On the extraction of spectral and spatial information from images

For the analysis of spatial and spectral correlations in RGB, multispectral, or hyperspectral images, different ways of combining Principal Component Analysis and multiresolution analysis are investigated and integrated into a unified framework. The integration of different frameworks is the purpose of this paper since a unified framework is necessary in the situation where the efficient analysis of both spatial and spectral correlations is needed at the same time. This unified framework also provides a remedy for a limitation of multivariate image analysis (MIA), namely the loss of spatial information in the images. The unified framework, multiresolutional multivariate image analysis (MR-MIA), is illustrated visually through the decomposition of a simple color image, and then used in a quantitative manner for color-textured image classification where the extraction of both spatial and spectral information is necessary. The performance of MR-MIA approaches are shown to be equal to or better than that of wavelet texture analysis, while employing a smaller number of features, and maintaining computational complexity at the same level. Wavelet texture analysis is shown to be a limiting case of one form of MR-MIA. The true advantages of MR-MIA will be most evident when analyzing hyperspectral images having a large number of spectral bands.