CN1162798C - Analysis method of tongue color, fur color and tongue thickness in traditional Chinese medicine based on multi-class support vector machine - Google Patents

Abstract

A TCM tongue color, fur color, and tongue thickness analysis method based on multi-class support vector machines is applied in the field of computer medical image processing. The method is to carry out image acquisition by a digital camera, and input the image into a computer processor through a USB interface, and segment the tongue body area from the tongue image in the processor, and it is characterized in that the method also includes the following in sequence Steps: On the basis of performing hierarchical clustering analysis to generate and adjust the clustering tree, the clustering tree-based support vector machine (CTSVM) method is used to identify the pixels of the tongue area, and the number of categories for tongue area pixel identification is set to 15 types; divide the tongue surface into five areas, count the pixel categories with the largest number in each area, and the number of pixels of each type; describe the tongue color; describe the color of the tongue coating; conduct quantitative analysis and description of the thickness of the tongue coating; finally use Text and pseudo-color plots show analysis results. The method is fast, accurate and practical, and conforms to the diagnostic principles of Chinese medicine experts.

Description

Analysis method of tongue color, fur color and tongue thickness in traditional Chinese medicine based on multi-class support vector machine

technical field

The present invention relates to the field of computer medical image processing, and designs a TCM tongue image analysis system based on multi-class support vector machines, which performs quantitative analysis and qualitative description of tongue color, fur color, tongue coating thickness, and assists TCM diagnosis .

Background technique

Tongue examination is an important part of inspection in the four examinations of traditional Chinese medicine. Tongue diagnosis is to rely on the doctor to observe the tongue, so as to make a judgment. This method is highly subjective and ambiguous, and the diagnosis results are closely related to the doctor's experience. At present, people strive to conduct objective and quantitative analysis of tongue image features, overcome the subjectivity of traditional visual observation, and improve the reliability of tongue diagnosis, which has great advantages.

Using computer image processing technology to analyze the characteristics of the tongue image can reflect the situation of the tongue image more comprehensively, conforms to the habit of tongue diagnosis in traditional Chinese medicine, facilitates the collection and preservation of tongue diagnosis data, and has good practical value. It collects the color tongue image of the subject under certain lighting conditions, digitizes it and inputs it into the computer, uses image analysis technology to automatically analyze the characteristics of the tongue image, and stores the result in the computer.

When existing image processing techniques are used for tongue image analysis, most of them only use simple statistical analysis of tongue image color values. After analysis, we found that the distribution of tongue color, fur color, and thickness is localized, and different areas on the tongue surface often show different characteristics. In TCM tongue diagnosis, the local area is first observed, and then the overall description is made. The description mainly emphasizes the features with dialectical significance.

In the field of image processing or analysis, due to the different conditions of image capture, the parameters in the processing or analysis method may be adjusted according to the actual situation. The method of adjustment is to take one or several images from the images taken under the same conditions, and determine the specific values of the parameters through experiments. Each parameter in this method is basically determined according to this method.

The support vector machine (SVM) method in the existing pattern recognition technology is a relatively effective method to solve the problem of small sample recognition. The basic idea can be illustrated by the two-dimensional situation in Figure 3. In Figure 3, solid points and hollow points represent two types of samples, H is the classification line, H ₁ and H ₂ are the straight lines parallel to the classification line that pass through the samples closest to the classification line in each category, and the distance between them is called Classification interval (margin). The so-called optimal classification line requires that the classification line can not only separate the two classes correctly (the training error rate is 0), but also maximize the classification interval.

It can be seen from Figure 3 that the classification interval in (b) is large, and the possibility of misclassification is smaller than that in (a), and the samples located on the optimal classification line are called support vectors. The above is an example of linear separability . For the nonlinear case, it is converted to a high-dimensional feature space (Feature Space) through nonlinear transformation, and a linear discriminant function is constructed in the high-dimensional space to realize the nonlinear discriminant function in the original space.

Tongue image samples include collected digital tongue images and clinical tongue diagnosis conclusions given by tongue diagnosis experts. The larger the number of samples, the more reliable the clinical judgment, the more realistic the model established on this basis, and the higher the accuracy of feature analysis for new tongue images.

However, some types of tongue appearance are very rare (such as gray and black coating, pale tongue, cyanotic tongue, etc.). Therefore, the distribution of various samples is uneven and the number is limited. If the general pattern recognition method is used, it will inevitably affect the accuracy and reliability of the automatic analysis of tongue image features. Therefore, the support vector machine (SVM) pattern recognition method is used for pixel recognition to improve the classification accuracy and reduce the classification error rate.

个识别结果，其中最占优势的类别属性为最终识别结果。这两种方法的缺点是训练后得到的支持向量数较多，影响应用时的识别速度。The classic SVM is generally aimed at two types of recognition problems, while tongue region pixel recognition is a multi-type problem. When used in multi-class pattern recognition, the original problem is usually decomposed into multiple two-class recognition problems. Existing methods include 1-aa (one-against-all) and 1-a-1 (one-against-one). Among them, 1-aa is composed of k two-class SVMs (f ¹ ,...,f ^k ), f ^j sets the category label of the j-th sample as +1, and sets the category labels of all other samples as -1, In 1-a-1, each type of sample and other types constitute two types of recognition problems, and a total of A two-class SVM. When applied, the samples go through all two types of SVMs, and get

recognition results, and the most dominant category attribute is the final recognition result. The disadvantage of these two methods is that the number of support vectors obtained after training is large, which affects the recognition speed in application.

Contents of the invention

In order to solve the foregoing problems, the features such as tongue color, fur color, and thickness of tongue fur can be analyzed automatically, accurately, and quickly. The present invention designs a new multi-category method—support vector machine based on clustering tree ( CTSVM) method. This method adopts the "two-step method" analysis step of first performing the pixel recognition of the tongue area (local analysis), and then performing the overall description. In the pixel recognition, combined with the support vector machine can effectively complete the small sample recognition problem, and the clustering tree adopts the classification method to convert multi-class problems into two-class problems, so it is fast and accurate to achieve practicality. The overall description is based on the main pixel categories, which is in line with the diagnostic principles of Chinese medicine experts. At the same time, the coating thickness index is introduced to quantify the thickness of tongue coating. Using this method, the qualitative description and quantitative analysis of tongue color, coating color and tongue thickness can be obtained by computer. Technical idea of the present invention is characterized in that:

1. When the image processing method is used to analyze the characteristics of the tongue image, it is different from the general method, but adopts the "two-step method" analysis of the tongue body area pixel recognition (local analysis) first, and then the overall description step. Because in TCM tongue diagnosis, the local part is observed first, and then the overall description is made. Therefore, this method is consistent with the diagnostic habits of traditional Chinese medicine, and can effectively identify the characteristics of the tongue image.

