A face recognition method based on semi-supervised dual-source face feature fusion

Abstract

This invention discloses a face recognition method based on semi-supervised dual-source face feature fusion, addressing the semi-supervised, multi-view, and cost-sensitive learning problems existing in real-world face recognition applications. The method includes: acquiring multiple face image data under preset dual views; inputting the face image data into a trained semi-supervised cost-sensitive canonical correlation analysis model; performing dual-source feature fusion using the cost-sensitive feature extraction matrix in the semi-supervised cost-sensitive canonical correlation analysis model to obtain the feature representation of the face image data; and classifying the face image data using a classification model based on the feature representation to obtain the face recognition result. This invention can obtain accurate and reliable face recognition results using only a small number of supervised dual-source face images, effectively improving the classification performance of the face recognition model.

Description

Technical Field

[0001] This invention relates to a face recognition method based on semi-supervised dual-source face feature fusion, belonging to the field of face image processing technology. Background Technology

[0002] In recent decades, facial recognition technology has developed rapidly and been widely applied. Existing facial recognition methods often use only single-source facial images to train facial recognition classification and prediction models. When facial images are "contaminated" by factors such as noise, it is difficult to learn robust facial recognition models based on single-source facial images. In fact, in real life, we easily obtain multi-source facial images, such as facial images from different camera perspectives, or facial images captured by natural light and infrared light respectively. Intuitively, facial images from different views contain complementary facial recognition information. If we can fuse features from multi-source facial images to extract additional complementary facial recognition information between views, the classification accuracy of facial recognition models can be improved.

[0003] Another drawback of existing facial recognition methods is their pursuit of minimizing the error rate, which assumes that the losses from different misclassifications are the same. However, in real-world facial recognition applications, this assumption often fails, as the losses from different misclassifications vary. For example, in a facial recognition access control system, assuming there are intruder and legitimate user classes, it is clear that the loss from misclassifying an intruder as a legitimate user is far greater than the loss from misclassifying a legitimate user as an intruder, while the loss from misclassifying a legitimate user as any other legitimate user is less than the loss from misclassifying a legitimate user as an intruder.

[0004] Furthermore, existing face recognition methods often assume supervised training datasets, meaning all training images have been manually labeled. In reality, manually labeling face images is difficult and requires significant human and financial resources. Therefore, in practical applications, more attention is often paid to semi-supervised face recognition models, where only a small number of face images in the training set are labeled, and the label information for the majority of the remaining face images is unknown.

[0005] To address the semi-supervised and multi-view learning problems in real-world face recognition applications, researchers have proposed numerous semi-supervised multi-view face recognition methods. However, these methods are cost-insensitive learning models, ignoring the unequal misclassification loss problem inherent in face recognition. Conversely, to address the semi-supervised and cost-sensitive learning problems in real-world face recognition applications, researchers have proposed numerous semi-supervised cost-sensitive face recognition models. However, these methods only use single-source face images for model construction, neglecting the complementary face recognition information existing between multi-source face images. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a face recognition method based on semi-supervised dual-source face feature fusion. This method can fuse features from dual-source face images to obtain complementary face recognition information between views. Moreover, the learned cost-sensitive face feature representation satisfies the minimum misclassification loss criterion, which can effectively improve the classification performance of the face recognition model.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical means:

[0008] This invention proposes a face recognition method based on semi-supervised dual-source face feature fusion, comprising the following steps:

[0009] Acquire multiple face image data under a preset dual-view configuration;

[0010] Input the face image data into the trained semi-supervised cost-sensitive canonical correlation analysis model;

[0011] We use the cost-sensitive feature extraction matrix in a semi-supervised cost-sensitive canonical correlation analysis model to perform dual-source feature fusion to obtain the feature representation of face image data;

[0012] Based on the feature representation of facial image data, a classification model is used to classify the data and obtain the facial recognition result.

