Efficient Underwater Image Super-Resolution via Learnable Color Conversion and Dual-Branch Mamba Network

Underwater image super-resolution plays an important role in marine exploration, as it aims to recover clearer and higher-resolution images from degraded low-resolution observations. However, in underwater environments, this task is particularly challenging due to non-uniform spectral attenuation and particle scattering. Underwater, red light fades quickly, leading to color distortion and loss of details. To address this while keeping the model lightweight, we propose a dynamic color-space network for single-image super-resolution. Traditional methods use fixed color conversion and cannot handle non-uniform attenuation well. We instead use a learnable module to adaptively obtain luminance (Y) and chrominance (CbCr) representations. Building upon this representation, we employ a highly efficient dual-branch architecture where the Y-channel features are processed using light Mamba blocks to capture spatial dependencies with linear complexity, while the CbCr-channel features are restored by a lightweight convolutional neural network. Finally, a lightweight residual model utilizing depth-wise separable convolutions (DSC), deformable convolutions, and pixel attention mechanisms is introduced to refine the features and suppress artifacts. Experimental results demonstrate that while achieving competitive restoration quality, the proposed method drastically reduces computational complexity and parameter count compared to other large-scale models. This balance between visual quality and computational efficiency makes the proposed method well-suited for real-time deployment on resource-constrained autonomous underwater vehicles (AUVs).