基于词典优化算法对舌象特征提取的研究
发布时间：2019-02-22 09:16
【摘要】：随着现代生活水平不断提高,医疗作为一项基本民生,需求越来越大。而中医是中华民族传统文化的精髓之一,在某些方面具有西医学不可取代的优势。因此,如何联系当代科学技术使中医诊断现代化具有重要研究意义。舌诊作为中医学望诊中的重要组成部分,舌诊的客观化不仅成为了国内学者的研究热点,并引起了国外数理医学界的关注。望舌诊断最为主要的判断依据就是舌色,不同颜色的舌象可以表现出人的体质差异与病症程度。现在已有一些学者用HSI色彩空间来对舌色做分量划分,但对于极相近的颜色特征边界划分不太明确。针对这个问题,文章主要做了以下工作:首先,为了更加准确地对舌色进行定量分析,本文基于稀疏表示和粒子滤波的思想提出了一种词典优化算法。该算法通过稀疏的线性表示建立了一个基础理论模型。分割图片时用取重叠块的方法建立词典集合,使选取的词典块能更大程度地包含特征信息。其次,论文处理特征信息时并未像现有的方法一样直接对颜色进行定性分析,而是基于蒙特卡洛粒子滤波的思想,以粒子作为特征信息,建立了一个KD树作为粒子的集合,通过粒子的相似度来确定舌象的颜色分类。在完善了程序设计的条件下,本课题选取了三种常见舌象进行了实验,验证了本算法的可行性和有效性。最后,由于妊娠期糖尿病的特殊性,本课题选定它作为案例进行分析。妊娠期糖尿病是一种初期病症,它的舌象以红舌和绛舌为主,在临床上并不容易辨别。依据论文已建立的词典优化模型对妊娠期糖尿病舌图像进行实验的过程中,因其舌色特征的难辨性,在KD树对匹配粒子查找时出现了工作量大、耗时长的问题。因此,本文在模型的求解上做了改进,引入加速近端梯度法,在模型的求解上做出了改进。该方法属于边界化重采样,可以在重要粒子区域进行二次收敛,淘汰权重度偏低的粒子,留下权重度较高的粒子,降低了计算的复杂度,同时也提高了特征提取的准确度。本课题所研究的舌色仅仅是中医四诊中望诊的一个小部分,随着计算机模式识别和图像处理技术的不断成熟,本课题采用的词典算法也需要不断的进行优化,从而进一步提高特征提取的精度和效率,这也将是我们以后研究学习的一个重要的目标。
[Abstract]:With the continuous improvement of modern living standards, medical care as a basic livelihood, the demand is growing. Traditional Chinese medicine is one of the quintessence of Chinese traditional culture and has irreplaceable advantages in some respects. Therefore, it is of great significance to study how to combine modern science and technology to modernize the diagnosis of traditional Chinese medicine. As an important part of traditional Chinese medicine (TCM) diagnosis, the objectification of tongue diagnosis has not only become the research hotspot of domestic scholars, but also attracted the attention of foreign medical circle. Tongue color is the most important basis of tongue diagnosis. Different tongue color can show the difference of physique and degree of disease. At present, some scholars have used HSI color space to divide tongue color components, but they are not clear about the very similar color feature boundaries. To solve this problem, the main work of this paper is as follows: firstly, in order to make quantitative analysis of tongue color more accurately, a dictionary optimization algorithm based on sparse representation and particle filter is proposed in this paper. The algorithm establishes a basic theoretical model by sparse linear representation. In the process of image segmentation, overlapping blocks are used to set up dictionary sets, so that the selected lexicon blocks can contain feature information to a greater extent. Secondly, when dealing with the feature information, the paper does not directly analyze the color qualitatively as the existing methods, but based on the Monte Carlo particle filter idea, taking the particle as the feature information, a KD tree is established as the set of particles. The color classification of tongue images is determined by the similarity of particles. Under the condition of perfect program design, three kinds of common tongue images are selected for experiments, and the feasibility and effectiveness of this algorithm are verified. Finally, due to the particularity of gestational diabetes mellitus, this topic selected it as a case study. Gestational diabetes mellitus (GDM) is an initial disease. Its tongue image is mainly red and crimson, which is not easy to distinguish clinically. During the experiment of tongue images of gestational diabetes mellitus (GDM) based on the dictionary optimization model established in this paper, due to the difficulty of distinguishing the tongue color features, it is difficult to find matching particles in the KD tree. Therefore, this paper improves the solution of the model, introduces the accelerated near-end gradient method, and improves the solution of the model. This method belongs to the boundary resampling and can converge twice in the important particle region, eliminate the particles with low weight degree, leave the particles with high weight degree, reduce the computational complexity and improve the accuracy of feature extraction. The tongue color studied in this paper is only a small part of the four diagnoses of TCM. With the development of computer pattern recognition and image processing technology, the dictionary algorithm used in this subject also needs to be optimized continuously. Thus, the accuracy and efficiency of feature extraction can be further improved, which will be an important goal for us to study in the future.
