Rede neural convolucional e padrão de metadados na classificação de grãos de soja

Abstract

The classification of grains in agriculture is done manually, where grain by grain must be analyzed so that the grains are allocated to their respective classes. The identification of the classes of the grains is done visually, where the analysis of the color and the state in which the grain is made. This classification is usually done by professionals specialized in the field, which can take a long time to classify a small amount of grains. The time to carry out the manual classification is essential for the grains to be evaluated and sent to consumers as soon as possible. In this work, a metadata standard model for grain classification was presented, an application to structure metadata in HTML or XML, as well as a computational method for the classification of soybeans of the Glycine max species where convolutional neural networks were used. The convolutional neural networks used were Resnet34, MobileNet and VGG19. The network was trained on the basis of grain defect data. Nine (9) classes of soybeans were considered, where each class contained 100 (one hundred) images. The result achieved by Resnet34 was 99.55% accurate with the use of data augmentation, and 80.85% without the use of this technique. The MobileNet and VGG19 networks achieved an accuracy of 97.33% and 97.22% with the use of data augmentation, respectively.