NNUIE-GAN: Near Natural Underwater Image Enhancement Based on Generative Adversarial Network

Abstract

Underwater images are often prone to many non-linear distortions due to different underwater light interaction phenomenon. This contributes to colour distortion and low contrast which severely affects visual perception of that scene. Now, in today's world many underwater expeditions rely on visual perception of underwater world, which makes underwater image enhancement techniques very important. In the present work, Generative Adversarial Network based model NNUIE-GAN is introduced for real time underwater image enhancement. In this work, generator is a U Net based architecture which is tuned to process fewer parameters and generate more natural looking images by enhancing the underwater image quality. In order to consider image features at both local level and global level perspective, dual discriminator is realized wherein patch level information is processed. To bring in this effect, perceptual quality of the image in terms of local texture, global content and illumination smoothness is considered for construction of loss function. The effectiveness of the loss function is tested on the ground of qualitative and quantitative analysis. The results validated the capability of NNUIE-GAN to improve visual perception of underwater images by enhancing primary characteristics such as colour correction and image clarity improvement. Thus, appropriate tailoring of the loss function can alleviate the visual perception of image in real time processing.