[0013] Furthermore, the objective function expression for the semi-supervised cost-sensitive canonical correlation analysis model is as follows:

[0014]

[0015] Where v represents the number of face views, v = 1, 2, W (1) W (2) Let F represent the cost-sensitive feature extraction matrices for view 1 and view 2, respectively, and let F represent the label matrix of the face image category predicted by the semi-supervised cost-sensitive canonical correlation analysis model, where F = [f1, ..., f2]. n ] T G 12 G represents the cost-sensitive cross-view covariance matrix. vv Let h(i) represent the data covariance matrix within the view, and h(i) represent the i-th face image. The importance of f i f represents the category label of the i-th face image predicted by the semi-supervised cost-sensitive canonical correlation analysis model. j Let represent the category label of the j-th face image predicted by the semi-supervised cost-sensitive canonical correlation analysis model. Represents the i-th face image in view v. and the j-th person's face image The similarity, where m represents the number of manually annotated supervised face images in each face view, and y i Represents the face image of the i-th person. Real label information, I n Let I be a diagonal matrix. n =diag(n1,…,n c ), n k This represents the number of samples of the k-th class of face images in each view, k = 1, ..., c, where c represents the total number of face image classes. λ, μ, and τ are the balancing parameters, and the superscript T denotes the matrix transpose operation. ||·|| F Let Frobenius norm be i,j=1,…,n, where n is the total number of face images in each view.

[0016] Furthermore, the cost-sensitive cross-view covariance matrix G 12 The expression is as follows:

[0017]

[0018] in, This represents the i-th face image in view 1. The function represents the j-th face image in the 2-view set. Used for calculation The category label, h(j), represents the j-th face image. The importance of X (1) X represents a dataset of face images in one view. (2) F represents a dataset of face images in two views. (k) Let H represent the k-th column of the label matrix F, and H be a diagonal matrix, H = diag(h(1), ..., h(n));

[0019] Data covariance matrix G within the view vv =X (v) X (v)T , where X (v) A dataset of face images representing the v-view.

[0020] Furthermore, in the semi-supervised cost-sensitive canonical correlation analysis model, the cost-sensitive feature extraction matrix W (1) W (2) The learning function is:

[0021]

[0022] Furthermore, in the semi-supervised cost-sensitive canonical correlation analysis model, the learning function of the label matrix F is:

[0023]

[0024] Through matrix transformation, the learning function of the label matrix F is transformed into the objective function of the label matrix F:

[0025]

[0026] in, `sum(HS,1)` sums the rows of the matrix HS, and `sum(SH,1)` sums the rows of the matrix SH. `S = S` (1) +S (2) S (1) S represents the sample similarity matrix of view 1. (2) Y represents the sample similarity matrix of two views. l Y represents the cost label information of m supervised samples. l =[y1,…,y m ].

[0027] Furthermore, the training method for the semi-supervised cost-sensitive canonical correlation analysis model is as follows:

[0028] Acquire multiple face image data from dual views as training samples for the model;

[0029] Perform vector normalization on the face image data;

[0030] Randomly initialize the cost-sensitive feature extraction matrix W (1) W (2) With the label matrix F, we obtain the initialized semi-supervised cost-sensitive canonical correlation analysis model;

[0031] The normalized face image data is input into the semi-supervised cost-sensitive canonical correlation analysis model, and the objective function value of the semi-supervised cost-sensitive canonical correlation analysis model in the t-th iteration is calculated.

[0032] Increment the iteration count by 1, fix the label matrix F, and extract the cost-sensitive feature matrix W. (1) W (2) Learning function update W (1) W (2) ;

[0033] Fixed updated cost-sensitive feature extraction matrix W (1) W (2) Update the label matrix F according to the objective function of the label matrix F;

[0034] According to the updated W (1) W (2) And F, we obtain the updated semi-supervised cost-sensitive canonical correlation analysis model, and calculate the objective function value of the (t+1)th iteration;

[0035] When the difference between the objective function values ​​of two consecutive iterations is less than a preset threshold T, or the number of iterations reaches the maximum number of iterations T. n If the iteration ends and the training of the semi-supervised cost-sensitive canonical correlation analysis model is complete, then proceed to the next iteration and continue updating W. (1) W (2) And F.

[0036] Furthermore, with the label matrix F fixed, the cost-sensitive feature extraction matrix W is used. (1) W (2) The learning function is obtained as W (1) W (2) Update rules:

[0037]

[0038]

[0039] in,

[0040] Furthermore, the updated cost-sensitive feature extraction matrix W is fixed. (1) W (2) Based on the objective function of the label matrix F, the elements F in the label matrix F are... ij Taking the partial derivative, we get:

[0041]

[0042] F ij The update rules are as follows:

[0043] F ij =(Δ2(F) ij ) / Δ1(F ij ))F ij (9)

[0044] Normalize the label matrix F, let F T F = I n .

