A method for age estimation in face images based on multi-view dimensionality reduction and ordinal regression

Abstract

The present invention discloses a method for estimating the age of a face image based on multi-view dimensionality reduction and ordered regression, comprising the following steps: S1, obtaining a face image data set; S2, classifying the face image data set into age categories; S3, preprocessing the face image data set after the category division; S4, extracting features and normalizing the preprocessed face image data set, thereby obtaining a training data set that can be used for classification; S5, establishing and training a multi-view dimensionality reduction ordered regression classifier; S6, inputting the face image whose age needs to be estimated into the trained multi-view dimensionality reduction ordered regression classifier, and finally obtaining the age estimation result of the face image. The present invention combines multi-view learning and dimensionality reduction, and incorporates the age sequence information between face images into the model training process, thereby greatly improving the accuracy of the ordered regression classifier for age estimation of such (high-dimensional multi-view) face images.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a method for estimating age of a facial image based on multi-view dimensionality reduction and ordered regression. Background Art

[0002] With the popularization of camera equipment, it is becoming more and more convenient to obtain facial images, and facial images have become an important medium for human-computer interaction. If age estimation can be performed based on facial images, it will be beneficial to the application of age-based control and assistance functions.

[0003] Age estimation based on facial images refers to the application of computer technology to extract features from facial images and build a model, so that the model can infer a person's approximate age or age range (age group) based on the facial image.

[0004] In real life, according to the sequence relationship of age labels, the age estimation problem can be regarded as an ordinal regression problem.

[0005] Ordered regression is usually defined as a multi-classification task of sorting input samples in order. Most existing methods treat it as a single-view learning problem. However, in the OR task of face image age estimation: a face image can often be represented by multiple separate feature sets (i.e., views), and when the dimension of multi-view data is too high, the classification performance of the OR classifier may be negatively affected, and the accuracy of face image age estimation will also decrease. Therefore, how to use high-dimensional multi-view face image features to construct an OR classifier and incorporate the order information of face image categories into the model training process to improve the classification accuracy (accuracy of face image age estimation) remains a challenge. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a facial image age estimation method based on multi-view dimensionality reduction and ordered regression for high-dimensional multi-view facial image features, so as to improve the accuracy of age estimation for such facial images.

[0007] In order to solve the above technical problems, the present invention provides a method for estimating age of a face image based on multi-view dimensionality reduction and ordered regression, comprising the following steps:

[0008] S1. Obtain a face image dataset;

[0009] S2, classify the face image dataset into age categories;

[0010] S3, preprocessing the face image dataset after category division;

[0011] S4, performing feature extraction and normalization processing on the preprocessed face image data set, so as to obtain a training data set that can be used for classification;

[0012] S5. Establish and train an ordered regression classifier for multi-view dimensionality reduction;

[0013] S6. Input the face image for which age estimation is required into the trained multi-view dimensionality reduction ordered regression classifier, and finally obtain the age estimation result of the face image.

[0014] Furthermore, in step S2, the categories are divided into four categories, the first category is 0-19 years old, the second category is 20-39 years old, the third category is 40-59 years old, and the fourth category is 60 years old and above.

[0015] Furthermore, in step S3, the preprocessing includes: performing face detection and alignment on each face image, and deforming and cutting the input image into a face image of a set size according to the size and position of the face.

[0016] Furthermore, in step S4, HOG features and LBP features of each face image are extracted.

[0017] Furthermore, the established multi-view dimensionality reduction ordered regression classifier includes:

[0018] For a test sample The prediction function is:

[0019]

[0020] In the above formula, α is the parameter for balancing view A and view B, For sample The prediction function in view A is as follows:

[0021]

[0022] For sample The prediction function in view B is as follows:

[0023]

[0024] In the above formula, w A and w B denote the weight vector of view A and the weight vector of view B, respectively, Q A and Q B is the corresponding projection matrix, represents transpose, represents the threshold corresponding to the jth hyperplane in view A; Represents the threshold corresponding to the j-th hyperplane in view B.

