The Compute Research Of Color Constancy Under Single Illumination

Recently,with the appearance and development of those technologies including big data?cloud computing?deep learning,the research for computer vision has been widely applied to build smart city and safety city. For instance,face recognition?license plate recognition and people-stream surveillance and statistics are the typical applications of them.Color have the characteristics of easily and directly perceived through the sense,which play an important role in above fields.However,color is also unreliable.because it can be affected easily by the changes of scene illumination.As a consequence, the accuracy of recognition may reduced.the ability of Humans' vision system to recognize the correct colors,independently of the color source present in a scene is known as color constancy. so how to make the computer vision system has the same ability is the main purpose of the research for color constancy.Currently, most of the compute research of color constancy focuses on the single illumination. its calculation can be decomposed two steps—the estimation of scene illuminant and correction of image color. In general,we can study it from aspects of both supervised and unsupervised methods.The former is to find the relationship of image color and scene illumination chrominance to estimate the Illumination chrominance.In contrast, the latter is directly to estimate scene illumination based on some assumptions. The paper to investigate color constancy based on selecting one of most excellent single algorithm from the set of five different algorithms is limited at single light source and the content are flowing:Firstly,we have deeply analyzed several classical supervised color constancy algorithms and compared it comprehensively.Those algorithms include Gamut Mapping?Bayesian based method?Color by Correlation and Neural Network.We have compared each of others from both algorithm procedures and performances and conclude a general features of supervised methods.Secondly,we propose a new algorithm based on nature image color?texture and contrast to select one of most excellent single algorithm for different images.The algorithm will firstly compute image color moment and color histogram and apply SVM classifier combining Weibull distribution which can represent image texture and contrast,together with the color features computing before to select a best single algorithm among five candidate algorithms for different images. Experiments show that this new method is better than single color constancy algorithm and is comparable with those similar algorithms.Further more,it improves the performance of illuminant estimation to some extent.Finally,some potential effected factors including the number and type of candidate algorithms on SVM classifier have been investigated. The experiment results are that illuminant estimation results are influenced slightly.In a word, the proposed algorithm effectively improved the shortcomings of narrow applications of single color constancy algorithm and proved that the algorithm is effective and reliable.