WO2019119505A1

WO2019119505A1 - Face recognition method and device, computer device and storage medium

Info

Publication number: WO2019119505A1
Application number: PCT/CN2017/119465
Authority: WO
Inventors: 严蕤; 牟永强
Original assignee: 深圳云天励飞技术有限公司
Priority date: 2017-12-18
Filing date: 2017-12-28
Publication date: 2019-06-27
Also published as: CN108197532B; WO2019120115A1; CN108197532A

Abstract

A face recognition method and device, a computer device and a storage medium. The method comprises: extracting feature vectors of any two samples in a preset training set according to a preset face feature extraction model (S101); normalizing the feature vectors of the any two samples respectively (S102); acquiring a fusion feature vector of the any two samples (S103); acquiring a reference similarity of the any two samples (S104); sequentially traversing every two samples that are different from each other in the preset training set, and obtaining the fusion feature vector and reference similarity of every two samples that are different from each other in the preset training set (S105); training a regression model according to the fusion feature vector and reference similarity of every two samples that are different from each other in the preset training set (S106); and recognizing a face picture to be recognized by using the trained regression model (S107). In the present invention, a regression model is trained according to all fusion feature vectors and reference similarities in the preset training set, the trained regression model may effectively distinguish samples having different category markers, and the effect and accuracy of face recognition when carrying out recognition on a face picture to be recognized is thus improved.

Description

Method and device for face recognition, computer device and storage medium

The present application claims priority to Chinese Patent Application No. 201711366133.0, entitled "Face Recognition Method, Apparatus and Computer Device", filed on Dec. 18, 2017, the entire contents of In this application.

Technical field

The invention belongs to the field of image processing, and in particular relates to a method and device for face recognition, a computer device and a storage medium.

Background technique

Face recognition is a biometric technology based on human facial feature information for identification. It is widely used in the fields of identity verification, security monitoring, access control and attendance systems, and judicial criminal investigation. Face recognition mainly includes processes such as face detection, face alignment, face feature extraction, and face similarity determination. Among them, the determination of face similarity is an important part of face recognition, which can directly affect the result of face recognition. At present, the existing methods for determining face similarity mainly include: (1) a method for determining face similarity based on distance, such as Euclidean distance, cosine distance or Mahalanobis distance, but the method is ineffective, It is difficult to distinguish samples that are closer in feature space distribution. (2) A method of determining face similarity based on classification, such as a classification method of a support vector machine. However, the model complexity of the method increases with the increase of training data, resulting in large computational complexity and low computational efficiency, which leads to poor performance and low efficiency of subsequent face recognition.

Therefore, the existing method of face recognition has a problem of poor effect and low efficiency.

Summary of the invention

The invention provides a method and device for face recognition, a computer device and a storage medium, and aims to solve the problem that the existing face recognition method has poor effect and low efficiency.

A first aspect of the present invention provides a method for face recognition, the method comprising:

Extracting feature vectors of any two samples in the preset training set according to the preset face feature extraction model;

Normalizing the feature vectors of any two samples;

Obtaining a fusion feature vector of any two samples;

Obtaining a reference similarity of any two samples;

And sequentially traversing all the two samples that are different from each other in the preset training set, and obtaining a fusion feature vector and a reference similarity of all the two samples that are different from each other in the preset training set;

Determining the trained regression model according to the fusion feature vector of the two different samples and the reference similarity training regression model in the preset training set;

The face image to be recognized is identified by using the trained regression model.

In a preferred embodiment, the acquiring the fused feature vector of the any two samples includes:

And multiplying the elements of the corresponding dimension of the feature vectors of the normalized arbitrary two samples respectively, and using the multiplied result as the element of the corresponding dimension of the fusion feature vector of the arbitrary two samples, obtaining the arbitrary two Fusion feature vector of samples.

In a preferred embodiment, the preset training set includes a category identifier corresponding to the sample, and the reference similarity for acquiring the any two samples includes:

Determining a cosine distance of a feature vector of the normalized two samples;

If the class labels of the two samples are the same, the reference similarity of the two samples is the sum of the cosine distance and the preset constant;

If the class labels of the two samples are different, the reference similarity of the two samples is the difference between the cosine distance and the preset constant.

In a preferred embodiment, the regression model is determined according to the fusion feature vector of the two samples that are different from each other in the preset training set and the reference similarity training model, and the regression model after the training is determined to include:

Obtaining any fusion feature vector of the preset training set;

Entering any of the fusion feature vectors into a regression model to obtain training similarity of two samples corresponding to any of the fusion feature vectors, wherein the regression model includes at least a first fully connected layer and a second fully connected a layer, and the first fully connected layer and the second fully connected layer respectively perform feature mapping transformation on the any of the fusion feature vectors by using an activation function;

Determining, by using the loss function, an error of a training similarity of two samples corresponding to any one of the fusion feature vectors and a reference similarity of two samples corresponding to the one of the fusion feature vectors;

If the error does not satisfy the preset convergence condition, the parameter of the first fully connected layer and the parameter of the second fully connected layer of the regression model are adjusted by a process of back propagation using a random gradient descent;

The above iterative process is repeated until the error satisfies the preset convergence condition, and the parameters of the first fully connected layer and the parameters of the second fully connected layer of the last iterative process before the preset convergence condition are used as the first full regression model The parameters of the connection layer and the parameters of the second fully connected layer determine the regression model after training.

In a preferred embodiment, the preset convergence condition includes:

The error is less than or equal to a preset error threshold or the error percentage corresponding to the error is less than or equal to a preset error percentage.

In a preferred embodiment, the using the trained regression model to identify the face image to be recognized includes:

Obtaining a fusion feature vector of the first face image and the second face image to be verified;

And inputting a fusion feature vector of the first face image and the second face image to the trained regression model, and acquiring a similarity between the first face image and the second face image;

If the similarity between the first face image and the second face image is greater than or equal to a preset similarity threshold, determining that the first face image and the second face image are the same person Face picture

If the similarity between the first face image and the second face image is less than a preset similarity threshold, determining that the first face image and the second face image are not the same person's face image .

Obtaining a target face image to be retrieved;

Extracting a feature vector of the target face image and a feature vector of a face image included in the preset search database by using the preset face feature extraction model;

Determining, respectively, a feature vector of the feature vector of the target face image and a feature vector of each face image included in the preset search database;

Inputting a feature vector of the target face image and a feature vector of the feature vector of each face image included in the preset search database to the trained regression model, and acquiring the target face image a similarity to each face image included in the preset retrieval database;

Arranging each face image included in the preset search database in descending order of the similarity between the target face image and each face image included in the preset search database, and arranging The face image after the search is used as the search result.

