CN107967461A - The training of SVM difference models and face verification method, apparatus, terminal and storage medium - Google Patents

Abstract

A kind of support vector machines difference model training method, the described method includes：Construct positive and negative sample set and training SVM difference models, the feature of wherein positive and negative sample set carries out the face characteristic normalization feature that simultaneously difference processing obtains for an action shot and certificate photograph, positive and negative sample training collection and positive and negative test sample collection are generated from the positive and negative sample set constructed, it is corresponding SVM difference models so as to which training obtains optimized parameter c, g combination.The present invention also provides a kind of support vector machines difference model training device, terminal and storage medium and face verification method, apparatus, terminal and storage medium.The present invention can train the support vector machines difference model of suitable face verification, obtain preferable face verification effect.

Description

SVM differential model training and face verification method, device, terminal and storage medium

technical field

The present invention relates to the technical field of image recognition, in particular to a SVM differential model training and face verification method, device, terminal and storage medium.

Background technique

Face Verification is a sub-field of Face Recognition. Face Verification is to judge whether two face pictures are the same person. The most commonly used scenario is to judge whether the certificate is the person. Given a face picture and then judging who the person is, it is essentially equivalent to multiple face verifications.

Face verification in a dynamic environment is very difficult due to the influence of illumination, posture, age, and clothing on face pictures in a dynamic environment. In recent years, many methods have been proposed to improve face verification in dynamic environments, and these methods can be roughly divided into two categories. One is the method based on traditional features, the goal is to extract distinguishing features, and then combine Euclidean distance or cosine angle distance for comparison. The classic face feature description operators include: Histogram of Oriented Gradient (Histogram of Oriented Gradient) , HOG), local binary patterns (Local Binary Patterns, LBP features), scale-invariant feature transform (Scale-invariant feature transform, SIFT), Gabor (Gabor) features, etc. The other is feature expression based on deep learning.

Although the traditional feature-based method can obtain faster speed, but the feature expression ability is weak, so the performance is poor. The feature expression based on deep learning requires a very large sample for model training. Secondly, to obtain a better expression effect requires a deeper network to maintain. In addition, since there are many differences in age, lighting, expression and posture between ID photos and photos collected arbitrarily under dynamic conditions, the feature space distribution between the two is quite different, even if the feature expression based on deep learning is used. Carry out model training.

Contents of the invention

In view of the above, it is necessary to propose a SVM differential model training and face verification method, device, terminal and storage medium, which can train a classification model suitable for face verification, and distinguish the features trained under the situation of a small number of samples. Strong ability, get better face verification effect.

The first aspect of the present application provides a kind of SVM difference model training method, described method comprises:

Construct positive and negative sample sets, including:

1) Extract the first facial feature of the face area on each ID photo, and extract the second facial feature of the face area on each dynamic photo;

2) carrying out normalization processing to described first human face feature to obtain the first normalized human face feature, carrying out normalization processing to described second human face feature to obtain the second normalized human face feature;

3) performing a difference to the first normalized face feature and the second normalized face feature to obtain the difference face feature;

4) Construct positive and negative sample pairs, wherein, the positive sample pair is the feature and the first category attribute obtained by normalizing and differentially processing facial features and first category attributes of the same person's dynamic photo and ID photo, and the negative sample pair is different people's The features and second category attributes obtained by normalizing face features and differential processing of dynamic photos and ID photos;

Train the SVM difference model, including:

1): Generate a positive and negative sample training set and a positive and negative sample test set from the constructed positive and negative sample set;

2): Find the optimal combination of the penalty parameter c and the parameter g of the kernel function;

3): Save the optimal parameter combination and the corresponding SVM difference model.

In a possible implementation, the generating a positive and negative sample training set and a positive and negative sample test set from the constructed positive and negative sample set includes:

In the generated positive and negative sample training set, randomly select the first preset number of positive and negative sample training sets to participate in training; in the generated positive and negative sample test set, randomly select the second preset number of positive and negative sample test sets to participate in the test.

In a possible implementation, the optimal combination of the parameter g of the search penalty parameter c and kernel function includes:

The positive and negative sample training sets of the first preset quantity are input in the SVM, and the first optimal combination parameters c, g are calculated;

Gradually expand the range of c and g and reduce the step size, and save the combined parameters c and g for each time;

Select the positive and negative sample test set to test on the SVM difference model corresponding to the saved combination parameters c and g, and the corresponding parameters with the highest accuracy rate are the second optimal combination parameters c and g.

The second aspect of the present application provides a kind of SVM difference model that described SVM difference model training method trains and carries out face verification method, and described method comprises:

Extract the third face feature of the face area in the ID photo of the person to be verified;

Extracting the fourth face feature of the face area in the scene photo of the person to be verified;

Carrying out normalization processing to described the 3rd human face feature obtains the 3rd normalized human face feature, carries out normalization processing to described 4th human face feature and obtains the 4th normalized human face feature;

The third normalized face feature and the fourth normalized face feature are differenced to obtain the difference face feature of the person to be verified; and

Calculate the similarity of the difference facial features according to the SVM difference model;

judging whether the similarity is greater than a preset threshold; and

When the similarity is greater than the preset threshold, it is determined that the ID photo and the scene photo are not the same person; or

When the similarity is less than or equal to a preset threshold, it is determined that the ID photo and the scene photo are of the same person.

A third aspect of the present application provides an SVM differential model training device, the device comprising:

Construction modules, used to construct positive and negative sample sets, including:

The feature extraction submodule is used to extract the first facial feature of the face area on each ID photo, and extract the second facial feature of the face area on each dynamic photo;

A normalization sub-module, configured to perform normalization processing on the first facial feature to obtain a first normalized facial feature, and perform normalization processing on the second facial feature to obtain a second normalization facial features;

Difference sub-module, is used to carry out difference to described first normalized human face feature and second normalized human face feature to obtain differential human face feature;

Constructing sub-modules for constructing positive and negative sample pairs, wherein the positive sample pairs are the features and first category attributes obtained by normalizing face features and differential processing of dynamic photos and ID photos of the same person, and the negative sample pairs The features and second category attributes obtained by normalizing face features and differential processing for dynamic photos and ID photos of different people;

The training module is used to train the SVM difference model, including:

Generate a submodule, which is used to generate a positive and negative sample training set and a positive and negative sample test set from the constructed positive and negative sample set;

The optimization sub-module is used to find the optimal combination of the penalty parameter c and the parameter g of the kernel function;

The saving sub-module is used to save the optimal parameter combination and the corresponding SVM difference model.

In a possible implementation, the generating submodule generates a positive and negative sample training set and a positive and negative sample test set from the constructed positive and negative sample sets including:

In the generated positive and negative sample training set, randomly select the first preset number of positive and negative sample training sets to participate in training; in the generated positive and negative sample test set, randomly select the second preset number of positive and negative sample test sets to participate in the test.

