Redes neurais convolucionais e ampliação de dados para detecção de antracnose em folhas de feijoeiro

Abstract

Deep learning models, especially Convolutional Neural Networks, are currently the protagonists of significant advances in computer vision tasks. The availability of robust computers, with great processing capacity, made it possible to expand the fields of work in this area and neural networks further increased the possibilities. However, deep learning models commonly require large databases to learn relevant characteristics of the images, and building plant disease databases requires the commitment of a considerable workforce and access to a vast collection of information. images, in addition to reasonable capture conditions. Data augmentation methods serve the purpose of increasing the number of items in a database through modifications to existing images and, sometimes, the generation of new images based on the characteristics observed in part of the set. This study undertook the collection of a set of images of bean leaflets for the elaboration of databases and the subsequent conduction of tests with different neural network architectures and configurations of data augmentation techniques for the purpose to evaluate the influence of such techniques in the simulation of images collected under different conditions, as well as their influence on the performance measures of accuracy, sensitivity, specificity and overfitting in order to distinguish images of parts of healthy leaflets from those containing structures affected by anthracnose fungal disease at different stages. The results show that the data expansion techniques allow the simulation of different conditions of image acquisition, and the classification models based on convolutional neural networks achieved results varying between 60 and 90% accuracy in the different experimental configurations. The bibliography consulted in conjunction with the results presented, gives rise to the suggestion of new works on the same theme, considering the expansion of the database and the use of other techniques for data augmentation and neural network architectures.