CLASSIFICATION OF APPLES WITH CONVOLUTIONAL NEURONAL NETWORKS

Abstract

Nowadays, in points of sale and in agro-industrial companies in Mexico, the classification of apples (Malus domestica) is carried out manually by people, which generates deficiencies in the quality of the product. These problems can be reduced with the implementation of in site vision equipment with machine learning algorithms. In this study, several convolutional neuronal network (CNN) architectures were analyzed and one of those was selected that allows apples to be classified into healthy and damaged in the postharvest process. The varieties used were Red Delicious, Granny Smith, Golden Delicious and Gala. The accuracy of the LeNet5 and VGG16 CNNs was compared. A series of treatments (combination of network with hyperparameters) was performed that were used for the classification of the object of study. As each treatment was tested, its performance was measured. At the end, the treatment with the best performance was LeNet5 trained from scratch with the RMSProp optimizer, which obtained an accuracy of 97 %.