In a possible implementation, the optimal combination of the optimization sub-module looking for the penalty parameter c and the parameter g of the kernel function includes:

The positive and negative sample training sets of the first preset quantity are input in the SVM, and the first optimal combination parameters c, g are calculated;

Gradually expand the range of c and g and reduce the step size, and save the combined parameters c and g for each time;

Select the positive and negative sample test set to test on the SVM difference model corresponding to the saved combination parameters c and g, and the corresponding parameters with the highest accuracy rate are the second optimal combination parameters c and g.

The fourth aspect of the present application provides a device for face verification using the SVM differential model trained by the SVM differential model training device, the device comprising:

The first extraction module is used to extract the third face feature of the face area in the ID photo of the person to be verified;

The second extraction module is used to extract the fourth facial feature of the human face area in the scene photo of the person to be verified;

A normalization module, configured to perform normalization processing on the third face feature to obtain a third normalized face feature, and perform normalization processing on the fourth face feature to obtain a fourth normalized face feature. facial features;

A difference module, used to make a difference between the third normalized face feature and the fourth normalized face feature to obtain the differential face feature of the person to be verified; and

A verification module, configured to calculate the similarity of the differential facial features according to the SVM differential model; and

The verification module is also used to judge whether the similarity is greater than a preset threshold; and

When the similarity is greater than the preset threshold, the verification module determines that the ID photo and the scene photo are not the same person; or

When the similarity is less than or equal to a preset threshold, the verification module determines that the ID photo and the scene photo are of the same person.

A fifth aspect of the present application provides a terminal, the terminal includes a processor, and the processor is configured to implement a support vector machine differential model training method or the face verification method when executing a computer program stored in a memory.

A sixth aspect of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, a support vector machine difference model training method or the face verification method is implemented.

The support vector machine difference model training, face verification method, device, terminal and storage medium described in the present invention combine the idea of support vector machine and difference model, apply it to face verification, and can solve the problem of certificate photos There is a difference between the feature space distribution and the feature space distribution of randomly collected photos under dynamic conditions; only a small number of samples are needed in the model training stage, which solves the problem of the algorithm’s demand for data volume, increases the practicability of the algorithm, and improves The accuracy of witness verification.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention, and those skilled in the art can also obtain other drawings according to the provided drawings without creative work.

FIG. 1 is a flowchart of a method for training a support vector machine difference model provided by Embodiment 1 of the present invention.

Fig. 2 is a flow chart of the face verification method provided by Embodiment 2 of the present invention.

FIG. 3 is a structural diagram of a support vector machine differential model training device provided by Embodiment 3 of the present invention.

Fig. 4 is a structural diagram of a face verification device provided in Embodiment 4 of the present invention.

FIG. 5 is a schematic diagram of a terminal provided in Embodiment 5 of the present invention.

The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings.

Detailed ways

In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

In the following description, many specific details are set forth in order to fully understand the present invention, and the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field of the invention. The terms used herein in the description of the present invention are for the purpose of describing specific embodiments only, and are not intended to limit the present invention.

Preferably, the face verification method of the present invention is applied to one or more terminals. The terminal is a device that can automatically perform numerical calculations and/or information processing according to preset or stored instructions, and its hardware includes but not limited to microprocessors, application specific integrated circuits (ASIC), Programmable gate array (Field-Programmable Gate Array, FPGA), digital processor (Digital Signal Processor, DSP), embedded devices, etc.

The terminal is not limited to a personal computer, a smart phone, a tablet computer, a desktop computer or an all-in-one computer equipped with a camera, or computing equipment such as a desktop computer, a notebook, a handheld computer, and a cloud server. The terminal can perform human-computer interaction with the user through keyboard, mouse, remote control, touch panel or voice control equipment.

The support vector machine difference model training method and the face verification method can also be applied to a hardware environment composed of a terminal and a server connected to the terminal through a network. Networks include, but are not limited to: Wide Area Networks, Metropolitan Area Networks, or Local Area Networks. The support vector machine difference model training method and the face verification method in the embodiment of the present invention can be executed by the server at the same time, and can also be executed by the terminal at the same time; it can also be executed by the server and the terminal, for example, the support vector machine difference The model training method is executed by the server, the face verification method is executed by the terminal, or the face verification method is executed by the server, and the support vector machine difference model training method is executed by the terminal. The application is not limited here.

For terminals that need support vector machine differential model training method and face verification method, the support vector machine differential model training function and face verification function provided by the method of this application can be directly integrated on the terminal, or installed to realize this application The client side of the application method. For another example, the method provided by this application can also run on devices such as a server in the form of a software development kit (Software Development Kit, SDK), and provide an interface for the support vector machine differential model training function and the face verification function in the form of an SDK. , terminals or other devices can realize face verification through the provided interface.

Embodiment one

FIG. 1 is a flowchart of a method for training a support vector machine difference model provided by Embodiment 1 of the present invention. According to different requirements, the execution sequence in the flowchart shown in Figure 1 can be changed, and some steps can be omitted.

101: Construct positive and negative sample sets.

In this preferred embodiment, prepare a photo of a certificate of different people and a photo randomly collected under a dynamic situation corresponding to the same person.

Described certificate photo refers to the photo that is used to prove identity on the certificate, can be ID card photo, can be passport photo, can also be various visa photos, or driver's license photo, graduation certificate photo etc.

In this preferred embodiment, the LFW data set can be used to construct positive and negative sample sets. The LFW data set is established to study the face recognition problem in an unrestricted environment, which contains more than 13,000 face images. The face images are all from the Internet, not a laboratory environment.

For the convenience of the following description, the photos randomly collected under dynamic conditions are called dynamic photos.

The construction of positive and negative sample sets specifically includes:

1) Extract the first facial feature of the face area on each ID photo, and extract the second facial feature of the face area on each dynamic photo;

In this preferred embodiment, the face feature corresponding to the ID photo of the i-th person, that is, the first face feature of the i-th person is denoted as _xi , and the face feature corresponding to the dynamic photo of the i-th person, that is, the first face feature of the i-th person The second face feature of person i is denoted as X _i .

In this preferred embodiment, before extracting the facial features, a pre-stored face detection algorithm can be used to detect the face area on the ID photo and the face area on the dynamic photo, and then use the pre-stored face feature The extraction algorithm performs face feature extraction on the detected face area.

The pre-stored face detection algorithm can be one or more combinations of the following algorithms: template-based face detection method, artificial neural network-based face detection method, model-based face detection method, skin color-based The face detection method or the face detection method based on the feature sub-face, etc.

The face feature extraction algorithm stored in advance can be one or more combinations of the following algorithms: Gabor feature, histogram of oriented gradient (Histogram of Oriented Gradient, HOG), local binary pattern (LocalBinary Patterns, LBP), Principal component analysis (Principal Component Analysis, PCA), linear discriminant analysis (Linear Discriminant Analysis, LDA) or independent component analysis (independent component analysis, ICA), etc. Preferably, because the deep learning feature has better feature expression ability, in this embodiment, the deep learning algorithm is used to extract the face features of the face area on the ID photo and the dynamic photo.

