AP-GAN: attention PatchGAN for low light underwater image enhancement

Abstract

Many pictures taken underwater suffer from low brightness, inadequate contrast or loss of fine details due to turbid waters and light absorption. Hence, this paper introduces AP-GAN (Attention PatchGAN), a deep learning framework which is specifically composed for the enhancement of underwater pictures. In this method, a U-Net generator is used to ensure enhanced structural fidelity and global colour correction. The adversarial feedback is delivered by a PatchGAN discriminator which again performs a local critic role, evaluating overlapping patches of the image to actively enforce the preservation of realistic textures and fine details. The proposed approach is applied on several underwater image databases and compared to classical enhancement strategies. Experimental results indicate that our model based on GAN is better than current methods, especially in colour preservation, contrast enhancement, and detail preservation. This paper focuses on the value of deep learning methods, specifically generative adversarial networks (GANs) with customised architectures such as U-Net and PatchGAN, in improving underwater images and recovering the fine details and structure of objects contained in them.