Underwater Image Enhancement Algorithm Based on Transformer

Underwater images suffer from color distortion, non-uniform illumination, and severe haze effects. Therefore, this paper proposes a cross-scale feature fusion Transformer by introducing the Channel-wise Multi-Dconv Head Attention module based on Swin Transformer. First, we design a Channel-Attention-based Swin Transformer module that enhances inter-channel dependencies by integrating channel attention mechanisms into Swin Transformer, while replacing the standard Multilayer Perceptron (MLP) with a Contextual Local Enhancement (CLE) module to improve detail and texture recovery. Second, nested dense skip connections are introduced to compensate for Swin Transformer's limitations in fine feature extraction. Finally, a cross-scale feature weighted fusion function optimizes model training. Extensive experiments demonstrate that the proposed algorithm effectively achieves color correction, detail restoration, and texture enhancement for underwater images.