Online depth estimation and application to underwater image dehazing

Abstract

Underwater images captured in a turbid medium often suffer from significant degradation of visibility. Conventional dehazing approaches focus on dehazing a single image by using multiple channels for color restoration and rarely consider computational efficiency. This paper proposes an online dehazing method with sparse depth priors using an incremental Gaussian Process (iGP). The main contribution of this paper is developing a practically usable dehazing method for underwater robots using incoming sparse depth priors (range measurements) from any calibrated depth sensors. To deal with incoming depth priors efficiently, we adopt iGP for incremental depth map estimation and dehazing. Because the input vector of the iGP model is easily reconfigured, we can use the same update method for both color and gray images. Our method also estimates color-balanced veiling light to compensate for the color attenuation problem. For the evaluation, we first verify the proposed method on a open RGBD dataset and test it on real underwater color and gray images, comparing the results with those of previous methods.