Illumination estimation using nonnegative matrix factorization and dominant chromaticity analysis

Human vision can perceive object colors as being the same as colors under a white illuminant. However, images captured by a camera are influenced by the chromaticity of the illumination. Therefore, various illumination estimation algorithms have already been proposed for removing the chromaticity of illuminations in an image to improve the image quality. Most recently, NMFsc (nonnegative matrix factorization with sparseness constraint) was introduced to extract the illumination and reflectance component using nonnegative matrix decomposition and sparseness constraints. However, if an image has a dominant object, the sparse constraint values include the dominant chromaticity of that object, thereby inducing color distortion. Therefore, this paper suggests illuminant estimation using a combination of the conventional NMFsc and a dominant chromaticity analysis. First, the dominant chromaticity region in an image is selected using a chromaticity histogram and the standard deviation. The non-negative matrix decomposition and sparseness constraints are then separately applied to the dominant chromaticity region and non-dominant chromaticity region. Finally, the illumination in an image is estimated by combining the low sparse constraint values that exclude the dominant chromaticity. The performance of the proposed method is evaluated using the angular error for the Ciurea 11,346 image data set, and the experimental results confirm that the proposed method reduces the angular error more than previous methods.