Research on Lightweight Models in the Identification of Tibetan Medicinal Materials

Objective: To address the issues of low efficiency and high error rates in the manual identification of Tibetan medicinal materials and the deployment of applications on mobile devices. Method: This study proposes an image identification method based on the lightweight model MobileNetV2 through comparative experiments for this task. The research dataset consists of 6,622 images of seven types of herbs for the training set and 1,656 images for the test set. The model's performance is evaluated using evaluation criteria, model parameters, and computational complexity. Results: The model achieves a test accuracy of 90.02%, precision of 90.70%, recall of 90.02%, F1 score of 88.70%, parameters of 2.24G, and FLOPs of 306.19M for the task of herb classification and identification. Conclusion: The model is compact and accurate, allowing for deployment on mobile devices to achieve high-precision identification of Tibetan medicinal materials, thereby promoting the intelligent development of the Tibetan medicine industry.