[0045] Furthermore, the cost-sensitive feature extraction matrix in the semi-supervised cost-sensitive canonical correlation analysis model is used for dual-source feature fusion to obtain the feature representation X of the face image data. The expression for X is as follows:

[0046]

[0047] Among them, W (1) W (2) Let X represent the cost-sensitive feature extraction matrices for view 1 and view 2, respectively. (1) X represents a dataset of face images in one view.(2) A dataset of face images representing two views.

[0048] Furthermore, the classification model employs a support vector machine model.

[0049] The following advantages can be obtained by adopting the above technical means:

[0050] This invention proposes a face recognition method based on semi-supervised dual-source face feature fusion. Under the condition of using only a small number of supervised dual-source face images, it uses a semi-supervised cost-sensitive canonical correlation analysis model for face recognition. This method can not only fuse features of dual-source face images to obtain complementary face recognition information between views, but also the learned cost-sensitive face feature representation satisfies the minimum misclassification loss criterion. This can improve the classification performance of the face recognition model, adapt to real-world face recognition scenarios, and obtain more accurate and reliable face recognition results, thus having high practical value. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the steps of a face recognition method based on semi-supervised dual-source face feature fusion according to the present invention.

[0052] Figure 2 This is a flowchart illustrating the training process of a semi-supervised cost-sensitive canonical correlation analysis model in an embodiment of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings:

[0054] This invention proposes a face recognition method based on semi-supervised dual-source face feature fusion, such as... Figure 1 As shown, the specific steps include the following:

[0055] Step A: Acquire multiple face image data under a preset dual-view system. There are many ways and methods to acquire face data from different views. This invention can select the face view according to specific scenarios and requirements, such as selecting a front view for one and a top view for the other. This embodiment of the invention takes a "natural light-wavelet" combination as an example, assuming there are already n face images captured by natural light. in, Let d1 represent the i-th face image in the natural light view, and d1 represent the face image space, i = 1, ..., n; the face image in the natural light image is processed using wavelet transform to obtain the corresponding wavelet face image. in, Let i represent the i-th face image in the wavelet view.

[0056] Step B: Input the face image data into the trained semi-supervised cost-sensitive canonical correlation analysis model.

[0057] To address the issues of semi-supervised, multi-view, and cost-sensitive learning in real-world face recognition applications, this invention combines the misclassification cost information h(i) of the samples with the inferred sample category label information. By embedding a canonical correlation analysis model, a semi-supervised cost-sensitive canonical correlation analysis model was designed to jointly learn the cost-sensitive feature extraction matrix of dual-source face images while fusing dual-source features. Label information for unlabeled face images

[0058] In this embodiment of the invention, the objective function expression of the semi-supervised cost-sensitive canonical correlation analysis model is as follows:

[0059]

[0060] Where v represents the number of face views, v = 1, 2, Indicates distribution at d v W contains n face images taken from view v within the space. (1) W (2) Let F represent the cost-sensitive feature extraction matrices for view 1 and view 2, respectively, and let F represent the label matrix of the face image category predicted by the semi-supervised cost-sensitive canonical correlation analysis model, where F = [f1, ..., f2]. n ] T G 12 G represents the cost-sensitive cross-view covariance matrix. vv Let h(i) represent the data covariance matrix within the view, and h(i) represent the i-th face image. The importance of f i f represents the category label of the i-th face image predicted by the semi-supervised cost-sensitive canonical correlation analysis model. j Let represent the category label of the j-th face image predicted by the semi-supervised cost-sensitive canonical correlation analysis model. Represents the i-th face image in view v. and the j-th person's face image The similarity, where m represents the number of manually annotated supervised face images in each face view, and y i Represents the face image of the i-th person. Authentic label information I n Let I be a diagonal matrix. n =diag(n1,…,n c ), n k This represents the number of samples of the k-th class of face images in each view, k = 1, ..., c, where c represents the total number of face image classes. λ, μ, and τ are the balancing parameters, and the superscript T denotes the matrix transpose operation. ||·|| FLet Frobenius norm be i,j=1,…,n, where n is the total number of face images in each view.

[0061] In this embodiment of the invention, considering the cost of data annotation, let m << n. The category of a face image refers to which person the face image belongs to. Assuming that a certain face image dataset contains 15 people, with 10 face images of each person, then the face image dataset contains a total of 150 face images, and the total number of categories c is 15.