2. We have adopted the analysis steps of performing the pixel recognition (local analysis) of the tongue body area first, and then carrying out the overall description. That is, local analysis is performed on the characteristics of the tongue quality and tongue coating. The pixels on the tongue are divided into different categories according to the color attribute, using a multi-class support vector machine. We set the number of categories for pixel recognition in the tongue body area to 15, including pale tongue, light red tongue, red tongue, dark red tongue, crimson tongue, dark purple tongue, thin white fur, white fur, and white thick fur , Thin yellow fur, yellow fur, yellow thick fur, gray fur, brown fur, black fur, etc. Among them, the first six are tongue types, and the last nine are tongue coating types. For pixel recognition, it is generally necessary to give pixels of known categories first, which are called samples. If the samples are relatively small, such recognition problems are called small-shot problems.

Aiming at the multi-class and small-sample classification problem of pixel recognition, a new clustering tree-based multi-class support vector machine (CTSVM) method is adopted. The main idea of CTSVM is to first perform hierarchical clustering analysis on various centers of the training sample set, generate a multi-level classifier called a clustering tree, and then form (k-1) two-class SVMs accordingly. It can be used to convert a complex multi-category classification problem into several simple classification problems to solve. Each non-leaf node of the clustering tree represents a support vector classifier. The structure of the clustering tree can be determined according to manual experience, and can also be obtained by using hierarchical clustering methods. The so-called clustering is to gather samples with large similarities into one type according to the principle of "like attracts like, and people are divided into groups". This method can effectively transform the multi-class recognition problem into a two-class classification problem with fewer numbers and faster recognition speed, and can overcome the influence of fewer samples on the recognition rate in tongue image analysis, so that it can be more accurate. Pixel recognition, and a clustering tree is obtained by hierarchical clustering combined with the experience of TCM experts, see Figure 4.

The form of the clustering tree is adaptively determined according to the distribution of the training sample set and the nearest distance criterion. Combining this clustering tree with two types of pattern recognition SVMs constitutes a CTSVM. In CTSVM, each non-leaf node (root node and intermediate node) is a two-class SVM.

According to the above method, we combined the diagnosis experience of Chinese medicine experts to determine the clustering tree shown in Figure 4. In the figure, H~M are tongue types, and N~V are tongue coating types. In the figure, X is called the root node, and A~G are called non-leaf nodes. The upper layer node connected to a node is called the parent node, and correspondingly, the node is called the child node of its parent node. In Figure 3, A has two sub-nodes C and D, and C has three sub-nodes H, I, and J. When identifying pixels, first divide the two categories with the largest distance, and then carry out finer divisions step by step. That is, start from the root node, use the classifier corresponding to the node to classify, and attribute it to a certain child node until it reaches a certain leaf node, that is, the final recognition result of the pixel is obtained. In the present invention, the classifier corresponding to each non-leaf node adopts a support vector machine. If a non-leaf node has two child nodes, two types of support vector machines can be used directly; A class SVM separates the classes represented by one of the child nodes from the classes represented by the other two.

3. Divide the tongue surface into five areas, and count the most numerous pixel categories in each area. The description of the tongue color and fur color in this area is mainly based on this category, combined with the experience of TCM experts. This is in line with the principle of emphasizing dialectical features when describing tongue diagnosis in TCM.

4. When classifying the thickness of the tongue coating, each pixel of the tongue body area is divided into four types: "no coating", "thin coating", "medium thick coating" and "thick coating". First, the thickness of the tongue coating is classified pixel by pixel, based on the fact that "the thickness of the tongue coating is based on the standard of bottoming out and not bottoming out". Then quantify the thickness of the tongue coating. On the basis of pixel-by-pixel recognition, the coating thickness index is introduced to quantitatively describe the thickness of the tongue coating in the entire tongue image.

5. In order to make the analysis results easy to understand and clinical application, text and pseudo-color graphics are used to display the analysis results. The pixel recognition results are expressed as tongue color and fur color distribution pseudo-color map. The pixel-by-pixel thickness classification results are expressed as a pseudo-color distribution map of tongue coating thickness. The tongue color and coating thickness were qualitatively described in words, and the tongue coating thickness index was given.

Technical scheme of the present invention sees Fig. 1, Fig. 2, Fig. 5, Fig. 6, Fig. 8, Fig. 9, Fig. 10, method step is to carry out image collection by digital camera, and image is input to computer processor by USB interface In the method, the general technique is adopted in the processor to segment the tongue body region from the tongue image, and it is characterized in that the method also includes the following steps in sequence:

1) On the basis of hierarchical clustering analysis to generate and adjust the clustering tree, the clustering tree-based support vector machine (CTSVM) method is used to identify the pixels of the tongue area;

The method steps of hierarchical cluster analysis to generate cluster tree are shown in Fig. 5:

(1) For k class samples, k class center sets are calculated: ( x ₁ , x ₂ ,…, x _k ),

(2) The first level of division is that each class center is divided into one class, that is, k leaf nodes,

(3) Find the distance D _ij between every two class centers, i=1,...k, j=1,...k, i≠j,

(4) Merge the two nearest centers to form a new class center, that is, the middle node,

(5) Repeat steps (3)-(4) until the k-1th level, all centers are merged into 1 category, namely the root node,

(6) According to the needs, adjust the structure of the decision tree so that the depth of the tree is as small as possible; the method and steps of tongue body area pixel recognition are shown in Figure 6:

(1) Calculate the mean value of the three-color chromaticity values R, G, and B in the 3×3 neighborhood of each pixel in the input image, as the pixel is another input feature,

(2) Read the R, G, and B values of a pixel in the tongue,

(3) Starting from the root node, according to the SVM classifier of the node, divide the pixel into a certain branch,

(4) If the branch node is not a leaf node, repeat the process in (3): divide the pixel into a certain branch according to the SVM classifier of the node, if the branch node is a leaf node, then the leaf node represents The class of is the category of the identified pixels, the classifier of each node is a two-class support vector machine classifier, and the classification process is implemented according to the conventional method.

