies in spatial and channel dimensions, respectively, to achieve accurate texture restoration and color modeling. Extensive experiments on multiple benchmarks demonstrate that MRU-Mamba outperforms recent UIE methods in both qualitative and quantitative experimental results.