Hyperspectral Imaging and Machine Learning Techniques for the Automatic Sorting of Horticultural Products

This work discusses the development of a procedure that can be used for the off-line assessment of the acquisition parameters needed for the in-line implementation of a quality control process, for horticultural production lines, based on machine vision. The procedure revolves around the near-infrared hyperspectral imaging inspection of the product to be sorted, with the intent of using machine learning techniques to find the best set of wavelengths to be observed for classification Once the optimal bands set is established, an evaluation of the resulting classification is performed, to decide if the bands are sufficient to classify the samples satisfying the application's requirements. Hyperspectral cameras are not suited for in-line applications due to low acquisition speed and the high quantity of data streamed each second, therefore, spectrum bands are selected to be observed (in-line) with the adoption of near-infrared cameras provided with band-pass filters. The last step of the procedure is the selection of the band-pass filters suited to catch the optimal selected bands. The objects of the analysis, used as a base during the development of the procedure, are tomatoes, which need to be sorted between fresh and rotten samples. Despite this specific case, the aim remains to have a flexible approach to dimensionality reduction, that can be applied to another product with minor adjustments.