(5) Repeat the process of (2)-(4) until all the pixels in the tongue body area are classified.

It can be seen that when identifying pixels, first divide the two categories with the largest distance, and then carry out finer divisions step by step until the leaf nodes, and then the category attributes of the pixels are obtained. Due to reducing the number of two types of discrimination and taking into account the distribution of samples, this method can accelerate the recognition speed and ensure the accuracy of recognition.

2) Count the number of pixels of various types in each partition;

3) To describe the tongue color, see Figure 8;

4) Carry out the moss color description, see Figure 9;

5) Carry out the quantitative analysis and description of tongue coating thickness, see Figure 10;

6) The analysis results are displayed in text and pseudo-color graphs. In order to make the analysis results easy to understand and clinical application, text and pseudo-color graphics are used to display the analysis results. The pixel recognition results are expressed as tongue color and fur color distribution pseudo-color map. The pixel-by-pixel thickness classification results are expressed as a pseudo-color distribution map of tongue coating thickness. The tongue color and coating thickness were qualitatively described in words, and the tongue coating thickness index was given. Figure 11 is an example showing the analysis results.

In addition, the description features of tongue color, fur color, and tongue thickness are that after the pixel recognition is completed, after counting the number of pixels of various types in each partition, let N _i ^j represent the jth class image in the i area, where i=1, …, 5 represent the five areas of tongue root, tongue middle, tongue tip, tongue left, tongue right, etc. j=1, …, 15 represent pale tongue, light red tongue, red tongue, dark red tongue, There are 15 types of tongue texture and tongue coating, such as crimson tongue, dark purple tongue, thin white coating, white coating, white thick coating, thin yellow coating, yellow coating, yellow thick coating, gray coating, brown coating, and black coating. According to the tongue diagnosis habit, describe the tongue color, fur color, and thickness of tongue fur.

The description method of the tongue color is to describe the color of the tongue according to the tongue type of the pixels in the tongue side area and the tongue tip area. Since the tongue quality is mainly distributed on the side and tip of the tongue, only the tongue tip, the left side of the tongue, and the right side of the tongue are considered to consider the type of tongue quality. Calculate the tongue tip, the left side of the tongue, and the right side of the tongue, which belong to each tongue type (that is, belong to the tongue type, which belongs to light tongue, light red tongue, red tongue, dark red tongue, crimson tongue, dark purple tongue) type), the tongue type with the largest number of pixels is the tongue feature of the tongue image. According to the experience of traditional Chinese medicine, it is very important whether the tongue is partially dark purple and whether the tip of the tongue is red, so these two conditions are treated separately. The method is: calculate the ratio of the number of pixels belonging to the "dark purple tongue" to the total number of pixels on the entire tongue in the three regions of the tongue tip, left side of the tongue, and right side of the tongue. If this ratio is greater than a certain preset value, it is considered " Partial dark purple"; Calculate the ratio of the number of pixels belonging to "tongue red" in the tongue tip area to the total number of pixels on the entire tongue surface. If this ratio is greater than a preset value, "tongue red" is considered. The specific process is:

1) Calculate the total area (the total number of pixels) A of the tongue surface;

2) Calculate the total number of the 6 tongue types in the tongue side and tongue tip area respectively

${\mathrm{<span\; class="notranslate">\; SUM</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 5</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; ;</span>$

3) order ${\mathrm{SUM}}^{j_{\max }}=\max ({\mathrm{SUM}}^j,j=1,\&CenterDot;\&CenterDot;\&CenterDot;,6),$ Then the tongue type represented by j _max is the tongue feature of the tongue image;

4) Calculate the area ratio of the dark purple tongue $R_{}$${}_6=\frac{{\mathrm{SUM}}^6}{A},$ If R ₆ >θ _p and j _max ≠6, add the description "local dark purple", where θ _p is the threshold value set according to the experiment;

5) Calculate the area ratio of "tongue red" at the tip of the tongue $R_{}$${}_3=\frac{{\mathrm{<figure-callout\; id="3"\; label="N"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">N</figure-callout>}}_3^3}{A},$ If R ₃ >θ _r and j _max ≠3, add the description "red tongue", where θ _r is the threshold value set according to the experiment.

Among them, the description of the color of the tongue coating is divided into three areas: the root of the tongue, the middle of the tongue, and the tip of the tongue, because there is very little tongue coating on the side of the tongue. Calculate the total number of pixels belonging to each tongue coating type (thin white coating, white coating, white thick coating, thin yellow coating, yellow coating, yellow thick coating, gray coating, brown coating, black coating, etc.) in the three areas of tongue base, tongue center, and tongue tip. In each region, determine whether the ratio of the number of pixels of the tongue coating type with the largest number of pixels to the total number of pixels on the entire tongue surface is less than a certain threshold, if less, then it is determined that the area has little coating, otherwise the type of tongue coating is the tongue coating feature of the area . These three thresholds are obtained by experiments in the present invention. According to the experience of traditional Chinese medicine, the base of the tongue and the center of the tongue often present a variety of tongue coatings, so the type of tongue coating with more pixels in the base of the tongue and the center of the tongue is also judged. The method is: calculate the ratio of the number of pixels of the fur type with the largest number of pixels in the base of the tongue and the center of the tongue to the number of pixels of the fur type with the largest number of pixels. The main types of tongue coating. The specific method is as follows:

1) Calculate the total area (the total number of pixels) A of the tongue surface, the area A _i of the tongue root, tongue center and tongue tip area, i=1,...,3;

2) For the base of the tongue, the middle of the tongue, and the tip of the tongue, calculate separately $N_i^{j_{\max }^i}=\max (N_i^j,j=7,\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;,15),$ where i = 1, 2, 3,

Then the type of tongue coating represented by j _max ⁱ is the tongue coating feature of this area;

3) Calculate the area ratio of j _max ⁱ tongue coating pixels in areas i=1, 2, and 3 $R(j_{\max }^i,i)=\frac{N_i^{j_{\max }^i}}{A_i};$

4) Describe the tongue root, tongue middle and tongue tip respectively. For tongue base area i=1, if $R({\mathrm{<figure-callout\; id="1"\; label="j\; max"\; filenames="C0210379500221.png"\; state="\{\{state\}\}">j</figure-callout>}}_{\max }^{},1)<{\mathrm{\&theta;}}_1$ but