[0062] In this embodiment of the invention, the cost-sensitive cross-view covariance matrix G 12 The expression is as follows:

[0063]

[0064] in, This represents the i-th face image in view 1. The function represents the j-th face image in the 2-view set. Used for calculation The category label, h(j), represents the j-th face image. The importance of X (1) X represents a dataset of face images in one view. (2) F represents a dataset of face images in two views. (k) Let H represent the kth column of the label matrix F, and H be a diagonal matrix, H = diag(h(1), ..., h(n)).

[0065] In this embodiment of the invention, the data covariance matrix G within the view... vv =X (v) X (v)T , where X (v) A dataset of face images representing the v-view.

[0066] In the semi-supervised cost-sensitive canonical correlation analysis model, the cost-sensitive feature extraction matrix W (1) W (2) The learning process not only relies on dual-source face images {X (1) ,X (2) It also effectively utilizes the predicted label matrix F of the face image and the cost-sensitive feature extraction matrix W. (1) W (2) The learning function is:

[0067]

[0068] In the cost-sensitive face feature space after feature fusion, the face image label matrix F is learned: by defining a cost-sensitive label extension term, the local manifold structure of the face image is effectively utilized. The cost label information Y of m supervised samples l =[y1,…,y m Furthermore, the nonnegativity of the label matrix F is constrained by defining a label nonnegativity regularization term. In the semi-supervised cost-sensitive canonical correlation analysis model, the learning function of the label matrix F is:

[0069]

[0070] Through matrix transformation, the learning function of the label matrix F is transformed into the objective function of the label matrix F:

[0071]

[0072] Among them, matrices B, Q, Y, They are defined as follows:

[0073]

[0074] Where sum(HS,1) represents the summation operation of matrix HS row by row, sum(SH,1) represents the summation operation of matrix SH row by row, and S = S (1) +S (2) S (1) S represents the sample similarity matrix of view 1. (2) Y represents the sample similarity matrix of two views. l Y represents the cost label information of m supervised samples. l =[y1,…,y m ].

[0075] Based on formulas (11), (13), and (15), this invention employs an alternating iterative method to solve for and update the cost-sensitive feature extraction matrix W. (1) W (2) And the label matrix F enables the training of a semi-supervised cost-sensitive canonical correlation analysis model using face images from different views, such as... Figure 2 As shown, the training method includes the following steps:

[0076] Step 201: Define the importance function h(i) and the sample similarity matrix S. (v) In this embodiment of the invention, taking a facial recognition access control system as an example, the definition of the importance function h(i), i = 1, ..., n, is given. Specifically, for the facial recognition access control system, it is assumed to contain an intruder class I and c-1 legitimate user classes G. k For k = 1, 2, ..., c-1, the misclassification cost matrix C is easily obtained as shown in Table 1:

[0077] Table 1

[0078] <![CDATA[G1]]> … <![CDATA[G c-1 ]]> I <![CDATA[G1]]> 0 … <![CDATA[C GG ]]> <![CDATA[C GI ]]> … <![CDATA[G c-1 ]]> <![CDATA[C GG ]]> … 0 <![CDATA[C GI ]]> I <![CDATA[C IG ]]> … <![CDATA[C IG ]]> 0

[0079] In Table 1, C GI C IG C GG The following three misclassification costs are represented respectively: (1) False rejection: incorrectly classifying a legitimate user as an intruder; (2) False acceptance: incorrectly classifying an intruder as a legitimate user; (3) False identification: incorrectly classifying a legitimate user as another legitimate user. Based on the misclassification cost matrix C shown in Table 1, the importance function h(i) can be defined as follows:

[0080]

[0081] The parameter δ is used to measure the importance of unlabeled training images.

[0082] This invention uses the k-nearest neighbor graph as an example to illustrate the sample similarity matrix S. (v) Definition method: Given any two samples and Where i,j=1,…,n, their similarity weights It can be defined as follows:

[0083]

[0084] Step 202: Obtain multiple face image data under dual views as training samples for the model.

[0085] Step 203: Perform vector normalization on the face image data. Assume a given face image x... (v) Using the formula Normalize the face image to a unit vector. In this invention, unless otherwise specified, x (v) This represents a normalized face image.

[0086] Step 204: Randomly initialize the cost-sensitive feature extraction matrix W (1) W (2) With the label matrix F, we obtain the initialized semi-supervised cost-sensitive canonical correlation analysis model.

[0087] Step 205: Input the normalized face image data into the semi-supervised cost-sensitive canonical correlation analysis model, and use formula (11) to calculate the objective function value of the semi-supervised cost-sensitive canonical correlation analysis model in the t-th iteration.

