The sparse least square support vector regression for estimating illumination chromaticity

The fusion of estimating algorithms for illumination chromaticity is an important strategy in computational color constancy and it has attracted the much attention of domestic and foreign scholars. Some effective approaches have been proposed to build the framework of the fusion, such as Committee-based Color Constancy, BP neural network, and machine learning regression. In this article, the framework of least square support vector regression (LS-SVR) is applied to illumination estimation fusion and provides a novel sparse method of LS-SVR. According to the experience of previous research, the complexity in training LS-SVR is obviously lower than it in training SVR, but due to all of the characters in images are chosen as the support vectors, the computation in LS-SVR takes more time. In order to improve accuracy and reduce time consumption, this study uses sparse strategy that only weighted samples that illumination chromaticity of image have higher weight than other images can be selected as the support vectors. Training via a part of the images with high-weight, the sparse LS-SVR can achieve a satisfactory result. Experiment with real images shows that this sparse LS-SVR method performs better than SVR, LS-SVR and current some other color constancy algorithms especially in the aspects of computing speed and accuracy.