Development of a multiview tomato classification system by computer vision using convolutional neural networks

Briones P.; Altamirano E.; Nieves A.; Carrizales J.; Klusmann M.; Sampen L. · International Conference on Electrical, Computer, Communications and Mechatronics Engineering, ICECCME 2025 (2025) · DOI: 10.1109/iceccme64568.2025.11277708

Resumen

The agri-food industry experiences considerable losses during the postharvest stage, largely due to manual tomato grading processes, which typically achieve less than 70% accuracy. This inefficiency contributes to an estimated 20-30% loss of total production, negatively impacting both profitability and quality standardization. The demand for accurate, robust, and automated grading systems is especially critical for mediumscale producers who cannot afford high-end industrial solutions. This project presents a cost-effective tomato grading system based on computer vision and deep learning, combining a mechanical module for controlled axial rotation and lightweight convolutional neural networks (CNNs) for classification. As the tomato rotates, multiple images are captured to ensure complete surface coverage, enhancing the reliability of the analysis. Two state-of-the-art CNN architectures-MobileNetV3 Small and EfficientNetB0-were trained and evaluated to classify tomatoes according to three USDA-defined maturity stages and four size categories. In a dataset comprising 570 tomato samples, the system achieved a classification accuracy of 87.4% for maturity and 98.7% for size. The average processing time per tomato was 4.5 seconds, and the solution was successfully deployed on a Raspberry Pi 5, demonstrating its feasibility for low-cost, realtime applications. The comparative analysis of MobileNetV3 and EfficientNetB0 underscores their suitability for embedded AI systems, offering a promising alternative to manual grading and paving the way for smarter Agriculture 4.0 practices.

Instituciones: Universidad Peruana de Ciencias Aplicadas