The undecimated wavelet transform–multivariate image analysis (UWT-MIA) for simultaneous extraction of spectral and spatial information

Abstract

Combining color (spectral) and textural information often leads to improved performance of image-based sensors developed for different applications, such as process monitoring and prediction of process or product quality variables. A new method based on the undecimated wavelet transform and multivariate image analysis (UWT-MIA) is proposed in this paper for simultaneous extraction of spatial and spectral information. The key advantages of this approach are: 1) it uses continuous wavelets which have a constant resolution over all scales compared to discrete wavelets, and are better suited for extracting size distribution features, 2) the wavelet detail and approximation sub-images have the same size and are spatially congruent which allows stacking the sub-images obtained at all scales and orientations (and wavelengths for multi- and hyperspectral images) as a separate channel in a new multivariate image from which a single MIA model can be built. The performance of UWT-MIA is illustrated using both synthetic and natural images. It is shown that variations in color, size and orientation of objects of interest can be tracked efficiently within the same MIA model. The method can also be cast within the multiresolutional multivariate image analysis (MR-MIA) framework proposed in the literature.