Research On Computational Color Constancy Under A Single Illumination

Color information is one of the most discriminative features of objects in the field of digital image processing and computer vision. The color of an object observed by image device depends on the illumination condition as the machine doesn't have the ability of color constancy, which is a well-known property of human vision system. To eliminate the influence of illumination and attain reliable color description, we must correct the image color as if the image of scene is taken under canonical illumination.This dissertation focuses on the research on computational color constancy under a single illumination.The author's main work is introduced as follows.Firstly, the basic concepts and principle of computational color constancy theory are described and many classical color constancy algorithms are introduced in details.Secondly, a color constancy algorithm based on Spatio-spectral model and image edge is proposed. This algorithm is an improvement of the algorithm based on Spatio-spectral model. The proposed algorithm only uses the edges of image and pixels in its neighborhood when calculating the parameters of Spatio-spectral model. The proposed method is analysed in benchmark dataset, and the experiment result shows the proposed method improves the accuracy of illumination estimation and significantly reduces the execution time of the algorithm, the storage space and training time at training stage.At last, a color constancy algorithm based on r chromaticity histogram and constrained illumination is proposed. We observe the r chromaticity histograms of benchmark dataset have high correlation with the ground-truth illumination's r chromaticity through experiments, and calculate the multiple linear regressions between image's r chromaticity and estimated illumination using this correlation. The proposed method is analysed in benchmark dataset, and the result shows the performance of new method is close to the original algorithm and significantly reduce the training time.