Light Attenuation Prior based Underwater Image Enhancement

Abstract

Images captured underwater usually suffer from color distortion, detail blurring, low contrast, and a bluish or greenish tone due to light scattering and absorption in the underwater medium, which in turn affects the visibility adversely. Underwater image processing schemes are broadly categorized into two groups, restoration methods and enhancement methods. The approach of restoration methods is to assume the effects of underwater environment as degradation but in enhancement, this environment is assumed to be natural and tries to enhance the visual information to next extent. Restoration schemes are proved to give better performance than enhancement schemes. These restoration schemes are further classified as optical imaging methods, polarization methods and prior knowledge methods. The key problems faced by these schemes are excessive optimization parameters, difficulty of recognizing artificial lighting, adapting to multi-scatter scenario, red artifacts and over exposure. A large number of schemes are proposed in the literature under these categories and most of them suffer from one or more of the above-mentioned issues. In this paper, a rapid and effective scene depth estimation model will be proposed based on underwater light attenuation prior for underwater images and train the model coefficients with learning-based supervised linear regression. With the correct depth map, the background light and transmission maps for R-G-B light are easily estimated to recover the true scene radiance under the water.