Análise do impacto do aumento de dados no reconhecimento de ferrugem em folhas de café

Abstract

Coffee rust is a very serious disease that affects plantations worldwide, and identifying it quickly and accurately can assist in its control. Currently, there are several ways to combat the rust disease, and the best choice of identification strategy will directly impact the process, potentially accelerating the time between identifying the problem and taking action to address it. This work aims to compare the performance of six Convolutional Neural Networks (CNN) in the context of detecting rust on coffee leaves, using a public image dataset called RoCoLe. The dataset consists of 6 classes of leaves (healthy, diseased with severity levels from 1 to 4, and leaves with damage caused by red mite). The networks were trained with and without data augmentation. The experiments showed that the use of data augmentation strategies improved the classification task in some situations.