基于颜色字典和Doublet特征优化的舌像分类方法研究

发布时间：2017-12-26 21:29

本文关键词：基于颜色字典和Doublet特征优化的舌像分类方法研究　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：自动化舌诊系统克服了传统的主观化和非量化等缺点,对舌诊具有重要意义和参考价值。但是舌诊系统往往受诸多因素影响,主要表现在两个方面,分别是特征提取方法和分类策略。针对以上两种问题,本文主要做了两方面工作:(1)在特征提取方面,本文提出了一种基于颜色字典特征提取方法和基于Doublet的特征优化方法;(2)在分类器方面，，采用了高效的GBDT分类器。基于颜色字典的特征提取方法包括颜色字典的定义和特征提取。分析了舌像在CIELab色彩空间的色域,并根据中医专家总结出的舌像主要颜色,定义了舌像的颜色字典;在特征提取过程中,通过对重叠分割方法得到的图像块与颜色字典之间进行相似性分析,提取分割块的类别直方图特征,并组合成为舌像特征。在特征提取过程中,由于采用了重叠分割法,保持了图像信息的完整性;并且提取的是图像局部特征,使得图像在缩放时能保持特征不变性。Doublet特征优化方法对基于颜色字典的特征进行了优化。首先利用舌像之间的相似性构建Doublet,定义了新的样本特征和类别;然后利用多项式核函数模型为Doublet定义新的核函数,并利用分类器训练得到关于样本Doublet的模型,由该模型重新构建原有未经过Doublet处理的样本,从而实现对原始样本特征的优化。在分类策略上,选择梯度提升决策树(Gradient Boosting Decision Tree,GBDT)分类器。GBDT是一种基于多个弱分类器组合的分类器,利用残差作为损失函数。其中,弱分类器采用的是分类回归树。为了证明方法的有效性,在舌像数据集上进行了验证。实验结果表明,基于颜色字典特征提取方法提取到的特征具有较强的鲁棒性,Doublet多核特征优化方法减少了特征的噪声干扰,而GBDT相比于支持向量机等分类器对舌像分类时具有较高的有较高的分类正确率和特异性。
[Abstract]:The automatic tongue diagnosis system overcomes the shortcomings of the traditional subjective and non quantified, which is of great significance and reference value for the diagnosis of the tongue. However, the tongue diagnosis system is often affected by many factors, mainly in two aspects, the feature extraction method and the classification strategy. In view of the above two problems, this paper mainly does two aspects: (1) in aspect of feature extraction, this paper proposes a feature extraction method based on color dictionary and Doublet based feature optimization. (2) in the aspect of classifier, an efficient GBDT classifier is adopted. The method of feature extraction based on color dictionary includes the definition and feature extraction of color dictionary. Analysis of tongue image in CIELab color space and color gamut, according to Chinese experts summed up the main tongue color, tongue color is defined in the dictionary; feature extraction process, similarity analysis between image blocks to get through the overlapping segmentation and color dictionary extraction category histogram segmentation block. And as the tongue image feature. In the process of feature extraction, the overlapped segmentation method is adopted to preserve the integrity of image information, and extract the local features of the image, so that the feature invariance can be maintained when the image is zoomed. The Doublet feature optimization method optimizes the feature based on the color dictionary. Firstly, the similarity between the construction of Doublet tongue, defines the features and categories of new samples; and then define the new kernel function for Doublet with polynomial kernel function model, and the use of a classifier training samples of Doublet model, the model constructs the original sample without Doublet treatment, so as to realize the optimization of the original sample the characteristics of the. In the classification strategy, the Gradient Boosting Decision Tree (GBDT) classifier is selected. GBDT is a classifier based on multiple weak classifier combination, which uses the residual as a loss function. Among them, the weak classifier uses the classified regression tree. In order to prove the validity of the method, it is verified on the data set of the tongue. The experimental results show that the characteristics of the color feature extraction method to extract the dictionary based on robust, Doublet kernel feature optimization method to reduce the noise interference characteristics, and GBDT compared to the support vector machine classifier for classification of tongue image has higher classification accuracy and high specificity.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】