CN104899607B - A kind of automatic classification method of traditional moire pattern - Google Patents

Abstract

The invention proposes an automatic classification algorithm for traditional moiré patterns. It mainly solves the problem of low efficiency of manual classification of moire patterns, and realizes automatic classification of moiré patterns through moiré pattern preprocessing, feature extraction, and clustering processing. The implementation process is as follows: (1) preprocessing the moiré image, including three steps of unifying the image size, removing background noise, and thinning the lines of the moiré image; The shape context descriptor (SC) algorithm is used to extract the features of the moiré image, and the initial similarity between the moiré images is obtained through the shape context distance; (3) the similarity matrix is optimized through the improved neighbor relation transfer algorithm; (4) The optimized similarity matrix is used as the input matrix of the MEAP algorithm, and the MEAP clustering process is performed to realize automatic classification. The clustering results show that the present invention has higher clustering accuracy than SIFT-MEAP and ED-MEAP algorithms, and the clustering effect is more ideal. At the same time, the automatic classification algorithm of moiré patterns proposed by the present invention has good reference significance for the cluster analysis of other traditional art patterns.

Description

An Automatic Classification Method of Traditional Moiré Patterns

technical field

The invention belongs to the technical field of cluster analysis and image classification, and relates to the preprocessing of moiré images, the extraction of shape context features, and the optimization of similarity matrices for neighbor relationship transfer. Specifically, it is an automatic classification method for clustering moiré images with the multi-subclass central neighbor propagation algorithm combined with neighbor relation transfer and shape context features.

Background technique

Among the traditional Chinese decorative patterns, moiré is a category with a long history, extremely rich shapes and unique oriental artistic charm. Moiré rheology is vivid, auspicious, and has various forms of expression, including changes of different monomers, and various grafted and continuous combination structures. It has been an important decorative element in various plane and three-dimensional shapes since ancient times. Today, it still has great reference value for contemporary art design and creation. For example, the well-known 2008 Beijing Olympic torch used the auspicious cloud pattern as a decorative pattern, and its shape originated from the single-curled cloud pattern that appeared in the pre-Qin period. Today, when promoting the excellent traditional culture of the nation, the study of traditional art forms is still very important. Among them, searching, collecting, categorizing, and analyzing traditional patterns and styles, and discovering the national wisdom contained in them and the artistic language that conforms to the oriental aesthetic concept, are of great value both for academic research and for enriching the vocabulary of modern art and design.

The starting point of the development of traditional Chinese cloud patterns can be traced back to the original swirling patterns in the painted pottery patterns of the Neolithic Age. The cloud and thunder pattern on the bronze wares of the Spring and Autumn and Warring States Periods is considered to be the earliest clearly formed cloud pattern, and then experienced rolling cloud pattern, cloud pattern, flowing cloud pattern, cloud pattern, wishful cloud pattern, group cloud pattern, overlapping pattern The reproduction and change of moiré and other styles, from the initial simple abstraction to imitation and freehand brushwork, the styles are extremely rich and changeable, even the same category of moiré, with different times, regions, and creators, there are many forms. Variety. Due to the extremely rich moiré art form and the vast amount of various patterns and styles, it is very inefficient to do a good job in this research and sorting work only by manual search, sorting and categorization. Therefore, by introducing an unsupervised clustering analysis method, It is undoubtedly of great significance to realize the automatic classification of moiré patterns based on the study of pattern features.

In 2007, Frey et al. proposed a new clustering algorithm based on cluster centers in science, that is, the AP clustering algorithm (Frey B J: "Clustering by passing messages between data points" [J].science,2007,315 (5814):972-976.), and combined the algorithm with the Euclidean distance between different image pixels to cluster face images, and achieved a better clustering effect than k-centers. Afterwards, Frey et al. used AP clustering algorithm combined with SIFT features for cluster analysis of Caltech101 images, and proved that using SIFT algorithm to extract image features combined with AP algorithm for image clustering has certain advantages (Dueck, Frey B J. " Non-metric Affinity Propagation for Unsupervised Image Categorization》[C]//Proc of 11thInternational Conference on IEEE Computer Vision. Toronto, Canada, 2007: 1-8). Wang Changdong and others proposed the MEAP clustering algorithm on PAMI in 2013, and expanded the model of the single-cluster cluster center of the AP algorithm to a clustering model of multiple sub-cluster centers, and combined SIFT features to cluster face images, Aaltech101 images and SceneClass13 The class has been studied, and the clustering accuracy of the algorithm to deal with multi-subclass images has been improved (Wang C D: "Multi-exemplar affinity propagation" [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2013, 35 (9): 2223-2237). At present, the research on the automatic classification of traditional moiré patterns in my country is still blank.

