Genuine identification for Saposhnikovia divaricata based on improved DenseNet

Saposhnikovia divaricata has been one of the most widely planted herbs in Northeast Asia nowadays. The current quality identification of medicinal materials in the market depends mainly on the phenotypic observation from the experience of experts, indicating the certain subjectivity and limitations in large-scale production. In this study, an accurate, efficient, and intelligent approach was proposed to identify the genuine medicinal materials of saposhnikovia divaricata using an improved DenseNet. A standard dataset was established with 18 543 images of medicinal materials of saposhnikovia divaricata from five main producing areas using deep learning. A dense connection network was also improved to distinguish the origin, properties, and quality of the medicinal material. A new neural network model was established as follows. Firstly, a residual module was optimized to embed in the Coordinate Attention (CA)mechanism for the high feature weight of the area to be identified in the feature map. As such, the interference of background information was reduced in the identification task, particularly for the complex images of medicinal materials with only a small difference of phenotype. Then, the improved residual module was integrated with the dense connection network, further reducing the operation parameters of the model for the enhanced utilization rate of feature information. Finally, the fully connected layer was reconstructed to identify the new dataset for better learning of the network. The network model presented much fewer training parameters and faster convergence speed in the training process, further effectively reducing the over-fitting. The coordinate attention mechanism was replaced to compare the network model on the dataset in the improved network. A series of ablation experiments were conducted on the data set of medicinal materials. The experimental results show that the improved network model performed better to identify the origin of medicinal materials, indicating the significant effect of the coordinate attention mechanism on the accuracy of the model. Consequently, the new model can reach the convergence state after about 48 rounds of training, which greatly improves the convergence speed and then evaluate the medicinal properties, with an average accuracy rate of 97.23%.The strong robustness of the improved model can greatly contribute to the quality identification and evaluation standard of medicinal materials. The finding can provide a strong reference to intelligently and automatically identify the quality of medicinal materials.