In a preferred embodiment, the method further includes:

Determining, respectively, a feature vector of the target face image and a cosine distance of a feature vector of each face image included in the preset search database;

The face images included in the preset search database are arranged in descending order of cosine distance, and the face images ranked in the top N are used as candidate sets, where N is a positive integer;

The fusion feature vector for determining the feature vector of the target face image and the feature vector of each face image included in the preset search database respectively includes:

Determining, respectively, a feature vector of the feature vector of the target face image and a feature vector of each face image included in the candidate set;

And inputting a feature vector of the target face image and a feature vector of the feature vector of each face image included in the preset search database to the trained regression model, and acquiring the target person The similarity between the face picture and each face picture included in the preset search database includes:

And respectively inputting a feature vector of the target face image and a feature vector of the feature vector of each face image included in the candidate set to the trained regression model, and acquiring the target face image and the location Comparing the similarity of each face picture included in the candidate set;

And arranging, according to the similarity degree of each of the face images included in the target face image and the preset search database, the face images included in the preset search database, and The arranged face images as search results include:

Arranging the face images included in the candidate set in descending order of the similarity between the target face picture and each face picture included in the candidate set, and arranging the arranged face pictures As a result of the search.

A second aspect of the present invention provides a device for recognizing a face, the device comprising:

a feature vector extraction module, configured to extract feature vectors of any two samples in the preset training set according to the preset facial feature extraction model;

a normalization module, configured to respectively normalize feature vectors of any two samples;

a fusion feature vector acquisition module, configured to acquire a fusion feature vector of any two samples

Referring to a similarity obtaining module, configured to acquire a reference similarity of the any two samples;

a traversing acquisition module, configured to sequentially traverse all the two samples that are different from each other in the preset training set, and obtain a fusion feature vector and a reference similarity of all the two samples that are different from each other in the preset training set;

a training module, configured to determine a regression model after training according to the fusion feature vector of the two samples that are different from each other in the preset training set and the reference similarity training regression model;

The identification module is configured to identify the face image to be recognized by using the trained regression model.

A third aspect of the present invention provides a computer apparatus, comprising: a processor, wherein the processor is configured to implement a method of face recognition according to any of the above embodiments when executing a computer program stored in a memory.

A fourth aspect of the present invention provides a computer readable storage medium having stored thereon a computer program, the computer program being executed by a processor to implement the method for face recognition according to any of the above embodiments.

In the present invention, the feature vectors of two mutually different samples in the preset training set are fused, and the regression model after training is determined according to the fusion feature vectors of the two samples that are different from each other and the reference similarity training regression model. model. The fusion feature vector includes the texture feature and the dynamic mode feature of the face image. Therefore, the trained regression model can effectively distinguish the samples of different category markers, and use the trained regression model to identify the face image to be recognized. Can effectively improve the effect and accuracy of face recognition.

DRAWINGS

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the present invention. Other drawings can also be obtained from those skilled in the art based on these drawings without paying any creative effort.

1 is a flowchart of an implementation of a method for face recognition according to an embodiment of the present invention;

2 is a flowchart of implementing step S106 in the method for face recognition according to an embodiment of the present invention;

3 is a flowchart of implementing step S107 in the method for face recognition according to an embodiment of the present invention;

4 is a flowchart of another implementation of step S107 in the method for face recognition according to the embodiment of the present invention;

FIG. 5 is a flowchart of still another implementation of step S107 in the method for face recognition according to the embodiment of the present invention;

FIG. 6 is a functional block diagram of a device for face recognition according to an embodiment of the present invention; FIG.

FIG. 7 is a structural block diagram of a training module 106 in a device for face recognition according to an embodiment of the present invention;

FIG. 8 is a structural block diagram of an identification module 107 in a device for recognizing a face according to an embodiment of the present invention;

FIG. 9 is a block diagram showing another structure of the identification module 107 in the device for recognizing a face according to an embodiment of the present invention;

FIG. 10 is a block diagram showing another structure of the identification module 107 in the device for recognizing a face according to an embodiment of the present invention;

FIG. 11 is a schematic structural diagram of a computer apparatus according to an embodiment of the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

FIG. 1 shows an implementation flow of a method for face recognition according to an embodiment of the present invention. The order of the steps in the flowchart may be changed according to different requirements, and some steps may be omitted. For the convenience of description, only the parts related to the embodiment of the present invention are shown, which are detailed as follows:

As shown in FIG. 1, a method for face recognition includes:

Step S101: Extract feature vectors of any two samples in the preset training set according to the preset facial feature extraction model.

The preset facial feature extraction model is a pre-trained facial feature extraction model. Specifically, a large number of facial images can be used to learn facial feature extraction through convolutional neural network, and the trained facial feature extraction is established. Model, no longer detailed here. The preset training set is a preset training set containing a large number of face images, and can be set. It is assumed that the preset training set includes M samples (ie, a face picture) and a category mark corresponding to the sample, where M is a positive integer greater than 1. The category mark of the sample is based on whether the sample belongs to the same person's pre-set category mark. If two samples are the face image of the same person, the two samples belonging to the same person are one category mark, if the two samples are different people For face images, two samples that are not the same person are marked with different categories, and one category label may correspond to one or more samples.

Here, for convenience of description, it is assumed that any two samples include a first sample and a second sample, and the first sample and the second sample are different two samples, and the first sample and The second sample is described as an example.

It is assumed that the feature vector of the first sample and the feature vector of the second sample extracted according to the preset face feature extraction model are x=(x ₁ , x ₂ , x ₃ ... x _d-3 , x _d-2 , respectively). x _d ) and y=(y ₁ , y ₂ , y ₃ ... y _d-3 , y _d-2 , y _d ), the class label of the first sample and the class label of the second sample are z _i and z, respectively _j .

The value of d is a dimension of the feature vector, and is a positive integer greater than 1. Specifically, it may be set when the preset facial feature extraction model is trained, and is not particularly limited herein.

Step S102, normalizing the feature vectors of the arbitrary two samples respectively.

In order to further improve the effect and accuracy of the face recognition, after acquiring the feature vector of the first sample and the feature vector of the second sample, the normalized processing is performed, and the elements of the normalized feature vector are The ratio of the elements of the corresponding dimension to the modulus length of the feature vector.

It is assumed that the eigenvector of the normalized first sample and the eigenvector of the second sample are respectively

with

Step S103: Acquire a fusion feature vector of any two samples.

After obtaining the normalized first sample eigenvector and the normalized second sample eigenvector, the eigenvector of the normalized first sample and the normalized second The feature vectors of the samples are fused to obtain a fused feature vector of the first sample and the second sample.

In order to further improve the effect and accuracy of the face recognition, in a preferred embodiment, in step S103, acquiring the fusion feature vector of the any two samples includes:

Multiplying the normalized first sample's feature vector and the normalized second sample's feature vector's corresponding dimension elements, respectively, and multiplying the result as the first sample and the An element of a corresponding dimension of the second sample of the fusion feature vector obtains a fusion feature vector of the first sample and the second sample.