Contents of the invention

The present invention draws on the research results of Frey and Wang Changdong et al., and proposes a multi-subcategory-centered neighbor propagation clustering algorithm (neighbor propagation based multi-exemplar affinity propagation, NP-MEAP) based on neighbor relation transfer, and realizes cloud Automatic classification of grain patterns. The purpose of the present invention is to overcome the shortcoming of very inefficient manual classification of moiré images, and to design an unsupervised automatic classification technology of moiré images.

The technical key to realize the present invention is: moiré image preprocessing, extracting the SC similarity matrix of the moiré image, optimizing the similarity matrix through the improved NP algorithm, and finally adopting the MEAP propagation clustering algorithm for automatic classification. The specific implementation steps include the following:

(1) Moiré image preprocessing, normalizing the size of the moiré image, removing background noise, and thinning the lines of the moiré image.

(1a) Normalize the size of the moiré image, which facilitates the unified processing of subsequent images and does not change the distribution of image lines

(1b) Remove background noise, and it is also convenient to use mathematical morphology method to refine the image.

(1c) Line thinning of moiré images, because the line shapes of different types of moiré images are different, and the line shapes of the same type of moiré images are basically similar, so we mainly focus on the line shapes of moiré images.

(2) Extract the shape context similarity matrix of the moiré image

(2a) The shape context algorithm considers that objects in each image can be approximately described by a finite number of discrete points evenly distributed on the graphic boundary, so it is necessary to extract discrete points on the boundary of the moiré image.

(2b) Calculate its shape context for each discrete point.

(2c) Compute the shape context difference between any two points in the two moiré images.

(2d) Compute the difference in tangent angle between any two points in the two moiré images.

(2e) Organically combine the shape context difference and tangent angle difference between any two points in two moiré images.

(2f) Calculate the shape context distance value between any two moiré images.

(3) The improved nearest neighbor transfer algorithm (NP) optimizes the S _sc matrix

(3a) Calculate the shape context distance matrix D.

(3b) Calculate the neighbor relation transmission threshold ε.

(3c) Calculate the similarity matrix S between the moiré images.

(3d) Calculate the neighbor relationship matrix N.

(3e) The neighbor relation transfer algorithm optimizes the similarity matrix.

(4) The similarity matrix after the above optimization is used as the input matrix of the MEAP algorithm, and the correct classification number is obtained by adjusting the reference value, so as to realize the automatic classification of moiré images.

The present invention preprocesses the moiré image, selects a suitable image feature extraction algorithm, based on the idea of manifold learning, adopts an improved neighbor relation transfer algorithm to optimize the similarity matrix, and finally adopts the latest multi-subclass central neighbor Propagation clustering algorithm realizes automatic classification, which fills the blank of automatic classification of moiré images, and at the same time ensures the accuracy of automatic classification of moiré images.

The present invention has the following advantages:

(1) By preprocessing the moiré image, the influence of image size, background noise, and line thickness is eliminated to ensure that the accuracy of automatic classification is not affected by it.

(2) Aiming at the main feature of the moiré image is its line shape, the current most superior shape feature extraction algorithm is selected, and the shape context descriptor ensures that the clustering result is ideal.

(3) The present invention adopts the neighbor relationship transfer algorithm to further optimize the shape context similarity matrix, so that the automatic classification accuracy obtained by the algorithm is higher, and the automatic classification of moiré images is realized.

