Integration of spatial information in hyperspectral imaging for real time quality control in an andalusite processing line

This paper presents an ANN hyperspectral classification system specifically developed to perform the quality control of an andalusite processing line. The main problem with these types of tasks is related with the way the ground truth is obtained, leading to labels that correspond to large areas with inhomogeneous contents. Thus, when any type of learning algorithm is used in order to train ANN based classifiers, one has to be sure that the samples presented to the networks really contain spectra that correspond to the labels. Therefore, a previous study on the size of the windows to be used by the ANNs as well as the way the information from the different pixels in these windows are combined must be carried out. The ANNs in the segmentation operator are based on Gaussian functions. The results obtained have shown that success rates, which were very poor when working with the spectral information of individual pixels, can be improved to better than 95%.