Assuming that the fusion feature vector is represented by m, then

Step S104: Acquire a reference similarity of any two samples.

Obtaining the eigenvector of the normalized first sample

And the normalized second sample eigenvector

Then, the reference similarity of the first sample and the second sample may be obtained according to the normalized first sample's feature vector and the normalized second sample's feature vector.

In order to further improve the effect and accuracy of the face recognition, in a preferred embodiment, in step S104, obtaining the reference similarity of the any two samples includes:

A cosine distance of the normalized first sample eigenvector and the normalized second sample eigenvector is determined.

The normalized feature vectors of the first sample and the second sample are respectively

with

Then the cosine distance of the two eigenvectors is:

Where x ⁱ · y ^j represents the dot product of the feature vector x ⁱ and the feature vector y ^j , and ||x ⁱ || ₂ and ||y ^j || ₂ represent the two vectors of the feature vector x ⁱ and the feature vector y ^j , respectively The number, the two-norm of the vector, refers to the sum of the square roots of the elements in the vector and the root number.

The cosine distance, also known as the cosine similarity, is the magnitude of the difference between two individuals using the cosine of the two vectors in the vector space, which can be used to characterize the first sample and the second sample. Similarity. In addition, the range of the cosine distance is [-1, +1], and the closer the distance is to 1, the closer the two vectors are to the same direction, that is, the positive correlation; the closer the distance is to -1, the more the direction of the two vectors Close to the opposite, that is, a negative correlation.

If the class labels of the two samples are the same, the reference similarity of the two samples is the sum of the cosine distance and the preset constant.

When the category mark of the first sample and the category mark of the second sample are marked as the same category mark, in order to increase the reference similarity of the samples of the same category mark, the reference of the first sample and the second sample is similar The degree is the sum of the cosine distance and a preset constant. The preset constant is a preset constant. In a preferred embodiment, the preset constant is 0.5.

Assuming that the reference similarity is represented by n, and the preset constant is represented by α, when the class mark of the first sample and the class mark of the second sample are the same, there is: n=cos(x ⁱ , y ^j )+α .

When the category mark of the first sample and the category mark of the second sample are marked as different category marks, in order to reduce the reference similarity of the samples of the different class marks, the reference of the first sample and the second sample is similar The degree is the difference between the cosine distance and the preset constant.

That is, when the category mark of the first sample and the category mark of the second sample are the same, there is: n=cos(x ⁱ , y ^j )−α.

Step S105: sequentially traverse all the two samples that are different from each other in the preset training set, and obtain a fusion feature vector and a reference similarity of all the two samples that are different from each other in the preset training set.

After acquiring the fusion feature vector of the first sample and the second sample and the reference similarity according to the above steps, the steps S101 to S104 are repeated to obtain the fusion features of all the two samples that are different from each other in the preset training set. The vector and the reference similarity, the two samples that are different from each other means that the two samples are different samples. It is assumed that the preset training set includes M samples, and each time two arbitrary samples are extracted from the preset training set until M*(M-1)/2 times are repeated, and the preset training set is completed. The extraction of any two mutually different samples, that is, repeating M*(M-1)/2 times steps S101 to S104, the fusion feature vectors of all the two samples different from each other in the preset training set can be obtained. And the reference similarity, and the obtained fusion feature vector and reference similarity of the two samples which are different from each other are used as the data of the training regression model. At this point, the construction of the regression model training data is completed, and the regression model is trained later. .

Step S106: Determine a regression model after training according to the fusion feature vector of the two samples that are different from each other in the preset training set and the reference similarity training regression model.

After obtaining the fused feature vector and the reference similarity of all the two different samples in the preset training set, the fused feature vector and the reference similarity of all the two different samples in the preset training set can be utilized. The regression model is trained, and after the training is terminated, the regression model after training is determined.

Step S107, using the trained regression model to identify the face image to be recognized.

After determining the regression model after training, the trained regression model can be used to identify the face image to be recognized. The recognition of the face image to be recognized mainly includes face verification and face retrieval. The face verification determines whether the two face images to be verified are face images of the same person, and the face search is based on the target face. The image is retrieved in the face database and the face image of the target face is the same person or the face image with the similarity of the target face image.

In the embodiment of the present invention, the feature vectors of all the two samples that are different from each other in the preset training set are merged, and the regression model is trained according to the fusion feature vectors of the two samples and the reference similarity training model to determine the training. Regression model. The fusion feature vector includes the texture feature and the dynamic mode feature of the face image. Therefore, the trained regression model can effectively distinguish the samples of different category markers, and use the trained regression model to identify the face image to be recognized. Can effectively improve the effect and accuracy of face recognition.

FIG. 2 shows an implementation flow of step S106 in the method for face recognition according to the embodiment of the present invention. The order of the steps in the flowchart may be changed according to different requirements, and some steps may be omitted. For the convenience of description, only the parts related to the embodiment of the present invention are shown, which are detailed as follows:

In a preferred embodiment, in order to further improve the effect and accuracy of the face recognition, as shown in FIG. 2, in step S106, according to the fusion feature vectors of the two samples that are different from each other in the preset training set. The regression model is trained with the reference similarity, and the regression model after the training is determined to include:

Step S1061: Acquire any fusion feature vector of the preset training set.

When the regression model is trained by using the fusion feature vector of the two samples different from each other in the preset training set and the reference similarity, any fusion feature vector of the preset training set is first acquired, and the fusion is performed. The feature vector is any one of the fusion feature vectors of the two different samples of the mutually different samples in the preset training set.

Step S1062: Input any of the fusion feature vectors into a regression model to obtain training similarity of two samples corresponding to any of the fusion feature vectors, wherein the regression model includes at least a first fully connected layer and a first The second fully connected layer, and the first fully connected layer and the second fully connected layer respectively perform feature mapping transformation on the any of the merged feature vectors by using an activation function.

In an embodiment of the present invention, the regression model includes at least a first fully connected layer and a second fully connected layer, and the first fully connected layer and the second fully connected layer both use an activation function for any of the fusions The feature vector is used for feature map transformation. In view of the correction of the linear unit (Rectified linear unit, Relu for short) activation function can accelerate the convergence of the regression model, improve the speed and efficiency of the regression model training, therefore, in a preferred embodiment, the first full connection Both the layer and the second fully connected layer perform feature mapping transformation on any of the fusion feature vectors using a Relu activation function. Alternatively, the first fully connected layer and the second fully connected layer may also adopt a variant function with a Relu activation function, such as a Leaky-Relu activation function or a P-Relu (English full name: Parametric-Relu) activation function or R-Relu. (English full name: Randomizied Relu) activation function and so on. After the fusion feature vector is input to the regression model, the first fully connected layer and the second fully connected layer of the regression model perform feature mapping transformation on the any fusion feature vector by using a Relu activation function. The training similarity of the two samples corresponding to any of the fusion feature vectors is outputted at the second fully connected layer.

