Uso de rede neural para desenvolvimento de sistema especialista para diagnose de doenças foliares em eucalipto

Abstract

After the growth of data creation and storage, which are the raw material of artificial intelligence, in recent years it has been noticed that almost every industry and health sector already works with artificial intelligence software, which are used mainly to aid automation, fraud analysis, diagnosis of human diseases, digital marketing, autonomous cars, social networks, among others. However, in the agroforestry sector, responsible for a large part of the Brazilian economic GDP, work, software and information related to artificial intelligence are scarce. The objective of this work is to create a system based on artificial neural networks (ANN) for detection of eucalyptus leaf diseases, capable of performing digital image processing using computer vision techniques and training a neural network. with the multilayer Perceptron architecture using the Backpropagation training algorithm, through the Python programming language. The present work was developed with the collection of leaves with Mycosphaerella leaf spots and eucalyptus rust (Austropuccinia psidii), as well as healthy leaves for the creation of the dataset for training the Artificial Neural Network (ANN) multilayer perceptron (MLP) with the algorithm of backpropagation. The sheets were scanned and submitted to the first process carried out by the expert system, transforming the color images into grayscale, reducing from three color dimensions (RGB) to just one dimension, standardizing the width of the sheet and resizing its height without loss of image proportion and finally binarization to extract only the object of interest, generating a histogram with grayscale frequencies that was used as input to the neural network for training and validation. Eight topologies of Artificial Neural Networks were proposed, containing four topologies with one hidden layer of neurons and four topologies with the one with two hidden layers of neurons. All topologies had an average of 92% hits, being considered the most suitable for the topology with only one layer with 86 neurons by the average of the best results obtained from the accuracy, precision, recall and F1 Score metrics above 93% and the low computational effort for leaf diagnosis which guarantees a better performance of the developed expert system.