Coffee disease classification at the edge using deep learning
João Vitor Yukio Bordin Yamashita, João Paulo Leite
Abstract
Brazil is the world’s largest producer and exporter of coffee and the second largest consumer of the beverage. The aim of this study is to embed convolutional networks in a low-cost microcontrolled board to classify coffee leaf diseases in loco, without the need for an internet connection. Early identification of diseases in coffee plantations is crucial for productivity and production quality. Two datasets were used, in addition to images taken with the development board itself, totaling more than 6000 images of six different types of diseases. The proposed architectures (cascade and single-stage), when embedded, presented accuracy values around 98% and 96%, respectively, demonstrating their ability to assist in the diagnosis of diseases in coffee farms, especially those managed by producers with less resources.