In this embodiment, the pre-stored face detection algorithm and the pre-stored face feature extraction algorithm are not limited to those listed above, any algorithm suitable for detecting the face area and extracting the face features of the face area algorithm can be applied to this. In addition, the pre-stored face detection algorithm and face feature extraction algorithm in this embodiment are all existing technologies, and will not be described in detail herein.

2) carrying out normalization processing to described first human face feature to obtain the first normalized human face feature, carrying out normalization processing to described second human face feature to obtain the second normalized human face feature;

In this preferred embodiment, the first face feature of the ith person is normalized to obtain the first normalized face feature as The second face feature of the i-th person is normalized to obtain the second normalized face feature as

It should be understood that ||x|| ₂ is the 1/2 power of the sum of the squares of the elements of the x vector, that is, the Euclidean distance in the sense of the L2 norm.

3) performing a difference to the first normalized face feature and the second normalized face feature to obtain the differential face feature;

In this preferred embodiment, the i-th person's differential face feature is denoted as

It should be noted that the purpose of normalizing face features is to make the processed data distributed between 0-1, and the statistical distribution characteristics of uniform samples can not only facilitate subsequent data processing, but also improve the performance of the algorithm so that Convergence is accelerated. Secondly, the difference processing of the normalized face features of the same person can represent the difference information caused by the changes of age, illumination, expression, posture, etc. between the photos arbitrarily collected by the person in dynamic situations and the ID photos.

4) Construct positive and negative sample pairs.

In this preferred embodiment, the face features of all the dynamic photos and ID photos are normalized and differentially processed to obtain differential face features. Then the feature of the kth positive sample is the feature obtained by normalizing face features and differential processing of the dynamic photo and ID photo of the same person, denoted as The category attribute of the positive sample is the first category attribute, denoted as 1. The hth negative sample feature is the feature obtained by normalizing face features and differential processing of dynamic photos and ID photos of different people, denoted as The category attribute of the negative sample is the second category attribute, denoted as 0. That is, the positive sample set is a data pair composed of multiple positive sample features and the first category attribute, and the negative sample set is a data pair composed of multiple negative sample features and the second category attribute.

102: Train the SVM difference model.

In this preferred embodiment, a support vector machine (Support Vector Machine, SVM) is used as a classifier for model training. SVM is a supervised learning model, usually used for pattern recognition, classification and regression analysis. The main idea of SVM: one is to analyze the case of linear separability. For the case of linear inseparability, the linear non-separable samples in the low-dimensional input space are transformed into high-dimensional feature spaces by using the nonlinear mapping algorithm to make them linearly separable, so that It is possible to linearly analyze the nonlinear characteristics of samples by using linear algorithms in high-dimensional feature spaces. The second is to construct the optimal hyperplane in the feature space based on the structural risk minimization theory, so that the learner can be globally optimized, and the expectation in the entire sample space meets a certain upper bound with a certain probability. In short, given a set of positive and negative training samples, a separating hyperplane is found in the feature space to classify the sample points into different classes. When and there is a unique classification hyperplane, so that the distance between the sample point and the classification hyperplane is the largest. After finding the hyperplane, for the point to be measured, it is classified by calculating the position of the point relative to the hyperplane. In the process of training the SVM model, the most important thing is to find the optimal combination value of the parameters cost and gamma. Among them, cost(-c) is the penalty parameter, that is, the tolerance to error. The higher the value of c, the lower the tolerance to error. Gamma(-g) is the parameter of the SVM kernel function, which implicitly determines the data mapping. The distribution after entering the new feature space. In this embodiment, a radial basis kernel function may be selected.

The training SVM differential model specifically includes:

1): Generate a positive and negative sample training set and a positive and negative sample test set from the constructed positive and negative sample set;

In this preferred embodiment, the idea of cross-validation (Cross Validation) can be used when training the SVM difference model, and the constructed positive and negative sample sets are divided into positive and negative sample training sets and positive and negative sample test sets according to appropriate proportions. The division ratio is 6:4.

The positive and negative sample training set is used to train the SVM differential model, and the positive and negative sample test set is used to test the performance of the trained SVM differential model. If the accuracy of the test is higher, it indicates that the trained SVM differential model The better the performance of the model is; if the test accuracy is lower, it indicates that the performance of the trained SVM difference model is poor.

Furthermore, if the total number of divided positive and negative sample training sets is still large, that is, all the positive and negative sample training sets are used to participate in the training of the SVM differential model, which will lead to finding the optimal parameters c, g corresponding to the SVM differential model The cost is high, therefore, the generating the training set of positive and negative samples may further include: randomly selecting a first preset number of training sets of positive and negative samples from the generated training set of positive and negative samples to participate in the training.

In this preferred embodiment, in order to increase the randomness of the training set of positive and negative samples participating in the training, a random number generation algorithm can be used for random selection.

In this preferred embodiment, the first preset number can be a preset fixed value, for example, 40, that is, 40 samples are randomly selected from the generated positive and negative sample training set to participate in the training of the SVM difference model. The first preset number can also be a preset ratio value, for example, 1/10, that is, randomly select samples with a ratio of 1/10 in the generated positive and negative sample training set to participate in the training of the SVM difference model.

2): Find the optimal combination of parameters c and g;

In this preferred embodiment, c is a penalty parameter, which represents the penalty weight for misclassified points, and g is the radius of the kernel function. The search for the optimal combination of parameters c and g may include: inputting the first preset number of positive and negative sample training sets into the SVM, and calculating the first optimal combination parameters c and g; On the basis of the optimal combination parameters c and g, gradually expand the range of c and g and reduce the step size, save the combination parameters c and g each time, and select the positive and negative sample test set corresponding to the saved combination parameters c and g Tested on the SVM difference model, the parameters corresponding to the highest accuracy rate are the second optimal combination parameters c and g.

3): saving the second optimal combination parameters c, g and the corresponding SVM difference model.

Further, it should be understood that when the penalty parameter c is too large, overfitting tends to occur. When the penalty parameter c is too small, the classification function of the trained model will be lost. Therefore, choosing an appropriate penalty parameter c will greatly improve the classification performance of the model. In order to ensure the high reliability of the second optimal combination parameters c and g, improve the verification effect of the corresponding SVM differential model, multiple optimizations can be carried out, then the selection of positive and negative sample test sets can also include: in the generated Randomly select a second preset number of positive and negative sample test sets from the positive and negative sample test set to participate in the test.

In this preferred embodiment, the second preset number can be a preset fixed value, for example, 30, that is, 40 samples are randomly selected from the generated positive and negative sample test set to participate in the test of the SVM difference model. The second preset number can also be a preset ratio value, for example, 1/5, that is, randomly select samples with a ratio of 1/5 in the generated positive and negative sample test set to participate in the test of the SVM difference model.

