Ghost-UNet: Lightweight model for underwater image enhancement

Light scattering and absorption during underwater propagation lead to colour distortion and fogging in underwater images, lacking reference for undegraded images, posing challenges in directly applying the underwater images to real-world scenes. We provide a novel approach for generating underwater images by utilizing a conditional diffusion model to address this issue. Besides, we present a novel approach called Ghost-UNet to address the issue of substantial memory and computational demands in implementations of deep convolutional neural networks. Compared with previous methods, our method exceeds 3.8% and 1.7% in the peak signal-to-noise ratio and underwater colour image quality evaluation respectively. Additionally, our method leads to a 0.016-second gain in test speed. The parameters of Ghost-UNet is 1/39 of a transformer model. Ghost-UNet is highly suitable for real-time underwater image enhancement network implementation.