Hyperspectral Imaging (HSI) combines the benefits of microscopy and spectroscopy to evaluate the spatial distribution of spectroscopically active compounds with widely varying applications in food quality control, pharmaceutical process, and waste sorting.
HSI datasets are extremely large to analyze and store, which makes pixel and variable selection to retain chemical information of the utmost importance to facilitate data storage/analysis. Such fast analysis is of the utmost importance, as the industrial throughput of waste plastic is enormous. Non-negative Matrix Factorization (NMF) recovers the pure contribution maps/spectral profiles of distributed compounds, extremely useful to base sorting decisions upon.
However, NMF can be considerably sped up further, by reducing the presented data to its essential information content. We propose a two-step procedure to remove uninformative and redundant pixels and wavelengths from the HSI data. Essential pixels/wavelengths can be selected from foreground objects, after removing the object background like a conveyor belt, using convex hull as a computational geometric tool. Such data selection may make waste plastics analyses both faster and more accurate than the traditional method. This enables rapid processing, e.g. plastic sorting for the circular economy, as shown by analyzing a range of simulated and real HSIs.