[0025] Furthermore, training an ordered regression classifier for multi-view dimensionality reduction includes:

[0026] Assume that the training set has r ordinal classes, and there are n ordinal classes in class k. k samples, k∈Y, Y={1,...,r} is the class label, then the total number of samples is For a dual-view face image dataset, the training samples are represents the i-th sample in the k-th class of view υ; υ is view A or view B; the k-th class represents the k-th age group;

[0027] For the two-view OR dataset, by using the projection matrix Project the original data

[0028] to two low-dimensional subspaces, where d υ is the dimension of the subspace; in the new subspace, ensure that the projected data can be classified by the ordered regression classifier in each view;

[0029] Support vector ordinal regression with implicit constraints is used as the basic ordinal regression classifier;

[0030] For each individual view, we try to obtain r-1 parallel hyperplanes to separate the r ordered classes. These parallel hyperplanes are represented as where w υ represents the weight vector, the threshold Determine the location of the hyperplane; by Introducing auxiliary hyperplane pass Introducing auxiliary hyperplane Samples belonging to the kth class should be located on the hyperplane and between;

[0031] In order to find two sets of hyperplanes and projection matrices Q in two views at the same time, the objective function is given:

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] In formula (1), w, b, Q, ξ, ξ * , η represent weight vector, threshold, projection matrix, first empirical error, and second empirical error slack variables respectively; w A and w B denote the weight vector of view A and the weight vector of view B respectively, represents transpose; C A >0, C B >0 and D>0 are regularization parameters, Indicates the distance that the i-th sample in the k-th category in view A deviates from the j-th category in the direction b0 equals negative infinity; represents the i-th sample in the k-th category in view A to b m Equal to the distance from the jth class in the positive infinity direction; η i represents the slack variable corresponding to the i-th sample; Indicates the distance that the i-th sample in the k-th category in view B deviates from the j-th category in the direction b0 equals negative infinity; represents the i-th sample in the k-th category in view B to b m Equal to the distance from the jth class in the positive infinity direction; b0 represents the auxiliary hyperplane The corresponding threshold; b m Denotes the auxiliary hyperplane The corresponding threshold value;

[0040] Formula (2) is used for the dual-view constraint, considering that the reduced-dimensional data from different views still correspond to the same face image; under this constraint, is a sample in view A The decision function value of Sample in View B The decision function value of ;∈ represents the similarity parameter; and The same sample Different representations in view A and view B; hence there are samples

[0041] is the difference between the decision function values ​​in views A and B; when this difference is greater than ∈, η>0; when the difference is not greater than ∈, η=0; based on the dual-view constraint, it can be ensured that different representations of a sample have similar decision function values;

[0042] Then optimize the objective function to find the optimal projection matrix Q A and QB , the weight vector w of view A A and the weight vector w of view B B , thus obtaining the trained multi-view dimensionality reduction ordered regression classifier.

[0043] Furthermore, the objective function is optimized to find the optimal projection matrix Q A and Q B , the weight vector w of view A A and the weight vector w of view B B ,include:

[0044] 1) Problem Reconstruction

[0045] Introduction and As shown below:

[0046] u A =Q A w A ,u B =Q B w B . (8)

[0047] By adding (8) as an equation constraint, equations (1)-(7) can be reformulated as:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056] 2) Calculate the projection matrix Q A and Q B :

[0057] First, solve the problem of and The constrained optimization problem (9) is used to initialize u A and u B ; Then, according to constraint (11), use the determined u A and uB Solve for the projection matrix Q A and Q B ;

[0058] 4) Calculate the weight vector w of view A A and the weight vector w of view B B :

[0059] The projection matrix Q A and Q B The optimal value of and

[0060] The optimal w A and w B It is to simplify the data and In order to obtain the optimal w A and w B We transform problem (1) into the following problem:

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] The above problems can be solved by using the SMO algorithm.