Step S1063: Determine, by using the loss function, an error of the similarity between the training similarity of the two samples corresponding to the any fusion feature vector and the reference similarity of the two samples corresponding to the any fusion feature vector.

After the fusion feature is passed through the trained regression model to obtain the training similarity of the two samples corresponding to any of the fusion feature vectors, the loss function may be used to determine any of the fusion features. The error between the training similarity of the two samples corresponding to the vector and the reference similarity of the two samples corresponding to any of the fusion feature vectors. In a preferred embodiment, the L2 loss function is used to determine the training similarity of the two samples corresponding to any one of the fusion feature vectors and the reference similarity of the two samples corresponding to the any of the fusion feature vectors. Error, wherein the L2 loss (English name: Squared hinge loss, L2 loss for short) function is used to evaluate the degree of inconsistency between the predicted value and the true value. In the embodiment of the present invention, the L2 loss function is used to evaluate the training similarity. The degree of inconsistency with the reference similarity.

If the error does not satisfy the preset convergence condition, step S1064 is performed, and the parameter of the first fully connected layer of the regression model and the second fully connected layer are adjusted by a process of backpropagation by using a random gradient descent parameter.

The preset convergence condition is a pre-set convergence condition. In order to improve the calculation efficiency of the training regression model and reduce the calculation amount of the training regression model, in a preferred embodiment, the preset convergence condition includes: The error is less than or equal to the preset error threshold or the error percentage corresponding to the error is less than or equal to the preset error percentage. The preset error threshold and the preset error percentage are preset error thresholds, and are not particularly limited herein.

The stochastic gradient descent is mainly used to perform weight update in the neural network model, and the parameters of the model are updated and adjusted in one direction to minimize the loss function. The stochastic gradient descent randomly selects one sample from the training set at a time (in the embodiment of the present invention) In the middle, it refers to the fusion feature vector) to learn. Backpropagation is to calculate the product of the input signal and its corresponding weight in the forward propagation, then apply the activation function to the sum of these products, and then return the relevant error in the back propagation of the network model, using a random gradient. The update weight value is decreased, and the weight parameter is updated in the opposite direction of the loss function gradient by calculating the gradient of the error function with respect to the weight parameter. Therefore, in the embodiment of the present invention, if the error does not satisfy the preset convergence condition, the parameters of the first fully connected layer and the second full of the regression model are adjusted by a process of backpropagation by using a random gradient descent. The parameters of the connection layer.

After adjusting the parameters of the first fully connected layer and the parameters of the second fully connected layer, the process goes to step S1061, and steps S1061 to S1063 are repeatedly performed until the error satisfies a preset convergence condition.

If the error satisfies the preset convergence condition, step S1065 is performed, and the parameters of the first fully connected layer and the parameter of the second fully connected layer of the last iteration process before the preset convergence condition are used as the first full of the regression model. The parameters of the connection layer and the parameters of the second fully connected layer determine the regression model after training.

After the error satisfies the preset convergence condition, the training regression model is stopped, and the parameters of the first fully connected layer and the parameters of the second fully connected layer of the last iterative process before the preset convergence condition are used as the regression model. The parameters of the first fully connected layer and the parameters of the second fully connected layer determine the regression model after training, and thus the training of the regression model is completed.

In the embodiment of the present invention, the fusion feature vector of the preset training set includes the texture feature and the dynamic mode feature of the face image, and the regression model is trained by using the fusion feature vector of the preset training set, and the random gradient is used to Adjusting the parameters of the regression model to the process of propagation, and determining the regression model after training, therefore, the trained regression model can effectively distinguish samples of different categories of markers, and use the trained regression model to treat the recognized face images When the recognition is performed, the effect and accuracy of face recognition can be effectively improved.

FIG. 3 shows an implementation flow of step S107 in the method for face recognition according to the embodiment of the present invention. The order of the steps in the flowchart may be changed according to different requirements, and some steps may be omitted. For the convenience of description, only the parts related to the embodiment of the present invention are shown, which are detailed as follows:

In a preferred embodiment, in order to further improve the effect and accuracy of the face recognition, as shown in FIG. 3, step S107 includes:

Step S201: Acquire a fusion feature vector of the first face image and the second face image to be verified.

In order to verify whether the first face image and the second face image are the same person's face image, firstly, the fusion feature vector of the first face image and the second face image is extracted, and the specific extraction method and the foregoing extraction are performed. The method for fusing the feature vector of the first sample and the second sample is the same, that is, the feature vector of the first face image and the second face image is first extracted according to the preset face feature extraction model, and then the first face is extracted. The feature vector of the picture and the second face image is normalized, and finally the fusion feature vector of the first face image and the second face image is obtained. For details, refer to the related content of step S101 to step S103 above. More details will be described.

Step S202, input a fusion feature vector of the first face image and the second face image to the trained regression model, and acquire the first face image and the second face image. Similarity.

When the similarity between the first face image and the second face image is obtained by using the trained regression model according to the fusion feature vector of the first face image and the second face image, The content of the above step S1062 can be referred to, and details are not described herein again.

If the similarity between the first face image and the second face image is greater than or equal to the preset similarity threshold, step S203 is performed to determine that the first face image and the second face image are The face image of the same person.

The preset similarity threshold is a preset similarity, and is not particularly limited herein. When the similarity between the first face image and the second face image is greater than or equal to a preset similarity threshold, the first face image and the second face image may be determined to be the same person. Face picture.

If the similarity between the first face image and the second face image is less than the preset similarity threshold, step S204 is performed to determine that the first face image and the second face image are not the same person. Face picture.

And determining that the first face picture and the second face picture are not the same person's face picture, where the similarity between the first face picture and the second face picture is less than a preset similarity threshold. At this point, the verification of the first face picture and the second face picture is completed.

In the embodiment of the present invention, the trained regression model can effectively distinguish the face images of different categories of markers, and use the trained regression model to verify the first face image and the second face image to be identified, which can effectively determine the first The similarity between the face image and the second face image, thereby determining whether the first face image and the second face image are face images of the same person, and thus, the face recognition can be further improved. Performance and accuracy.

FIG. 4 shows another implementation flow of step S107 in the method for face recognition according to the embodiment of the present invention. The order of the steps in the flowchart may be changed according to different requirements, and some steps may be omitted. For the convenience of description, only the parts related to the embodiment of the present invention are shown, which are detailed as follows:

In a preferred embodiment, in order to further improve the effect and accuracy of the face recognition, as shown in FIG. 4, step S107 includes:

Step S301: Acquire a target face image to be retrieved.

In order to retrieve the face image, in order to return the face image with the similarity of the target face image to be retrieved in a certain range, it is first necessary to obtain the target face image to be retrieved. The target face image to be retrieved may be acquired by an image acquisition device, such as a camera or a camera, or the target face image to be retrieved may be obtained through a network, where the target face image to be retrieved is obtained. The route is not subject to special restrictions.

