Deep learning for fine-grained classification of jujube fruit in the natural environment

Jujube is a popular fruit with a long cultivation history and numerous varieties in China. It is necessary to develop an automatic visual identification system of jujube classification in the natural environment. However, practical success in this area is still limited. In this paper, we propose a deep convolutional neural network model for the fine-grained classification of jujube, which exploits a two-stream network to effectively learn discriminative features for each image from both shape level and fine-grained level simultaneously. Specifically, it can also learn the contrastive discrepancies from jujube image pairs. To further facilitate the research, we create a rich jujube image dataset in the natural environment. The dataset consists of more than 1700 images of 20 jujube varieties, and these images have a large degree of variations including angles, background and illumination conditions. The proposed model achieves an average accuracy of 84.16% on this dataset, which outperforms the other four models, including SVM, AlexNet, VGGNet-16 and ResNet-18. The feasibility of this method is demonstrated by the experiment results.