[0068] Further, calculate the projection matrix Q A and Q B The specific process includes:

[0069] Solving for u A and u B :

[0070] Without the projection matrix Q A and Q B Under the relevant constraints (10-11), u A and u B is the solution to the OR problem, which is as follows:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] The above learning problem is a support vector ordinal regression problem with two views, which is solved in its dual form using the sequential minimization algorithm;

[0078] Solving the projection matrix Q A and Q B :

[0079] For the best Projection matrix Q A and Q B , we can get the following equation:

[0080]

[0081] Formula (13) has DA×d A +D B ×d B variables and d A ×d A +d B ×d B +D A +D B If the number of equations is greater than the number of variables, there is no guarantee of a solution; considering this situation, due to

[0082] make Close to its transformation Therefore, the projection matrix Q is obtained by optimizing the following formula A and Q B :

[0083]

[0084] By applying several standard operations in linear algebra, problem (14) can be transformed as follows:

[0085]

[0086] Formula (15) is a closed-form solution with two and The trace maximization problem, where the optimal solution is a matrix The eigenvector of d υMaximum eigenvalue.

[0087] The beneficial effects of implementing the present invention are:

[0088] The original data is mapped to a low-dimensional subspace. After mapping, the redundant information in the original features is removed, thereby obtaining the projection data in the subspace, which is more discriminative.

[0089] In addition, the data of view A and view B are associated with the same face image and should be classified into the same age group. In order to ensure that samples from different views are assigned to the same age group, a dual-view Figure 1 Consistency constraints maximize the mutual consistency across different views.

[0090] In addition, unlike ordinary dimensionality reduction methods, the present invention combines multi-view learning and dimensionality reduction, and incorporates the age sequence information between facial images into the model training process, thereby greatly improving the accuracy of the ordered regression classifier for age estimation of such (high-dimensional multi-view) facial images. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is a principle flow chart of a method for estimating age of a face image based on multi-view dimensionality reduction and ordered regression according to the present invention;

[0092] Figure 2 A working principle diagram of a multi-view dimensionality reduction ordered regression classifier used in a method for estimating age of a face image based on multi-view dimensionality reduction and ordered regression according to the present invention;

[0093] Figure 3 Schematic diagram of the principle of a hyperplane in a single view. DETAILED DESCRIPTION

[0094] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. It is hereby stated that the directional terms such as up, down, left, right, front, back, inside, outside, etc. that appear or will appear in the text of the present invention are only based on the accompanying drawings of the present invention, and are not specific limitations of the present invention.

[0095] like Figure 1 As shown, the present invention provides a method for estimating age of a face image based on multi-view dimensionality reduction and ordered regression, comprising the following steps:

[0096] S1. Obtain a face image dataset.

[0097] S2, classify the face image dataset into age categories;

[0098] The categories are divided into four categories: the first category is 0-19 years old, the second category is 20-39 years old, the third category is 40-59 years old, and the fourth category is 60 years old and above.

[0099] S3, preprocessing the face image dataset after category division;

[0100] Preprocessing includes: face detection and alignment for each face image, and deformation and cutting of the input image into face images of a set size according to the face size and position.

[0101] S4. Perform feature extraction and normalization on the preprocessed face image dataset to obtain a training dataset that can be used for classification. In this step, the HOG features and LBP features of each face image are mainly extracted.

[0102] S5. Establish and train Figure 2 The ordinal regression classifier for multi-view dimensionality reduction shown;

[0103] Among them, the established multi-view dimensionality reduction ordered regression classifier includes:

[0104] For a test sample The prediction function is:

[0105]

[0106] In the above formula, α is the parameter for balancing view A and view B, For sample The prediction function in view A is as follows:

[0107]

[0108] For sample The prediction function in view B is as follows:

[0109]

[0110] In the above formula, w A and w B denote the weight vector of view A and the weight vector of view B, respectively, Q A and Q B is the corresponding projection matrix, represents transpose, represents the threshold corresponding to the jth hyperplane in view A; Represents the threshold corresponding to the j-th hyperplane in view B.