The support vector machine difference model training method of embodiment 1 first constructs positive and negative sample sets, and then trains the SVM difference model according to the positive and negative sample sets. When constructing the positive and negative sample sets, normalize the facial features of all the dynamic photos and ID photos and perform differential processing to obtain differential facial features, which can effectively represent the facial features caused by changes in age, illumination, expression, posture, etc. Face difference information. When training the SVM difference model, the support vector machine (Support Vector Machine, SVM) is used as the classifier for model training. The support vector machine algorithm has a good performance effect on the binary classification problem, and the number of training samples is not high. In order to obtain the optimal combination of parameters c and g, a cross-validation method is used to randomly select the first preset number of positive and negative sample training sets in the generated positive and negative sample training set to participate in training, and in the generated positive and negative sample test set Randomly select a second preset number of positive and negative sample test sets to participate in the test. Therefore, the support vector machine differential model training method in the first embodiment combines the ideas of support vector machine, cross-validation and random numbers, and applies it to face verification. Under the premise of limited sample size, it can train suitable SVM difference model for face verification.

Embodiment two

Fig. 2 is a flow chart of the face verification method provided by Embodiment 2 of the present invention.

As shown in FIG. 2 , the face verification method specifically includes the following steps. According to different requirements, the order of the steps in the flow chart can be changed, and some steps can be omitted.

201: Extract the third face feature of the face area in the ID photo of the person to be verified.

In some embodiments, a card reader may be used to first read the face area in the ID photo, and then detect the face area in the ID photo of the person to be verified.

202: Extract the fourth face feature of the face area in the scene photo of the person to be verified.

For the convenience of description below, the scene photos collected under dynamic conditions are referred to as scene photos.

In this preferred embodiment, the pre-stored face detection algorithm can be used to detect the face area on the certificate photo of the person to be verified and the face area in the scene picture, and the pre-stored face feature extraction algorithm is used to detect The detected face area is subjected to face feature extraction. Preferably, because the deep learning feature has better feature expression ability, in this embodiment, the deep learning algorithm is used to extract the face features of the face area on the certificate photo and the scene photo. This application does not repeat in detail the detection of the face area and the feature extraction process of the face area.

In some embodiments, the steps 201 and 202 may be processed in parallel. In some embodiments, before using the face detection algorithm to detect the face area in the scene picture, the face verification method of the present invention may also include: preprocessing the scene picture.

In this embodiment, the preprocessing of the scene picture may include, but not limited to, image denoising, illumination normalization, posture calibration, grayscale normalization, etc. For example, a Gaussian filter can be used to filter the scene picture to remove noise in the scene picture; use quotient image technology to remove the impact of high-brightness illumination on the scene picture; face pose calibration.

203: Perform normalization processing on the third face feature to obtain a third normalized face feature, and perform normalization processing on the fourth face feature to obtain a fourth normalized face feature.

In this preferred embodiment, the Euclidean distance under the L2 norm sense is used to normalize the third facial feature and the fourth facial feature respectively. For details, refer to the corresponding description of 2) in step 101.

204: Perform a difference between the third normalized face feature and the fourth normalized face feature to obtain the differential face feature of the person to be verified.

In this preferred embodiment, the acquisition process of the differential face features of the person to be verified can refer to the corresponding description of 3) in step 101 for details.

205: Calculate the similarity of the differential facial features according to the SVM differential model.

In this preferred embodiment, the SVM differential model is a classifier model trained according to the support vector machine differential model training method provided in Embodiment 1 of the present invention. The difference face feature of the person to be verified is input into the SVM difference model, and the similarity of the difference face feature can be calculated through the mapping of the SVM difference model.

206: Determine whether the similarity is greater than a preset threshold.

When the similarity is greater than the preset threshold, perform step 207; when the similarity is less than or equal to the preset threshold, perform step 208. The preset threshold may be, for example, 0.5.

207: Determine that the ID photo and the scene photo are not of the same person.

208: Determine that the ID photo and the scene photo are of the same person.

When the similarity of the calculated differential face features is greater than the preset threshold, it is determined that the ID photo of the person to be verified is the same person as the scene picture collected under the dynamic situation of the person to be verified, that is, the ID photo is The person to be verified, that is, when the similarity of the calculated differential facial features is smaller than or equal to the preset threshold, it is determined that the ID photo of the person to be verified is not the same as the scene picture collected under the dynamic situation of the person to be verified Personal, that is, the photo of the said certificate is not of the person to be verified.

It should be noted that the face verification method of the present invention can extract the third face feature of the face area in the ID photo in advance, and store the third face feature in a preset database. The facial features and ID number of the ID photo are associated and stored in the database. When it is necessary to verify whether the ID photo and the scene picture are the same person, obtain the ID number on the ID photo, search for the corresponding face feature from the preset database according to the ID number, and then use the SVM differential model The similarity of the difference facial features is calculated, thereby shortening the time for extracting the facial features on the ID photo, and further improving the comparison efficiency.

In summary, the face verification method of the present invention uses a deep learning algorithm to extract the facial features (referred to as deep learning features) of the face area on the ID photo and the dynamic photo for model training, and the trained model It has a strong feature area capability; the present invention normalizes the deep learning feature features again and performs differential processing to obtain differential face features, which can effectively express the difference information of the feature space distribution of ID photos and scene photos; in the SVM differential model In the training phase, a small number of ID photos and dynamic photos of different people are used, thereby reducing the demand for the sample size of the algorithm and increasing the practicability of the algorithm. The face verification method of the present invention can obtain a better face verification effect according to the SVM differential model.

Embodiment three

The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. For those of ordinary skill in the art, without departing from the inventive concept of the present invention, it is also possible to make Improvements, but these all belong to the protection scope of the present invention.

The functional modules and hardware structure of the terminal implementing the above-mentioned support vector machine differential model training method and face verification method are introduced below in conjunction with Figures 3 to 5.

It should be understood that the described embodiment is only for illustration, and is not limited by this structure on the scope of the patent application.

Referring to Fig. 3, it is a functional block diagram in a preferred embodiment of the support vector machine differential model training device 40 of the present invention.

In some embodiments, the support vector machine differential model training device 40 runs in the terminal 6. The support vector machine differential model training device 40 may include a plurality of functional modules composed of program code segments. The program codes of each program segment in the support vector machine differential model training device 40 can be stored in the memory 61, and executed by the at least one processor 62 to execute (see Figure 1 for details) to support Training of the vector machine difference model.

In this embodiment, the support vector machine differential model training device 40 of the terminal 1 can be divided into multiple functional modules according to the functions it performs. The functional modules may include: a construction module 400 and a training module 402, wherein the construction module 400 further includes: a feature extraction submodule 4001, a normalization submodule 4002, a difference submodule 4003 and a construction submodule 4004. The training module 402 also includes: a generating submodule 4021 , an optimization submodule 4022 and a saving submodule 4023 . The module referred to in the present invention refers to a series of computer program segments that can be executed by at least one processor 62 and can complete fixed functions, which are stored in the memory 61. In some embodiments, the functions of each module will be described in detail in subsequent embodiments.

A construction module 400, configured to construct positive and negative sample sets.

In this preferred embodiment, prepare a photo of a certificate of different people and a photo randomly collected under a dynamic situation corresponding to the same person.

