A multitask model based on MobileNetV3 for fine-grained classification of jujube varieties

Jujube is one of the important members of the Rhamnaceae family, which has high economic and nutritional value. The automatic classification of jujube varieties is of great importance for the development of the jujube industry. The task of jujube variety classification based on image data belongs to fine-grained visual classification, which is relatively more difficult. Recently, some deep learning approaches have been available, but capabilities remain limited. In this paper, we present a fine-grained classification model for jujube varieties, named JLN. It is based on MobileNetV3 and multi-task learning. JLN performs shape classification and fine-grained classification simultaneously through a multitask framework and learn more phenotypic differences among different varieties by mining the relationship between the two tasks. In order to further improve the classification performance of JLN, we let the fine-grained features involved in the task of shape classification but prohibited this process in the model's backward passing stage. Finally, we used Reduce Focal Loss to train the shape classification task and jointly optimize the variety classification task through Arcface and Softmax Loss. To further facilitate the study, we established an image dataset of 20 jujube varieties and divided them into 7 categories according to their shapes under the advice of experts. The proposed method achieves 91.79% accuracy on the jujube variety data, and the number of model parameters is only 3.56 M, which is better than the other four models. The experimental results show that the method is more suitable for jujube variety classification tasks in the natural environment.