Step S302: Extract the feature vector of the target face image and the feature vector of the face image included in the preset search database by using the preset face feature extraction model.

The preset retrieval database is a preset retrieval database, which includes a large number of face images. For details, refer to the content of step S101 above, and details are not described herein again.

In a preferred embodiment, in order to further improve the effect and accuracy of the face recognition, step S107 further includes: separately performing a feature vector of the target face image and a face image included in the preset search database. The feature vector is normalized.

For the normalization process of the feature vector, the content of the above step S102 can be specifically referred to, and details are not described herein again.

Step S303, respectively determining a feature vector of the target face image and a feature vector of the feature vector of each face image included in the preset search database.

For example, when the feature vector of the target face image and the feature vector of the feature vector of each face image included in the preset search database are determined, the content of step S103 may be specifically referred to, and details are not described herein again.

Step S304, respectively input a feature vector of the target face image and a feature vector of a feature vector of each face image included in the preset search database to the trained regression model, and acquire the target. The similarity between the face picture and each face picture included in the preset search database.

When the similarity between the target face image and each face image included in the preset search database is obtained, the content of step S1062 may be specifically referred to, and details are not described herein again.

Step S305, arranging, according to the similarity between the target face image and each face image included in the preset search database, each face image included in the preset search database. The arranged face images are used as search results.

In order to display the face image most similar to the target face image to be retrieved in the front, the similarity between the target face image and each face image included in the preset search database is obtained. And each face image included in the preset search database may be arranged in descending order according to the similarity between the target face image and each face image included in the preset search database. The arranged face image is used as a retrieval result to return the retrieval result, for example, displayed on the display screen.

In the embodiment of the present invention, the image is identified by using the fusion feature vector and the trained regression model, and the similarity of each face image included in the target face image and the preset retrieval database is from large to small. The order of each face image included in the preset search database is arranged, and the arranged face image is used as a retrieval result, which can improve the accuracy of face retrieval, thereby improving the effect and accuracy of face recognition. .

FIG. 5 shows still another implementation flow of step S107 in the method for face recognition according to the embodiment of the present invention. The order of the steps in the flowchart may be changed according to different requirements, and some steps may be omitted. For the convenience of description, only the parts related to the embodiment of the present invention are shown, which are detailed as follows:

In a preferred embodiment, in order to further improve the effect and accuracy of the face recognition, as shown in FIG. 5, based on the steps shown in FIG. 4, step S107 further includes:

Step S306, respectively determining a cosine distance of the feature vector of the target face image and the feature vector of each face image included in the preset search database.

In view of the fact that the preset search database contains a large number of face images, all the face images in the preset search database are calculated by the above method, which generates a large amount of calculation. Since the cosine distance can also be used to characterize the similarity between the two vectors, in order to improve the efficiency of the face search and the subsequent face recognition, in the embodiment of the present invention, the target face picture may be first determined. And a cosine distance of the feature vector and the feature vector of each face image included in the preset retrieval database, to initially represent the similarity between the target face image and each face image included in the preset search database degree. For determining the cosine distance of the feature vector of the target face image and the feature vector of each face image included in the preset search database, refer to the determining the normalized first sample in step S104. The method of the feature vector and the cosine distance of the feature vector of the second sample will not be described in detail herein.

Step S307, the face images included in the preset search database are arranged in descending order of cosine distance, and the face images ranked in the top N are used as candidate sets, where N is a positive integer.

After determining the cosine distance of the feature vector of the target face image and the feature vector of each face image included in the preset search database, the preset search may be performed according to the cosine distance from the largest to the smallest. The face images included in the database are arranged, and the top N face images are used as candidate sets, so as to narrow the search range, reduce the calculation amount of face search and subsequent face recognition, and improve face search and follow-up persons. The efficiency of face recognition. The positive integer N can be set. For example, in a preferred embodiment, the positive integer N is 100, that is, a face image ranked in the top 100 is used as a candidate set, so as to subsequently determine the target face. The similarity between the picture and the 100 face pictures in the candidate set.

Correspondingly, in step S303, the determining a fusion feature vector of the feature vector of the target face image and the feature vector of each face image included in the preset retrieval database respectively includes:

Step S3031: respectively determine a feature vector of the target face image and a feature vector of the feature vector of each face image included in the candidate set.

After the search range is reduced and the candidate set meets certain conditions and requirements, the fusion feature vector of the feature vector of the target face image and the feature vector of each face image included in the candidate set may be determined. For details, refer to the content of step S303 above, and details are not described herein again.

Correspondingly, in step S304, the feature vector of the target face image and the feature vector of the feature vector of each face image included in the preset search database are respectively input to the trained regression model. And acquiring the similarity between the target face image and each face image included in the preset search database includes:

Step S3041: respectively input a feature vector of the target face image and a feature vector of a feature vector of each face image included in the candidate set to the trained regression model, and acquire the target face. The similarity of the picture and each face picture included in the candidate set.

Similarly, the content of the above step S304 can be referred to in step S3041, and details are not described herein again.

Correspondingly, in step S305, the face included in the preset search database is in descending order of the similarity of each face image included in the target face picture and the preset search database. The images are arranged and the arranged face images are included as search results:

Step S3051: Arrange the face images included in the candidate set in descending order of the similarity between the target face picture and each face picture included in the candidate set, and arrange the arranged faces. The face image is used as a search result.

Similarly, the content of step S305 can be referred to in step S3051, and details are not described herein again.

In the embodiment of the present invention, first, determining a cosine distance of a feature vector of the target face image and a feature vector of each face image included in the preset search database, in the order of cosine distance from large to small The face images included in the preset search database are arranged, and the face images ranked in the top N are used as candidate sets, and then the target face images are similar to each face image included in the candidate set. The face images included in the candidate set are arranged in descending order, and the arranged face images are used as search results. Therefore, since the cosine distance can preliminarily characterize the similarity between the pictures, by calculating the cosine distance, the plurality of face pictures ranked in the forefront most similar to the target face picture are first screened out as a candidate set for subsequent retrieval, therefore, The scope of the search can be narrowed, the retrieval speed can be improved, and the efficiency of face recognition can be improved.

FIG. 6 is a functional block diagram of a device for face recognition according to an embodiment of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown, which are described in detail as follows:

Referring to FIG. 6 , each module included in the apparatus 10 for face recognition is used to perform various steps in the corresponding embodiment of FIG. 1 . For details, refer to the related description in the corresponding embodiment of FIG. 1 and FIG. 1 . Narration. In the embodiment of the present invention, the device 10 for face recognition includes a feature vector extraction module 101, a normalization module 102, a fusion feature vector acquisition module 103, a reference similarity acquisition module 104, a traversal acquisition module 105, a training module 106, and Identification module 107.

The feature vector extraction module 101 is configured to extract feature vectors of any two samples in the preset training set according to the preset face feature extraction model.

