A New Data Preparation Methodology in Machine Learning-based Haze Removal Algorithms

Haze removal is an intellectually challenging object of scientific study. There is a myriad of methods has been proposed hitherto, ranging from histogram-based, contrast-based to machine learning-based. Haze removal approaches founded upon machine learning require a large and reliable training database. Researchers are currently using the synthetic database due to the complexity of real database acquisition. To introduce the synthetic haze into the clear images, they assume that the depth map is drawn from the standard uniform distribution. In this paper, we present a new methodology for preparing the synthetic training database, in which the proposed equidistribution is employed instead of standard uniform distribution. The effectiveness of the proposed method has been verified by a large number of experiments.