Identificação do estádio de Maturação de frutos de café com a utilização de visão computacional

Abstract

This work proposes and presents the results of a system based on computer vision for iden- tifying different maturation stages of coffee fruits. This identification plays a crucial role at the time of harvest, as the degree of maturation directly influences the quality of the final product. The system was developed using client-server architecture, in which a mobile application (client) is res- ponsible for capturing photos and sending them to a web application (server), where the image is processed by a Deep Learning model (Mask RCNN or YOLOv8 ). After processing, the informa- tion obtained is returned to the mobile application for user viewing. All communication was done via HTTPS protocol. In order to deal with network and asynchrony problems, on both sides of the system there is a database storing all photos and information. Results related to image processing were obtained in three different contexts: (i) for green, sugarcane, cherry and raisin/dried fruits, (ii) for green and ripe fruits, and (iii) fruit identification (counting) regardless of the stage of matu- ration. The Mask RCNN model presented the best performance when the results were compared, presenting an average APE in context (iii) of 9.61%, in context (i) the lowest APE obtained was 9.12% for the green-cane class and in the context (ii) obtaining its lowest APE for the mature class, in the amount of 8.22%. The web system was developed, with its server implemented in Python, where the best performing Deep Learning model (Mask RCNN) was integrated. The mobile appli- cation was implemented using the React Native framework, being able to send photos and receive reports from the server.