Described certificate photo refers to the photo that is used to prove identity on the certificate, can be ID card photo, can be passport photo, can also be various visa photos, or driver's license photo, graduation certificate photo etc.

In this preferred embodiment, the LFW data set can be used to construct positive and negative sample sets. The LFW data set is established to study the face recognition problem in an unrestricted environment, which contains more than 13,000 face images. The face images are all from the Internet, not a laboratory environment.

For the convenience of the following description, the photos randomly collected under dynamic conditions are called dynamic photos.

The constructing module 400 constructing positive and negative sample sets specifically includes:

1) feature extraction sub-module 4001, for extracting the first face feature of the face area on each ID photo, and extracting the second face feature of the face area on each dynamic photo;

In this preferred embodiment, the face feature corresponding to the ID photo of the i-th person, that is, the first face feature of the i-th person is denoted as _xi , and the face feature corresponding to the dynamic photo of the i-th person, that is, the first face feature of the i-th person The second face feature of person i is denoted as X _i .

In this preferred embodiment, before extracting the facial features, a pre-stored face detection algorithm can be used to detect the face area on the ID photo and the face area on the dynamic photo, and then use the pre-stored face feature The extraction algorithm performs face feature extraction on the detected face area.

The pre-stored face detection algorithm can be one or more combinations of the following algorithms: template-based face detection method, artificial neural network-based face detection method, model-based face detection method, skin color-based The face detection method or the face detection method based on the feature sub-face, etc.

The face feature extraction algorithm stored in advance can be one or more combinations of the following algorithms: Gabor feature, histogram of oriented gradient (Histogram of Oriented Gradient, HOG), local binary pattern (LocalBinary Patterns, LBP), Principal component analysis (Principal Component Analysis, PCA), linear discriminant analysis (Linear Discriminant Analysis, LDA) or independent component analysis (independent component analysis, ICA), etc. Preferably, because the deep learning feature has better feature expression ability, in this embodiment, the deep learning algorithm is used to extract the face features of the face area on the ID photo and the dynamic photo.

In this embodiment, the pre-stored face detection algorithm and the pre-stored face feature extraction algorithm are not limited to those listed above, any algorithm suitable for detecting the face area and extracting the face features of the face area algorithm can be applied to this. In addition, the pre-stored face detection algorithm and face feature extraction algorithm in this embodiment are all existing technologies, and will not be described in detail herein.

2) The normalization sub-module 4002 is used to perform normalization processing on the first facial feature to obtain the first normalized facial feature, and to perform normalization processing on the second facial feature to obtain the second facial feature. Normalized facial features;

In this preferred embodiment, the first face feature of the ith person is normalized to obtain the first normalized face feature as The second face feature of the i-th person is normalized to obtain the second normalized face feature as

It should be understood that ||x|| ₂ is the 1/2 power of the sum of the squares of the elements of the x vector, that is, the Euclidean distance in the sense of the L2 norm.

3) difference sub-module 4003, used to make a difference between the first normalized face feature and the second normalized face feature to obtain the difference face feature;

In this preferred embodiment, the i-th person's differential face feature is denoted as

It should be noted that the purpose of normalizing face features is to make the processed data distributed between 0-1, and the statistical distribution characteristics of uniform samples can not only facilitate subsequent data processing, but also improve the performance of the algorithm so that Convergence is accelerated. Secondly, the difference processing of the normalized face features of the same person can represent the difference information caused by the changes of age, illumination, expression, posture, etc. between the photos arbitrarily collected by the person in dynamic situations and the ID photos.

4) Construction sub-module 4004, used to construct positive and negative sample pairs.

In this preferred embodiment, the face features of all the dynamic photos and ID photos are normalized and differentially processed to obtain differential face features. Then the feature of the kth positive sample is the feature obtained by normalizing face features and differential processing of the dynamic photo and ID photo of the same person, denoted as The category attribute of the positive sample is the first category attribute, denoted as 1. The feature of the hth negative sample is the feature obtained by normalizing face features and differential processing of dynamic photos and ID photos of different people, denoted as The category attribute of the negative sample is the second category attribute, denoted as 0. That is, the positive sample set is a data pair composed of multiple positive sample features and positive sample category attributes, and the negative sample set is a data pair composed of multiple negative sample features and negative sample category attributes.

The training module 402 is used for training the SVM difference model.

In this preferred embodiment, a support vector machine (Support Vector Machine, SVM) is used as a classifier for model training.

SVM is a supervised learning model, usually used for pattern recognition, classification and regression analysis. The main idea of SVM: one is to analyze the case of linear separability. For the case of linear inseparability, the nonlinear mapping algorithm is used to convert the linearly inseparable samples of the low-dimensional input space into a high-dimensional feature space to make them linearly separable, so that It is possible to linearly analyze the nonlinear characteristics of samples by using linear algorithms in high-dimensional feature spaces. The second is to construct the optimal hyperplane in the feature space based on the structural risk minimization theory, so that the learner can be globally optimized, and the expectation in the entire sample space can meet a certain upper bound with a certain probability. In short, given a set of positive and negative training samples, a separating hyperplane is found on the feature space to classify the sample points into different classes. If and there is a unique classification hyperplane, the distance between the sample point and the classification hyperplane is the largest. After finding the hyperplane, for the points to be measured, they are classified by calculating the position of the point relative to the hyperplane. In the process of training the SVM model, the most important thing is to find the optimal combination value of the parameters cost and gamma. Among them, cost(-c) is the penalty parameter, that is, the tolerance to errors. The higher the value of c, the lower the tolerance to errors. Gamma(-g) is the parameter of the SVM kernel function, which implicitly determines the data mapping. The distribution after entering the new feature space. In this embodiment, a radial basis kernel function may be selected.

The training module 402 training SVM difference model specifically includes:

1) Generate a sub-module 4021 for generating a positive and negative sample training set and a positive and negative sample test set;

In this preferred embodiment, the idea of cross-validation (Cross Validation) can be used when training the SVM difference model, and the constructed positive and negative sample sets are divided into positive and negative sample training sets and positive and negative sample test sets according to appropriate proportions. The division ratio is (6:4).

The positive and negative sample training set is used to train the SVM differential model, and the positive and negative sample test set is used to test the performance of the trained SVM differential model. If the accuracy of the test is higher, it indicates that the trained SVM differential model The better the performance of the model is; if the test accuracy is lower, it indicates that the performance of the trained SVM difference model is poor.

Furthermore, if the total number of divided positive and negative sample training sets is still large, that is, all the positive and negative sample training sets are used to participate in the training of the SVM difference model, which will lead to finding the optimal parameter cost c corresponding to the SVM difference model, g is relatively large, therefore, the generating the training set of positive and negative samples may further include: randomly selecting a first preset number of training sets of positive and negative samples from the generated training set of positive and negative samples to participate in training.

In this preferred embodiment, in order to increase the randomness of the training set of positive and negative samples participating in the training, a random number generation algorithm can be used for random selection.