Description of drawings

Fig. 1 is the realization flowchart of the present invention;

Figure 2 is the clustering model of MEAP;

Figure 3 is an example of a part of the moiré image;

Figure 4 is an example of a moiré image after size normalization;

Fig. 5 is the moiré image after binarization;

Fig. 6 is the moiré image after thinning;

Fig. 7 is a schematic diagram of discrete points extracted on the moiré image lines;

Fig. 8 is a vector schematic diagram from a certain point to other points in the moiré image;

Fig. 9 is an example diagram of six kinds of cirrus moiré images;

Figure 10 is a schematic diagram of the comparison of clustering accuracy of NP-MEAP, SIFT-MEAP, and ED-MEAP moiré images;

Figure 11 is a schematic diagram of the partial clustering results of the NP-MEAP algorithm on the moiré image;

detailed description

1. Introduction to basic theory

1. Multi-subclass central nearest neighbor propagation clustering algorithm

The MEAP algorithm is a clustering algorithm with a two-layer structure. As shown in Figure 2, the algorithm assigns all data objects to the most appropriate sub-cluster center, and assigns each sub-cluster center to the most appropriate super-cluster center, thus realizing The purpose of modeling multi-subclass problems.

Similar to the AP algorithm, the MEAP algorithm establishes similarity information s(i,j) and connectivity information l(i,j) for each data object with other data objects. The algorithm sets bias parameters p=s(k,k) and pp=l(k,k) values for each data object. The larger the value of p and pp, the corresponding data object is used as the candidate subclass center and super cluster The greater the possibility of the center, the more the number of clusters obtained, usually set p and pp as the median of the similarity matrix and connectivity matrix respectively. The core steps of the MEAP algorithm are the alternate update process of 4 types of 7 formulas, and the update formulas are as follows:

The specific meaning of the relevant parameters in the above formula can be found in the document multi-exemplar affinity propagation, and all new variables are initialized to 0. During the entire iterative update process of the MEAP algorithm, each data object competes to automatically generate the corresponding sub-class center and super-cluster center, and assigns other data objects to the nearest sub-class center, and the sub-class center is combined by the super-cluster center. together to form the final clustering result.

Two, the present invention is a kind of traditional moiré image automatic classification method

With reference to Fig. 1, the concrete implementation process of the present invention comprises the following steps:

Step 1. Moiré image preprocessing

Figure 3 is an example of multiple moiré patterns. It can be seen from Figure 3 that the sizes of different moiré patterns collected from various materials are different, and the thickness of the lines is different. At the same time, some of the collected moiré images contain gray backgrounds noise. Therefore, it is necessary to preprocess the moiré image to obtain more accurate clustering accuracy.

(1.1) The size of moiré images is different. First, all images are normalized to the size of 85*45, which is convenient for the unified processing of subsequent images and does not change the distribution of image lines. The moire image normalized to 85*45 is shown in Figure 4.

(1.2) Aiming at the problem that the moiré image contains background noise, the image is binarized to eliminate background noise, and it is also convenient to use mathematical morphology methods to refine the image. The Otsu method is used to calculate the binarization threshold, and the moiré image after binarization is shown in Figure 5.

(1.3) Because the line shapes of different types of moiré images are different, but the line shapes of the same type of moiré images are basically similar, so the present invention mainly focuses on the line shapes of moiré patterns. The different thickness of the lines is not related to the category of moiré patterns, but may affect the accuracy of clustering. Therefore, the mathematical morphology method is used to thin the lines to the width of one pixel. The thinned moiré image is shown in Figure 6.

Step 2. Compute the shape-context similarity matrix of the moiré image set

(2.1) The SC algorithm believes that objects in each image can be approximately described by a finite number of discrete points, and these discrete points do not need to be key points such as inflection points and extreme points in the graph, but are uniformly distributed on the graph boundary The discrete points of . Figure 7 is a schematic diagram of the discrete points extracted on the lines of the moiré image after preprocessing, wherein the small picture a in Figure 7 corresponds to the middle small picture in the upper row in Figure 6, and the small picture b in Figure 7 corresponds to the uppermost right side in Figure 6 small picture. It can be seen from Figure 7 that the discrete points on the lines of the moiré image can more accurately describe the shape of the corresponding moiré pattern lines, and the more boundary points are extracted, the more accurate the approximate description of the pattern is. However, when there are too many boundary points extracted, the running time of the algorithm will be too long. Generally, 100-150 boundary points can be selected to describe the line shape more accurately. The present invention uses n=100 boundary discrete points to describe the line shape.