The normalization module 102 is configured to normalize the feature vectors of the any two samples separately.

The fusion feature vector obtaining module 103 is configured to acquire a fusion feature vector of the arbitrary two samples.

The reference similarity obtaining module 104 is configured to obtain a reference similarity of the any two samples.

The traversal obtaining module 105 is configured to sequentially traverse all the two samples that are different from each other in the preset training set, and obtain a fusion feature vector and a reference similarity of all the two samples that are different from each other in the preset training set.

The training module 106 is configured to determine a regression model after training according to the fusion feature vector of the two samples that are different from each other in the preset training set and the reference similarity training regression model.

The identification module 107 is configured to identify the face image to be recognized by using the trained regression model.

In the embodiment of the present invention, the fused feature vector obtaining module 103 fuses feature vectors of two samples that are different from each other in the preset training set, and the training module 106 combines the eigenvectors of the two samples that are different from each other and the reference similarity. The regression model is trained to determine the regression model after training. The fusion feature vector includes the texture feature and the dynamic mode feature of the face image. Therefore, the trained regression model can effectively distinguish the samples of different category markers, and use the trained regression model to identify the face image to be recognized. Can effectively improve the effect and accuracy of face recognition.

In a preferred embodiment, in order to further improve the effect and accuracy of the face recognition, the fusion feature vector obtaining module 103 is specifically configured to:

And multiplying the elements of the corresponding dimension of the normalized feature vectors of the two samples respectively, and using the multiplied result as the element of the corresponding dimension of the fusion feature vector of the arbitrary two samples, obtaining the A fused feature vector of any two samples.

FIG. 7 shows the structure of the training module 106 in the device for face recognition according to the embodiment of the present invention. For the convenience of description, only the parts related to the embodiment of the present invention are shown, which are as follows:

In a preferred embodiment, with reference to FIG. 7, each unit included in the training module 106 is used to perform various steps in the corresponding embodiment of FIG. 2. For details, refer to the related embodiments in FIG. 2 and FIG. Description, no longer repeat here. In the embodiment of the present invention, the training module 106 includes a first obtaining unit 1061, a second obtaining unit 1062, an error determining unit 1063, a parameter adjusting unit 1064, and a regression model determining unit 1065.

The first acquiring unit 1061 is configured to acquire any fusion feature vector of the preset training set.

The second obtaining unit 1062 is configured to input any of the fusion feature vectors to a regression model, and obtain training similarity of two samples corresponding to any of the fusion feature vectors, wherein the regression model includes at least a first fully connected layer and a second fully connected layer, and the first fully connected layer and the second fully connected layer both perform feature mapping transformation on the any of the merged feature vectors by using an activation function.

The error determining unit 1063 is configured to determine, by using the loss function, an error of the similarity between the training similarity of the two samples corresponding to the any of the fusion feature vectors and the reference similarity of the two samples corresponding to the any of the fusion feature vectors. .

The parameter adjustment unit 1064 is configured to adjust a parameter of the first fully connected layer of the regression model and the first step by using a process of back propagation by using a random gradient descent if the error does not satisfy a preset convergence condition. The parameters of the two fully connected layers.

The regression model determining unit 1065 is configured to: when the error satisfies a preset convergence condition, use a parameter of the first fully connected layer and a parameter of the second fully connected layer that meet the last iteration process before the preset convergence condition The parameters of the first fully connected layer of the regression model and the parameters of the second fully connected layer are determined to determine the regression model after training.

In a preferred embodiment, the preset convergence condition includes:

FIG. 8 shows the structure of the identification module 107 in the device for face recognition according to the embodiment of the present invention. For the convenience of description, only parts related to the embodiment of the present invention are shown, which are described in detail as follows:

In a preferred embodiment, with reference to FIG. 8, each unit included in the identification module 107 is used to perform various steps in the corresponding embodiment of FIG. 3. For details, refer to the related embodiments in FIG. 3 and FIG. Description, no longer repeat here. In the embodiment of the present invention, the identification module 107 includes a fusion feature vector acquisition unit 201, a first similarity acquisition unit 202, and a determination unit 203.

The fused feature vector obtaining unit 201 is configured to acquire a fused feature vector of the first face image and the second face image to be verified.

The first similarity obtaining unit 202 is configured to input the fusion feature vector of the first facial image and the second facial image to the trained regression model to obtain the first facial image. The similarity with the second face picture.

The determining unit 203 is configured to determine the first face image and the second if the similarity between the first face image and the second face image is greater than or equal to a preset similarity threshold The face image is the face image of the same person.

The determining unit 203 is further configured to: if the similarity between the first face image and the second face image is less than a preset similarity threshold, determine the first face image and the second person The face image is not the face image of the same person.

In the embodiment of the present invention, the fusion feature vector obtaining unit 201 acquires a fusion feature vector of the first face image and the second face image to be verified, and the first similarity acquisition unit 202 is configured according to the fusion feature vector. Acquiring the similarity between the first face image and the second face image, the determining unit 203 compares the similarity with a preset similarity threshold, and further determines the first face image and the location Whether the second face image is a face image of the same person. In the embodiment of the present invention, the first similarity acquiring unit 202 determines the similarity between the first facial image and the second facial image according to the fusion feature vector, and the determining unit 203 determines the first facial image and the location. Whether the second face image is a face image of the same person, therefore, the effect and accuracy of the face recognition can be further improved.

FIG. 9 shows another structure of the identification module 107 in the device for face recognition according to the embodiment of the present invention. For convenience of description, only parts related to the embodiment of the present invention are shown, which are described in detail as follows:

In a preferred embodiment, with reference to FIG. 9, each unit included in the identification module 107 is used to perform various steps in the corresponding embodiment of FIG. 4. For details, refer to the related embodiments in FIG. 4 and FIG. Description, no longer repeat here. In the embodiment of the present invention, the identification module 107 includes a target face image acquisition unit 301, a feature vector extraction unit 302, a fusion feature vector determination unit 303, a second similarity acquisition unit 304, and a retrieval result determination unit 305.

The target face image obtaining unit 301 is configured to acquire a target face image to be retrieved.

The feature vector extracting unit 302 is configured to separately extract a feature vector of the target face image and a feature vector of a face image included in the preset search database by using the preset face feature extraction model.

The fused feature vector determining unit 303 is configured to respectively determine a fused feature vector of the feature vector of the target face image and the feature vector of each face image included in the preset search database.

The second similarity obtaining unit 304 is configured to input the feature vector of the target face image and the feature vector of the feature vector of each face image included in the preset search database to the training Regression model, and obtain similarity between the target face image and each face image included in the preset search database.

The search result determining unit 305 is configured to select, according to the similarity between the target face image and each face image included in the preset search database, the preset search database Each face image is arranged, and the arranged face image is used as a search result.

