Classificação do nível de maturação e contagem de tomates com redes neurais convolucionais usando dados de colorimetria

Abstract

Tomato maturity classification is an important task in precision agriculture and is traditionally performed manually and subjectively. This work presents an approach that combines colorimetric measurements in the CIELAB color space with YOLO models for automatic tomato detection, counting, and classification into green, intermediary, and red classes. A dataset containing 1,040 images and 8,748 annotated fruits was built, with maturity classes defined from measurements obtained from 150 samples and grouped using the K-means algorithm. YOLOv10, YOLO11, YOLO12, and YOLO26 models, in both Small and Medium variants, were evaluated, achieving precision values above 95%, recall above 94%, and mAP@0.5 up to 97.5%. YOLO26m achieved the best overall performance. The trained models were integrated into a web application developed using Streamlit for image and video processing. The results highlight the potential of combining colorimetry and deep learning techniques for automated tomato maturity monitoring.