Described as "little coating on the base of the tongue"; otherwise, describe the base of the tongue as j _max ¹ type of tongue coating. For the middle tongue region i=2, if $R({\mathrm{<figure-callout\; id="2"\; label="j\; max"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">j</figure-callout>}}_{\max }^{},2)<{\mathrm{\&theta;}}_2,$ It is described as "little coating on the tongue"; otherwise, the tongue coating is described as j _max ² type. For the tongue tip area i=3, if $R({\mathrm{<figure-callout\; id="3"\; label="j\; max"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">j</figure-callout>}}_{\max }^{},3)<{\mathrm{\&theta;}}_3,$ It is described as "little coating on the tip of the tongue"; otherwise, the tip of the tongue is described as j _max ³ type of tongue coating. Among them, θ ₁ , θ ₂ , and θ ₃ are thresholds set according to experiments;

5) For the base of the tongue and the center of the tongue, calculate the second main type of tongue coating $j_{\sec }^i,i=\mathrm{1,2},$ And calculate the area ratio of j _sec ⁱ type pixel area to j _max ⁱ type $R_{\mathrm{ms}}^i=\frac{N_i^{j_{\sec }^i}}{N_i^{j_{\max }^i}}.$ like $R_{\mathrm{ms}}^i>{\mathrm{\&theta;}}_{\sec },$ Then the i-th area is added to describe the second main type of tongue coating. Where θ _sec is the threshold set according to the experiment.

The analysis and description of the thickness of the tongue coating is divided into two parts: (1) The classification of the thickness of the tongue coating. Each pixel in the tongue body area is divided into four types: "no coating", "thin coating", "medium thick coating" and "thick coating". The basis of pixel-by-pixel tongue coating thickness classification is "the thickness of the tongue coating is based on the standard of not seeing the bottom". (2) Quantification of tongue coating thickness. On the basis of pixel-by-pixel recognition, the tongue coating thickness of the whole tongue image is quantitatively described. Among them, "seeing the bottom but not seeing the bottom" is judged by the number of tongue pixels in the pixel neighborhood. If there are many tongue pixels in the pixel neighborhood, the degree of bottoming out is large. The tongue coating type of the pixel also reflects the thickness of the tongue coating. Belonging to thin white fur and thin yellow fur indicates that the pixel is thin fur; belonging to white fur, yellow fur, gray fur, and brown fur indicates that the pixel is medium thick fur; belonging to white thick fur, yellow thick fur, and black fur indicates that the pixel is thick fur. The specific steps are:

1) Read the category number of the tongue body pixel and its 5×5 neighborhood;

2) Let c ₁ and c ₂ be the threshold value c ₁ >c ₂ set according to the experiment, and. Calculate the number of pixels S _body belonging to the tongue type in the 5×5 neighborhood of the pixel. If S _body > c ₁ , the tongue coating thickness type at this point is defined as "no coating";

3) If c ₁ ≥ S _body > c ₂ , then the type of tongue coating thickness at this point is defined as "thin coating",

4) If S _body ≤ c ₂ , then determine the type of thickness according to the type j of the tongue quality and coating at this point. If it is the tongue quality (j<7), it is "thin coating". Thin yellow fur), it is "thin fur", if j=8 (white fur), 11 (yellow fur), 13 (gray fur), 14 (brown fur), it is "medium thick fur"; if j=9 ( White thick fur), 12 (yellow thick fur), 15 (black fur), it is "thick fur";

5) Repeat steps (1)-(3) until all pixels are processed;

6) According to the result of pixel recognition, calculate the coating thickness index T of the whole image, as the quantitative result of tongue coating thickness analysis, the calculation method is:

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}$

表示对舌体区域中所有象素求和，j^k表示第k个象素的厚度类型，对应于“无苔”、“薄苔”、“中厚苔”、“厚苔”分别取0、1、2、3，w_jk为权值，根据实验和中医专家的经验确定，苔厚指数能够描述舌苔的总体厚度。In the formula, A represents the total number of pixels in the tongue body area,

Indicates the summation of all pixels in the tongue body area, j ^k represents the thickness type of the kth pixel, corresponding to "no fur", "thin fur", "medium thick fur", and "thick fur" respectively take 0, 1, 2, 3, w _jk are weights, determined according to experiments and experience of Chinese medicine experts, the coating thickness index can describe the overall thickness of the tongue coating.

It can be seen from Figure 11 that this method can obtain quantitative analysis and qualitative description results obtained by computer, and the results are consistent with the visual inspection results of traditional Chinese medicine, which is conducive to the objectification and standardization of tongue diagnosis.

Description of drawings

Figure 1 is a block diagram of the TCM tongue color, fur color, and fur thickness analysis system

1. Digital camera, 2. USB interface, 3. Computer processor, 4. Output cache, 5. Tongue image analysis, 6. Display, 7. Analysis result;

Fig. 2 is a flow chart of the main program of the tongue image analysis method in traditional Chinese medicine;

Figure 3 Schematic illustration of support vector machine principle

(a) classification surface with small classification interval, (b) optimal classification surface with largest classification interval;

Figure 4. Schematic diagram of clustering tree generated for hierarchical clustering analysis

H, dark red tongue, I, light red tongue, J, pale tongue, K, purple tongue, L, dark purple tongue, M, red tongue, N, white coating, O, white thick coating, P, Thin white coating, Q, yellow coating, R, thin yellow coating, S, yellow thick coating, T, brown coating, U, gray coating, V, black coating;

Fig. 5 is the subroutine flowchart of the generating method of Fig. 4 clustering tree;

Fig. 6 is a subroutine flow chart of the tongue region pixel recognition method;

Fig. 7 is a schematic diagram of tongue surface partition;

Fig. 8 is a subroutine flowchart of tongue color description method;

Fig. 9 is a subroutine flow chart of the moss color description method;

Fig. 10 is a subroutine flow chart of tongue coating thickness analysis and description method;

Figure 11 is the analysis result of the tongue image

(a) tongue body image, (b) pseudo-color map of tongue color and fur color, (c) pseudo-color map of tongue coating thickness, (d) automatic analysis results of text description;

Fig. 12 is the program main flow chart of the tongue image analysis that runs on the computer;

Fig. 13 is a flow chart of the tongue region pixel recognition subroutine running on the computer.

