ZeaScan: ferramenta computacional para classificação e tipificação de grãos de milho utilizando redes neurais convolucionais

Abstract

Maize, an essential crop for global food security and the agri-food industry, requires effective quality control methods to minimize losses and ensure compliance with technical standards. In this context, this study introduces ZeaScan, a mobile application based on convolutional neural networks (CNNs), designed to automate the classification of maize kernels into healthy, mold-damaged, and weevil-damaged categories. The methodology involved image preprocessing to enhance morphological and chromatic features, the application of data augmentation techniques to improve the model’s generalization capacity, and the training of a CNN architecture optimized for efficient execution on mobile devices. The results demonstrated high system performance. In the three-class classification, global accuracy, precision, recall, and F1-score values of 97.36% were achieved. In the binary configuration, which differentiates healthy from defective kernels, the model reached 97.50% accuracy, 97.62% precision, 97.50% recall, and 97.53% F1-score. These findings highlight the robustness of ZeaScan, capable of reducing the subjectivity of manual methods and providing a practical, innovative, and efficient solution for maize quality control. This study represents a significant advancement in the application of computer vision to the production chain, offering a tool that combines technical rigor with operational practicality, with potential positive impacts on the sustainability and efficiency of agribusiness.