Research On Computational Color Constancy For Natural Images

Color is an important and effective feature, which has been widely used in computer vision. However, it is not quite robust and easily influenced by the color of the light source. Computational color constancy is to remove the effect of the color of the light source, get the original objects'color under canonical light source, in order to make the color feature much more robust to the color of the light sources. This dissertation focuses on computational color constancy for natural images under one or multiple illuminations. Our work is carried out from the following four aspects.For color constancy under uniform illumination, which driven by low-level image information, as not all pixels are helpful to illuminant estimation, a color constancy algorithm using effective regions is proposed. This algorithm makes use of natural image statistics to select the proper color constancy algorithm from the committee algorithms and corresponding effective regions for unseen images. Then estimate the color of the light source on the effective regions using the selected color constancy algorithm. This algorithm is proved to be able to improve accuracy of illuminant estimation.For the fusion or selection of color constancy algorithms under uniform illumination, which is driven by low-level image information, a simple but effective machine learning approach ridge regression is used in two ways:one is to get the contribution of color constancy algorithms'estimations to the final illuminant estimation; the other one is to get the relationship between color constancy algorithm selection and natural image statistics.For color constancy under uniform illumination, which is supervised by high-level visual information, there is little work on this direction. Now only indoor-outdoor information is made use of. To overcome such shortcoming, a much more generalized scene classification method is introduced. Specifically, a novel color constancy algorithm using 3D scene geometry is proposed. In this algorithm,3D scene geometry is used to determine which color constancy method to use for the different geometrical regions found in images. Our algorithm opens the possibility to estimate the remote scene illumination color, by distinguishing nearby light source from distant illuminations.For color constancy under multiple illuminations, first we collect three different datasets, i.e. hyper-spectral dataset, dataset under laboratory setting and real-world dataset; after that a novel color constancy algorithm framework for multiple light sources. Compared with the algorithms available, this algorithm is easy to carry out, not specific for any imaging devices and no human intervention needed. In order to evaluate the performances of algorithms, a new measurement is proposed.