Detailed ways

In the block diagram of the TCM tongue image analysis system in Fig. 1, both digital cameras and USB interfaces are commercially available, mainly to complete the collection of tongue images, and convert the optical signals of the tongue body and color code into image electrical signals and input them to the computer, which is convenient for the computer Processing, transmission and other operations; computer processing is mainly to read/write the tongue image through the USB interface software, and the processed tongue image is output to the buffer for easy display. The monitor is an image output device, through which the human eye can view the original tongue image and analysis results. Tongue body analysis is to carry out quantitative analysis and qualitative description of tongue color, fur color and fur thickness on the tongue image read by computer, and output the analysis results.

The original tongue image can be an image collected in real time by a digital camera, or it can be an image collected by a digital camera and stored in a computer hard disk.

Using support vector machine for pixel recognition, the key issues involved include selection of input feature space, composition of learning set, selection of kernel function, selection of penalty factor, etc. A learning set refers to a collection of samples whose categories are known.

In this system, the input feature space is RGB color space. To avoid the influence of noise, for each pixel, the RGB mean value of its 3×3 neighborhood is taken as the feature vector. A series of image sub-blocks are selected from several typical images, and the learning set is formed after the categories are determined block by block by experts in traditional Chinese medicine. Samples (image sub-blocks) are determined according to the specific conditions of the image, and there is no uniform size.

The penalty factor C of the SVM in the present invention reflects the degree of trust in the learning set. The larger C is, the higher the degree of trust is, and the larger C is, the higher the degree of trust is. In the research of tongue pixel recognition, due to the large number of categories and poor separability between categories, it is necessary to select an appropriate penalty factor through experiments. In the present invention, C=400 is selected through experiments.

Different kernel functions correspond to learning machines with different types of nonlinear decision surfaces in the input space. We experimented with commonly used polynomial, radial basis function (RBF), neural network and other kernel function forms. The final applied kernel function is RBF, which has the form:

$\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}$

In the formula, σ is an important parameter, which determines the specific form of the RBF kernel function, and σ=50 is selected through experiments.

In the present invention, in order to make the analysis results easy to understand and clinically applied, qualitative and quantitative descriptions of the tongue quality and tongue coating are performed. The tongue surface is divided into 5 regions: tongue root, tongue middle, tongue tip, tongue left side, and tongue right side. The method of division is to divide the tip of the tongue to the midpoint of the herringbone-shaped boundary groove into 5 equal parts. In addition, a line is drawn between the midline of the tongue and the midpoint of the edge of the tongue, and the part outside the line is called the side of the tongue. As shown in Figure 7. Count the number of pixels of various types in each partition, and according to the tongue diagnosis habits of TCM experts, carry out a computer analysis method based on multi-class support vector machine for TCM tongue color, fur color, and tongue thickness.

Tongue image analysis is mainly realized through software. Complete the following programs in the computer (the main program is shown in Figure 12):

1. Read in the image data of the tongue area, read in the support vectors and corresponding coefficients of CTSVM, and initialize the parameters θ _p , θ _r , θ ₁ , θ ₂ , θ ₃ , θ _sec , c ₁ , c ₂ , w ₀ , w ₁ , w ₂ , w ₃ . Among them, θ _p and θ _r are the thresholds for judging "partial dark purple" and "red tip of the tongue" respectively in tongue color analysis. In the present invention, θ _p =0.01 and θ _r =0.06. The threshold values of θ ₁ , θ ₂ , and θ ₃ tongue base, tongue center, and tongue tip are taken as θ ₁ =0.5, θ ₂ =0.2, and θ ₃ =0.4 in the present invention. c ₁ and c ₂ are the thresholds used for judging the presence or absence of tongue coating and the type of tongue coating in the coating thickness analysis. In the present invention, c ₁ =20 and c ₂ =8. w ₀ , w ₁ , w ₂ , and w ₃ are the weight coefficients when calculating the moss thickness index, which are determined as w ₀ =0, w ₁ =0.2, w ₂ =0.7, and w ₃ =1.3 in the present invention.

2. Enter the subroutine of pixel identification, and adopt the clustering tree-based support vector machine (CTSVM) method to identify the pixels of the tongue area. Here it is assumed that the clustering tree shown in FIG. 4 has been obtained through training. Each node in the tree is a support vector machine classifier, and the support vectors and corresponding coefficients have been obtained through training. For a pixel, the classification process of the support vector machine is as follows:

1) Calculate the RGB mean x=(R, G, B) in the 3×3 neighborhood of the pixel;

2) Starting from the root node, calculate the kernel function inner product of the pixel and each support vector corresponding to the node, multiply the corresponding coefficients and labels, sum all these products, and then calculate the result after adding the constant b symbol. That is to say:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; sign</span>}<span\; class="notranslate">\; (</span>\underset{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; sv</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; \&alpha;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span>$

In the formula, x _i is the support vector of the classifier, α _i is the coefficient corresponding to the support vector, y _i is the label of the support vector sample, b is a constant, and b is 40 in the present invention, K(x _i , x) is the inner product of the kernel function,

$\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}$

where ‖.‖ is the norm of the vector, and σ is 50.

Sign() is a sign function. If the value of the argument is greater than 0, the function value is 1, and if it is less than 0, it is -1.

If f(x) is greater than 0, it is classified into the left branch, otherwise it is classified into the right branch;

3) Repeat the process in 2) until reaching a certain branch node. For example, the recognition process of a certain pixel is: X→A→D→K, then the pixel is judged as "red tongue";

4) For each pixel, perform 2) and 3) until all pixels in the tongue body area are identified, and each pixel is assigned a category number.

