A study on design of object sorting algorithms in the industrial application using hyperspectral imaging

Many industrial object-sorting applications leverage benefits of hyperspectral imaging technology. Design of object sorting algorithms is a challenging pattern recognition problem due to its multi-level nature. Objects represented by sets of pixels/spectra in hyperspectral images are to be allocated into pre-specified sorting categories. Sorting categories are often defined in terms of lower-level concepts such as material or defect types. This paper illustrates the design of two-stage sorting algorithms, learning to discriminate individual pixels/spectra and fusing the per-pixel decisions into a single per-object outcome. The paper provides a case-study on algorithm design in a real-world industrial sorting problem. Four groups of algorithms are studied varying the level of prior knowledge about the sorting problem. Apart of the sorting accuracy, the algorithm execution speed is estimated assuming an ideal implementation. Relating these two performance criteria allows us to discuss the accuracy/speed trade-off of different algorithms.