Color Constancy Computation Based On Bayesian Inference

Color constancy computation is a subject that covers technology of computer vision, signal processing, artificial intelligence, cognitive science etc. It is used for object recognition, object tracking, video surveillance, image retrieval and so on. Though exiting algorithms perform well on certain scenes, none of them can be considered as universal.For different scenes and different images, one of important research direction is how to select appropriate algorithm or combination of different algorithms according to the physical characteristics of images. For this, the paper proposes a new algorithm which is based on Bayesian inference using natural image statistics to identify the most important physical characteristics. My work is as follows:(1)Some theory of chromaticity and several classical color constancy algorithms have been introduced. And have done some comparison of them.(2)A new algorithm which is based on Bayesian inference is proposed. The algorithm takes Weibull-distribution of image statistics to identify the physical characteristics. Then compute the probability of each algorithm chosen by using Bayesian inference. By the result, two strategies for illumination estimation are proposed: one is to choose the algorithm that has the biggest probability (single algorithm); the other is to re-compute the new illumination estimation according to the probability of each algorithm chosen (the combination of algorithms).(3)At last, detailed steps and the flow charts of the algorithm have been shown. The experiment has been done on the data set for the research of color constancy computation. Then done some comparisons with other algorithms according to the results and corrected images have been shown.Experiments show that, on a large data set, the approach performs well and achieves selection and combining of color constancy algorithms according to different scenes. The algorithm has low computation and can be considered as universal. The algorithm performs better than some other algorithms.