3. Tongue surface division. Scan the tongue area from left to right and from top to bottom to obtain the circumscribed rectangle of the tongue area. Suppose the coordinates of the left and right sides of the rectangle in the tongue image are l, r, and the coordinates of the upper and lower sides in the tongue image are t, b, then according to the aforementioned principles of tongue surface partitioning (the tongue surface is divided into tongue root, Tongue center, tongue tip, tongue left side, and tongue right side are divided into 5 equal parts from the tongue tip to the midpoint of the herringbone boundary groove. The front 1/5 is called the tongue tip, and the middle 2/5 is called the tongue center. The rear 2/5 is called the root of the tongue. In addition, draw a line with the midpoint of the tongue midline and the tongue edge, and the part outside the line is called the tongue side.), obtain the following partition results: Let (x, y) represent the vertical and vertical dimensions of the pixel in the tongue image Abscissa, the left side of the tongue: if $x<\frac{3l+\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4},$ And (x, y) is in the tongue body area, on the right side of the tongue: if $x\&Greater\; Equal;\frac{l+3\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4},$ And (x, y) is in the tongue body area, tongue root area: if $\frac{3l+\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4}\&le;x<\frac{l+34}{4},$ $\mathrm{the\; y}<\frac{3t+2\mathrm{<figure-callout\; id="4"\; label="b"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">b</figure-callout>}}{4},$ And (x, y) is in the tongue body area, the middle area of the tongue: if $\frac{3l+\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4}\&le;x<\frac{l+3\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4},$ $\frac{3t+2\mathrm{<figure-callout\; id="4"\; label="b"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">b</figure-callout>}}{4}\&le;\mathrm{the\; y}<\frac{t+4\mathrm{<figure-callout\; id="4"\; label="b"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">b</figure-callout>}}{4}$ Tongue tip area: if $\frac{3l+\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4}\&le;x<\frac{l+3\mathrm{<figure-callout\; id="4"\; label="r"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">r</figure-callout>}}{4},$ $\mathrm{the\; y}\&Greater\; Equal;\frac{t+4\mathrm{<figure-callout\; id="4"\; label="b"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">b</figure-callout>}}{4},$ And (x, y) is in the tongue area.

4. Count the number of pixels N _i ^j of various types in each partition.

5. Tongue color analysis and description. Describe the tongue quality according to the tongue quality on the side and tip of the tongue. The specific method is as follows:

1) Calculate the total area (the total number of pixels) A of the tongue surface;

2) Calculate the total number of the 6 tongue types in the tongue side and tongue tip area respectively:

${\mathrm{<span\; class="notranslate">\; SUM</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; N</span>}}_{\mathrm{<span\; class="notranslate">\; 5</span>}}^{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; ;</span>$

3) order ${\mathrm{SUM}}^{j_{\max }}=\max ({\mathrm{SUM}}^j,j=1,\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;,6),$ Then the tongue type represented by j _max is the tongue feature of the tongue image, for example, when j _max = 2, it is described as "tongue light red";

4) Calculate the area ratio of the dark purple tongue $R_{}$${}_6=\frac{{\mathrm{SUM}}^6}{A},$ If R ₆ >0.01 and j _max ≠6, add the description "local dark purple";

5) Calculate the area ratio of "tongue red" at the tip of the tongue $R_{}$${}_3=\frac{{\mathrm{<figure-callout\; id="3"\; label="N"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">N</figure-callout>}}_3^3}{A},$ If R ₃ >0.05 and j _max ≠3, add the description "red tongue".

6. Analysis and description of moss color. The description of the tongue coating is divided into three areas: the root of the tongue, the middle of the tongue, and the tip of the tongue. The specific steps are as follows:

1) Calculate the total area (the total number of pixels) A of the tongue surface, the area A _i of the tongue root, tongue center and tongue tip area, i=1,...,3;

2) For the base of the tongue, the middle of the tongue, and the tip of the tongue, calculate separately $N_i^{j_{\max }^i}=\max (N_i^j,j=7,\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;,15),$ where i = 1, 2, 3,

Then the type of tongue coating represented by j _max ⁱ is the tongue coating feature of this area;

3) Calculate the area ratio of j _max ⁱ tongue coating pixels in areas i=1, 2, and 3 $R(j_{\max }^i,i)=\frac{N_i^{j_{\max }^i}}{A_i};$

4) For tongue root region i=1, if $R(j_{\max }^i,i)<0.5,$ It is then described as "little coating on the root of the tongue"; for the middle area of the tongue i=2,

like $R(j_{\max }^i,i)<0.4,$ It is described as "little coating in the tongue"; for the tip of the tongue area i=3, if $R(j_{\max }^i,i)<0.2,$

It is described as "little coating on the tip of the tongue";

5) If $R(j_{\sec }^i,i)\&Greater\; Equal;0.5,$ Then describe the base of the tongue as j _max ⁱ type of tongue coating. For example $j_{\max }^{}$${}_1^{}=11$ It is described as "yellow coating on the root of the tongue". like $R({\mathrm{<figure-callout\; id="2"\; label="j\; max"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">j</figure-callout>}}_{\max }^{},2)\&Greater\; Equal;0.4,$ Then describe the tongue coating of j _max ² type in the tongue, for example $j_{\max }^{}$${}_2^{}=8$ Sometimes, it is described as "white coating in the tongue". like $R({\mathrm{<figure-callout\; id="3"\; label="j\; max"\; filenames="C0210379500201.png"\; state="\{\{state\}\}">j</figure-callout>}}_{\max }^{},3)\&Greater\; Equal;0.2,$ Then describe the tongue tip as j _max ³ type of tongue coating, for example $j_{\max }^{}$${}_3^{}=7$ , described as "thin white coating on the tip of the tongue";

6) For the base of the tongue and the center of the tongue, calculate the second main type of tongue coating $j_{\sec }^i,i=\mathrm{1,2},$ And calculate the area ratio of j _sec ⁱ type pixel area to j _max ⁱ type $R_{\mathrm{ms}}^i=\frac{N_i^{j_{\sec }^i}}{N_i^{j_{\max }^i}}.$ like $R_{\mathrm{ms}}^i>0.3,$ Then the i-th area is added to describe the second main type of tongue coating.

