A Novel Approach to Color Constancy using Vector Filtered Edge Weighting

Abstract

Computer vision systems like human computer interaction, color object identification, image retrieval and image classification use the color features of the real world scene for computation. But the objects in real world appear to have variation in colors due to different colored illuminants, which place a serious challenge for computer vision systems to detect and characterize the objects. To solve this problem color constancy methods are incorporated in computer vision systems. Color constancy algorithms are used to estimate the color of illuminant and color correction is used to eliminate the illuminant effect. In this work, various state-of-art algorithms have been studied with their pros and cons. These algorithms are compared using the angular distance between estimate illuminant and natural white illuminant. Effect of the various parameters on these algorithms has been studied to increase accuracy. All the color constancy algorithms are based on assumptions about the real world scenes. To obtain the optimal solution for the problem of color constancy is impossible. Because real world scene change on run time, information present in image also change, so assumption working for one kind of scene may fail in other kind of scene. Therefore we cannot get complete optimal solution for this problem. In this research work an effort has been made to improve the results to decrease the angular error. In this thesis, we proposed a new algorithm using vector filtered edge weighting, which is tested on sample images from a publicly available image data set. The experiments show that the proposed color constancy algorithm obtains better results than the current state-of-the-art color constancy methods. The proposed algorithm combines the weighted gray edge and the gray edge method to give 68.45% improvement in existing results.