Analysis of hyperspectral images of apple fruits for automated sorting

Abstract

One non-invasive method for detecting fruit damage involves monitoring them using hyper-spectrometers. Hyperspectral images comprise a set of spatially resolved radiation spectra of a reflected object. By using these data, it is possible to identify the characteristics and parameters of fruits that may indicate their damage. In this study, the spectral and spatial components of hyperspectral images of apple fruits were analysed. Random forest classifiers were used to detect objects in the images, with reflection spectra, vegetation indices and spatial texture descriptors (local binary patterns) used as input data for classifiers. Classifiers based on spectral characteristics proved to be more reliable than those trained without spectral data. Using spectral information about fruits signifi cantly improved the classifi cation results under the conditions of uneven lighting interference. By combining spectral data with machine learning methods, fruit sorting efficiency can be improved. This approach offers advanced development of methods for processing data from hyperspectral sensors installed on sorting lines in order to detect damage to apple fruits reliably