北京生物医学工程

基于TripletLoss损失函数的舌象分类方法研究

Tongue image classification based on TripletLoss metric

作者：孙萌张新峰

单位：北京工业大学信息学部（北京 100124）

关键词：肿瘤；舌象；分类；深度学习； Tripletloss； FaceNet

分类号：R318.04

出版年·卷·期（页码）：2020·39·2（131-137）

摘要：

目的舌象体质分类对后续肿瘤患者舌象的客观化辨证具有重要意义，但对于中医舌图像而言，部分类型的舌图像样本较难采集，达不到目前流行的深度学习方法需要的样本数量，且基于传统分类的深度学习只注重寻找具有相似特征，导致模型在中医舌图像这种类间样本特征差异较小的问题上，分类性能不佳。因此，本文提出一种基于Tripletloss的度量分类方法，在最大化非同类样本的特征距离同时缩小类间样本特征的间距。方法首先通过建立卷积神经网络Inception-ResNet-V1提取对应的高维抽象特征。然后使用L2范数进一步约束高维特征的分布，同时引入降维压缩后的高维特征，最后使用TripletLoss得到有效的映射空间。因此可以根据舌象间的特征向量距离计算相似度以实现分类。结果经过本文方法得到的特征空间，不同类型舌象之间的距离较大，同一类型的舌象距离较小，可以更好地对类间差异较小的舌图像进行分类，且分类速度更快。与现有方法比较，本论文方法在分类精确度上提升了18.34%，并且所需时间最短。结论该方法可以很好地实现舌象体质分类，具有一定的应用价值。

Objective Constitution classification based on tongue image is of great significance to the objective differentiation of tongue image of subsequent tumor patients. However, for tongue images of traditional Chinese medicine, certain types of tongue image samples are difficult to collect, which can not meet the number of samples required by the current popular deep learning methods. Moreover, deep learning based on traditional classification only focuses on finding similar features, resulting in poor classification performance of the model on the problem of small differences in sample features between classes, such as tongue images of traditional Chinese medicine. Therefore, this paper proposes a metric classification method based on tripletloss, which maximizes the feature distance of different classes of samples while reducing the distance between classes of samples. Methods Firstly, the corresponding high-dimensional abstract features are extracted by establishing convolution neural network Inception-ResNet-V1. Then L2 norm is used to further constrain the distribution of high-dimensional features. Meanwhile, the high-dimensional features after data dimensionality reduction and compression are introduced. Finally, a valid mapping space is obtained by using Tripletloss. Therefore, the similarity can be calculated according to the feature vector distance between tongue images to realize classification. Results According to the feature space obtained by the method in this paper, the distance between different types of tongue images is larger, and the distance between the same type of tongue images is smaller, which can better classify tongue images with smaller differences between classes, and the classification speed is faster. Compared with the existing methods, the classification accuracy of the method in this paper is improved by 18.34%, and the required time is the shortest. Conclusions This method can well realize the constitution classification based on tongue image and has certain application value.

参考文献：

[1]杨晓蕾,杨超,张钦婷,孙权,葛杰.恶性肿瘤患者中医体质类型相关研究[J].辽宁中医药大学学报,2015,17(8):164-166. Yang XL, Yang C, Zhang QT, et al. Malignant tumor patients with traditional chinese medicine constitution type of related research[J].Journal of Liaoning University of Traditional Chinese Medicine,2015,17(8): 164-166.

[2]中国中西医结合研究会肿瘤专业委员会中医诊断协作组.4417例癌症患者舌象临床观察[J].浙江中医杂志, 1992, 37 (8): 368-369.

[3] 钱峻, 刘沈林.消化系恶性肿瘤舌象辨治探微[J].吉林中医药, 2005, 25 (12) :1-2.

[4]Litjens G, Kooi T, Bejnordi BE , et al. A survey on deep learning in medical image analysis[J]. Medical Image Analysis, 2017, 42: 60-88.

[5]Rawat W, Wang Z. Deep convolutional neural networks for image classification: a comprehensive review[J]. Neural Computation, 2017, 29(9): 2352-2449.

