Training of a Convolutional Neural Network for the Classification of Coffee Fruit State

¹ FI, Universidad Tecnológica Centroamericana (UNITEC), San Pedro Sula, Honduras
² Dagestan State Technical University, Makhachkala, Russia
³ Kadyrov Chechen State University, Grozny, Russia

^* Corresponding author: matygov.movsar@gmail.com

Abstract

Convolutional neural networks are recognized for their high artificial intelligence capacity, mostly for their identification of objects either in images or videos. Honduras has a strong dependence on coffee cultivation, as it plays a crucial role in both its culture and economy. However, Honduras has endured difficulties linked to aspects such as quality, contamination, processing problems, illnesses, or lack of training. The root of problems often arises during the coffee harvest, and since Honduras is a less developed country, the use of advanced technology is not common. The recurring research aims to train a cnn model by using images and object detection to classify the quality of the coffee fruit. The training of a convolutional neural network was developed by using Robo Flow for the classification of the state of the coffee fruit. Firstly, the coffee farm was visited to see firsthand the fruit in its various stages. Due to the seasonal cuts established on the farms, a second visit was made to collect the coffee fruit. After that, the fruit was taken into a controlled environment to start the photo session. The session concluded with a total of 1702 photographs. The fruits within the images were individually annotated by classifying them into two categories: good and bad. A model result gave a precision, recall and mpa of 97.7%, 94.1% and 98.3%, respectively. The configurations were the use of 600 images, using a ratio of 80:10:10 for the division of training, validation, and testing tasks. Only one adjustment was made in the preprocessing, which consisted of changing the size of the images from 3000x4000 to 640x640. This is considered the best result after carrying out more training to test the alteration of the results of other models by making changes such as the number of images and the types of augmentations. Therefore, the results conclude the fulfillment of the objective.