MambaUNet: An Efficient U-Shaped State-Space Network for Underwater Image Enhancement

Underwater images are frequently degraded by wavelength-dependent absorption and scattering, resulting in color casts, low contrast, blurred textures, and loss of structural details. Existing enhancement networks may struggle to balance global context modeling, local detail recovery, and computational efficiency. To address this problem, we propose MambaUNet, an efficient U-shaped state-space network for underwater image enhancement. Its core VMEC pipeline integrates visual state-space scanning to capture long-range spatial dependencies, multi-scale alignment and adaptive aggregation to improve skip-feature coherence, efficient channel attention to recalibrate feature responses, and cross-channel state-space modeling to represent channel-dependent degradation. These components are assigned stage-specific roles within the U-shaped network, forming a spatial–scale–response–channel restoration pipeline. By coordinating these components within an encoder–decoder architecture, MambaUNet improves global tone consistency and structural recovery without relying on computationally expensive self-attention. Experiments on the full-reference LSUI and UIEB benchmarks and the no-reference C60 and S16 test sets show that the proposed network achieves competitive or superior restoration quality compared with representative conventional, CNN- or GAN-based, Transformer-based, and recent Mamba-based methods. Ablation and complexity analyses further demonstrate the complementary roles of the VMEC components and the favorable balance between enhancement quality, model size, and inference speed. MambaUNet can therefore serve as a lightweight preprocessing component for underwater imaging and vision-based applications.