An evaluation of plastic classification techniques using Hyperspectral Near-Infrared Imaging

This work provides an evaluation of hyperspectral image-based classification techniques. All considered classification techniques can be separated into two groups, namely neural networks and non-neural networks. The latter group can also be seen as the more conventional approach and comprises the quotient-based classifier and the spectral angle mapper (SAM). Neural networks on the other side originate from the field of machine learning and are conceptually motivated by the human brain. Their goal is to automatically learn and recognize important relationships or patterns in data. Most notably, not only the classification of pure spectra was evaluated but also the task of multi-material detection in mixed spectra was considered.