7. Quantitative analysis and qualitative description of fur thickness. The specific steps are:

1) Read the category number of the tongue body pixel and its 5×5 neighborhood;

2) Let c ₁ and c ₂ be the threshold value c ₁ >c ₂ set according to the experiment, and. Calculate the number of pixels S _body belonging to the tongue type in the 5×5 neighborhood of the pixel. If S _body > c ₁ , the tongue coating thickness type at this point is defined as "no coating";

3) If c ₁ ≥ S _body > c ₂ , the type of tongue coating thickness at this point is defined as "thin coating";

4) If S _body ≤ c ₂ , then determine the thickness type according to the type j of the tongue body and coating at this point. If it is tongue quality (j<7), it is "thin fur"; if j=7 (thin white fur), 10 (thin yellow fur), it is "thin fur"; if j=8 (white fur), 11 (yellow fur ), 13 (gray fur), 14 (brown fur), it is "medium thick fur"; if j=9 (white thick fur), 12 (yellow thick fur), 15 (black fur), it is "thick fur";

5) Repeat steps 1) to 4) until all pixels are processed;

6) Calculate the coating thickness index T of the entire image according to the pixel recognition result, and use it as the quantitative result of tongue coating thickness analysis. The calculation method is:

$\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{{\mathrm{<span\; class="notranslate">\; j</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; .</span>$

In the formula, A represents the total number of pixels in the tongue body area, Indicates the summation of all pixels in the tongue body area, j ^k represents the thickness type of the kth pixel, corresponding to "no fur", "thin fur", "medium thick fur", and "thick fur" respectively take 0, 1, 2, 3, w _jk are weights, which are determined according to experiments and experience of Chinese medicine experts. In the present invention, w ₀ =0, w ₁ =0.2, w ₂ =0.7, and w ₃ =1.3 are determined. The coating thickness index can describe the overall thickness of the tongue coating.

8. Display of analysis results. In order to make the analysis results easy to understand and clinical application, text and pseudo-color graphics are used to display the analysis results. The pixel recognition results are expressed as tongue color and fur color distribution pseudo-color map. The pixel-by-pixel thickness classification results are expressed as a pseudo-color distribution map of tongue coating thickness. Qualitative description of tongue color and fur color and thickness of tongue fur was given in words, and the index value of tongue thickness was given.

Figure 11 is an example showing the analysis results.

Claims (2)

1, a kind of traditional Chinese medical science tongue color, coating colour, thickness of the tongue coating analytical approach based on multi-class support vector machine, be to carry out image acquisition by digital camera, and image is input in the computer processor by USB interface, in processor, adopt current techique, the tongue body zone is split from tongue image, the invention is characterized in that this method also comprises the steps: successively

1) on the basis of carrying out hierarchical clustering analysis generation clustering tree and adjusting, adopt support vector machine method to carry out the identification of tongue body area pixel based on clustering tree, below will abbreviate CTSVM as based on the support vector machine of clustering tree;

The method step that the hierarchical clustering analysis generates clustering tree is:

(1), calculates k class centralization: (x for k class sample ₁, x ₂..., x _k),

(2) the 1st grades of divisions are that each class center respectively becomes a class, i.e. k leaf node,

(3) ask distance D between per two class centers _Ij, i=1 ... k, j=1 ... k, i ≠ j,

(4) nearest two centers are merged, form new class center, i.e. intermediate node,

(5) repeat (3)-(4) step, up to the k-1 level, 1 class is merged at all centers, i.e. root node,

(6) as required, adjust the structure of clustering tree, make that the degree of depth of tree is the least possible, the clustering tree of generation combines with two quasi-modes identification support vector machine, has just constituted CTSVM;

The method of tongue body area pixel identification and step be:

(1) calculate three chroma color value R in each pixel 3 * 3 neighborhood, the average of G, B in the input imagery, as the input feature vector of pixel identification,

(2) read R, G, the B value of a pixel in the tongue body,

(3) from the root node of described clustering tree, the CTSVM sorter according to this node is divided into a certain branch with this pixel, and concrete grammar is:

Calculate the kernel function inner product of this pixel each support vector corresponding, and multiply by corresponding coefficient and label,, ask the symbol that adds constant b result afterwards then all these product summations with this node.Promptly ask:

$f\left(x\right)=\mathrm{sign}(\underset{x_i\&Element;\mathrm{sv}}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_i{y}_{i}K(x_i,x)+b)$

In the formula, x _iBe the support vector of this sorter, α _iBe the coefficient corresponding with support vector, y _iBe the label of support vector sample, b is a constant, and b gets 40 among the present invention, K (x _i, x) be the kernel function inner product,

$K(x,y)=e^{\frac{{\left|\right|x-y\left|\right|}^2}{{2\mathrm{\&sigma;}}^2}}$

‖ x-y ‖ is the norm of vector x-y in the formula, and σ gets 50,

Sign () is-symbol function, if the value of independent variable greater than 0, then functional value is 1, less than 0, then is-1.

According to the value of f (x) and the classifying rules of this CTSVM, this pixel is divided into a certain branch then; (4) if branch node is not a leaf node, then repeat the process in (3): according to the CTSVM sorter of this node, this pixel is divided into a certain branch, if branch node is a leaf node, then the class of this leaf node representative is the classification of the pixel of being discerned, and the sorter of each node is two class support vector machines sorters, and assorting process realizes according to the method for routine, (5) process of repetition (2)-(4), the whole pixels classification up to the tongue body zone finish;

2) add up various types of number of picture elements in each subregion;

3) carrying out tongue color describes, color according to the tongue nature type specification tongue nature of tongue lateral areas and the tip of the tongue district pixel, calculate the sum of all pixels that belongs to each tongue nature type in 3 zones such as the tip of the tongue, tongue left side, tongue right side, the tongue nature type that pixel count is maximum is the tongue nature feature of this tongue image;

4) carrying out the tongue fur look describes, calculate in the root of the tongue, the tongue, belong to the sum of all pixels of each tongue fur type in 3 zones such as the tip of the tongue, in three zones, whether the pixel count of judging the tongue fur type that pixel count is maximum respectively and the ratio of whole lingual surface sum of all pixels determine the tongue fur feature in this district less than a certain threshold value;

5) carry out thickness of the tongue coating quantitative test and description, tongue color, tongue fur look type and quantity according to pixel and neighborhood interior pixel thereof, be divided into " no tongue ", " thin tongue ", " middle thick coating ", " thick coating " 4 types, thickness type and corresponding quantity according to whole tongue body zone interior pixel, calculate the thickness of the tongue coating index, describe thickness of the tongue coating according to the thickness index;

6) adopt literal and pseudocolour picture display analysis result.