Mark a point in the two small pictures in Figure 7 with a small box. For a certain discrete point in the contour point set p={p ₁ ,p ₂ ,...,p _n },n=100 in Figure 7, consider the vector starting from this point to other n-1 points , these n-1 vectors can more accurately describe the shape information of the moiré image. Figure 8 shows the vector illustration from the discrete point marked in Figure 7 to all other points.

(2.2) Each point in Figure 8 can be described by n-1 vectors starting from this point and ending at other points, and each moiré line is described by n n-1 dimensional vectors, thus we can get A rich feature description matrix for each moiré image. However, to calculate all these vectors to describe the moiré image, the amount of calculation will be very large, which is not realistic. For the shape, it is only necessary to know and calculate the positional relationship of all discrete points on the contour of the moiré image relative to the point. Therefore, the Cartesian coordinate system where the moiré lines are located is converted to the logarithmic polar coordinate system, and the discrete point to be calculated is the logarithmic polar coordinate system circle point, and the polar coordinate system is divided into 12 parts on average from 0 to 2π in the direction , the radius starts from the polar coordinate point outward to 2r and is divided into 5 parts by logarithmic space function conversion, where r is the average value of the Euclidean distance of the data set, so that the entire polar coordinate system is divided into 60 parts (bin) . Calculate the number of discrete points where the contour points of the moiré image are scattered into each bin to form a 60-dimensional vector. This 60-dimensional vector is called the shape context of the corresponding discrete point, that is, the logarithmic polar coordinate histogram of the discrete point. The formula for calculating the histogram is as follows:

h _i (k)＝#{q≠p _i :(qp _i )∈bin(k)}

Where k represents the kth bin in the polar coordinate system, with a value from 1 to 60, p _i is the boundary point of the moiré image of the histogram to be calculated, and q is the other n-1 boundary points except p _i point , qp _i is the number of boundary points in the kth bin.

(2.3) Calculate the shape context difference between any two points in the two moiré images. For a boundary point p _i in the moiré image P and a boundary point q _j in the moiré image Q, use mark the shape context difference of these two points, then The calculation formula of is as follows, where h _i (k) and h _j (k) represent the number of boundary points in the kth bin in the histograms of p _i and q _j respectively.

(2.4) Calculate the tangent angle difference between any two points in the two moiré images. The shape context difference better captures the overall difference of the discrete points on different moiré shapes. In order to make the difference between the discrete points of the moiré shape more accurate, the tangent angle difference of the discrete points is added. The formula is as follows, where θ _i and θ _j are respectively It is the tangent angle between p _i and q _j points.

(2.5) Combining the shape context difference and tangent angle difference between any two points in two moiré images, the shape context distance between any two points on different moiré images can be measured more accurately. The formula is as follows:

(2.6) Calculate the shape context distance between two moiré images. By calculating the shape context distance between any boundary point p _i in the moire image P and any boundary point q _j in the moire image Q according to the above formula, a distance matrix of n*n (n=100) is obtained, and the distance The average value of the horizontal and vertical minimum values of the matrix is summed to obtain the shape context distance value between two moiré images. Calculated as follows:

The smaller the value obtained by the above formula, the smaller the difference between the two images and the greater the similarity, and vice versa. Invert this value as the shape context similarity measure between two images, denoted as S _sc (P,Q)=-D _sc (P,Q), and calculate the shape context similarity measure between all images to obtain The similarity matrix S _sc of the moiré image set.

Step 3. The improved nearest neighbor transfer algorithm (NP) optimizes the S _sc matrix

(3.1) Calculate the shape context distance matrix D=[d _ij ] _n×n , which is used to initialize the neighbor relationship matrix N described below, and the element d _ij in the matrix is the shape context distance between moiré images i and j, The inverse of this value is used to update the similarity matrix S after the neighbor relationship is passed successfully.

