Underwater Image Restoration Using color distance and Image Formation Models

Abstract

One of the biggest issues with cameras in underwater systems is image degradation caused by the environment. It leads to severe changes of colour fidelity and to structural contrast, feature visibility, and optical clarity. If marine monitoring and autonomous underwater vehicles are to be successful, image restoration should be the first thing on the agenda as it can indicate possible huge losses of information and damages of high-level computer vision tracking activities. Old-fashioned physical restoration techniques are usually computationally complex or rely very much on external depth sensors and large learning-based training datasets. This is why in this work an Underwater Image Restoration System Using Colour Distance and an Image Formation Model is introduced to attempt a hybrid underwater image restoration method which takes degraded raw images as input and retrieves the original scene radiance through wavelength-dependent physical restoration methods. Various stages of digital image processing techniques like pre-channel compensation, Grey World white balancing, unsharp masking, and local CLAHE in the Lab colour space are performed to achieve pixel-level colour correction and multi-scale weight map features. Satellite-type optical underwater images go through preprocessing steps like channel scaling, intensity adjustment, and noise reduction first, and then optimized luminance maps and fused restored maps are produced. System's performance has been checked by using such metrics as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Underwater Colour Image Quality Evaluation (UCIQE). Experimental tests have been conducted on the famous UIEB dataset and very efficient restoration, and clarity enhancement results have been attained compared with traditional methods using contrast maps, colour-balanced outputs, and visual analysis reports that are helping subsea systems to make the right decisions at the right time.