[6]刘飞, 张俊然, 杨豪. 基于深度学习的医学图像识别研究进展[J].中国生物医学工程学报,2018,37(1):86-94. Liu F,ZhangJR,YangH. Research progress of medical image recognition based on deep learning[J]. Chinese Journal of Biomedical Engineering,2018,37(1):86-94.

[7]LeCunY, BengioY, Hinton G. Deep learning[J]. Nature,2015,521(7553):436-444.

[8] Brown GW. On small-sample estimation[J].The Annals of Mathematical Statistics,1947, 18(4): 582-585.

[9] Koch G, Zemel R, Salakhutdinov R. Siamese neural networks for one-shot image recognition[C]// International Conference on Machine Learning. Lille France:JMLR, W CP, 2015, 37.

[10]Snell J, Swersky K, Zemel R. Prototypical networks for few-shot learning[C]//Advances in Neural Information Processing Systems. Long Beach,America: NIPS, 2017: 4077-4087.

[11] Ravi S,LarochelleH. Optimization as a model for few-shot learning[C]//International Conference on Learning Representations (ICLR).Toulon, France: ICLR, 2017.

[12]张新峰, 沈兰荪. 加权SVM在中医舌象分类与识别中的应用研究[J]. 中国生物医学工程学报, 2006, 25(2):230-233. Zhang XF, Shen LS. Application of weighted SVM on the classification and recognition of tongue images[J]. Chinese Journal of Biomedical Engineering,2006, 25(2):230-233.

[13] 胡继礼, 阚红星. 基于卷积神经网络的舌象分类[J]. 安庆师范大学学报(自然科学版),2018,24(4):44-49. Hu JL, Kan HX.Tongue classification based on convolutional neural network[J]. Journal of Anqing Normal University(Natural Science Edition),2018, 24(4): 44-49.

[14] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, USA:IEEE Press,2016: 770-778.

[15]Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition. Boston, Massachusetts,USA: IEEE Press, 2015: 1-9.

[16]Lin, Min, Chen, Qiang, Yan, Shuicheng. Network In Network[J]. Computer Science, 2013.

[17]Amato G, Falchi F. kNN based image classification relying on local feature similarity[C]// Third International Workshop on Similarity Search and Applications. Istanbul, Turkey:SISAP, 2010: 101-108.

[18]Bottou L. Stochastic gradient learning in neural networks[J]. Proceedings of Neuro-N?mes, 1991,91(8): 12.

[19]Wold S, Esbensen KH, Geladi P. Principal component analysis[J]. Chemometrics and Intelligent Laboratory Systems, 1987, 2(1-3):37-52.

[20] 王琦.中医体质学[M].北京：中国医药科技出版社, 1995.

[21]Yosinski J, Clune J, Bengio Y, et al. How transferable are features in deep neural networks?[C]//Advances in Neural Information Processing Systems(NIPS). Montreal, Canada：NIPS,2014: 3320-3328.

[22]Schroff F, Kalenichenko D, Philbin J. Facenet: a unified embedding for face recognition and clustering[C]// IEEE Conference on Computer Vision and Pattern Recognition. Boston, MA, USA: IEEE Press, 2015: 815-823.

[23]LeCun Y, Boser BE, Denker J, et al. Handwritten digit recognition with a back-propagation network[C]//Advances in Neural Information Processing Systems. 1990: 396-404.

[24]杨晶东, 张朋. 基于迁移学习的全连接神经网络舌象分类方法[J]. 第二军医大学学报, 2018, 39(8): 897-902. Yang JD, Zhang P. Tongue image classification method based on transfer learning and fully connected neural network[J]. Academic Journal of Second Military Medical University, 2018, 39(8): 897-902.

[25] Parkhi OM, Vedaldi A, Zisserman A. Deep face recognition[C]//British Machine Vision Conference. Swansea, UK: BMVC, 2015.

[26] Taigman Y, Yang M, Ranzato MA, et al. DeepFace: closing the gap to human-level performance in face veri?cation[C]// Conference on Computer Vision and Pattern Recognition. Columbus, USA：IEEE Press, 2014: 1701–1708.

服务与反馈：

【文章下载】【加入收藏】

提示：您还未登录，请登录！点此登录