Method for the detection and classification of quinoa seeds via computer vision

To solve the problem that there are many human factors, great difficulty, and low efficiency in distinguishing quinoa seeds from weed seeds and distinguishing the quality of quinoa seeds by appearance, a method of quinoa seed detection and classification based on computer vision is proposed. In this study, convolutional neural network and Vision Transformer (ViT) were used to quickly and nondestructively classify different quinoa seeds and weed seeds. The dataset used in this experiment was 1440 sample images containing quinoa seeds and weed seeds, which were divided into training set, test set, and validation set at a ratio of 8:1:1. The training set was 1152 pieces, test set was 144, and the validation set was 144 pieces. The convolutional neural network model and ViT model based on deep learning were established. The results show that the average classification accuracies of MobileNet, VGG16, ResNet50, and ViT models used in the experiment are 93.75%, 90.97%, 93.75%, and 98.61% respectively. The accuracy of ViT classification is much higher than that of convolutional neural networks, establishing a benchmark for quinoa seed classification. This study provides a reproducible dataset construction method and a dual-imaging strategy, and demonstrates practical deployment value for automated seed grading and purity testing.