Classificação de defeitos em grãos de soja com o uso de visão computacional e redes neurais artificiais

Abstract

Soybean is one of Brazil's main agricultural commodities, playing a fundamental role in both the national and global economy. Grain quality classification is an essential process for its commercialization and industrial application, currently performed manually and subject to subjectivities and inconsistencies. In this context, this study proposes the development of an automatic defect classification system for soybean grains using computer vision and artificial intelligence techniques, aiming to increase the accuracy and efficiency of the process. The methodology employed consisted of training and evaluating different convolutional neural network (CNN) architectures. Custom networks were tested, along with the pre-trained models ResNet50 and InceptionV3, to compare their performance in terms of accuracy and practical applicability. The results demonstrated that pre-trained networks significantly outperformed custom networks. The InceptionV3 model stood out as the best alternative, achieving 95.9% accuracy and demonstrating high generalization capability, making it a viable option for implementation in industrial environments. The ResNet50 model, with 63.1% accuracy, proved suitable for applications requiring a balance between performance and computational cost. Meanwhile, the custom CNNs showed unsatisfactory performance, reaching a maximum accuracy of 16.6%, indicating the need for improvements in architecture and training. At this stage, it is concluded that the application of convolutional neural networks for soybean grain classification has great potential to modernize and optimize the agro-industrial sector. It is also worth highlighting the creation of a dataset containing images of soybean grains with various defects, which may support future research in the field of computer vision applied to grain analysis.