LGT: Luminance-guided transformer-based multi-feature fusion network for underwater image enhancement

When light propagates through water, absorption and scattering effects lead to uneven brightness distribution and significant color deviation in underwater images. Existing methods struggle to compensate for light attenuation at each wavelength while simultaneously balancing luminance variations across different regions. To address these challenges, we propose the Luminance-Guided Transformer (LGT)-based multi-feature fusion network for underwater image enhancement. Specifically, we design an Underwater Scene Enhancement Curve (USEC) based on the retinal response to light, which adaptively balances luminance variations indifferent regions and compensates for the attenuation of different wavelengths in water. We introduce Luminance- Guided Self-Attention (LGSA) as a core component of LGT, which uses USEC to achieve dynamic weighted fusion of different features, effectively capturing spatial and channel correlations among different features. We construct an underwater image enhancement network with LGT as the core module of the backbone network. Two sub-networks-Edge Feature Enhancement Network (EFEN) and Luminance Feature Extraction Network (LFEN)-are designed to extract edge and luminance features, respectively. These features are then fused through multi-scale, layer-by-layer fusion to progressively enhance image details and ensure global consistency. Experimental results on multiple underwater image datasets with various scenes demonstrate that the proposed LGT has significant advantages in enhancing image quality.