Design of a sugarcane diseases recognition system based on GoogLeNet for a web application

• Autores: Barroso Maza, Cristian Leoncio, Lucas Cordova, Juan Carlos, Sotomayor Beltran, Carlos Alberto
• Formato: artículo
• Fecha de Publicación: 2022
• Institución: Universidad Tecnológica del Perú
• Repositorio: UTP-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:repositorio.utp.edu.pe:20.500.12867/5993
• Enlace del recurso: https://hdl.handle.net/20.500.12867/5993
• Nivel de acceso: acceso abierto
• Materia: Artificial neural networks; Sugarcane; Plant diseases; https://purl.org/pe-repo/ocde/ford#2.02.00

Sugarcane diseases in Peru occur due to the agricultural community's lack of understanding of these, which means a slow response to the application of methods of control and eradication of these diseases; thus, causing economic losses and underproduction. Due to the aforementioned, a web application for sugarcane diseases recognition is proposed. The five types of sugarcane diseases that will be recognized using this system are: Pineapple Sett Rot, Ring Spot, Mosaic, Brown Rust and Leaf Scorch. This system was developed using GoogLeNet, which is a 22 layers convolutional neural network (CNN), and also the Matlab software and its App Designer extensions (for the web application creation); additionally, Matlab Web App Server was used to host the application on the web. The pre-trained neural network developed in Matlab based on the GoogLeNet architecture allowed the creation and configuration of the training parameters (supervised learning) that were evaluated, and it was considered convenient to split the data between training, validation and testing (70%, 20% and 10%, respectively). A total of 250 images composed of 50 images for each disease were used. The web application was designed in App Designer which provided us with a set of tools and a programming interface for the insertion of the trained CNN, with a validation percentage of 94.67% obtained by varying the number of epochs, reaching a maximum of 6000 iterations. Finally, the web application supported by the Matlab Web App Server was generated and tests were performed on a local network, resulting in a web application capable of identifying images within the established guidelines, with an accuracy rate of 96%.