融合改进残差网络和Transformer的易混淆中药饮片识别研究

Research on identification of confusing TCM decoction pieces by integrating of improved residual network and transformer

目的:基于改进的残差卷积神经网络和Transformer混合架构对5组常见的根茎类易混淆中药饮片构建分类模型,探索人工智能对复杂场景中药饮片的分类价值。方法:以玉竹与知母;大血藤与鸡血藤;桔梗与黄芪;防风、前胡与板蓝根;怀牛膝、川牛膝与续断12类5组易混淆中药饮片为研究对象,构建包含8720张图片的中药饮片数据集。设计基于残差模块和注意力机制的TyNet103 CNN+Transformer模型,对TyNet103自建模型进行消融实验并与3个经典模型的准确率进行比较,结合集成学习思想进行基分类器投票得到中药饮片分类结果。结果:4个模型在测试集上的初始准确率分别为78.41%、77.83%、72.57%和80.13%,模型改进并应用组合策略优化后准确率分别达到95.35%、96.65%、96.33%、97.56%,应用集成学习方法得到的强分类器在测试集上准确率可达99.41%。结论:通过模型改进,组合优化策略和集成学习实现的中药材分类模型具有智能化、客观化、准确率高且高效易用的特点,可辅助鉴别易混淆的中药饮片。

Objective Based on the improved residual convolutional neural network and Transformer hybrid architecture,a classification model of 5 easily confused rhizomatic TCM decoction pieces was constructed to explore the value of AI in classifying TCM decoction pieces in complex scenarios.Methods A total of 5 groups of 12 kinds of easily confused TCM decoction pieces(Rhizoma Polygonati Odorati and Rhizoma Anemarrhee,Caulis Sargentodoxae and Caulis Millettiae,Radix Platycodi and Radix Astragali,Radix Saposhnikoviae divaricatae and radix Peucedani and Radix Isatidis seu Baphicacanthi,Radix Achyranthis bidentatae,Radix Cyathulae and Radix Dipsaci)were selected as the research objects and a dataset of 8,720 pictures of TCM decoction pieces was constructed.A TyNet103 CNN+Transformer model based on residual module and attention mechanism was designed.The classification accuracy of this self-built model was compared with the three classical models through an ablation experiment.With the idea of integrated learning,the classification results of TCM decoction pieces were obtained by base classifier voting.Results The initial accuracy of the four models on the test set was 78.41%,77.83%,72.57%and 80.13%,respectively.After model improvement and combinatorial strategy optimization,the accuracy of the model was improved to 95.35%,96.65%,96.33%and 97.56%,respectively.The strong classifier obtained by integrated learning method has an accuracy of 99.41%on the test set.Conclusion The proposed classification model of TCM decoction pieces achieved by model improvement,combinatorial optimization strategy and integrated learning has the characteristics of intelligent,objective,high accuracy and efficiency and easy to use,which can provide technical support for the identification of TCM decoction pieces.