WO2016150240A1 - Identity authentication method and apparatus - Google Patents

Abstract

The present invention relates to the field of digital image processing and mode recognition. Disclosed are an identity authentication method and apparatus. The method comprises: acquiring a face image sample pair, wherein one face image is an identity card picture and the other face image is an image collected on site; processing the acquired face image sample pair by using a face image pyramid algorithm, so as to obtain a face image pyramid structure; calculating the gray scale and the gradient orientation of a face image of each size in the obtained face image pyramid structure of the face image sample pair, so as to obtain a gradient orientation pyramid (GOP) characteristic; and performing calculation according to the obtained GOP characteristic of the face image sample pair, so as to determine whether the face image sample pair belongs to a same person. The present invention effectively avoids interferences from factors such as illumination, an age and an expression, and also remarkably improves the accuracy of identity authentication.

Description

Identity authentication method and device Technical field

The present invention relates to digital image processing and pattern recognition, and more particularly to an identity authentication method and apparatus.

Background technique

The second-generation ID card is a valid certificate for the residents of China, and records the basic information of the holder. In most cases, ID card registration is still manual, which is very inconvenient. With the maturity of computer technology, image processing and character recognition algorithms, it has become possible to use computers to automatically identify and enter ID cards. The identity of the person can be determined by comparing the photo on the ID card with the video photo taken on site. However, due to the low pixel and large age span of the second generation photo, the video photos collected on site are subject to interference caused by illumination, posture, expression, glasses, etc., as well as the single sample problem of the ID card, which makes the current based on The face authentication system for ID cards faces many challenges.

China's second-generation resident ID card is made by non-contact IC card technology. The machine ID can be obtained by machine reading, and compared with the face photo collected by the on-site camera imaging device to determine whether it is the same person and belongs to face authentication. category. At present, the ID authentication system based on ID card adopts the existing face recognition algorithm. From the perspective of face feature extraction, there are mainly feature subspace method, local feature based method and machine learning based method. All three face authentication methods have the disadvantages of poor anti-interference and low accuracy.

Summary of the invention

The technical problem to be solved by the present invention is to provide an identity authentication method and apparatus with strong anti-interference and high accuracy.

In order to solve the above technical problem, the present invention provides the following technical solutions:

An authentication method includes:

Obtaining a pair of face image samples, wherein one of the face images is an ID card photo, and the other face image is an image acquired in the field;

The face image pyramid pair is processed by the face image pyramid algorithm to obtain a pyramid structure of the face image;

Calculating a grayscale gradient direction on the face image of each scale in the pyramid image structure of the obtained face image sample pair to obtain a gradient direction pyramid GOP feature;

According to the obtained GOP feature of the face image sample pair, it is calculated whether the face image sample pair belongs to the same person.

An identity authentication device comprising:

The acquisition module is configured to obtain a pair of face image samples, wherein one face image is an ID card photo, and another face image is an image acquired in the field;

Sampling module: used to process the acquired face image sample pairs by using the face image pyramid algorithm to obtain a pyramid structure of the face image;

The calculation module is configured to calculate a gray gradient direction of the face image of each scale in the pyramid image structure of the obtained face image sample pair to obtain a gradient direction pyramid GOP feature;

The judging module is configured to calculate, according to the GOP feature of the obtained face image sample pair, whether the pair of face image samples belong to the same person.

The invention has the following beneficial effects:

Compared with the prior art, the identity authentication method of the present invention first acquires a pair of face image samples, wherein one face image is an ID card photo, and the other face image is an image acquired in the field; then the face image pyramid algorithm is used. The obtained face image sample pairs are processed to obtain a pyramid structure of the face image. The algorithm uses a multi-resolution method to represent the face image, which can comprehensively describe the face image feature, and avoids adopting only one in the prior art. The resolution of the image indicates that the image features are not comprehensively expressed, which effectively improves the accuracy of the identity authentication of the present invention, and the number of pixels of each layer of the pyramid structure of the face image is continuously reduced from bottom to top. The calculation amount of the invention in the data processing process is reduced, and the identity authentication process of the present invention is made easier to implement; then the grayscale is calculated for the face image of each scale in the face image pyramid structure of the obtained face image sample pair. Gradient direction, get the gradient direction pyramid GOP feature, Since the GOP feature of the face image is robust to external factors such as illumination, age and expression, the anti-interference of the invention is enhanced, and the accuracy of the face authentication result is further improved; finally, the obtained face image is obtained. The GOP feature of the sample pair is calculated to determine whether the face image sample pair belongs to the same person. The identity authentication method of the invention effectively avoids interference of factors such as illumination, age and expression, and also significantly improves the accuracy of identity authentication.