[0088] Step 206: Increment the iteration count by 1, fix the label matrix F, and extract the cost-sensitive feature matrix W. (1)W (2) Learning function update W (1) W (2) .

[0089] This invention extracts the cost-sensitive feature matrix W. (1) W (2) The learning function is obtained by using singular value decomposition to obtain W. (1) W (2) Update rules:

[0090]

[0091]

[0092] in,

[0093] Step 207: Fix the updated cost-sensitive feature extraction matrix W (1) W (2) The label matrix F is updated according to the objective function of the label matrix F.

[0094] This invention, based on the objective function of the label matrix F, processes the elements F in the label matrix F. ij Taking the partial derivative, we get:

[0095]

[0096] Therefore, F is obtained. ij The update rules are as follows:

[0097] F ij =(Δ2(F) ij ) / Δ1(F ij ))F ij (twenty two)

[0098] Step 208: Normalize the label matrix F, let F T F = I n .

[0099] Step 209: Update W according to steps 206-208 (1) W (2) And F, we obtain the updated semi-supervised cost-sensitive canonical correlation analysis model, and calculate the objective function value of the (t+1)th iteration.

[0100] Step 210: When the difference between the objective function values ​​of two adjacent iterations is less than a preset threshold T, or the number of iterations reaches the maximum number of iterations T. n If the iteration ends and the semi-supervised cost-sensitive canonical correlation analysis model training is complete, then return to step 206 to begin the next iteration and continue updating W. (1) W(2) And F.

[0101] Step C: Use the cost-sensitive feature extraction matrix in the semi-supervised cost-sensitive canonical correlation analysis model to perform dual-source feature fusion to obtain a robust feature representation of the face image data.

[0102] In this embodiment of the invention, the feature representation X of the face image data is expressed as follows:

[0103]

[0104] Step D: Based on the feature representation of the face image data, classify it using a classification model to obtain the face recognition result.

[0105] This invention employs a support vector machine (SVM) model for face recognition, training a face recognition classification decision model. After training, a classification model Y is obtained. pred =f svm (X), where Y pred This indicates that the input X is processed by the support vector machine model f. svm (·) Classification prediction, resulting in a category label matrix.

[0106] This invention combines the misclassification cost information h(i) of the sample with the inferred sample category label information f. i By embedding a canonical correlation analysis model, a semi-supervised cost-sensitive canonical correlation analysis model was designed to jointly learn the cost-sensitive feature extraction matrix W of dual-source face images while fusing dual-source features. (v) Label information F of unlabeled face images. Dual-source face feature fusion and cost-sensitive feature extraction matrix W. (v) The learning processes of the label matrix F and the label learning process are interdependent and updated alternately. This not only establishes a link between label learning and cost-sensitive multi-view learning, but also facilitates the acquisition of a robust cost-sensitive feature extraction matrix W. (v) Furthermore, by defining cost-sensitive label extension terms, the local manifold structure of face images is effectively utilized. The cost label information Y of a small number of supervised samples in the training set l =[y1,…,y m This improves the accuracy of predicting label F. Furthermore, the cost-sensitive feature extraction matrix W learned in this invention... (v) It minimizes the overall misclassification loss of face recognition, thereby improving the classification performance of the face recognition model.

[0107] Compared with existing technologies, this invention takes into account semi-supervised learning, multi-view learning, and cost-sensitive learning, which can effectively improve the classification performance of face recognition models, adapt to real face recognition scenarios, and obtain more accurate and reliable face recognition results, thus having high practical value.

[0108] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A face recognition method based on semi-supervised dual-source face feature fusion, characterized in that, Includes the following steps: Acquire multiple face image data under a preset dual-view configuration; Input the face image data into the trained semi-supervised cost-sensitive canonical correlation analysis model; We use the cost-sensitive feature extraction matrix in a semi-supervised cost-sensitive canonical correlation analysis model to perform dual-source feature fusion to obtain the feature representation of face image data; Based on the feature representation of facial image data, a classification model is used to classify the data and obtain the facial recognition result. The objective function expression for the semi-supervised cost-sensitive canonical correlation analysis model is as follows: ； Where v represents the number of face views, , Let represent the cost-sensitive feature extraction matrices for view 1 and view 2, respectively, and let F represent the label matrix of the face image category predicted by the semi-supervised cost-sensitive canonical correlation analysis model. , Represents the cost-sensitive cross-view covariance matrix. This represents the data covariance matrix within the view. Indicates the first Personal facial images The importance of The semi-supervised cost-sensitive canonical correlation analysis model predicts the first... Category labels for individual facial images. Let represent the category label of the j-th face image predicted by the semi-supervised cost-sensitive canonical correlation analysis model. express The first in the view Personal facial images and the j-th person's face image Similarity, This represents the number of manually labeled supervised face images in each face view. Indicates the first Personal facial images Real label information, It is a diagonal matrix. , Indicates the first view in each view The number of samples of human-like face images, c represents the total number of categories in the face image. , , These are the balancing parameters, and the superscript T indicates the matrix transpose operation. Denotes the Frobenius norm. , where n is the total number of face images in each view. .