(3.2) Calculate the threshold value of the neighbor relationship transmission, record the distance between the moiré image x _i and its kth neighbor point as d _ik , take the average value of the distance between all moire images and its kth neighbor point as the threshold value, the threshold value can be to a certain extent Reducing the influence of noisy data and selecting different k values for different data sets can make the neighbor relationship transfer more accurate. The new threshold formula is defined as follows:

(3.3) Calculate the similarity matrix between the moiré images, the similarity matrix S=[s _ij ] _n×n , the calculation formula of the i row j column element s _ij in the matrix is defined as follows:

d _ij is the shape context distance. Here, the distance between all moires is enlarged by exponential transformation. The main purpose is to enlarge the distance between moiré images on different manifolds, thereby reducing their similarity.

(3.4) Calculate the neighbor relationship matrix N. If the element d _ij in the distance matrix D is less than the neighbor relationship transfer threshold ε, then the data object x _i and x _j are considered to be neighbors to each other, expressed as ( _xi , x _j )∈R, From this definition, the neighbor relationship matrix of all moiré images is obtained. That is, when the data objects x _i and x _j are close neighbors to each other, the value of the corresponding element n _ij in the matrix is 1, otherwise the value is 0, and the diagonal elements are 0.

(3.5) Neighbor relationship transfer optimizes the similarity matrix, that is, if n _ij =0, and n _ik =1, n _kj =1, then set n _ij =1, n _ji =1, and update s _ij =s _ji =- min(d _ik ,d _kj ).

Step 4. Use the above-mentioned optimized similarity matrix as the input matrix of the MEAP algorithm to run the algorithm, and obtain the correct classification number by adjusting the reference value, so as to realize the automatic classification of moiré images.

The effect of the present invention can be further illustrated by the following experiments.

1. Simulation conditions

The curly cloud pattern produced in the pre-Qin period expanded from the abstract curl shape of the cloud and thunder pattern to a combination of curls with different structures. From the Spring and Autumn Period and the Warring States Period to the Qin and Han Dynasties, the cirrus cloud pattern was widely used in the surface decoration of various utensils. Various shapes of cirrus cloud patterns can be seen on bronze ware, lacquerware, jade ornaments, tile carvings, brocade embroidery, etc. There are many types of moiré shapes and various decorative effects, and it plays an important role in the history of decorative art development and in the field of contemporary design applications.

Here, the cirrus moiré pattern of the pre-Qin period is selected as a sample pattern for testing the method of the present invention. The dataset contains 230 moiré patterns of 6 types. According to the shape of the cirrus curve, there are single hook cirrus, restrained cirrus, divergent cirrus, comprehensive cirrus, S-shaped cirrus, and Ruyi cirrus. Figure 9 shows example patterns of these 6 types of cirrus moiré.

In order to verify the feasibility and effectiveness of the proposed algorithm of the present invention, the present invention, that is, the MEAP algorithm (NP-MEAP) based on neighbor transfer and SC feature and the MEAP algorithm (SIFT-MEAP) based on SIFT feature and based on Euclidean distance The MEAP algorithm (ED-MEAP) for comparison. Throughout the experiment process, the initial value of p and pp of the algorithm is set to the median value of the similarity matrix, the damping coefficient lam=0.9, the number of convergence iterations convits=50, the maximum number of cycles maxits=1000, and γ=3. The experimental operating environment is as follows: the processor is Core(TM) i5-3470, the main frequency is 3.2GHz, the memory is 4GB, the hard disk is 500GB, the operating system is Windows 7 Ultimate 64-bit operating system, and the programming language is Matlab R2013a. The present invention adopts commonly used clustering result evaluation indexes NMI index and FMI index.

The calculation formula of the standardized total information NMI index is as follows:

Among them, π is the class label of the cluster class obtained by the clustering algorithm, ζ is the class label of the real classification of the data set, and n _i (h) represents the number of data objects in the cluster class l and the real class h. H(π) is the Shannon entropy of the cluster label π, H(ζ) is the Shannon entropy of the true classification label ζ, and n _i and n ^(j) are the number of sample points in the cluster i and the true classification j respectively. The larger the value of NMI, the closer the clustering result is to the real classification.

