CN106886754B - Object identification method and system under a kind of three-dimensional scenic based on tri patch - Google Patents

Abstract

The invention discloses an object recognition method in a three-dimensional scene based on triangular facets. Based on the traditional two-dimensional image recognition method, this method takes into account the triangular grid storage method of the three-dimensional scene reconstruction information, divides the three-dimensional image into triangular patches, and uses the color neighborhood histogram feature statistics to divide the spatial pyramid into multi-scale. The information and the neighborhood space information are fused, and the support vector machine is used for training to obtain a classification model to identify the grass patches in the three-dimensional scene. The present invention also provides an object recognition system in a three-dimensional scene based on triangular facets. The technical solution of the present invention solves the problem that the recognition area is difficult to determine in the recognition problem in the three-dimensional reconstruction scene, and the local information in the image cannot fully express the image features. The problem is of great significance for surface reconstruction and scene understanding.

Description

A method and system for object recognition in three-dimensional scenes based on triangular facets

technical field

The invention belongs to the technical field of computer vision, and more particularly, relates to an object recognition method and system in a three-dimensional scene based on triangular facets.

Background technique

With the development of human society, image recognition has rapidly developed into an extremely important scientific and technological means. Its research goal is to give computers the visual ability of human beings, so that they can recognize the surrounding environment information through images, including recognizing the geometry of three-dimensional objects in the environment. shape, position, and attitude, etc. Image recognition requires a combination of disciplines such as computer science, pattern recognition, machine vision, and image understanding, and advances with the development of these disciplines.

Generally speaking, an image recognition system consists of the following parts. The first part is the acquisition of image information. For image recognition, it is to convert pictures, negatives and text graphics into electrical signals for subsequent processing. The second part is image preprocessing, the main purpose is to eliminate irrelevant features and strengthen interesting features. The third step is feature extraction. Commonly used features include color, texture, Scale-invariant feature transform (SIFT), Histogram of Oriented Gradient (HOG), etc. The fourth part is judgment or classification. The training method may adopt a support vector machine method (Support Vector Machine, SVM), an improved iterative training method (Adaptive Boostint, AdaBoost) and the like.

At present, the three most mainstream 3D object recognition research ideas are: 1) method based on model or geometry; 2) method based on appearance and view; 3) method based on local feature matching.

The principle of model-based or geometry-based methods is to use sensors to obtain three-dimensional information of real targets and analyze and process the information to obtain a description of surfaces, boundaries and connections. The advantage of this method is that it is relatively intuitive and easy to understand, but the computational load of the commonly used algorithm is relatively large, and the model needs to be manually generated by software such as Computer Aided Design (CAD).

View-based methods identify objects by visual similarity, and the design of the recognition system is relatively simple, without the need to explicitly calculate the 3D model of the object. The method is generally divided into two steps: first, automatically learn the representation of the object or train the system through different lighting conditions and different poses of the 3D object in the 2D image; then, in an unknown 2D image to determine whether there is a target. The limitation of this method is that the object of interest must be well differentiated from the background, so it is sensitive to overlap between objects and requires better image segmentation. But when geometric modeling of objects is difficult or impossible, view-based methods can be used to identify three-dimensional objects.

The method based on local feature matching mainly includes the local features of points, lines and surfaces. At the same time, because all local features do not need to be matched, this method has better robustness in the case of overlapping objects and complex backgrounds. Because the similarity between regions is established, object recognition is also localized; in two-dimensional image recognition, in the preprocessing stage of the image, it is often divided into different regions by superpixels, and then the regions are extracted and classified. Scenes cannot be identified by dividing superpixel regions due to their structural complexity.

SUMMARY OF THE INVENTION

In view of the above defects or improvement requirements of the prior art, the present invention provides a method and system for object recognition in a three-dimensional scene based on triangular patches, the purpose of which is to divide the three-dimensional scene to be identified into triangular patches of different sizes, using The feature of the target in the triangular patch is identified, thus the two-dimensional image recognition method is extended, and a grass recognition method based on the triangular patch is proposed.

In order to achieve the above object, according to an aspect of the present invention, a method for object recognition in a three-dimensional scene based on a triangular patch is provided, and the method includes the following steps:

(1) 3D reconstruction is performed on the 3D scene to be identified, and triangular patches are divided according to the triangular mesh of the surface of the 3D scene;

(2) Extract the HSI features of all triangular patch units;

(3) Perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and combine the eigenvectors of the triangular facets with its eigenvalues. The eigenvectors of adjacent triangular patches are superimposed (detailed in the specification), and the superposition result is used as the eigenvectors of the selected triangular patch;

(4) Input the feature vector of the selected triangular facet into the SVM recognition model to obtain the recognition result.