[0111] Training an ordinal regression classifier for multi-view dimensionality reduction involves:

[0112] Assume that the training set has r ordinal classes, and there are n ordinal classes in class k. k samples, k∈Y, Y={1,...,r} is the class label, then the total number of samples is For a dual-view face image dataset, the training samples are represents the i-th sample in the k-th class of view υ; υ is view A or view B; the k-th class represents the k-th age group;

[0113] For the two-view OR dataset, by using the projection matrix Project the original data

[0114] to two low-dimensional subspaces, where d υ is the dimension of the subspace; in the new subspace, ensure that the projected data can be classified by the ordered regression classifier in each view;

[0115] Support vector ordinal regression with implicit constraints is used as the basic ordinal regression classifier;

[0116] For each individual view, we try to obtain r-1 parallel hyperplanes to separate the r ordered classes. These parallel hyperplanes are represented as where w υ represents the weight vector, the threshold Determine the location of the hyperplane; by Introducing auxiliary hyperplane pass Introducing auxiliary hyperplane Samples belonging to the kth class should be located on the hyperplane and between;

[0117] In order to find two sets of hyperplanes and projection matrices Q in two views at the same time, the objective function is given:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] In formula (1), w, b, Q, ξ, ξ* , η represent weight vector, threshold, projection matrix, first empirical error, and second empirical error slack variables respectively; w A and w B denote the weight vector of view A and the weight vector of view B respectively, represents transpose; C A >0, C B >0 and D>0 are regularization parameters, Indicates the distance that the i-th sample in the k-th category in view A deviates from the j-th category in the direction b0 equals negative infinity; represents the distance that the i-th sample in the k-th category in view A deviates from the j-th category in the direction where bm is equal to positive infinity; η i represents the slack variable corresponding to the i-th sample; Indicates the distance that the i-th sample in the k-th category in view B deviates from the j-th category in the direction b0 equals negative infinity; Indicates the distance of the i-th sample in the k-th category in view B from the j-th category in the direction where bm is equal to positive infinity; b0 represents the auxiliary hyperplane The corresponding threshold; b m Denotes the auxiliary hyperplane The corresponding threshold value;

[0126] Figure 3 An example of a hyperplane in a single view is briefly shown, with respect to the empirical errors ξ and ξ * The detailed discussion of is as follows. Assume that a face dataset has three classes (age groups). In this figure, the solid line represents The dotted line indicates the hyperplane classifier According to the constraints (3-6), and Should be on the hyperplane and Because they belong to class 2. Sample In the negative direction of the hyperplane , then and sample On the hyperplane and In between, there are and sample In the positive direction of the hyperplane , so there is and

[0127] Formula (2) is used for the dual-view constraint, considering that the reduced-dimensional data from different views still correspond to the same face image; under this constraint, is a sample in view A The decision function value of Sample in View B The decision function value of ;∈ represents the similarity parameter; and The same sample Different representations in view A and view B; hence there are samples

[0128] is the difference between the decision function values ​​in views A and B; when this difference is greater than ∈, η>0; when the difference is not greater than ∈, η=0; based on the dual-view constraint, it can be ensured that different representations of a sample have similar decision function values; therefore, they can be classified into the same category. In the method described in this embodiment, an OR classifier is learned for each view, while minimizing the difference in the decision function outputs in different views, so that the information of different views can be integrated into the OR classifier, thereby improving its accuracy in estimating the age of face images.

[0129] The functions of the constraints in equations (3)-(7) are as follows:

[0130] For formula (3), its function is to transform the sample Constrained to the hyperplane on the left.

[0131] For formula (5), its function is to transform the sample Constrained to the hyperplane on the right.

[0132] At this time, the combined effect of formula (3) and (5) is to transform the sample Constrained to the middle of two hyperplanes, that is, constrained to a certain category.

[0133] Specifically, in equations (3) and (5), ignoring and In the case of and

[0134] For example: Figure 3 In For example. At first I didn’t know It is not known which category it is in, nor whether it has been misclassified.

[0135] When the general Substitute f(x) = w T Q T After x, we get Therefore, it is known that it is located on the left side of the hyperplane b2-1.

