HAZE REMOVAL BASED ON CHROMINANCE AND SPARSITY PRIORS OF NATURAL IMAGES

Abstract

This thesis addresses the problem of haze removal from a single image. Like all image restoration tasks, haze removal is an underdetermined inverse problem whose solution hinges on valid image priors or models. In this work, two new physically based models are proposed for image dehazing: 1. Gaussian mixture model of chrominance distribution of outdoor scenes; 2. Piecewise linear model of transmittance map. The first model is learnt using a large training set; the second model is derived from observations that object surfaces outdoors are either planar or of small curvatures. These two models can be combined themselves or used in conjunction with other dehazing techniques, such as the dark channel method, to recover the clear image via a sparsity-based image restoration process.