In this preferred embodiment, the first preset number can be a preset fixed value, for example, 40, that is, 40 samples are randomly selected from the generated positive and negative sample training set to participate in the training of the SVM difference model. The first preset number can also be a preset ratio value, for example, 1/10, that is, randomly select samples with a ratio of 1/10 in the generated positive and negative sample training set to participate in the training of the SVM difference model.

2) The optimization sub-module 4022 is used to find the optimal combination of parameters c and g;

In this preferred embodiment, c is a penalty parameter, which represents the penalty weight for misclassified points, and g is the radius of the kernel function. The optimization sub-module 4022 searching for the optimal combination of parameters c and g may include: inputting the randomly selected positive and negative sample training set into SVM, and calculating the first optimal combination of parameters c and g; On the basis of the first optimal combination parameters c and g, gradually expand the range of c and g and reduce the step size, save the combination parameters c and g each time, and select positive and negative sample test sets in the saved combination parameters c and g The corresponding SVM difference model is tested until the second optimal combination parameters c and g are obtained.

It should be understood that the parameters corresponding to the highest accuracy obtained are the second optimal combination parameters c, g, and the second optimal combination parameters c, g and the corresponding SVM difference model are saved.

Further, it should be understood that when the penalty parameter c is too large, overfitting tends to occur. When the penalty parameter c is too small, the classification function of the trained model will be lost. Therefore, choosing an appropriate penalty parameter c will greatly improve the classification performance of the model. In order to ensure the high reliability of the second optimal combination parameters c, g, and improve the verification effect of the corresponding SVM differential model, multiple optimizations can be performed, and the selection of positive and negative sample test sets by the generating submodule 4021 can also include : Randomly select a second preset number of positive and negative sample test sets from the generated positive and negative sample test sets to participate in the test.

In this preferred embodiment, the second preset number can be a preset fixed value, for example, 30, that is, 40 samples are randomly selected from the generated positive and negative sample test set to participate in the test of the SVM difference model. The second preset number can also be a preset ratio value, for example, 1/5, that is, randomly select samples with a ratio of 1/5 in the generated positive and negative sample test set to participate in the test of the SVM difference model.

The SVM differential model training device 40 of Embodiment 1 first constructs the module 400 to construct positive and negative sample sets, and then the training module 402 trains the SVM differential model according to the positive and negative sample sets. When the construction module 400 constructs the positive and negative sample sets, the facial features of all the dynamic photos and ID photos are normalized and differentially processed to obtain differential facial features, which can effectively characterize the differences due to age, illumination, expression, posture, etc. The difference information of the face caused by the change. When described training module 402 trains SVM differential model, adopt support vector machine (Support Vector Machine, SVM) to carry out model training as classifier, support vector machine algorithm has good performance effect to binary classification problem, simultaneously to the quantity of training samples Not demanding. In order to obtain the optimal combination of parameters c and g, a cross-validation method is used to randomly select the first preset number of positive and negative sample training sets in the generated positive and negative sample training set to participate in training, and in the generated positive and negative sample test set Randomly select a second preset number of positive and negative sample test sets to participate in the test. Therefore, the support vector machine difference model training device 40 of Embodiment 1 combines the ideas of support vector machine, cross-validation and random numbers, and applies it to face verification. Under the premise of limited sample size, it can train SVM difference model suitable for face verification.

Referring to FIG. 4 , it is a functional block diagram of a preferred embodiment of the face verification device 50 of the present invention.

In some embodiments, the face verification device 50 runs in the terminal 1 . The face verification device 50 may include a plurality of functional modules made up of program code segments. The program codes of each program segment in the face verification device 50 can be stored in the memory 61, and executed by the at least one processor 62 to perform (see Figure 2 for details) verification of the face .

In this embodiment, the face verification device 50 of the terminal 1 can be divided into multiple functional modules according to the functions it performs. The functional modules may include: a first extraction module 500, a second extraction module 502, a preprocessing module 504, a normalization module 506, a difference module 508, and a verification module 510. The module referred to in the present invention refers to a series of computer program segments that can be executed by at least one processor 62 and can complete fixed functions, and are stored in the memory 61. In some embodiments, the functions of each module will be described in detail in subsequent embodiments.

The first extraction module 500 is used to extract the third facial feature of the facial area in the ID photo of the person to be verified.

In some embodiments, a card reader may be used to first read the face area in the ID photo, and then detect the face area in the ID photo of the person to be verified.

The second extraction module 502 is used to extract the fourth facial feature of the facial area in the scene photo of the person to be verified.

For the convenience of description below, the scene photos collected under dynamic conditions are referred to as scene photos.

In this preferred embodiment, the pre-stored face detection algorithm can be used to detect the face area on the certificate photo of the person to be verified and the face area in the scene picture, and the pre-stored face feature extraction algorithm is used to detect The detected face area is subjected to face feature extraction. Preferably, because the deep learning feature has better feature expression ability, in this embodiment, the deep learning algorithm is used to extract the face features of the face area on the certificate photo and the scene photo. This application does not repeat in detail the detection of the face area and the feature extraction process of the face area.

In some embodiments, the first extraction module 500 and the second extraction module 502 may be processed in parallel. In some embodiments, before using the face detection algorithm to detect the face area in the scene picture, the face verification device 50 of the present invention may further include a preprocessing module 504, configured to: pre-process the scene picture deal with.

In this embodiment, the preprocessing performed by the preprocessing module 504 on the scene picture may include, but is not limited to, image denoising, illumination normalization, pose calibration, grayscale normalization, and the like. For example, a Gaussian filter can be used to filter the scene picture to remove noise in the scene picture; use quotient image technology to remove the impact of high-brightness illumination on the scene picture; face pose calibration.

A normalization module 506, configured to perform normalization processing on the third facial feature to obtain a third normalized facial feature, and perform normalization processing on the fourth facial feature to obtain a fourth normalized facial feature facial features.

In this preferred embodiment, the Euclidean distance under the L2 norm sense is used to normalize the third facial feature and the fourth facial feature respectively. For details, please refer to the corresponding description of step normalization sub-module 4002.

The difference module 508 is used to make a difference between the third normalized face feature and the fourth normalized face feature to obtain the difference face feature of the person to be verified.

In this preferred embodiment, the acquisition process of the difference face feature of the person to be verified can refer to the corresponding description of the difference sub-module 4003 for details.

Verification module 510, is used for calculating the similarity of described difference face feature according to described SVM difference model.

In this preferred embodiment, the SVM difference model is a classifier model trained by the support vector machine difference model training device 40 provided according to Embodiment 1 of the present invention. The differential facial features of the people to be verified are input into the SVM differential model, and the similarity of the differential facial features can be calculated for classification through the mapping of the SVM differential model.

The verification module 510 is further configured to determine whether the similarity is greater than a preset threshold.

When the similarity is greater than a preset threshold, the verification module determines that the ID photo and the scene photo are not the same person; when the similarity is less than or equal to a preset threshold, the verification module determines that the The ID photo is of the same person as the scene photo. The preset threshold may be, for example, 0.5.

