Humanlike-GAN: a two-stage asymmetric CycleGAN for underwater image enhancement
摘要
Exploration of the underwater world is still the direction we are heading. Underwater imaging remains challenging due to blurred visibility, colour deviation, and low contrast. Existing datasets and methods often suffer from limitations in realism and synthetic biases. This paper proposes Humanlike-GAN, a two-stage CycleGAN-based method for underwater image enhancement, to effectively solve the problem of difficult image enhancement or overexposure in different underwater environments, and provide preprocessing operations for subsequent target detection. Firstly, we create an unpaired Underwater Realistic Captured (URC) dataset to address the lack of realistic underwater images. Secondly, a two-stage training strategy is employed: unsupervised learning on the unpaired URC dataset followed by supervised learning on the paired UIEB dataset. Thirdly, to mitigate the discriminator’s over-optimization which leads to the gradient disappearance of generator, we adopt a G1:D2 asymmetric training ratio, stimulating the generator’s capabilities. Furthermore, we improve the loss functions, replacing adversarial loss with Wasserstein Distance (WD) and cycle consistency loss with Learned Perceptual Image Patch Similarity (LPIPS), enhancing perceptual quality. Experimental results demonstrate significant improvements in image quality, colour aberration and perceptual effects compared to the original images. Quantitative metrics show enhancements of 22.45