[0136] When the general Substitute f(x) = w T Q T After x, we get Therefore, it is known that it is located to the right of the hyperplane b1+1.

[0137] This sample (that is, ) is correctly classified between the hyperplanes b2-1 and b1+1, that is, it is correctly classified into the second category.

[0138] For such correctly classified samples, since they have been correctly classified, there is no need to constrain them to the correct class.

[0139] The empirical error corresponding to this sample So they are all 0 (because there is no mistake, there is no error).

[0140] The same is true for equations (4) and (6), except that the perspective B is used, so we will not go into details here.

[0141] Formula (7) is a linear algebra knowledge, similar to a theorem or axiom. Formula (7) is the condition that the projection matrix must satisfy. If a matrix does not satisfy formula (7), it cannot be called a projection matrix.

[0142] Then optimize the objective function to find the optimal projection matrix Q A and Q B , the weight vector w of view A A and the weight vector w of view B B , thus obtaining the trained multi-view dimensionality reduction ordered regression classifier, the process includes:

[0143] 1) Problem reconstruction (Since constraint (7) is non-convex, the optimization problem is non-convex for the projection matrices QA and QB, and the problem needs to be reconstructed)

[0144] Introduction and As shown below:

[0145] u A =Q A w A ,u B =Q B w B . (8)

[0146] By adding (8) as an equation constraint, equations (1)-(7) can be reformulated as:

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155] Before solving the above problems, the following theorem and its corresponding proof are given first.

[0156] Theorem: For w∈R d , u∈R D and Q∈R D×d , due to the existence of u = Qw and , it can be inferred that:

[0157] 1.

[0158] 2.

[0159] 3.

[0160] Proof: The derivation process is as follows:

[0161] 1.

[0162] 2.

[0163] 3.

[0164] Due to the solution of Q A and Q B Constraint (7) is generated, so the theorem is applied to transform problems (1-7) into problems (9-11). The specific transformation steps are as follows. First, according to formula 3 in the theorem, the objective function can be rewritten as (9). Then, based on the additional equation (8), the constraints (2)-(6) are Replace with Finally, according to Theorem 2, the additional equation (8) is rewritten as (11).

[0165] 2) Calculate the projection matrix Q A and Q B :

[0166] First, solve the problem of and The constrained optimization problem (9) is used to initialize u A and u B ; Then, according to constraint (11), use the determined u A and u B Solve for the projection matrix Q A and Q B ;

[0167] Among them, the projection matrix Q is calculated A and Q B The specific process includes:

[0168] Solving for u A and u B :

[0169] Without the projection matrix Q A and Q B Under the relevant constraints (10-11), u A and u B is the solution to the OR problem, which is as follows:

[0170]

[0171]

[0172]

[0173]

[0174]

[0175]

[0176] The above learning problem is a support vector ordinal regression problem with two views, which is solved in its dual form using the sequential minimization algorithm;

[0177] Solving the projection matrix Q A and Q B :

[0178] For the best Projection matrix Q A and Q B , we can get the following equation:

[0179]

[0180] Formula (13) has D A ×d A +D B ×d B variables and dA ×d A +d B ×d B +D A +D B If the number of equations is greater than the number of variables, there is no guarantee of a solution; considering this situation, due to

[0181] make Close to its transformation Therefore, the projection matrix Q is obtained by optimizing the following formula A and Q B :

[0182]

[0183] By applying several standard operations in linear algebra, problem (14) can be transformed as follows:

[0184]

[0185] Formula (15) is a closed-form solution with two and The trace maximization problem, where the optimal solution is a matrix The eigenvector of d υ Maximum eigenvalue.

[0186] 3) Calculate the weight vector w of view A A and the weight vector w of view B B :

[0187] The projection matrix Q A and Q B The optimal value of and

[0188] The optimal w A and w B It is to simplify the data and In order to obtain the optimal w A and w B We transform problem (1) into the following problem:

[0189]

[0190]

[0191]

[0192]

[0193]

[0194]

[0195] The above problems can be solved by using the SMO algorithm.

