Classificação de grãos de soja utilizando redes neurais convolucionais: uma análise comparativa entre modelos customizados e pré-treinados

Abstract

This study investigates the application of convolutional neural networks (CNNs) in soybean grain classification, with an emphasis on distinguishing between healthy, defective, and damaged grains. Considering the importance of soybeans as one of the most relevant agricultural crops, widely used in human and animal nutrition as well as in various industrial sectors, ensuring grain quality is essential to maintaining the efficiency and competitiveness of production chains. Several CNN architectures were analyzed, including customized models and pre-trained networks such as ResNet50 and InceptionV3. For image pre-processing, segmentation was performed using the YOLO model, followed by cropping, normalization, and data augmentation techniques to enhance the robustness of the models. Experimental results showed that pre- trained models performed significantly better than customized models, with InceptionV3 achieving an accuracy of 95.9%. In contrast, customized models achieved a maximum accuracy of only 16.6%. These findings highlight the effectiveness of pre-trained models, especially in image classification tasks involving subtle variations, such as identifying different types of grains. Based on the results obtained, this study provides a solid foundation for the future development of automated real-time grain classification systems, aiming to optimize inspection processes and improve the quality of agricultural monitoring.