Título: Automatic Detection of Spots on Coffee Leaves using Deep Learning
Autores: Caio Dias, Valdivino Alexandre de Santiago Junior & Elcio H. Shiguemori
Resumo: Coffee cultivation is an agricultural activity of great importance to the global economy, but it faces challenges due to diseases and pests that compromise crop productivity. Therefore, the automatic detection of spots on coffee leaves, which may indicate diseases and pests such as rust and leaf miner, is extremely relevant for increasing productivity in plantations. The use of deep learning (DL) models has become increasingly common for this problem, but robust evaluations are still necessary to understand the suitability of these models, as well as the importance of including both laboratory and field images. This study evaluates the performance, in terms of accuracy and computational cost, of seven DL-based object detectors considering two datasets of coffee leaves. Models such as YOLOv10-M and YOLOv10-X offer a good balance between accuracy and latency but still present high inference times for embedded applications, especially when compared to lighter models like YOLOv10-Nano. Among all the approaches analyzed, YOLOv10-Nano stands out as the most efficient option, offering a good balance between accuracy (0.332 mAP@50) and speed, with the lowest latency (8 ms) and a reduced training time (40 minutes). In addition, the Eigen-CAM technique, an explainable artificial intelligence (XAI) approach, was applied to the YOLOv10-Nano model with the aim of improving its interpretability, allowing for a better understanding of how the network focuses on different regions of the leaf during the detection process. The YOLOv10-Nano model has started to be deployed on mobile devices, aiming at future developments that will make this technology viable for small producers.
Palavras-chave: Coffee; Object Detection; Leaves; Deep Learning; YOLO.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1175669
Artigo em PDF: CBIC_2025_paper1175669.pdf
Arquivo BibTeX:
CBIC_2025_1175669.bib