Towards Quality Advancement of Underwater Machine Vision with Generative Adversarial Networks

Underwater machine vision attracts more attention, but the terrible quality prohibits it from a wide range of applications. There exist many algorithms to solve this problem, but the real-time and adaptive method is frequently deficient. In this paper, based on filtering and Generative Adversarial Networks (GANs), two approaches are proposed for the aforementioned issue, i.e., a filtering-based restoration scheme (FRS) and a GANbased restoration scheme (GAN-RS). Distinct from the previous methods, FRS restores underwater image in the frequency domain, which is composed of parameter search, filtering, and enhancement. Aiming to further improve the image quality, the GAN-RS is able to adaptively restore underwater machine vision in real time without any pretreatment. In particular, the information in Lab color space and the dark channel is developed as the loss function, namely underwater index loss and dark channel prior loss. More specifically, learning from the underwater index, the discriminator is equipped with a carefully crafted underwater branch to predict the underwater probability of an image. Multi-stage loss strategy, then, is developed to guarantee the effective training of GANs. Through extensive comparisons on image quality and applications, the superiority of the proposed approaches is confirmed. As a result, the GANRS is much faster and achieves the state-of-the-art performance on color correction, contrast stretch, dehazing, and feature restoration of various underwater scenes. Note that the source code will be made available.