Detecção de ferrugem de folhas vegetais por meio de imagens aplicando redes neurais convolucionais

Abstract

Leaf rust is a fungal pathology that is defined by the presentation of pustules with dark yellow to brown spores on the face of the leaves, from the incident to the stage of maturation, causing losses in grain production that can exceed 50%, so cool the photosynthetic area and then increase the breath. The classification of plant leaves and the identification of fungal pathology as a rust is a process that until then has been done manually by taxonomists and this ends up taking a lot of time to be carried out, it is then necessary for the producer to monitor the development of the infection to determine the level infection and apply the correct medication and even then it can bring a low success rate. This work proposes an algorithm to solve the classification of leaves with rust using Convolutional Neural Networks (Deep Learning algorithms) based on 440 images of 12 different species, with the objective of classifying them into rust carrier or not. It was transferred according to the convolutional neural network architectures: AlexNet, GoogleNet, Resnet-18 and VGG-19. The best results obtained were used with Resnet-18 Convolutional Neural Networks and VGG-19, both with 20- season training. Resnet-18 achieved an accuracy of 96.6%, sensitivity of 97.7% and specificity of 95.5%, while VGG-19 reached an accuracy of 96.6%, sensitivity of 95.5% and specificity of 97.7%. Thus, the use of convolutional neural networks has great potential for use in the detection of fungal pathology in the case of rust.