Multibranch underwater image reconstruction algorithm integrating attention features

Underwater imaging frequently suffers from reduced clarity, loss of texture, and color inaccuracies, primarily caused by the interaction of light with water and the presence of various suspended particles. These challenges pose significant difficulties for underwater applications. Existing enhancement methods struggle to fully address the varying scattering intensities across different spatial locations and lack efficient integration of attention mechanisms with convolutional features at multiple scales. To tackle these issues, we present a new multi-branch reconstruction framework for underwater images, which blends attention-driven features. Our design incorporates a Swin Transformer module to adjust convolutional outputs through attention mechanisms at multiple scales and employs a dark channel prior for position encoding, reducing the impact of scattering variations in different spatial regions. Additionally, we leverage a large visual language model to boost the enhancement process by utilizing its image encoding and semantic insights to guide the network. Experimental evaluations show that our network excels in maintaining accurate color representation and preserving texture details.