In the embodiment of the present invention, the image is identified by using the fusion feature vector and the trained regression model, and the retrieval result determination unit 305 is similar to each face image included in the target face image and the preset retrieval database. The face images included in the preset search database are arranged in descending order, and the arranged face images are used as search results. Therefore, the effect and accuracy of face recognition can be further improved.

FIG. 10 shows still another structure of the identification module 107 in the device for face recognition according to the embodiment of the present invention. For the convenience of description, only parts related to the embodiment of the present invention are shown, which are described in detail as follows:

In a preferred embodiment, referring to FIG. 10, each unit or subunit included in the identification module 107 is used to perform various steps in the corresponding embodiment of FIG. 5. For details, refer to FIG. 5 and FIG. The related descriptions are not repeated here. In the embodiment of the present invention, the identification module 107 further includes a cosine distance determining unit 306 and a candidate set determining unit 307 on the basis of the structure shown in FIG. 9 . Correspondingly, the fusion feature vector determining unit 303 includes a fusion feature vector determining subunit 3031, the second similarity acquiring unit 304 includes a similarity acquiring subunit 3041, and the retrieval result determining unit 305 includes a retrieval result determining subunit. 3051.

The cosine distance determining unit 306 is configured to respectively determine a cosine distance of a feature vector of the target face image and a feature vector of each face image included in the preset search database.

The candidate set determining unit 307 is configured to arrange the face images included in the preset search database according to the cosine distance from the largest to the smallest, and use the face images ranked in the top N as the candidate set. Where N is a positive integer.

The fused feature vector determining sub-unit 3031 is configured to respectively determine a fused feature vector of a feature vector of the target face image and a feature vector of each face image included in the candidate set.

The similarity acquisition sub-unit 3041 is configured to respectively input a feature vector of the target face image and a feature vector of a feature vector of each face image included in the candidate set to the trained regression model. And acquiring the similarity of the target face picture and each face picture included in the candidate set.

The search result determining sub-unit 3051 is configured to select a face image included in the candidate set according to the similarity degree of the target face picture and each face picture included in the candidate set in descending order Arrange and arrange the arranged face images as the search results.

In the embodiment of the present invention, the cosine distance determining unit 306 determines a cosine distance of the feature vector of the target face image and a feature vector of each face image included in the preset search database, and the candidate set determining unit 307 follows the cosine distance. The face images included in the preset search database are arranged in descending order, and the face images ranked in the top N are used as candidate sets, and the search result determining subunit 3051 follows the target face image. And matching the face images included in the candidate set with the similarity degree of each face image included in the candidate set, and arranging the arranged face images as a retrieval result, in view of the cosine distance It is possible to preliminarily characterize the similarity between pictures, and by calculating the cosine distance, the plurality of face pictures ranked in the forefront most similar to the target face picture are first screened out as a candidate set for subsequent retrieval, thereby narrowing the scope of the search. Improve search speed and improve the efficiency of face recognition.

FIG. 11 is a schematic structural diagram of a computer apparatus 1 according to a preferred embodiment of a method for implementing face recognition according to an embodiment of the present invention. As shown in FIG. 11, the computer device 1 includes a memory 11, a processor 12, and an input/output device 13.

The computer device 1 is a device capable of automatically performing numerical calculation and/or information processing according to an instruction set or stored in advance, and the hardware includes, but not limited to, a microprocessor, an application specific integrated circuit (ASIC). ), Field-Programmable Gate Array (FPGA), Digital Signal Processor (DSP), embedded devices, etc.

The computer device 1 can be any electronic product that can interact with a user, such as a personal computer, a tablet computer, a smart phone, a personal digital assistant (PDA), a game machine, an interactive network television ( Internet Protocol Television (IPTV), smart wearable devices, etc. The computer device 1 may be a server, including but not limited to a single network server, a server group composed of a plurality of network servers, or a cloud computing-based cloud composed of a large number of hosts or network servers, wherein the cloud Computation is a type of distributed computing, a super-virtual computer consisting of a cluster of loosely coupled computers. The network in which the computer device 1 is located includes, but is not limited to, the Internet, a wide area network, a metropolitan area network, a local area network, a virtual private network (VPN), and the like.

The memory 11 is used to store programs of various methods of face recognition and various data, and realizes high-speed, automatic completion of access of programs or data during the operation of the computer device 1. The memory 11 may be an external storage device and/or an internal storage device of the computer device 1. Further, the memory 11 may be a circuit having a storage function in a physical form, such as a RAM (Random-Access Memory), a FIFO (First In First Out), or the like, or the memory 11 It may be a storage device having a physical form, such as a memory stick, a TF card (Trans-flash Card), or the like.

The processor 12 can be a Central Processing Unit (CPU). The CPU is a very large-scale integrated circuit, which is the computing core (Core) and the Control Unit of the computer device 1. The processor 12 can execute an operating system of the computer device 1 and various installed applications, program codes, and the like, such as an operating system in each module or unit in the device 10 for performing face recognition, and various installed applications and programs. Code to implement face recognition methods.

The input/output device 13 is mainly used to implement an input/output function of the computer device 1, such as transceiving input digital or character information, or displaying information input by a user or information provided to a user and various menus of the computer device 1.

The modules/units integrated by the computer device 1 can be stored in a computer readable storage medium if implemented in the form of a software functional unit and sold or used as a stand-alone product. Based on such understanding, the present invention implements all or part of the processes in the foregoing embodiments, and may also be completed by a computer program to instruct related hardware. The computer program may be stored in a computer readable storage medium. The steps of the various method embodiments described above may be implemented when the program is executed by the processor. Wherein, the computer program comprises computer program code, which may be in the form of source code, object code form, executable file or some intermediate form. The computer readable medium may include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a removable hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM). , random access memory (RAM, Random Access Memory), electrical carrier signals, telecommunications signals, and software distribution media. It should be noted that the content contained in the computer readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in a jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, computer readable media Does not include electrical carrier signals and telecommunication signals.

The above-described characteristic means of the present invention can be realized by an integrated circuit and control the function of the living body detecting method described in any of the above embodiments. That is, the integrated circuit of the present invention is mounted in the computer device 1 such that the computer device 1 functions as follows:

Normalizing the feature vectors of any two samples;

Obtaining a fusion feature vector of any two samples;

Obtaining a reference similarity of any two samples;

In any of the embodiments, the functions of the living body detecting method can be installed in the computer device 1 by the integrated circuit of the present invention, so that the computer device 1 can perform the living body detecting method in any of the embodiments. The functions implemented are not detailed here.

In the several embodiments provided by the present invention, it should be understood that the disclosed system, apparatus, and method may be implemented in other manners. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and the actual implementation may have another division manner.

The modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of hardware plus software function modules.