Further, in the step (3), the histogram is specifically set as:

Set the number of groups of H eigenvalue histogram and S eigenvalue histogram to 8, and the group spacing is 1/8 of the value range of H eigenvalue and 1/8 of the value range of S eigenvalue.

Further, putting the statistical data into the feature vector in the step (3) is specifically:

Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H feature values and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector.

Further, in the step (3), the feature vector is superimposed as follows:

Among them, {H ₁₁ , H ₁₂ , H ₁₃ ,...,H ₁₈ ,S ₁₁ ,S ₁₂ ,...,S ₁₈ } and {H ₂₁ ,H ₂₂ ,H ₂₃ ,...,H ₂₈ , S ₂₁ , S ₂₂ ,...,S ₂₈ } represent 16-dimensional feature vectors.

Further, the training process of the SVM recognition model in the step (4) is specifically:

(41) collecting and setting positive samples and setting negative samples as sample sets, the positive samples are selected as two-dimensional pictures of objects to be identified, and the negative samples are selected as two-dimensional pictures without objects to be identified;

(42) adjusting the sizes of all samples to be consistent, selecting a plurality of preset scales to divide the samples into triangular patches, and extracting the HSI features of all triangular patches;

(43) Perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and combine the eigenvectors of the triangular facets with its eigenvalues. The eigenvectors of adjacent triangular patches are superimposed, and the superposition result is used as the eigenvectors of the selected triangular patch;

(44) Input the feature vectors of all samples into SVM for training; the decision function of SVM is:

Among them, L represents the total number of classifications, and the value is 2; _ci represents the function coefficient, _ci ∈ R; γ represents the parameter of cross-checking, γ ∈ R ₊ ; represents the support vector selected for training; h ⁰ represents the non-support vector for training; the chi-square distance of the histogram is used As a kernel function, it is defined as:

Among them, N represents the total number of feature vectors in the sample set; represents the support vector corresponding to the kth eigenvector; h ⁰ (k) represents the kth eigenvector;

(45) After training, the parameters of the SVM recognition model are obtained, and the SVM recognition model is constructed.

According to another aspect of the present invention, an object recognition system based on a triangular patch in a three-dimensional scene is provided, and the system includes the following modules:

The triangular patch division module is used to perform 3D reconstruction of the 3D scene to be identified, and divide triangular patches according to the triangular mesh of the surface of the 3D scene;

The HSI feature extraction module is used to extract the HSI features of all triangular patch units;

The eigenvector calculation module is used to perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and calculate the The eigenvectors and the eigenvectors of its adjacent triangular patches are superimposed, and the superimposed result is used as the eigenvectors of the selected triangular patch;

The recognition module is used to input the feature vector of the selected triangular facet into the SVM recognition model to obtain the recognition result.

Further, the specific setting of the histogram in the feature vector calculation module is:

Set the number of groups of H eigenvalue histogram and S eigenvalue histogram to 8, and the group spacing is 1/8 of the value range of H eigenvalue and 1/8 of the value range of S eigenvalue.

Further, putting the statistical data into the feature vector in the feature vector calculation module is specifically:

Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H feature values and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector.

Further, in the step (3), the feature vector is superimposed as follows:

Among them, {H ₁₁ , H ₁₂ , H ₁₃ ,...,H ₁₈ ,S ₁₁ ,S ₁₂ ,...,S ₁₈ } and {H ₂₁ ,H ₂₂ ,H ₂₃ ,...,H ₂₈ , S ₂₁ , S ₂₂ ,...,S ₂₈ } represent 16-dimensional feature vectors.

