Redes neurais convolucionais para classificação e avaliação da maturação de frutos de café

Abstract

In this work, we investigated deep learning methods to classify coffee trees maturation and assist farmers in the decision-making of the coffee harvest. These are little explored problems in the literature and they are difficult because the fruits ripen unevenly in the plant. In this sense, a database was created with more than 10,000 images of coffee fruits in their five stages of maturation, which was labeled by experts in the field. In addition, we tripled the classified images using data augmentation methods and trained a convolutional neural network over this data set. As these models are intended to be available in smartphones, we focused on the investigation of convolutional neural networks appropriate to this scenario, such as MobileNet, MobileNet V2, and NasNetMobile, which were trained and had their predictive performance compared. The results show that our best model obtained an accuracy of 93 \% and F1 of 92 \% in a separate set of test images. A prototype that implemented this model was created to allow coffee farmers to test our system.