Identifying Chinese Herbal Medicine by Image with Three Deep CNNs

Chinese herbal medicine (CHM) plays a key role in restoring the balance of the body for sick people and maintaining health for common people. Automated identification of CHM using images is a challenging task owing to the fine-grained variability in the appearance of CHM. This paper presented an application study of transfer learning using deep convolutional neural networks (CNNs) for image classification to identify CHM. 72,026 images of 7 categories of easily confusing CHM were collected. Fine-tuning the pre-training model of AlexNet, VGGNet, and ResNet based on transfer learning. The verification results of the three models with 10,806 images show, that the recognition accuracy of the ResNet model is the highest at 98.17%. Compared with other previous CHM image recognition methods, this model has reached a comparable level of CHM experience identification experts. Also, the key factors influencing the fine-tuning of pre-training models are evaluated. We describe a general artificial intelligence (AI) platform for the identification of CHM, and this system can be extended to other varieties of CHM.