Further, the training system of the SVM recognition model in the recognition module includes the following units:

The sample collection unit is used to collect multiple positive samples and multiple negative samples as a sample set, the positive samples are selected as two-dimensional pictures of the objects to be recognized, and the negative samples are selected as two-dimensional pictures without the objects to be identified;

The triangular patch division unit is used to adjust the size of all samples to be consistent, select multiple preset scales to divide the samples into triangular patches, and extract the HSI features of all triangular patches;

The eigenvector calculation unit is used to perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and calculate the The eigenvectors and the eigenvectors of its adjacent triangular patches are superimposed, and the superimposed result is used as the eigenvectors of the selected triangular patch;

The training unit is used to input the feature vectors of all samples into the SVM for training; the decision function of the SVM is:

Among them, L represents the total number of classifications, and the value is 2; _ci represents the function coefficient, _ci ∈ R; γ represents the parameter of cross-checking, γ ∈ R ₊ ; represents the support vector selected for training; h ⁰ represents the non-support vector for training; the chi-square distance of the histogram is used As a kernel function, it is defined as:

Among them, N represents the total number of feature vectors in the sample set; represents the support vector corresponding to the kth eigenvector; h ⁰ (k) represents the kth eigenvector;

The model building unit is used to obtain the parameters of the SVM recognition model after training, and construct the SVM recognition model.

In general, compared with the prior art, the above technical solutions conceived by the present invention have the following technical features and beneficial effects:

(1) The technical scheme of the present invention utilizes the triangular mesh in the reconstructed three-dimensional scene, and divides the three-dimensional scene into a plurality of two-dimensional triangular patches, so as to avoid the problem that superpixel extraction is too complicated in the three-dimensional scene;

(2) The technical solution of the present invention adopts the fusion of multi-scale feature information and neighborhood information, which is simple and effective, has a high recognition rate, and can be further used for three-dimensional scene surface reconstruction and scene understanding and analysis.

Description of drawings

1 is a flowchart of a grass identification method according to an embodiment of the present invention;

Fig. 2 is the flow chart of the SVM training of grass field recognition according to the embodiment of the present invention;

Fig. 3a is a statistical histogram of grass sample H data according to an embodiment of the present invention;

FIG. 3b is a statistical histogram of sky sample H data according to an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

As shown in FIG. 1, the process flow of the grass recognition method of the three-dimensional scene using the technical solution of the present invention includes the following steps:

(1) 3D reconstruction is performed on the 3D scene to be recognized, and triangular patches are divided according to the triangular mesh of the surface of the 3D scene.

(2) Extract the HSI features of all triangular patch units.

(3) Perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and combine the eigenvectors of the triangular facets with its eigenvalues. The eigenvectors of adjacent triangular patches are superimposed, and the superposition result is used as the eigenvectors of the selected triangular patch.

(4) Input the feature vector of the selected triangular facet into the SVM recognition model to obtain the recognition result.

As shown in Figure 2, the process of adopting the technical solution of the present invention to construct the grass SVM recognition model of the three-dimensional scene:

(1) Collect 50 positive samples and 200 negative samples, and select all grass images as positive samples to facilitate training set sampling and feature extraction. Since the types of negative samples to be covered are complex, the Pascal dataset (PASCALVOC) is selected here. Various types of non-grass targets in are used as negative samples.

(2) Adjust the size of the positive and negative training samples to 162*162, and divide the training samples into regular right-angled triangular patches. There are three sizes of triangular patches, which are 3 pixels and 6 pixels on a right-angled side. point and 9 pixels; because the sample image is in RGB mode, but because the RGB mode is under different conditions, the measured RGB color value distribution is scattered, and the Euclidean distance and color distance between two points in the RGB space are not linear, which is not conducive to the Classification of color features. In order to improve the recognition effect, the color representation is converted from the RGB color space to the HSI color space that is closer to the way the human eye observes, thereby improving the robustness of the recognition system under different lighting environments. In the HSI color space: H stands for chroma; S stands for saturation; I stands for lightness. Among the three components, the I component is most affected by light, while H and S are usually determined by the light absorption and reflection characteristics of the raw materials that constitute the object. In order to reduce the influence of illumination on the recognition accuracy, when recognizing the target, H and S are used as the extracted features. The conversion formula from RGB to HSI is as follows:

I=(R+G+B)/3

S=1-3min(R,G,B)/I

If B>G, H=360-H

(3) Perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and combine the eigenvectors of the triangular facets with its eigenvalues. The eigenvectors of adjacent triangular patches are superimposed, and the superposition result is used as the eigenvectors of the selected triangular patch. Set the number of groups of H eigenvalue histogram and S eigenvalue histogram to 8, and the group spacing is 1/8 of the value range of H eigenvalue and 1/8 of the value range of S eigenvalue. Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H feature values and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector. Save the feature vector of the triangular patch as a mat file for use by SVM. Figure 3a shows the statistical result of the H histogram of a triangular patch in the sky sample, and Figure 3b shows the statistical result of the H histogram of a triangular patch in the grass sample.

