Classification of objects on hyperspectral images — further method development

Classification of objects (such as tablets, cereals, fruits, etc.) is one of the very important applications of hyperspectral imaging and image analysis. Quite often, a hyperspectral image is represented and analyzed just as a bunch of spectra without taking into account spatial information about the pixels. This makes classification of the objects inefficient. Recently, several methods, which combine spectral and spatial information, has been also proposed and this approach becomes more and more wide-spread. The methods use local rank, topology, spectral features calculated for separate objects and other spatial characteristics.

In this work we would like to show several improvements to the classification method, which utilizes spectral features calculated for individual objects [1]. The features are based (in general) on descriptors of spatial patterns of individual object’s pixels in a common principal component space. In the original method score histograms were used as the descriptors. The method modifications include both the way the principal component space is built as well as the use of new descriptors for the patterns.

The comparison of the modified method with its previous version and competitors will be shown based on several real cases.

References:
[1] S. Kucheryavskiy. A new approach for discrimination of objects on hyperspectral images. Chemometrics and Intelligent Laboratory Systems 120, 126 (2013).