When the similarity of the calculated differential face features is greater than the preset threshold, it is determined that the ID photo of the person to be verified is the same person as the scene picture collected under the dynamic situation of the person to be verified, that is, the ID photo is The person to be verified, that is, when the similarity of the calculated differential facial features is smaller than or equal to the preset threshold, it is determined that the ID photo of the person to be verified is not the same as the scene picture collected under the dynamic situation of the person to be verified Personal, that is, the photo of the said certificate is not of the person to be verified.

It should be noted that the face verification device 50 of the present invention can extract the third face feature of the face area in the ID photo in advance, and store the third face feature in a preset database. The face features and the ID number of the ID photo are associated and stored in the above-mentioned database. When it is necessary to verify whether the ID photo and the scene picture are the same person, obtain the ID number on the ID photo, search for the corresponding face feature from the preset database according to the ID number, and then according to the SVM difference model The similarity of the difference facial features is calculated, thereby shortening the time for extracting the facial features on the ID photo, and further improving the comparison efficiency.

In summary, the face verification device 50 of the present invention uses a deep learning algorithm to extract the facial features of the face area on the ID photo and the dynamic photo (referred to as deep learning features) for model training, and the trained The model has a strong feature area capability; the present invention normalizes the deep learning feature features again and performs differential processing to obtain differential face features, which can effectively express the difference information of the feature space distribution of ID photos and scene photos; In the model training stage, a small number of ID photos and dynamic photos of different people are used, thereby reducing the sample size requirement of the algorithm and increasing the practicability of the algorithm. The face verification device 50 of the present invention can obtain a better face verification effect according to the SVM difference model.

The above-mentioned integrated units implemented in the form of software function modules can be stored in a computer-readable storage medium. The above-mentioned software function modules are stored in a storage medium, and include several instructions to enable a computer device (which may be a personal computer, a dual-screen device, or a network device, etc.) or a processor (processor) to execute the functions described in various embodiments of the present invention. method part.

As shown in Figure 5, it is a schematic diagram of the hardware structure of a terminal that implements the SVM differential model training method and/or the face verification method of the present invention.

In a preferred embodiment of the present invention, the terminal 6 includes a memory 61 , at least one processor 62 , at least one communication bus 63 and a camera 64 .

Those skilled in the art should understand that the structure of the terminal shown in Figure 5 does not constitute a limitation of the embodiment of the present invention, it can be a bus structure or a star structure, and the terminal 6 can also include more terminals than shown in the figure. More or less additional hardware or software, or a different arrangement of components.

In some embodiments, the terminal 6 includes a dual-screen device that can automatically perform numerical calculation and/or information processing according to preset or stored instructions, and its hardware includes but not limited to microprocessors, application-specific integrated circuits , programmable gate arrays, digital processors, embedded devices, etc. The terminal 6 may also include user equipment, which includes but is not limited to any electronic product that can interact with the user through a keyboard, mouse, remote control, touch pad, or voice control device, for example, a personal Computers, tablet computers, smart phones, personal digital assistants (Personal Digital Assistant, PDA) and any other electronic products with dual screens.

It should be noted that the terminal 6 is only an example, and other existing or future electronic products that may be adapted to the present invention should also be included in the protection scope of the present invention, and are included here by reference.

In some embodiments, the memory 61 is used to store program codes and various data, such as the information storage system 10 installed in the terminal 6, and realize high-speed, automatic completion of programs or Data access. Described memory 61 comprises read-only memory (Read-Only Memory, ROM), random access memory (RandomAccess Memory, RAM), programmable read-only memory (Programmable Read-Only Memory, PROM), erasable programmable read-only memory ( Erasable Programmable Read-Only Memory, EPROM), One-time Programmable Read-Only Memory (OTPROM), Electrically-Erasable Programmable Read-Only Memory, EEPROM , Compact Disc Read-Only Memory (CD-ROM) or other optical disk storage, magnetic disk storage, tape storage, or any other computer-readable medium that can be used to carry or store data.

In some embodiments, the at least one processor 62 may be composed of an integrated circuit, for example, may be composed of a single packaged integrated circuit, or may be composed of multiple integrated circuits with the same function or different functions packaged, including a Or a combination of multiple central processing units (Central Processing unit, CPU), microprocessors, digital processing chips, graphics processors, and various control chips. The at least one processor 62 is the control core (Control Unit) of the terminal 6, using various interfaces and lines to connect the various components of the entire terminal 6, by running or executing programs or modules stored in the memory 61, And call the data stored in the memory 61 to execute various functions of the terminal 6 and process data, such as executing the SVM difference model training device 40 and/or the face verification device 50 .

In some embodiments, the at least one communication bus 63 is configured to realize connection and communication between the memory 61, the at least one processor 62, the camera 64, and the like.

Although not shown, the terminal 6 can also include a power supply (such as a battery) for supplying power to various components. Preferably, the power supply can be logically connected to the at least one processor 62 through a power management system, so as to realize management through the power management system Charging, discharging, and power management functions. The power supply may also include one or more DC or AC power sources, recharging systems, power failure detection circuits, power converters or inverters, power status indicators, and any other components. The terminal 6 may also include various sensors, bluetooth modules, Wi-Fi modules, cameras, etc., which will not be repeated here.

It should be understood that the described embodiment is only for illustration, and is not limited by this structure on the scope of the patent application.

In a further embodiment, in conjunction with FIG. 1 and FIG. 2, the at least one processor 62 can execute the operating system of the terminal 6 and installed various applications (such as the SVM differential model training device 40), Program code, etc., for example, each of the above-mentioned modules includes: construction module 400, construction submodule 4004 and feature extraction submodule 4001, normalization submodule 4002, difference submodule 4003, generation submodule 4021 and optimization submodule 4022 . The at least one processor 62 can also execute the operating system of the terminal 6 and installed various applications (such as the SVM differential model training device 40), program codes, etc., for example, the above-mentioned modules include: A first extraction module 500 , a second extraction module 502 , a preprocessing module 504 , a normalization module 506 , a difference module 508 , and a verification module 510 .

Specifically, for the specific implementation method of the above instruction by the at least one processor 62, reference may be made to the description of relevant steps in the embodiments corresponding to FIG. 1 and FIG. 3 , and details are not repeated here.

In the several embodiments provided by the present invention, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

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, they may be located in one place, or may be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to realize the purpose of the solution of this embodiment.

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

It will be apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, but that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Accordingly, the embodiments should be regarded in all points of view as exemplary and not restrictive, the scope of the invention being defined by the appended claims rather than the foregoing description, and it is therefore intended that the scope of the invention be defined by the appended claims rather than by the foregoing description. All changes within the meaning and range of equivalents of the elements are embraced in the present invention. Any reference sign in a claim should not be construed as limiting the claim concerned. Furthermore, it is clear that the word "comprising" does not exclude other elements or the singular does not exclude the plural. A plurality of units or devices stated in the system claims may also be realized by one unit or device through software or hardware. The words first, second, etc. are used to denote names and do not imply any particular order.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation, although the present invention has been described in detail with reference to preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements can be made without departing from the spirit and scope of the technical solutions of the present invention.