2, the traditional Chinese medical science tongue color based on multi-class support vector machine according to claim 1, coating colour, thickness of the tongue coating analytical approach, the description of its tongue color, coating colour, thickness of the tongue coating is characterised in that, after pixel identification is finished, add up in each subregion behind various types of number of picture elements, make N _i ^jRepresent j class pixel number in the i district, i=1 wherein, 5 represent in the root of the tongue, the tongue respectively, 5 zones such as the tip of the tongue, tongue left side, tongue right side, j=1,15 expression tongue natures are light, tongue nature is light red, tongue nature is red, dimly red tongue, tongue nature deep red red, tongue nature is dark violet, 15 kinds of tongue natures such as thin and whitish fur, white tongue, white thick coating, thin and yellowish fur, yellow tongue, yellow thick coating, grayish fur, brown tongue, black tongue and tongue fur type, on this basis, according to the lingual diagnosis custom of Traditional Chinese Medicine experts, carry out the description of tongue color, coating colour, thickness of the tongue coating;

The describing method of tongue color is the color according to the tongue nature type specification tongue nature of tongue lateral areas and the tip of the tongue district pixel, calculate the sum of all pixels that belongs to each tongue nature type in 3 zones such as the tip of the tongue, tongue left side, tongue right side, the tongue nature type that pixel count is maximum is the tongue nature feature of this tongue image, and detailed process is:

1) calculate the total area (sum of pixel) A,

2) calculate the sum of 6 kinds of tongue nature types of tongue side and the tip of the tongue district respectively

${\mathrm{SUM}}^j=N_3^j+N_4^j+N_5^j,j=1,\&CenterDot;\&CenterDot;\&CenterDot;,6,$

3) order ${\mathrm{SUM}}^{j_{\max }}=\max ({\mathrm{SUM}}^j,j=1,\&CenterDot;\&CenterDot;\&CenterDot;,6),$ J then _MaxThe tongue nature type of representative is the tongue nature feature of this tongue image,

4) the area ratio of the dark violet tongue nature of calculating $R_6=\frac{{\mathrm{SUM}}^6}{A},$ If R ₆＞θ _pAnd j _Max≠ 6, then increase and describe " local dark violet ", wherein θ _pFor according to the experiment preset threshold,

5) the area ratio in calculating the tip of the tongue district " tongue nature is red " $R_3=\frac{N_3^3}{A},$ If R ₃＞θ _rAnd j _Max≠ 3, then increase and describe " red tip of tongue ", wherein θ _rFor according to the experiment preset threshold;

The describing method of coating colour is for calculating in the root of the tongue, the tongue, belonging to the sum of all pixels of each tongue fur type in 3 zones such as the tip of the tongue, in three zones, whether the pixel count of judging the tongue fur type that pixel count is maximum respectively and the ratio of whole lingual surface sum of all pixels be less than a certain threshold value, determine the tongue fur feature in this district, concrete grammar is as follows:

1) calculates the total area (sum of the pixel) A of lingual surface, in the root of the tongue, the tongue, the area A in the tip of the tongue district ₁, i=1 ..., 3,

2) in the root of the tongue, the tongue, the tip of the tongue district, calculate respectively $N_i^{j_{\max }^i}=\max (N_i^j,j=7,\&CenterDot;\&CenterDot;\&CenterDot;,15),$ I=1 wherein, 2,3,

J then _Max ⁱThe tongue fur type of representative is the tongue fur feature in this district,

3) calculate i=1, j in 2,3 districts _Max ⁱThe area ratio that the tongue fur pixel is shared $R(j_{\max }^i,i)=\frac{N_i^{j_{\max }^i}}{A_i},$

4) respectively in the root of the tongue, the tongue, the tip of the tongue district is described, for root of the tongue district i=1, if $R(j_{\max }^1,1)<{\mathrm{\&theta;}}_1$ Then being described as " the few tongue of the root of the tongue ", is j otherwise describe the root of the tongue _Max ¹The tongue fur of type, for distinguishing i=2 in the tongue, if $R(j_{\max }^2,2){<\mathrm{\&theta;}}_2,$ Then being described as " few tongue in the tongue ", is j otherwise describe in the tongue _Max ²The tongue fur of type, for the tip of the tongue district i=3, if $R(j_{\max }^3,3){<\mathrm{\&theta;}}_3,$ Then being described as " the few tongue of the tip of the tongue ", is j otherwise describe the tip of the tongue _Max ³The tongue fur of type, wherein θ ₁, θ ₂, θ ₃For according to the experiment preset threshold, 5) for distinguishing in the root of the tongue, the tongue, calculate the second main type tongue fur j respectively _Sec ⁱ, i=1,2, and calculate j _Sec ⁱType pixels area and j _Max ⁱThe area ratio of type $R_{\mathrm{ms}}^i=\frac{N_i^{j_{\sec }^i}}{N_i^{j_{\max }^i}},$ If $R_{\mathrm{ms}}^i{>\mathrm{\&theta;}}_{\sec },$ Then the i district increases description second main type tongue fur, the wherein θ _SecFor according to the experiment preset threshold;

The analysis and the description of thickness of the tongue coating are divided into two parts: thickness of the tongue coating classification and thickness of the tongue coating quantification, concrete steps are: 1) read the classification number of tongue body pixel and 5 * 5 neighborhoods thereof, 2) establish c ₁, c ₂For according to experiment preset threshold, and c ₁＞c ₂, calculate the number of picture elements S that belongs to the tongue nature type in these pixel 5 * 5 neighborhoods _Body, if S _Body＞c ₁, then the thickness of the tongue coating type of this point is decided to be " no tongue ", 3) if c ₁〉=S _Body＞c ₂, then the thickness of the tongue coating type of this point is decided to be " thin tongue ", 4) if S _Body≤ c ₂, then the tongue body and tongue coating type j according to this point determines the thickness type, if tongue nature is j＜7, then be " thin tongue ", if j=7,10 then is " a thin tongue ", if j=8,11,13,14 then is " a middle thick coating ", if j=9,12,15, then be " thick coating " 5) repeat 1)～3) step, finish up to all processes pixel, 6) according to the result of pixel identification, calculate the thick index T of tongue of view picture image, as the quantification result that thickness of the tongue coating is analyzed, computing method are:

$T=\frac{\underset{k}{\mathrm{\&Sigma;}}\left(w_{j^k}\right)}{A}$

A represents the pixel sum in tongue body district in the formula, Expression is to all pixel summations in the tongue body zone, j ^kThe thickness type of representing k pixel gets 0,1,2,3 respectively corresponding to " no tongue ", " thin tongue ", " middle thick coating ", " thick coating ",

Be weights, determine that according to the experience of experiment and Traditional Chinese Medicine experts the thick index of tongue can be described the general thickness of tongue fur.

Images