2. The face recognition method based on semi-supervised dual-source face feature fusion according to claim 1, characterized in that, Cost-sensitive cross-view covariance matrix The expression is as follows: ； in, Represents the first view in view 1 Personal facial images The function represents the j-th face image in the 2-view set. Used for calculation Category tags, Represents the j-th person's face image The importance of A dataset of face images representing a 1-view view. This represents a dataset of face images in two views. Let H represent the k-th column of the label matrix F, where H is a diagonal matrix. ; Data covariance matrix within the view ,in, express A dataset of human face images from the view.

3. The face recognition method based on semi-supervised dual-source face feature fusion according to claim 1, characterized in that, In a semi-supervised cost-sensitive canonical correlation analysis model, the cost-sensitive feature extraction matrix The learning function is: 。 4. The face recognition method based on semi-supervised dual-source face feature fusion according to claim 1, characterized in that, In the semi-supervised cost-sensitive canonical correlation analysis model, the learning function of the label matrix F is: ； Through matrix transformation, the learning function of the label matrix F is transformed into the objective function of the label matrix F: ； in, , This indicates that the HS matrix is ​​summed row by row. This indicates that the SH matrix is ​​summed row by row. , This represents the sample similarity matrix of view 1. This represents the sample similarity matrix of two views. express Cost label information for each supervised sample. .

5. A face recognition method based on semi-supervised dual-source face feature fusion according to claim 3 or 4, characterized in that, The training method for a semi-supervised cost-sensitive canonical correlation analysis model is as follows: Acquire multiple face image data from dual views as training samples for the model; Perform vector normalization on the face image data; Randomly initialize cost-sensitive feature extraction matrix With the label matrix F, we obtain the initialized semi-supervised cost-sensitive canonical correlation analysis model; The normalized face image data is input into the semi-supervised cost-sensitive canonical correlation analysis model, and the objective function value of the semi-supervised cost-sensitive canonical correlation analysis model in the t-th iteration is calculated. Increment the iteration count by 1, fix the label matrix F, and extract the cost-sensitive feature matrix. Learning function update ; Fixed updated cost-sensitive feature extraction matrix Update the label matrix F according to the objective function of the label matrix F; According to the updated And F, we obtain the updated semi-supervised cost-sensitive canonical correlation analysis model, and calculate the objective function value of the (t+1)th iteration; When the difference between the objective function values ​​of two consecutive iterations is less than a preset threshold T, or the number of iterations reaches the maximum number of iterations. If the iteration ends and the training of the semi-supervised cost-sensitive canonical correlation analysis model is complete, then proceed to the next iteration and continue updating. And F.

6. The face recognition method based on semi-supervised dual-source face feature fusion according to claim 5, characterized in that, With a fixed label matrix F, extract the cost-sensitive feature matrix. The learning function is obtained Update rules: ； ； in, .

7. A face recognition method based on semi-supervised dual-source face feature fusion according to claim 5, characterized in that, Fixed updated cost-sensitive feature extraction matrix Based on the objective function of the label matrix F, the elements in the label matrix F are... Taking the partial derivative, we get: ； The update rules are as follows: ； Normalize the label matrix F, let .

8. The face recognition method based on semi-supervised dual-source face feature fusion according to claim 1, characterized in that, We use the cost-sensitive feature extraction matrix in a semi-supervised cost-sensitive canonical correlation analysis model to perform dual-source feature fusion, obtaining the feature representation X of the face image data. The expression for X is as follows: ； in, These represent the cost-sensitive feature extraction matrices for view 1 and view 2, respectively. A dataset of face images representing a 1-view view. A dataset of face images representing two views.

9. A face recognition method based on semi-supervised dual-source face feature fusion according to claim 1, characterized in that, The classification model uses a support vector machine model.

Images