Restoring Underwater Images using Axial Depthwise Convolutions and Channel Attention modules

The absorption and dispersion of light impede underwater imaging, resulting in a severe blue-green color cast, reduced contrast, and loss of detail. These degradations pose challenges to marine research and other domains dependent on underwater imagery. Existing enhancement techniques have made a lot of progress but struggle with limitations such as inaccurate color correction (especially in the red channel), over-enhancement, and loss of structural consistency. To address these limitations, this paper presents ADCCA-GAN, an innovative GAN-based architecture for underwater picture restoration. The model integrates a lightweight generator that incorporates axial depth-wise convolutions and channel attention modules. A Patch-GAN discriminator is used to enhance global color fidelity and enforce local texture realism. In addition, a composite loss (RGB and VGG perceptual loss) is used to ensure robust color fidelity and structural coherence. Experiments on the UIEB dataset show that ADCCA-GAN makes colors look more natural, textures look sharper, and the overall quality of the images look better than those generated by other models. The proposed framework effectively addresses limitations unhandled by existing methods by achieving a strong balance between color restoration, contrast recovery, and detail preservation.