Nondestructive Control of Fruit Quality via Millimeter Waves and Classification Techniques: Investigations in the Automated Health Monitoring of Fruits
Flora Zidane, Jérôme Lanteri, Julien Marot, L. Brochier, Nadine Joachimowicz, Hélène Roussel, Claire Migliaccio
Abstract
Fast and efficient nondestructive evaluation (NDE) methods for food control is still an ongoing field of research. We have recently proposed to combine W-band imaging with nonlinear support vector machine (SVM) classifiers to sort out healthy from damaged fruits for a single variety of fruit. We have tested it on apples and peaches separately with a mean accuracy of 96%. We have also shown the limitation of a biclass SVM since it has failed to sort healthy from damaged fruits when the set of fruits was composed of a mix of apples and peaches. In this article, we continue to explore the capability of SVM associated with millimeter-wave (mm-wave) low-terahertz (THz) measurements. First, we tackle the problem of classifying a mix of fruits with a multiclass SVM using the Digital Binary Tree architecture. With this method, the error rate does not exceed 2%. Secondly, we move from the W- to D-band (lowTHz). The main reason is the increase of the lateral resolution and the possibility to have more compact systems in the view of an industrial deployment. We start our D-band investigations with range measurements to estimate the average permittivity of the apple in this frequency bandwidth. We have found a drastic decrease compared to the microwave region. It is consistent with the behavior of the water, which is one of the main components of the apple. Then we trained the SVM with the D-band database and finally performed the classification on unknown samples and obtained an accuracy of 100%.