Color Image Processing Using Principal Component Analysis

It is known since 1988 that utilizing linear dimension reduction gives appropriate lower dimensional representation of homogenous swatches of color images of the nature. Though, it is common to see serious research projects, even dated 2005, which are based on the fixed color space paradigm. In this thesis, first experimental evaluations are performed to compare the principal component analysis (PCA)–based representation of color images of the nature with respective color space–based representations. According to the promising results of the experiments and the theoretical anticipation of the properness of local PCA–based approach, we turn into applications. The PCA model for the color vectors of homogenous swatches defines one principal direction and two less important ones. This framework is utilized for designing outperforming color image processing algorithms such as colorizing, recoloring, compression, segmentation, watermarking, deliberate distortion, and so on. For every algorithm numerous experiments are performed and the results are discussed. Also, wherever possible, the proposed methods are compared to the literature. All algorithms are incorporated into a unique color image processing toolbox for MATLAB.