UEnhancer: Spatially Enriched Transformer with Dual-color Information for Underwater Image Enhancement

The underwater environment is complex and continuously changing, and underwater images usually suffer from visual degradation issues like color distortion, low contrast, and blurred details. Existing underwater image enhancement methods often use deep models designed initially for in-air image enhancement, ignoring the challenges present in underwater environments, such as scattering of lights, non-homogeneous information flow in different color channels, etc. To address these intricate problems, we propose UEnhancer, a Transformer-based framework that employs a dual-color fusion hierarchical encoder–decoder architecture to effectively enhance underwater images. To better restore fine spatial details, we introduce a Spatial Attention Enhancer (SAE) module within the Transformer block to explicitly capture local dependencies. This module mitigates the Transformer's limitation in modeling fine local structures, enabling improved restoration of textures and structural details degraded by scattering and wavelength-dependent attenuation. Furthermore, a dual-color interaction block is designed to facilitate effective information exchange between RGB and HSV representations, allowing the network to learn correlated features across color spaces. To combine the latent space of different encoders, a non-linear fusion method is incorporated into the proposed model, resulting in capturing a broader range of features, leading to an overall performance gain. Additionally, we propose a multi-scale frequency-based SSIM loss to preserve local structures, maintain contrast in high-frequency regions, and retain fine details. A comprehensive set of experiments reveals that the proposed UEnhancer outperforms other state-of-the-art (SOTA) methods on several publicly accessible underwater image enhancement datasets (e.g., \(~{}\color{#E90000}{4.74\%}\) , \(~{}\color{#E90000}{3.25\%}\) , \(~{}\color{#E90000}{5.13\%}\) , and \(~{}\color{#E90000}{5.82\%}\) in PSNR on EUVP, UIEB, LSUI, and SUIM-E datasets, respectively), demonstrating its superiority. The code is available at: https://github.com/Alik033/UEnhancer .