Near infrared spectrum medicine identification method based on combined strategy

• Main Authors: JIN SONGLIN, ZHENG YING, LI ZEXU, ZHOU LING, ZHAO WENYI, TIAN YONGQIN, ZHANG WEIDONG, LI GUOHOU, QU PEIXIN
• Format: Patent
• Language: Chinese; English
• Published: 18.07.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; ELECTRIC DIGITAL DATA PROCESSING; INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTEDFOR SPECIFIC APPLICATION FIELDS; INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIRCHEMICAL OR PHYSICAL PROPERTIES; MEASURING; PHYSICS; TESTING

The invention provides a near infrared spectrum drug identification method based on a combined strategy, which comprises the following steps of: firstly, constructing the advantage of a deep convolutional network by utilizing a residual mechanism to improve the feature extraction capability of a spectrum, and then, carrying out model transfer on small sample data by utilizing the advantage of overcoming the data sensitivity of transfer learning to obtain the drug identification result. According to the method, the model lifting rate is higher and the convergence effect is better in the training process, and finally, a Focal Loss loss function is proposed for the imbalance of a source data set, so that a sample with a smaller expected scale can also make a greater contribution to the loss, and the tendency of the model to most samples is solved. According to the method, after the network depth is deepened, the drug spectrum feature extraction capacity is higher, model cross-drug type transfer is achieved, and