Soybean Seed Classification with Fusion of Convolutional Neural Networks and ViT Technique

Abstract; Soybeans are a nutritious legume cultivated worldwide and are a staple in the human diet. They also play a vital role in animal feed production. Furthermore, soybean oil, extracted from soybeans, is widely used in the food industry. Several different convolutional neural network (CNN) models, along with image transformation (ViT) models, were used to classify soybean seeds. The seed classification comprised five categories: broken, immature, intact, shell-damaged, and defective. To enhance the generalizability of the dataset, data augmentation techniques were applied during the training process. Seventeen pre-trained CNN and ViT models with varying architectures were tested. Model performance was evaluated using performance metrics such as accuracy, fine-tuning, recall, and F1. The proposed model was constructed by combining feature extraction layers from the top-performing ViT-L/16, DenseNet201, and DenseNet169 models. The proposed model outperformed all other models, achieving an accuracy of 96.02% in the classification process. With this proposed model, we have obtained a hybrid approach to classifying soybean seeds, and it contributes to the use of artificial intelligence in the agricultural sector.