Development of a System for Classification of Rice Grains Using Convolutional Neural Networks

Abstract

The growing demand for quality in the rice industry has driven innovative solutions for classifying rice varieties, preventing mixtures that impact the final product. This study introduces a convolutional neural network (CNN) for automatic rice grain classification using digital images. A dataset of 75,000 images, divided into five categories ('Ipsala,' 'Arborio,' 'Jasmine,' 'Karacadag,' and 'Basmati'), was used. The model was trained with 80% of the data (56,000 images) and validated with the remaining 20% (14,000 images), using 5,000 new images for final evaluation. The CNN achieved 99.2% accuracy, demonstrating high performance even among visually similar varieties. This approach modernizes traditional methods, improving efficiency and ensuring higher-quality products in the Colombian rice industry.