A Vision Technology Fusion Framework for Sorting Panax notoginseng

In the automated processing of traditional Chinese medicine, the efficiency and accuracy of sorting are key factors affecting its quality. Panax notoginseng’s various parts must be clearly distinguished before pharmaceutical preparation due to the different content of saponins and heavy metals. However, the various parts have significant differences among individuals, which makes it difficult for existing methods to distinguish them accurately and efficiently. Therefore, a multiclass detection method net integrating detector, tracker, and classifier (DTC-Net), which integrated multiple vision techniques, was proposed to aggregate the time domain characteristics with more refined part discrimination. Then, make a comprehensive judgment on the categories for each part. Compared to other algorithms, DTC-Net can track multicategory targets end-to-end in real time and has higher inference efficiency by jointly training the detector and classifier. On the video (or image) dataset taken in actual application scenarios, DTC-Net achieved excellent performance in industrial-grade scenarios under good and harsh lighting conditions, while efficiently focusing on multicategory targets. This can later guide pharmaceutical quality control.