A Novel Multi-Scale Fused Generative Adversarial Network for Underwater Image Enhancement

Robotic vision plays an important role in autonomous underwater vehicles (AUV) and hybrid aerial underwater vehicle (HAUV) perception and task execution in underwater environment. However, due to the special underwater imaging environment, images often have color distortion, texture blur and other problems. Underwater image enhancement (UIE) can effectively solve the problem of underwater image distortion and improve the quality of imaging and vision. In this paper, a novel Generative Adversarial Network (GAN)-based approach for underwater image restoration is proposed which can effectively addresses challenges such as color distortion and blurriness. The generator employs an enhanced U-Net model, incorporating multi-scale fusion to capture features in different layers. Additionally, the discriminator fuses branch information at multiple resolutions in parallel to consider both local and global information, where the single discriminator is a PatchGAN with spectrum normalization in each layer. Experimental results demonstrate that the proposed model can significantly improves both visual quality and quantitative performance, such as PSNR, SSIM and UIQM, improved by 0.55%, 0.93%, and 1.84%, compared to other UIE methods.