Recovering Intrinsic Image Using Illumination Decomposition

In human visual system,eyes can perceive the color and distinguish the reflectance variance form illumination changes.From instance,human eyes can recognize the black and white color easily.In research of intrinsic images,we denote the object material as reflectance and illumination information as shading.The task of intrinsic images is recovering the reflectance and shading images form the input image.Intrinsic image is a based problem in computer vision,and has various application,such as shape from shading,material edition,and texture edition.Because of its ill-posed property,it is a challenging problem,and until now it is an open problem in computer vision.All current algorithms use an assumption that shading is smooth all over the image,but it is broken down by shading and finally lead to inexactly decomposition.In this paper,we present a novel approach that focuses on recovering the intrinsic image using illumination decomposition.Our method is based on the significant observation that the shading component of an intrinsic image contains step and drift(or smoothness)properties simultaneously.Hence,we segment the shading into two components.One component,step-lighting,indicates the shading step property and the other component,drift-lighting,represents the smoothness property.Under this shading assumption,which accounts for the formation of shading,we can handle images that contain shadows by decomposing the shading field into two components.Mathematically,we describe the intrinsic image problem using a two-parameter energy function and give the minimizing solution.By constraints on shading and reflectance,we can obtain the intrinsic energy function and it can be solved by gradient descend.Experiments were done on the MIT benchmark dataset,as well as real world images to evaluate our approach.The results verify that our approach outperforms other single image methods that only use a global shading smoothness prior constraint.Our illumination decomposition method can be used in other single image algorithm as a framework to improve decomposition accuracy.