The calculation formula of FMI (Fowlkes-Mallows Index) index is as follows:

Suppose the clustering result of the clustering algorithm is represented by C={c ₁ ,c ₂ ,...,c _m }, and the real classification of the data set is represented by P={p ₁ ,p ₂ ,...,p _l } . x _i and x _j are any two data objects in the dataset. where a is the number of x _i and x _j belonging to the same cluster in C and P; b is the number of x _i and x _j belonging to the same cluster in C but different clusters in P; c is the number of x _i and x _j belongs to different clusters in C, but belongs to the same cluster number in P; d is the number of x _i and x _j belong to different clusters in C and P, where a+b+c+d=n(n -1)/2. It can be seen that the value range of FMI is [0,1], and the larger the value, the higher the clustering accuracy of the algorithm.

The MEAP algorithm based on the SIFT feature of the moiré image, after the preprocessing of the moiré image, appropriately expands the moiré line with a width of one pixel to ensure that the SIFT algorithm can extract the appropriate SIFT more effectively while keeping the shape of the line unchanged. feature. After the preprocessing of the moiré image, the MEAP algorithm based on Euclidean distance also appropriately expands the moiré lines with a width of one pixel, and at the same time restores the image to the gray value of the image pixel before binarization, so as to ensure the Euclidean distance It can more accurately reflect the distance between moiré images. The discrete points extracted by the MEAP algorithm based on shape context and neighbor relationship transfer have a certain degree of randomness, so the discrete points detected each time have certain differences, which makes the shape context distance obtained by the algorithm run once fluctuate to a certain extent, so the The shape context algorithm runs 20 times, and the sum of the 20 shape context distances is taken as the input of the neighbor relation transfer algorithm, and the similarity matrix is optimized to ensure the stability of the algorithm.

2. Simulation results

The method of the present invention (NP-MEAP) was compared with the ED_MEAP and SIFT_MEAP methods.

Figure 10 is a schematic diagram of the comparison of moiré image clustering accuracy between NP-MEAP, SIFT-MEAP, and ED-MEAP. It can be seen from the figure that the clustering accuracy of the two algorithms of SIFT-MEAP and ED-MEAP is basically the same. The clustering accuracy can reach about 40%. It can be seen that the matching number of moiré image features extracted based on SIFT and the similarity of moiré images based on negative Euclidean distance cannot reflect the similarity between moiré images well. In contrast, the clustering accuracy of NP-MEAP can reach more than 80%. For very complex moiré pattern clustering problems, such accuracy is quite good, which can greatly reduce the work intensity and efficiency of manual classification. At the same time, it also shows that the similarity matrix based on the shape context feature optimized by the neighbor relationship transfer algorithm can better reflect the similarity between moiré images, so the clustering effect is better.

Figure 11 is a schematic diagram of the clustering results of the NP-MEAP algorithm on the moiré image, where each rounded rectangle represents a cluster, and the moiré image marked by a square in each rounded rectangle is the error clustering of the corresponding cluster. class image.

Claims (3)