(4) Input the feature vectors of all samples into SVM for training; the decision function of SVM is:

Among them, L represents the total number of classifications, and the value is 2; _ci represents the function coefficient, _ci ∈ R; γ represents the parameter of cross-checking, γ ∈ R ₊ ; represents the support vector selected for training; h ⁰ represents the non-support vector for training; the chi-square distance of the histogram is used As a kernel function, it is defined as:

Among them, N represents the total number of feature vectors in the sample set; represents the support vector corresponding to the kth eigenvector; h ⁰ (k) represents the kth eigenvector.

Computes the covariance matrix space from the chi-square distance of the histogram. Mirror transformation is performed on this covariance matrix space to obtain the corresponding hyperplane matrix, that is to say, transforming a vector into a mirror reflected by a hyperplane is a linear transformation, and SVM constructs the optimal hyperplane in the feature space. The covariance matrix and its corresponding hyperplane matrix are obtained, the characteristic coefficients of each feature are calculated according to these two matrices, and the characteristic coefficients are scaled to the covariance matrix. After inverting the scaled covariance matrix, the model parameters are calculated. The parameters of the trained SVM model are stored in the matrix file for subsequent 3D scene grass recognition calls.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (6)

1. a grass recognition method based on the three-dimensional scene of triangular facet, it is characterised in that the method comprises the following steps: (1) 3D reconstruction is performed on the 3D scene to be identified, and triangular patches are divided according to the triangular mesh of the surface of the 3D scene; (2) Extract the HSI features of all triangular patch units; (3) Perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and combine the eigenvectors of the triangular facets with its eigenvalues. The eigenvectors of adjacent triangular patches are superimposed, and the superposition result is used as the eigenvectors of the selected triangular patch; The statistical data is put into the feature vector as follows: Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H feature values and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector; The superposition of the feature vector of the triangular facet and the feature vector of its adjacent triangular facets is specifically: Among them, {H ₁₁ , H ₁₂ , H ₁₃ ,...,H ₁₈ ,S ₁₁ ,S ₁₂ ,...,S ₁₈ } and {H ₂₁ ,H ₂₂ ,H ₂₃ ,...,H ₂₈ , S ₂₁ , S ₂₂ ,...,S ₂₈ } represents a 16-dimensional feature vector; (4) Input the feature vector of the selected triangular facet into the SVM recognition model to obtain the recognition result. 2. a kind of grass recognition method based on triangular facet three-dimensional scene according to claim 1, is characterized in that, in described step (3), the histogram is specifically set as: Set the number of groups of H eigenvalue histogram and S eigenvalue histogram to 8, and the group spacing is 1/8 of the value range of H eigenvalue and 1/8 of the value range of S eigenvalue. 3. a kind of grass recognition method based on the three-dimensional scene of triangular face piece according to claim 1, is characterized in that, the training process of SVM recognition model in described step (4) is specifically: (41) Collecting multiple positive samples and multiple negative samples as a sample set, the positive samples are all two-dimensional pictures of objects to be recognized, and the negative samples are two-dimensional pictures that do not contain objects to be recognized; (42) adjusting the sizes of all samples to be consistent, selecting a plurality of preset scales to divide the samples into triangular patches, and extracting the HSI features of all triangular patches; (43) Perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and combine the eigenvectors of the triangular facets with its eigenvalues. The eigenvectors of adjacent triangular patches are superimposed, and the superposition result is used as the eigenvectors of the selected triangular patch; The statistical data is put into the feature vector as follows: Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H feature values and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector; The superposition of the feature vector of the triangular facet and the feature vector of its adjacent triangular facets is specifically: Among them, {H ₁₁ , H ₁₂ , H ₁₃ ,...,H ₁₈ ,S ₁₁ ,S ₁₂ ,...,S ₁₈ } and {H ₂₁ ,H ₂₂ ,H ₂₃ ,...,H ₂₈ , S ₂₁ , S ₂₂ ,...,S ₂₈ } represents a 16-dimensional feature vector; (44) Input the feature vectors of all samples into SVM for training; the decision function of SVM is: Among them, L represents the total number of classifications, and the value is 2; _ci represents the function coefficient, _ci ∈ R; γ represents the parameter of cross-checking, γ ∈ R ₊ ; represents the support vector selected for training; h ⁰ represents the non-support vector for training; the chi-square distance of the histogram is used As a kernel function, it is defined as: Among them, N represents the total number of feature vectors in the sample set; represents the support vector corresponding to the kth eigenvector; h ⁰ (k) represents the kth eigenvector; (45) After training, the parameters of the SVM recognition model are obtained, and the SVM recognition model is constructed. 