It is apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, and the present invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the present embodiments are to be considered as illustrative and not restrictive, and the scope of the invention is defined by the appended claims instead All changes in the meaning and scope of equivalent elements are included in the present invention. Any accompanying drawings in the claims should not be construed as limiting the claim. In addition, it is to be understood that the word "comprising" does not exclude other elements or steps. A plurality of modules or devices recited in the system claims may also be implemented by a module or device by software or hardware. The first, second, etc. words are used to denote names and do not denote any particular order.

It should be noted that the above embodiments are only for explaining the technical solutions of the present invention and are not intended to be limiting, and the present invention will be described in detail with reference to the preferred embodiments. Modifications or equivalents are made without departing from the spirit and scope of the invention.

Claims

A method for face recognition, characterized in that the method comprises:

Extracting feature vectors of any two samples in the preset training set according to the preset face feature extraction model;

Normalizing the feature vectors of any two samples;

Obtaining a fusion feature vector of any two samples;

Obtaining a reference similarity of any two samples;

And sequentially traversing all the two samples that are different from each other in the preset training set, and obtaining a fusion feature vector and a reference similarity of all the two samples that are different from each other in the preset training set;

Determining the trained regression model according to the fusion feature vector of the two different samples and the reference similarity training regression model in the preset training set;

The face image to be recognized is identified by using the trained regression model.
The method according to claim 1, wherein the acquiring the fused feature vector of the arbitrary two samples comprises:

And multiplying the elements of the corresponding dimension of the normalized feature vectors of the two samples respectively, and using the multiplied result as the element of the corresponding dimension of the fusion feature vector of the arbitrary two samples, obtaining the A fused feature vector of any two samples.
The method according to claim 1, wherein the preset training set includes a category mark corresponding to the sample, and the reference similarity for acquiring the any two samples comprises:

Determining a cosine distance of a feature vector of the normalized two samples;

If the class labels of the two samples are the same, the reference similarity of the two samples is the sum of the cosine distance and the preset constant;

If the class labels of the two samples are different, the reference similarity of the two samples is the difference between the cosine distance and the preset constant.
The method according to claim 1, wherein the regression model is determined according to the fusion feature vector of the two samples that are different from each other in the preset training set and the reference similarity training model, and the regression model after the training is determined to include:

Obtaining any fusion feature vector of the preset training set;

Entering any of the fusion feature vectors into a regression model to obtain training similarity of two samples corresponding to any of the fusion feature vectors, wherein the regression model includes at least a first fully connected layer and a second fully connected a layer, and the first fully connected layer and the second fully connected layer respectively perform feature mapping transformation on the any of the fusion feature vectors by using an activation function;

Determining, by using the loss function, an error of a training similarity of two samples corresponding to any one of the fusion feature vectors and a reference similarity of two samples corresponding to the one of the fusion feature vectors;

If the error does not satisfy the preset convergence condition, the parameter of the first fully connected layer and the parameter of the second fully connected layer of the regression model are adjusted by a process of back propagation using a random gradient descent;

The above iterative process is repeated until the error satisfies the preset convergence condition, and the parameters of the first fully connected layer and the parameters of the second fully connected layer of the last iterative process before the preset convergence condition are used as the first full regression model The parameters of the connection layer and the parameters of the second fully connected layer determine the regression model after training.
The method of claim 4, wherein the predetermined convergence condition comprises:

The error is less than or equal to a preset error threshold or the error percentage corresponding to the error is less than or equal to a preset error percentage.
The method according to claim 1, wherein the identifying the face image to be recognized by using the trained regression model comprises:

Obtaining a fusion feature vector of the first face image and the second face image to be verified;

And inputting a fusion feature vector of the first face image and the second face image to the trained regression model, and acquiring a similarity between the first face image and the second face image;

If the similarity between the first face image and the second face image is greater than or equal to a preset similarity threshold, determining that the first face image and the second face image are the same person Face picture

If the similarity between the first face image and the second face image is less than a preset similarity threshold, determining that the first face image and the second face image are not the same person's face image .
The method according to claim 1, wherein the identifying the face image to be recognized by using the trained regression model comprises:

Obtaining a target face image to be retrieved;

Extracting a feature vector of the target face image and a feature vector of a face image included in the preset search database by using the preset face feature extraction model;

Determining, respectively, a feature vector of the feature vector of the target face image and a feature vector of each face image included in the preset search database;

Inputting a feature vector of the target face image and a feature vector of the feature vector of each face image included in the preset search database to the trained regression model, and acquiring the target face image a similarity to each face image included in the preset retrieval database;

Arranging each face image included in the preset search database in descending order of the similarity between the target face image and each face image included in the preset search database, and arranging The face image after the search is used as the search result.
The method of claim 7 wherein the method further comprises:

Determining, respectively, a feature vector of the target face image and a cosine distance of a feature vector of each face image included in the preset search database;

The face images included in the preset search database are arranged in descending order of cosine distance, and the face images ranked in the top N are used as candidate sets, where N is a positive integer;

The fusion feature vector for determining the feature vector of the target face image and the feature vector of each face image included in the preset search database respectively includes:

Determining, respectively, a feature vector of the feature vector of the target face image and a feature vector of each face image included in the candidate set;

And inputting a feature vector of the target face image and a feature vector of the feature vector of each face image included in the preset search database to the trained regression model, and acquiring the target person The similarity between the face picture and each face picture included in the preset search database includes:

And respectively inputting a feature vector of the target face image and a feature vector of the feature vector of each face image included in the candidate set to the trained regression model, and acquiring the target face image and the location Comparing the similarity of each face picture included in the candidate set;

And arranging, according to the similarity degree of each of the face images included in the target face image and the preset search database, the face images included in the preset search database, and The arranged face images as search results include:

Arranging the face images included in the candidate set in descending order of the similarity between the target face picture and each face picture included in the candidate set, and arranging the arranged face pictures As a result of the search.
A device for recognizing a face, characterized in that the device comprises:

And a feature vector extraction module, configured to extract feature vectors of any two samples in the preset training set according to the preset face feature extraction model, where the preset training set includes a category tag corresponding to the sample;

a normalization module, configured to respectively normalize feature vectors of any two samples;

a fusion feature vector acquisition module, configured to acquire a fusion feature vector of any two samples

Referring to a similarity obtaining module, configured to acquire a reference similarity of the any two samples;

a traversing acquisition module, configured to sequentially traverse all the two samples that are different from each other in the preset training set, and obtain a fusion feature vector and a reference similarity of all the two samples that are different from each other in the preset training set;

a training module, configured to determine a regression model after training according to the fusion feature vector of the two samples that are different from each other in the preset training set and the reference similarity training regression model;

The identification module is configured to identify the face image to be recognized by using the trained regression model.
A computer apparatus, comprising: a processor, the processor for performing a face recognition method according to any one of claims 1-8 when executing a computer program stored in a memory.
A computer readable storage medium having stored thereon a computer program, wherein the computer program is executed by a processor to implement the method of face recognition according to any one of claims 1 to 8.