Abstract: Wheat grains under different angles have different feature information, resulting in differences in classification results. In this paper, we adopt the same wheat grain multi-angle pictures, and use the three angle pictures of wheat grain groove upward, groove downward and groove forward to construct the wheat grain variety dataset. Six wheat varieties with large planting area in Huanghuai wheat area were selected as test materials to compare the accuracy of different models in wheat grain recognition. VGG-16, ResNet-50, and Inception-V3 convolutional neural networks are used to build a classification model for wheat seed variety recognition by transfer learning, and the highest recognition accuracy of the validation set is 99.35%, which is higher than that of the recognition method without transfer learning and the traditional machine learning recognition method. Under the same test conditions, the test set accuracies of the three models for wheat seed grain recognition using migration learning reached 99.55%, 99.77%, and 99.22%, respectively, which were better than single-sided feature modeling recognition. Based on the selection of the optimal test among each of the three models, their three angles were recognized separately. The results showed that the recognition rate of ventral groove downward was the best among the three models, ventral groove toward the front was the second best, and ventral groove upward was poor. It was found that the use of multi-angle pictures of the same wheat grain can extract wheat grain features more accurately and help the classification model to improve the accuracy of variety identification.