4. A grass recognition system based on a three-dimensional scene of a triangular patch, characterized in that the system comprises the following modules: The triangular patch division module is used to perform 3D reconstruction of the 3D scene to be identified, and divide triangular patches according to the triangular mesh of the surface of the 3D scene; The HSI feature extraction module is used to extract the HSI features of all triangular patch units; The eigenvector calculation module is used to perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and calculate the The eigenvectors and the eigenvectors of its adjacent triangular patches are superimposed, and the superimposed result is used as the eigenvectors of the selected triangular patch; The statistical data is put into the feature vector as follows: Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H eigenvalues and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector; The superposition of the feature vector of the triangular facet and the feature vector of its adjacent triangular facets is specifically: Among them, {H ₁₁ , H ₁₂ , H ₁₃ ,...,H ₁₈ ,S ₁₁ ,S ₁₂ ,...,S ₁₈ } and {H ₂₁ ,H ₂₂ ,H ₂₃ ,...,H ₂₈ , S ₂₁ , S ₂₂ ,...,S ₂₈ } represents a 16-dimensional feature vector; The recognition module is used to input the feature vector of the selected triangular facet into the SVM recognition model to obtain the recognition result. 5. The grass recognition system of a three-dimensional scene based on a triangular patch according to claim 4, wherein the histogram in the feature vector calculation module is specifically set as: Set the number of groups of H eigenvalue histogram and S eigenvalue histogram to 8, and the group spacing is 1/8 of the value range of H eigenvalue and 1/8 of the value range of S eigenvalue. 6. a kind of grass recognition system based on the three-dimensional scene of triangular face piece according to claim 4, is characterized in that, the training system of SVM recognition model in described recognition module comprises following unit: The sample collection unit is used to collect multiple positive samples and multiple negative samples as a sample set, the positive samples are selected as two-dimensional pictures of the objects to be recognized, and the negative samples are selected as two-dimensional pictures without the objects to be identified; The triangular patch division unit is used to adjust the size of all samples to be consistent, select multiple preset scales to divide the samples into triangular patches, and extract the HSI features of all triangular patches; The eigenvector calculation unit is used to perform histogram statistics according to the H eigenvalues and S eigenvalues of the selected triangular facet and its adjacent triangular facets with shared edges, put the statistical data into the eigenvectors, and calculate the The eigenvectors and the eigenvectors of its adjacent triangular patches are superimposed, and the superimposed result is used as the eigenvectors of the selected triangular patch; The statistical data is put into the feature vector as follows: Perform histogram statistics on the H eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors, and perform histogram statistics on the S eigenvalues of the triangular facets to obtain a set of 8-dimensional vectors. The 8-dimensional vectors of the H eigenvalues and the S feature The 8-dimensional vector of values is connected end to end to form a 16-dimensional feature vector; The superposition of the feature vector of the triangular facet and the feature vector of its adjacent triangular facets is specifically: Among them, {H ₁₁ , H ₁₂ , H ₁₃ ,...,H ₁₈ ,S ₁₁ ,S ₁₂ ,...,S ₁₈ } and {H ₂₁ ,H ₂₂ ,H ₂₃ ,...,H ₂₈ , S ₂₁ , S ₂₂ ,...,S ₂₈ } represents a 16-dimensional feature vector; The training unit is used to input the feature vectors of all samples into the SVM for training; the decision function of the SVM is: Among them, L represents the total number of classifications, and the value is 2; _ci represents the function coefficient, _ci ∈ R; γ represents the parameter of cross-checking, γ ∈ R ₊ ; represents the support vector selected for training; h ⁰ represents the non-support vector for training; the chi-square distance of the histogram is used As a kernel function, it is defined as: Among them, N represents the total number of feature vectors in the sample set; represents the support vector corresponding to the kth eigenvector; h ⁰ (k) represents the kth eigenvector; The model building unit is used to obtain the parameters of the SVM recognition model after training, and construct the SVM recognition model.