Authors:Mateus Coelho Silva1;2;Jonathan Cristóvão Ferreira da Silva2 andRicardo Augusto Rabelo Oliveira1
Affiliations:1Departamento de Computação, Instituto de Ciências Exatas e Biológicas, Universidade Federal de Ouro Preto, Brazil;2Instituto Federal de Educação, Ciência e Tecnologia de Minas Gerais, Campus Avançado Itabirito, Brazil
Keyword(s):Edge Computing, Embedded Systems, Computer Vision, Machine Learning.
Abstract:Orange and citrus agriculture has a significant economic role, especially in tropical countries. The use of edge systems with machine learning techniques presents a perspective to improve the present techniques, with faster tools aiding the inspection diagnostics. The usage of cost- and resource-restrictive devices to create these solutions improves this technique’s reach capability and reproducibility. In this perspective, we propose a novel edge-computing-based intelligent diagnosis support system performing a pseudospectral analysis to improve the orange inspection processes. Our results indicate that traditional machine learning methods reach over 92% accuracy, reaching 99% on the best performance technique with Artificial Neural Networks in the binary classification stage. For multiple classes, the accuracy varies from 97% up to 98%, also reaching the best performance with Artificial Neural Networks. Finally, the Random Forest and Artificial Neural Network obtained the best results, considering algorithm parameters and embedded hardware performance. These results enforce the feasibility of the proposed application.(More)
Orange and citrus agriculture has a significant economic role, especially in tropical countries. The use of edge systems with machine learning techniques presents a perspective to improve the present techniques, with faster tools aiding the inspection diagnostics. The usage of cost- and resource-restrictive devices to create these solutions improves this technique’s reach capability and reproducibility. In this perspective, we propose a novel edge-computing-based intelligent diagnosis support system performing a pseudospectral analysis to improve the orange inspection processes. Our results indicate that traditional machine learning methods reach over 92% accuracy, reaching 99% on the best performance technique with Artificial Neural Networks in the binary classification stage. For multiple classes, the accuracy varies from 97% up to 98%, also reaching the best performance with Artificial Neural Networks. Finally, the Random Forest and Artificial Neural Network obtained the best results, considering algorithm parameters and embedded hardware performance. These results enforce the feasibility of the proposed application.