1. an automatic classification method of a traditional moiré pattern, comprising the steps of: (1) Moiré image preprocessing, normalizing the size of the moiré image, removing background noise, and thinning the lines of the moiré image: (1a) Normalize the size of the moiré image, which will facilitate the unified processing of subsequent images and will not change the distribution of image lines; (1b) Remove background noise, and it is also convenient to use mathematical morphology method to refine the image; (1c) Line thinning of moiré images, because the line shapes of different types of moiré images are different, and the line shapes of similar moiré images are basically similar, so the main focus is on the line shapes of moiré images; (2) Extract the shape context similarity matrix of the moiré image: (2a) The shape context algorithm considers that objects in each image can be approximately described by a finite number of discrete points uniformly distributed on the graphic boundary, so 100 discrete points on the boundary of the moiré image are extracted to approximately describe the line contour of the moiré image ; (2b) Calculate its shape context for each discrete point, convert the Cartesian coordinate system where the moiré lines are located to the logarithmic polar coordinate system, take the discrete point to be calculated as a logarithmic polar coordinate system circle point, and convert the polar coordinate system In the direction from 0 to 2π, it is divided into 12 parts on average, and in the radius, it is divided into 5 parts from the polar coordinate point outward to 2r through the logarithmic space function conversion, where r is the average value of the Euclidean distance of the data set, so that the whole The polar coordinate system is divided into 60 parts (bins), and the number of discrete points scattered in each bin is calculated from the contour points of the moiré image to form a 60-dimensional vector, which is called the shape context of the corresponding discrete point , that is, the logarithmic polar coordinate histogram of discrete points, the formula for calculating the histogram is as follows: h _i (k)=#{q≠p _i :(qp _i )∈bin(k)}, Where k represents the kth bin in the polar coordinate system, with a value from 1 to 60, p _i is the boundary point of the moiré image of the histogram to be calculated, and q is the other n-1 boundary points except p _i point , qp _i is the number of boundary points in the kth bin; (2c) Calculate the shape context difference between any two points in the two moiré images. For a boundary point p _i in the moiré image P and a boundary point q _j in the moiré image Q, use mark the shape context difference of these two points, then The calculation formula of is as follows, where h _i (k) and h _j (k) represent the number of boundary points in the kth bin in the histograms of p _i and q _j respectively: <mrow><msubsup><mi>C</mi><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>s</mi><mi>c</mi></mrow></msubsup><mo>=</mo><mi>C</mi><mrow><mo>(</mo><msub><mi>p</mi><mi>i</mi></msub><mo>,</mo><msub><mi>q</mi><mi>j</mi></msub><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mn>2</mn></mfrac><munderover><mo>&amp;Sigma;</mo><mrow><mi>k</mi><mo>=</mo><mn>1</mn></mrow><mi>K</mi></munderover><mfrac><msup><mrow><mo>&amp;lsqb;</mo><msub><mi>h</mi><mi>i</mi></msub><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>-</mo><msub><mi>h</mi><mi>j</mi></msub><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>&amp;rsqb;</mo></mrow><mn>2</mn></msup><mrow><msub><mi>h</mi><mi>i</mi></msub><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow><mo>+</mo><msub><mi>h</mi><mi>j</mi></msub><mrow><mo>(</mo><mi>k</mi><mo>)</mo></mrow></mrow></mfrac><mo>;</mo></mrow> (2d) Calculate the tangent angle difference between any two points in the two moiré images. In order to make the difference between the moiré shapes more accurate, add the tangent angle difference of the contour points. The formula is as follows, where θ _i and θ _j are p The tangent angle between _i and q at point _j : <mrow><msubsup><mi>C</mi><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>t</mi><mi>a</mi><mi>n</mi></mrow></msubsup><mo>=</mo><mn>0.5</mn><mrow><mo>(</mo><mn>1</mn><mo>-</mo><mi>c</mi><mi>o</mi><mi>s</mi><mo>(</mo><mrow><msub><mi>&amp;theta;</mi><mi>i</mi></msub><mo>-</mo><msub><mi>&amp;theta;</mi><mi>j</mi></msub></mrow><mo>)</mo><mo>)</mo></mrow><mo>;</mo></mrow> (2e) Combining the shape context difference and tangent angle difference between any two points in two moiré images organically can measure the shape context distance between any two points on different moiré images more accurately. The formula