Red Channel-Enhanced GAN for Underwater Image Restoration

Underwater images typically exhibit a bluish-green tone due to the absorption and scattering effects of water, especially in deep-water environments, where red wavelengths are easily absorbed, leading to the loss of red channel information. This, in turn, affects the image's color and detail. To address this issue, this paper proposes a GAN-based model enhanced by the red channel. The model utilizes a U-Net autoencoder to extract features from degraded underwater images. Additionally, considering the unique characteristics of the underwater environment, a red channel feature enhancement network is designed. This network takes the red channel of the degraded image as input and, through an attention module, focuses on severely degraded regions where the red channel prior cannot be satisfied, thereby assisting the autoencoder in image reconstruction. Experimental results demonstrate that, when tested on the UIEB dataset, the proposed algorithm outperforms the compared models in terms of PSNR and SSIM, achieving 25.26 dB and 0.90, respectively. The model effectively removes the bluish-green background from degraded images and restores their color. Further, the model's practical feasibility is verified through generalization testing, feature point matching, and edge detection application analysis.