Small Sample Succulent Image Classification Based on Transfer Learning

Abstract: Succulents encompass a wide range of species, each with unique shapes and different colors. Despite their widespread cultivation in recent years, the study of succulents lags behind that of flowers. Furthermore, there is currently no public data-set for succulent plants. As a result, we have utilized a self-built data-set and employed deep learning methods to implement succulent plant image classification. While CNN (Convolutional Neural Networks) excels at extracting local features, it requires numerous layers to capture global features, and the improvement in model performance is limited. On the other hand, the Transformer model is adept at capturing long-range dependencies in sequences due to its self-attention mechanism. However, a pure Transformer model requires a significant amount of data and parameters to surpass convolutional models. Therefore, we have focused on implementing the classification of 31 succulent plant species using the lightweight network MobileViT, which combines both convolution and Transformer. In the context of limited sample data, we have employed data augmentation and transfer learning methods to train the model for 150 epochs, achieving a test accuracy of 96.90%. Compared to other convolutional neural networks, such as Resnet50, MobilenetV2, Xception, and ShufflenetV2, our model and method require fewer parameters and achieve higher accuracy.