Research On Color Constancy Of A Single Illumination

Color constancy originates from research on human vision system. Human can recognizethe researched objects, which keeps still under the changing illuminations. The machinevision is employed commonly in modern life. The images captured by machine vision systemvary with the changing of environment illumination, which is unbeneficial machine vision. Inorder to make the images really reflect the characteristics of the interested objects, it’snecessary to make a research on the images to realize color constancy.This paper mainly studies and analyzes images captured under a single illumination. Tomake images’ color constant, the paper’s job is shown as follows:The captured images vary with the change of environment illumination. The colorcalibration is performed on the distorted images captured under nonstandard illuminations.Based on the theory of mapping, the standard color checker is used as the reference to realizethe transformation of color images from the RGB color space to sRGB color space, withinwhich images’ color is independent of devices. LS-SVR algorithm is adopted to calibrate theimages captured under the illuminations of D65and LED. The performance of the adoptedalgorithm is compared to that of polynomial regression. The results show the LS-SVRperforms better than polynomial regression in color constancy and has a less color error.Fusion algorithm is adopted to perform color constancy on collected images. As a singlealgorithm lacks universality in color constancy, five well-developed algorithms are chosen asthe basic candidate of fusion algorithm. LS-SVR is employed as the fusion algorithm toestimate the illumination. Compared to the candidate algorithms, the experimental resultsshow illumination chromaticity estimated by LS-SVR achieves a less error. LS-SVR performsbetter in color constancy and furthermore has a good universality than the candidatealgorithms.Based on the theory of texture similarity, the best algorithm is chosen from the candidatealgorithms of color constancy. This paper structured a co-occurrence matrix based on thecolor information to extract the texture of color images. The Euclidean distance of texture isused to choose the best similar image to obtain the best algorithm of color constancy. Theexperimental results show the chromaticity of74.8%estimated illumination is within10%relative error despite of a matching rate of48.14%. Compared to the candidate algorithms,this illumination estimated by this algorithm achieves the best precision in chromaticity errorand angular error, which proves the validity of this proposed algorithm.