is as follows: <mrow><msub><mi>C</mi><mrow><mi>i</mi><mi>j</mi></mrow></msub><mo>=</mo><mrow><mo>(</mo><mn>1</mn><mo>-</mo><mi>&amp;beta;</mi><mo>)</mo></mrow><msubsup><mi>C</mi><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>s</mi><mi>c</mi></mrow></msubsup><mo>+</mo><msubsup><mi>&amp;beta;C</mi><mrow><mi>i</mi><mi>j</mi></mrow><mrow><mi>t</mi><mi>a</mi><mi>n</mi></mrow></msubsup><mo>,</mo><mi>&amp;beta;</mi><mo>=</mo><mn>0.1</mn><mo>;</mo></mrow> (2f) Calculate the shape context distance between any two moiré images to obtain the shape context similarity matrix of the moiré image, by calculating any boundary point p _i in the moiré image P and any boundary point q in the moiré image Q The shape context distance between _j is obtained as a distance matrix of n*n (n=100), and the average value of the horizontal and vertical minimum values of the distance matrix is summed to obtain the shape context distance value between two moiré images, and the calculation The formula is as follows: <mrow><msub><mi>D</mi><mrow><mi>s</mi><mi>c</mi></mrow></msub><mrow><mo>(</mo><mi>P</mi><mo>,</mo><mi>Q</mi><mo>)</mo></mrow><mo>=</mo><mfrac><mn>1</mn><mi>n</mi></mfrac><munderover><mo>&amp;Sigma;</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><munder><mi>argmin</mi><mrow><mi>j</mi><mo>=</mo><mn>1</mn><mo>,</mo><mn>...</mn><mi>n</mi></mrow></munder><mi>C</mi><mrow><mo>(</mo><msub><mi>p</mi><mi>i</mi></msub><mo>,</mo><msub><mi>q</mi><mi>j</mi></msub><mo>)</mo></mrow><mo>+</mo><mfrac><mn>1</mn><mi>n</mi></mfrac><munderover><mo>&amp;Sigma;</mo><mrow><mi>j</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><munder><mi>argmin</mi><mrow><mi>i</mi><mo>=</mo><mn>1</mn><mo>,</mo><mn>...</mn><mi>n</mi></mrow></munder><mi>C</mi><mrow><mo>(</mo><msub><mi>p</mi><mi>i</mi></msub><mo>,</mo><msub><mi>q</mi><mi>j</mi></msub><mo>)</mo></mrow><mo>;</mo></mrow> (3) The improved neighbor transfer algorithm (NP) optimizes the S _sc matrix: (3a) Calculate the shape context distance matrix D=[d _ij ] _n×n , the element d _ij in the matrix is the shape context distance between moiré image i and j, and the inverse of this value is used to update the neighbor relationship after the transfer is successful Similarity matrix S; (3b) Calculate the neighbor relation transmission threshold ε, record the distance between the moiré image x _i and its kth neighbor point as d _ik , take the average value of the distances between all moiré images and their kth neighbor point as the threshold, and the threshold formula is defined as follows: <mrow><mi>&amp;epsiv;</mi><mo>=</mo><mi>m</mi><mi>e</mi><mi>a</mi><mi>n</mi><mrow><mo>(</mo><munderover><mo>&amp;Sigma;</mo><mrow><mi>i</mi><mo>=</mo><mn>1</mn></mrow><mi>n</mi></munderover><msub><mi>d</mi><mrow><mi>i</mi><mi>k</mi></mrow></msub><mo>)</mo></mrow><mo>;</mo></mrow> (3c) Calculate the similarity matrix S between the moiré images, the similarity matrix S=[s _ij ] _n×n , the calculation formula of the i row, j column element s _ij in the matrix is defined as follows: s _ij =-d _ij ^γ , Where d _ij is the shape context distance. Here, the distance between all moires is enlarged by exponential transformation. The main purpose is to enlarge the distance between moiré images on different manifolds, thereby reducing their similarity; (3d) Calculate the neighbor relationship matrix N, if the element d _ij in the distance matrix D is less than the neighbor relationship transmission threshold ε, then the data object x _i and x _j are considered to be neighbors to each other, and the value of the corresponding matrix element n _ij is set to 1, Otherwise, the value is 0, and the diagonal elements are 0; (3e) The neighbor relationship transfer algorithm optimizes the similarity matrix, that is, if n _ij =0, and n _ik =1, n _kj =1, then set n _ij =1, n _ji =1, and update s _ij =s _ji = -min(d _ik ,d _kj ); (4) The similarity matrix after the above optimization is used as the input matrix of the MEAP algorithm, and the correct classification number is obtained by adjusting the reference value, so as to realize the automatic classification of moiré images. 2. the automatic classification method of traditional moiré pattern according to claim 1, wherein step (1a) is carried out by following process: adopt Otsu method to calculate the binarization threshold value of moiré image, in the moiré image pixel point is greater than this threshold value, The pixel value at this point is set to 0, otherwise it is set to 1. 3. The automatic classification method of traditional moiré patterns according to claim 1, wherein step (1b) is carried out as follows: the image is thinned into one-pixel moiré lines by a mathematical morphology method.