Classification of Chinese Herbal Medicines by deep neural network based on orthogonal design

Chinese herbal medicines (CHMs) play an important role in the clinical efficacy of TCM. It is necessary to establish an intelligent identification method for CHMs to assist in Chinese medicine dispensing. In the present study, 160 kinds of CHMs were collected from the real world, and an image database containing 44467 pictures was constructed. Based on the L18 (3^6,6^1) hybrid orthogonal table, an orthogonal experiment was carried out on seven influencing factors, namely architecture, learning rate, optimizer, weight decay, batch size, gradual unfreezing, and discriminative fine-tuning. Five-fold cross-validation was used to calculate the accuracy rate, recall rate, and other evaluation indicators, and in this way, the optimal model and hyperparameter combination were screened. It was found that the optimal model was ResNeXt-152. When the batch size was 64, the learning rate was 0.00005, weight decay was 0.0005, gradual unfreezing was 7, discriminative fine-tuning was 3, and optimizer was AdaMax, which was the optimal combination of hyperparameters. At this time, the highest accuracy rate was 95.36%. This shows that deep learning model training based on orthogonal design can reduce the number of experiments while improving classification accuracy.