【学位授予单位】：长沙理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
[Abstract]:With the continuous improvement of modern living standards, medical care as a basic livelihood, the demand is growing. Traditional Chinese medicine is one of the quintessence of Chinese traditional culture and has irreplaceable advantages in some respects. Therefore, it is of great significance to study how to combine modern science and technology to modernize the diagnosis of traditional Chinese medicine. As an important part of traditional Chinese medicine (TCM) diagnosis, the objectification of tongue diagnosis has not only become the research hotspot of domestic scholars, but also attracted the attention of foreign medical circle. Tongue color is the most important basis of tongue diagnosis. Different tongue color can show the difference of physique and degree of disease. At present, some scholars have used HSI color space to divide tongue color components, but they are not clear about the very similar color feature boundaries. To solve this problem, the main work of this paper is as follows: firstly, in order to make quantitative analysis of tongue color more accurately, a dictionary optimization algorithm based on sparse representation and particle filter is proposed in this paper. The algorithm establishes a basic theoretical model by sparse linear representation. In the process of image segmentation, overlapping blocks are used to set up dictionary sets, so that the selected lexicon blocks can contain feature information to a greater extent. Secondly, when dealing with the feature information, the paper does not directly analyze the color qualitatively as the existing methods, but based on the Monte Carlo particle filter idea, taking the particle as the feature information, a KD tree is established as the set of particles. The color classification of tongue images is determined by the similarity of particles. Under the condition of perfect program design, three kinds of common tongue images are selected for experiments, and the feasibility and effectiveness of this algorithm are verified. Finally, due to the particularity of gestational diabetes mellitus, this topic selected it as a case study. Gestational diabetes mellitus (GDM) is an initial disease. Its tongue image is mainly red and crimson, which is not easy to distinguish clinically. During the experiment of tongue images of gestational diabetes mellitus (GDM) based on the dictionary optimization model established in this paper, due to the difficulty of distinguishing the tongue color features, it is difficult to find matching particles in the KD tree. Therefore, this paper improves the solution of the model, introduces the accelerated near-end gradient method, and improves the solution of the model. This method belongs to the boundary resampling and can converge twice in the important particle region, eliminate the particles with low weight degree, leave the particles with high weight degree, reduce the computational complexity and improve the accuracy of feature extraction. The tongue color studied in this paper is only a small part of the four diagnoses of TCM. With the development of computer pattern recognition and image processing technology, the dictionary algorithm used in this subject also needs to be optimized continuously. Thus, the accuracy and efficiency of feature extraction can be further improved, which will be an important goal for us to study in the future.
【学位授予单位】：长沙理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41
【参考文献】
相关期刊论文 前10条
1 张国军;郑丽华;孙锡红;;妊娠期糖尿病研究进展[J];河北医科大学学报;2015年07期
2 徐杰;许家佗;朱蕴华;陶枫;林sダ，
本文编号：2428070
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2428070.html