Claims (10)

1. A kind of 1. SVM difference models training method, it is characterised in that the described method includes：

   Positive and negative sample set is constructed, including：

   1) the first face characteristic of the human face region on every certificate photograph is extracted, extracts the human face region on every an action shot The second face characteristic；

   2) first face characteristic is normalized to obtain the first normalization face characteristic, it is special to second face Sign is normalized to obtain the second normalization face characteristic；

   3) the described first normalization face characteristic and the second normalization face characteristic are made the difference to obtain difference face characteristic；

   4) positive negative sample pair is constructed, wherein, the positive sample carries out face to an action shot for same person and certificate photograph The feature and first category attribute that feature normalization and difference processing obtain, negative sample is to an action shot and certificate for different people Photo carries out the face characteristic normalization feature that simultaneously difference processing obtains and second category attribute；

   Training SVM difference models, including：

   1)：Positive and negative sample training collection and positive and negative test sample collection are generated from the positive and negative sample set constructed；

   2)：Find the optimum combination of the parameter g of penalty parameter c and kernel function；

   3)：Preserve best parameter group and corresponding SVM difference models.
2. 2. the method as described in claim 1, it is characterised in that described to generate positive negative sample from the positive and negative sample set constructed Training set and positive and negative test sample collection include：

   The positive and negative sample training collection for concentrating the default quantity of random selection first in the positive and negative sample training generated participates in training； The positive and negative test sample generated concentrates the positive and negative test sample collection of the default quantity of random selection second to participate in test.
3. 3. method as claimed in claim 2, it is characterised in that the parameter g's for finding penalty parameter c and kernel function is optimal Combination includes：

   The positive and negative sample training collection of described first default quantity is input in SVM, calculates the first optimum combination parameter c, g；

   Progressively expand the scope of c, g and reduce step-length, preserve combination parameter c, g each time；

   Positive and negative test sample collection is selected to be tested on the SVM difference models corresponding to combination parameter c, the g preserved, accurately Corresponding parameter is the second optimum combination parameter c, g during rate highest.
4. 4. the SVM difference models that a kind of method using described in 3 any one of claims 1 to 3 trains carry out face verification side Method, it is characterised in that the described method includes：

   Extract third party's face feature of the human face region in the certificate photograph of witness to be tested；

   Extract the 4th face characteristic of the human face region in the scene photo of witness to be tested；

   Third party's face feature is normalized to obtain the 3rd normalization face characteristic, to the 4th face characteristic It is normalized to obtain the 4th normalization face characteristic；

   Described 3rd normalization face characteristic and the 4th normalization face characteristic are made the difference to obtain the witness's to be tested Difference face characteristic；And

   The similarity of the difference face characteristic is calculated according to the SVM difference models；

   Judge whether the similarity is more than predetermined threshold value；And

   When the similarity is more than the predetermined threshold value, determine the certificate photograph with the scene photo not to be same People；Or

   When the similarity is less than or equal to predetermined threshold value, determine the certificate photograph and the scene photo to be same People.
5. 5. a kind of SVM difference models training device, it is characterised in that described device includes：

   Constructing module, for constructing positive and negative sample set, including：

   Feature extraction submodule, for extracting the first face characteristic of the human face region on every certificate photograph, extracts every and moves Second face characteristic of the human face region on state photo；

   Submodule is normalized, for first face characteristic to be normalized to obtain the first normalization face characteristic, Second face characteristic is normalized to obtain the second normalization face characteristic；

   Difference submodule, for being made the difference the described first normalization face characteristic and the second normalization face characteristic to obtain Difference face characteristic；

   Submodule is constructed, for constructing positive negative sample pair, wherein, the positive sample is to an action shot and certificate for same person Photo carries out the face characteristic normalization feature that simultaneously difference processing obtains and first category attribute, and negative sample for different people to moving State photo and certificate photograph carry out the face characteristic normalization feature that simultaneously difference processing obtains and second category attribute；

   Training module, for training SVM difference models, including：

   Submodule is generated, for generating positive and negative sample training collection and positive and negative test sample collection from the positive and negative sample set constructed；

   Optimizing submodule, the optimum combination of the parameter g for finding penalty parameter c and kernel function；

   Submodule is preserved, for preserving best parameter group and corresponding SVM difference models.
6. 6. device as claimed in claim 5, it is characterised in that the generation submodule is raw from the positive and negative sample set constructed Include into positive and negative sample training collection and positive and negative test sample collection：

   The positive and negative sample training collection for concentrating the default quantity of random selection first in the positive and negative sample training generated participates in training； The positive and negative test sample generated concentrates the positive and negative test sample collection of the default quantity of random selection second to participate in test.
7. 7. device as claimed in claim 6, it is characterised in that the optimizing submodule searching penalty parameter c and kernel function The optimum combination of parameter g includes：

   The positive and negative sample training collection of described first default quantity is input in SVM, calculates the first optimum combination parameter c, g；

   Progressively expand the scope of c, g and reduce step-length, preserve combination parameter c, g each time；

   Positive and negative test sample collection is selected to be tested on the SVM difference models corresponding to combination parameter c, the g preserved, accurately Corresponding parameter is the second optimum combination parameter c, g during rate highest.
8. 8. the SVM difference models that a kind of device using described in claim 5 to 7 any one trains carry out face verification dress Put, it is characterised in that described device includes：

   First extraction module, third party's face feature of the human face region in certificate photograph for extracting witness to be tested；

   Second extraction module, the 4th face characteristic of the human face region in scene photo for extracting witness to be tested；

   Module is normalized, it is right for third party's face feature to be normalized to obtain the 3rd normalization face characteristic 4th face characteristic is normalized to obtain the 4th normalization face characteristic；

   Difference block, for being made the difference the described 3rd normalization face characteristic and the 4th normalization face characteristic to obtain State the difference face characteristic of witness to be tested；And

   Authentication module, for calculating the similarity of the difference face characteristic according to the SVM difference models；And

   The authentication module, is additionally operable to judge whether the similarity is more than predetermined threshold value；And

   When the similarity is more than the predetermined threshold value, the authentication module determines the certificate photograph and the scene photo It is not same person；Or

   When the similarity is less than or equal to predetermined threshold value, the authentication module determines that the certificate photograph shines with the scene Piece is same person.
9. A kind of 9. terminal, it is characterised in that：The terminal includes processor, and the processor is used to perform what is stored in memory The SVM difference models training method as described in any one in claims 1 to 3 is realized during computer program or realizes such as right It is required that the 4 face verification methods.
10. 10. a kind of computer-readable recording medium, is stored thereon with computer program, it is characterised in that：The computer program The SVM difference models training method as described in any one in claims 1 to 3 is realized when being executed by processor or is realized as weighed Profit requires the 4 face verification methods.