基于改进VGG19的中医背部痧象特征分类研究  

Research on Classification of Traditional Chinese Medicine Sha Features Based on Improved VGG19

作　　者：李斌[1] 李霄[1] 胡广芹[1] 张新峰 LI Bin;LI Xiao;HU Guangqin;ZHANG Xinfeng(Department of Environment and Life,Beijing University of Technology,Beijing 100124,China;Department of Information Science,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学环境与生命学部,北京100124 [2]北京工业大学信息学部,北京100124

出　　处：《中国医疗设备》2023年第9期12-16,共5页China Medical Devices

基　　金：北京市中医药科技发展基金(QN-2020-08;QYSF-2020-06);国家重点研发计划中医现代化研究专项(2018YFC1707705)。

摘　　要：目的比较3种网络模型[VGG19网络、多任务学习、卷积注意力模块(Convolutional Block Attention Module,CBAM)]在中医背部痧象特征分类的可行性与有效性,进而说明该模块的有效性。方法首先通过对痧象图片数据的分析,将图片数据按照颜色特征划分为暗红、红、淡红3类,形状特征划分为点状和片状两类,确定实验包含二分类和三分类两个任务,将图片按照训练集∶验证集∶测试集=8∶1∶1的比例进行划分,然后采用VGG19网络模型对两个任务分别进行训练,并以该网络模型为主干进行改进,引入多任务学习模型的思想,添加CBAM。采用训练准确率以及测试准确率进行评价,并设置颜色与形状准确率的均值,以判断网络模型的性能高低,同时通过消融实验比较最终的分类准确率,以及分析中医痧象特征与证型间的对应关系。结果以VGG19网络模型为主干,采用多任务学习并加入CBAM的改进网络取得了最高的分类准确率,当缩减率为1/8、batch_size为8时得到的训练结果最好,颜色分类准确率为93.90%,形状分类准确率为95.12%,平均准确率为94.51%。结论以VGG19网络模型为主干,采用多任务学习并加入CBAM的改进网络在中医痧象特征自动分类识别上可取得较好的效果,能够结合传统中医的经验知识准确完成对于人体证型的判断。Objective To compare the feasibility and effectiveness of the three network models[VGG19 network,multi-task learning,convolutional block attention module(CBAM)]in the feature classification of traditional Chinese medicine Sha images,and to illustrate the effectiveness of the module.Methods Firstly,through the analysis of the image data of Sha,the image data was divided into dark red,red and light red according to the color characteristics,and the shape characteristics were divided into point and sheet.The experiment included two classification tasks and three classification tasks,and the pictures were divided according to the ratio of training set∶verification set∶test set=8∶1∶1.Then,VGG19 network model was used to train the two tasks respectively,and the network model was improved as the main task,the idea of multi-task learning model was introduced,and CBAM was added.The training accuracy and test accuracy were evaluated,and the mean value of color and shape accuracy was set to judge the performance of the network model.Meanwhile,the final classification accuracy was compared by ablation experiment,and the corresponding relationship between traditional Chinese medicine Sha image characteristics and syndrome types was analyzed.Results Using the VGG19 network model as the backbone,an improved network using multi-task learning and incorporating CBAM achieved the highest classification accuracy.When the reduction rate was 1/8,the batch_size was 8,the best training results were obtained.The accuracy of color classification was 93.90%,and the accuracy of shape classification was 95.12%.The average accuracy was 94.51%.Conclusion Based on VGG19 network model,the improved network with multi-task learning and CBAM can achieve good results in the automatic classification and recognition of traditional Chinese medicine Sha image features,and can accurately judge the human syndrome type combined with the experience and knowledge of traditional Chinese medicine.

关 键 词：中医痧象 特征分类 注意力机制 VGG19 多任务学习 

分 类 号：R197.39[医药卫生—卫生事业管理]