Enhancing Unpaired Underwater Images With Cycle Consistent Network

Abstract

Visual perception of underwater imagery is affected greatly due to poor visibility, scattering, absorption and refraction of light despite using high-end expensive cameras. The optical hindrances create a non linear distortion that affects the performance of tasks involving object detection, object tracking, classification and object segmentation.Underwater exploration is carried out by Autonomous Underwater Vehicles (AUV) and Underwater robots. Several tasks that require visually perceptive imagery for performing marine trash identification, monitoring the species and their migration, inspection of the health of coral reefs, tracking the cables of submarines and ship wreckage, seabed mapping. This can also be used for enhancing underwater images of user pictures. Marine life is under an increasing threat from human activities which can be monitored using the proposed idea.The paper successfully implements automatic color enhancement, dehazing, and contrast adjustment using deep learning techniques that perform image-to-image translations. The setup uses Unsupervised learning approach on unpaired dataset and Cycle Consistent Generative Adversarial Network. The paper provides comparison of the Unsupervised learning approach with the Classical Image processing approach that uses color correction adaptatively and the application of dehazing theory.