[0196] After solving problems (12), (15) and (16), we can obtain the projection matrix Q A and Q B , and classification for dimensionality reduction data and Based on these, we can predict the label of an unknown sample. Assume is a test example. In view A, the jth class is represented by the hyperplane and is bounded. Hyperplane is located in the middle of the jth class. (Sample X t HOG features), which to the hyperplane The distance is If the sample have That is Then this sample lies on the hyperplane and Therefore, it belongs to the jth category (age group). have That is and Then the sample is on the hyperplane and It is outside the range of , so it does not belong to the jth category.

[0197] S6. Input the face image for which age estimation is required into the trained multi-view dimensionality reduction ordered regression classifier, and finally obtain the age estimation result of the face image.

[0198] Although the description of the present disclosure has been quite detailed and specifically describes several described embodiments, it is not intended to be limited to any of these details or embodiments or any particular embodiment, but should be regarded as providing a broad possible interpretation of these claims by reference to the appended claims in view of the prior art, thereby effectively covering the intended scope of the present disclosure. In addition, the above description of the present disclosure is based on the embodiments foreseeable by the inventor, the purpose of which is to provide a useful description, and those non-substantial changes to the present disclosure that have not yet been foreseen may still represent equivalent changes to the present disclosure.

Claims

1. A method for age estimation of face images based on multi-view dimensionality reduction and ordinal regression, characterized in that: The following steps are involved: S1. Obtain a face image dataset; S2, classify the face image dataset into age categories; S3, preprocessing the face image dataset after category division; S4, performing feature extraction and normalization processing on the preprocessed face image data set, so as to obtain a training data set that can be used for classification; S5. Establish and train an ordered regression classifier for multi-view dimensionality reduction; S6, inputting the face image whose age needs to be estimated into the trained multi-view dimensionality reduction ordered regression classifier, and finally obtaining the age estimation result of the face image; In step S4, HOG features and LBP features of each face image are extracted; The established ordinal regression classifier for multi-view dimensionality reduction includes: For a test sample The prediction function is: In the above formula, α is the parameter for balancing view A and view B, For sample The prediction function in view A is as follows: For sample The prediction function in view B is as follows: In the above formula, w A and W B denote the weight vector of view A and the weight vector of view B, respectively, Q A and Q B is the corresponding projection matrix, represents transpose, represents the threshold corresponding to the jth hyperplane in view A; Represents the threshold corresponding to the j-th hyperplane in view B.

2. The method for estimating age of a face image based on multi-view dimensionality reduction and ordinal regression according to claim 1, characterized in that: In step S2, the categories are divided into four categories, the first category is 0-19 years old, the second category is 20-39 years old, the third category is 40-59 years old, and the fourth category is 60 years old and above.

3. The method for estimating age of a face image based on multi-view dimensionality reduction and ordinal regression according to claim 1, characterized in that: In step S3, the preprocessing includes: performing face detection and alignment on each face image, and deforming and cutting the input image into a face image of a set size according to the size and position of the face.

