Unsupervised Color Constancy Computation Algorithm Research

Color is one of the most basic characteristics for an image, at the same time it is alsoa very important characteristics in computer vision. The stability of thecolor is directlyrelated to the stability of computer vision, however, it is not a very stable characteristicsand it is affected by illumination variation easily. The purpose of color constancycomputation is to eliminate the illumination influence on color features and to get the truecolour of surface.The paper studies color constancy computation with the single light, themain completed content is in the following:Firstly,because some color constancy algorithms ignored space feature information ofpixel and the effective area of image,the paper propose an new color constancy algorithmwhich based on effective area combined with discrete wavelet transform. The paperintroduces detailedly image decomposition process of choice by using discrete wavelettransform and how to choose effective area of an image and it is used on image illuminationestimation. Error results show that this algorithm has a good experimental results.Secondly, Under different weighting scheme, the light estimate results have obviousdifference, so the paper study a iterative weighted color constancy algorithm based onedge. The algorithm makes three different edge weighting schemes under three differenttypes：Specular edge weighted scheme、material edge weighted scheme、shadow edgeweighted scheme.it estimate the illumination value on the three schemes, and find thatspecular edge weighted scheme is more valuable than others.Finally, the above algorithm do not involve the prior knowledge, the paper studies ancolor constancy algorithm based on the maximum likelihood estimation. The prioriknowledge of image is obtained by training process, and we use prior knowledge tocalculate light in the process of maximum likelihood estimation. Error results in thealgorithm compared with the results of other algorithms，it has obtained the goodexperimental results.