Dual-branch underwater image enhancement approach combining CNN and transformer architectures

In underwater environments, imaging devices face numerous challenges, including turbid water, light attenuation, and scattering. These factors collectively degrade image quality, reduce contrast, and cause color distortion, posing significant challenges to underwater vision tasks. To address these issues, this study proposes a dual-branch underwater image enhancement approach that combines CNN and transformer architectures. First, a color correction module (CCM) is designed to address color bias. Additionally, a multi-level cascaded subnetwork (MCSNet) is designed to effectively perform context modeling, enabling the accurate fusion of color and contextual information. By progressively extracting and integrating color and context information from the image at each level, MCSNet enhances the ability to understand complex scenes. Finally, a frequency-domain and spatial-domain fusion transformer module (FSTM) is proposed to process information in both domains, effectively supplementing detailed information. Experimental results on the UIEB, LSUI, and EUVP datasets show that the PSNR, SSIM, and MSE reach 24.444/0.917/425, 29.354/0.929/155, and 30.786/0.929/79, respectively. Compared to several state-of-the-art networks, certain improvements have been achieved.