4. The method for estimating age of a face image based on multi-view dimensionality reduction and ordinal regression according to claim 1, characterized in that: Training an ordinal regression classifier for multi-view dimensionality reduction involves: Assume that the training set has r ordinal classes, and there are n ordinal classes in class k. k samples, k∈Y, Y={1,...,r} is the class label, then the total number of samples is For a dual-view face image dataset, the training samples are represents the i-th sample in the k-th class of view υ; υ is view A or view B; the k-th class represents the k-th age group; For the two-view OR dataset, by using the projection matrix The original data is projected into two low-dimensional subspaces, where d v is the dimension of the subspace; in the low-dimensional subspace, the projected data can be classified by the ordinal regression classifier in each view; Support vector ordinal regression with implicit constraints is used as the basic ordinal regression classifier; For each individual view, we try to obtain r-1 parallel hyperplanes to separate the r ordered classes. These parallel hyperplanes are represented as j = 1, ..., r-1, where w v represents the weight vector, the threshold Determine the location of the hyperplane; by Introducing auxiliary hyperplane pass Introducing auxiliary hyperplane Samples belonging to the kth class should be located on the hyperplane and between; In order to find two sets of hyperplanes and projection matrices Q in two views at the same time, the objective function is given: In formula (1), w, b, Q, ξ, ξ*, η represent the weight vector, threshold, projection matrix, first empirical error, and second empirical error relaxation variables respectively; W A and w B Represent the weight vector of view A and the weight vector of view B respectively, represents transpose; C A >0,C B >0 and D>0 are regularization parameters, Indicates the distance that the i-th sample in the k-th category in view A deviates from the j-th category in the direction where b0 is equal to negative infinity; represents the i-th sample in the k-th category in view A to b m Equal to the distance from the jth class in the positive infinity direction; η i represents the slack variable corresponding to the i-th sample; Indicates the distance that the i-th sample in the k-th category in view B deviates from the j-th category in the direction where b0 is equal to negative infinity; represents the i-th sample in the k-th category in view B to b m Equal to the distance from the jth class in the positive infinity direction; b0 represents the auxiliary hyperplane The corresponding threshold; b m Denotes the auxiliary hyperplane The corresponding threshold value; Formula (2) is used for the dual-view constraint, considering that the reduced-dimensional data from different views still correspond to the same face image; under this constraint, is a sample in view A The decision function value of Sample in View B The decision function value of ;∈ represents the similarity parameter; and The same sample Different representations in view A and view B; hence there are samples is the difference between the decision function values ​​in views A and B; when this difference is greater than ∈, η>0; when the difference is not greater than ∈, η=0; based on the dual-view constraint, it can be ensured that different representations of a sample have similar decision function values; Then optimize the objective function to find the optimal projection matrix Q A and Q B , the weight vector w of view A A and the weight vector w of view B B , thus obtaining the trained multi-view dimensionality reduction ordered regression classifier.

5. The method for estimating age of a face image based on multi-view dimensionality reduction and ordinal regression according to claim 4, characterized in that: Optimize the objective function and find the optimal projection matrix Q A and Q B , the weight vector w of view A A and the weight vector w of view B B ,include: 1) Problem Reconstruction Introduction and As shown below: u A =Q A w A ,u B =Q B w B . (8) By adding (8) as an equation constraint, equations (1)-(7) can be reformulated as: 2) Calculate the projection matrix Q A and Q B : First, solve the problem of and The constrained optimization problem (9) is used to initialize u A and u B ; Then, according to constraint (11), use the determined u A and u B Solve for the projection matrix Q A and Q B ; 3) Calculate the weight vector w of view A A and the weight vector w of view B B : The projection matrix Q A and Q B The optimal value of and The optimal w A and w B It is to simplify the data and In order to obtain the optimal w A and w B We transform problem (1) into the following problem: The above problems can be solved by using the SMO algorithm.

6. The method for estimating age of a face image based on multi-view dimensionality reduction and ordinal regression according to claim 5, characterized in that: Calculate the projection matrix Q A and Q B The specific process includes: Solving for u A and u B : Without the projection matrix Q A and Q B Under the relevant constraints (10-11), u A and u B is the solution to the OR problem, which is as follows: The above learning problem is a support vector ordinal regression problem with two views, which is solved in its dual form using the sequential minimization algorithm; Solving the projection matrix Q A and Q B : For the best Projection matrix Q A and Q B , we can get the following equation: Formula (13) has D A ×d A +D B ×d B variables and d A ×d A +d B ×d B +D A +D B If the number of equations is greater than the number of variables, there is no guarantee of a solution; considering this situation, due to make Close to its transformation Therefore, the projection matrix Q is obtained by optimizing the following formula A and Q B : By applying several standard operations in linear algebra, problem (14) can be transformed as follows: Formula (15) is a closed form with two solutions and The trace maximization problem, where the optimal solution is a matrix The eigenvector of , corresponding to d v Maximum eigenvalue.

Images