DRAWINGS

1 is a schematic flowchart 1 of an identity authentication method according to the present invention;

2 is a schematic flow chart of extracting a face gradient direction pyramid GOP feature according to the present invention;

3 is a schematic diagram of a hyper-classification interval hyperplane of the SVM classification algorithm of the present invention;

4 is a schematic flowchart 2 of an identity authentication method according to the present invention;

Figure 5 is a schematic structural diagram 1 of the identity authentication apparatus of the present invention;

6 is a schematic flowchart 2 of an identity authentication method according to the present invention;

FIG. 7 is a schematic diagram of a GOP+SVM identity authentication process according to the present invention.

detailed description

The technical problems, the technical solutions, and the advantages of the present invention will be more clearly described in the following description.

In one aspect, the present invention provides an identity authentication method, as shown in FIG. 1, including:

Step S101: Acquire a pair of face image samples, wherein one face image is an ID card photo, and another face image is an image collected in the field;

In this step, the images collected on the spot can be obtained by intercepting the collected video samples, or by taking pictures on the spot.

Step S102: processing the acquired face image sample pair by using a face image pyramid algorithm to obtain a face image pyramid structure;

In this step, the face image pyramid structure is used to represent the face image in multi-scale and multi-resolution. The algorithm can comprehensively describe the face image feature, and the face image pyramid structure is from bottom to top. The number is continuously reduced, reducing the invention in the data processing process The amount of calculation.

Step S103: Calculating a grayscale gradient direction on the face image of each scale in the pyramid image structure of the obtained face image sample pair to obtain a gradient direction pyramid GOP feature;

In this step, since the face gradient direction pyramid GOP feature is robust to factors such as illumination, expression and age, the present invention preferably uses the GOP feature of the face image sample pair to enhance the anti-interference of the present invention. It also enhances the accuracy of identity authentication.

Step S104: Calculate, according to the GOP feature of the obtained face image sample pair, whether the pair of face image samples belongs to the same person.

The identity authentication method of the present invention first acquires a pair of face image samples, wherein one face image is an ID card photo, and the other face image is an image acquired in the field; then the face image pyramid algorithm is used to acquire the face image sample pair. The processing is performed to obtain a pyramid structure of the face image. The method uses a multi-resolution method to represent the face image, and can comprehensively describe the face image feature, thereby avoiding the prior art that only one resolution is used to represent the image. The phenomenon that the image features are not comprehensively expressed effectively improves the accuracy of the identity authentication of the present invention, and the number of pixels of each layer of the face image pyramid structure is reduced from bottom to top, which reduces the data processing of the present invention. The amount of calculation in the process makes the identity authentication process of the present invention easier to implement; nextly, the grayscale gradient direction is calculated for the face image of each scale in the face image pyramid structure of the obtained face image sample pair, and the gradient direction pyramid is obtained. GOP features, due to external factors such as illumination, age, and expression of GOP features of face images The robustness enhances the anti-interference of the present invention, and further improves the accuracy of the face authentication result. Finally, according to the GOP feature of the obtained face image sample pair, it is calculated whether the face image sample pair is the same. personal. The identity authentication method of the invention effectively avoids interference of factors such as illumination, age and expression, and also significantly improves the accuracy of identity authentication.

The algorithm in the above step S103 may specifically be: a gray gradient direction pyramid GOP feature for calculating a pair of face image samples, and the specific calculation process of the GOP feature may be referred to as follows:

For a two-dimensional face image I(x,y), the gradient direction at each pixel (x,y) at scale s is defined as

Where τ is an artificially set threshold such that the gradient direction at which the gradation is almost constant is zero, so that the gradient direction pyramid of the original two-dimensional image I(x, y) is defined as all pixel points at each scale Cascading of gradient directions, expressed as

G(I) is represented as a GOP feature of a face image, which is a d×2 matrix, where d represents the sum of the number of images on a face pyramid.

The GOP feature extraction process in the present invention may be as shown in FIG. 2, wherein (a) is the original face image I(x, y), and (b) is the face image pyramid P(x, y), (c) represents a gradient direction pyramid, and (d) represents a GOP feature G(I).

As an improvement of the present invention, step S102 may further be: obtaining a face image pyramid structure by using Gaussian kernel convolution smoothing and downsampling on the acquired face image sample pairs.

In this step, the face image pyramid is a structure for interpreting a face image with multiple resolutions. The pyramid structure of an image is a collection of images whose resolution is gradually reduced in a pyramid shape. The pyramid is an image representation that combines Gaussian kernel convolution smoothing and downsampling operations. The bottom of the pyramid is the original high resolution representation of the image to be processed, while the top is a low resolution approximation. As you move toward the upper level of the pyramid, the size and resolution are reduced, resulting in images at different scales. The specific calculation process of the face image gold tower is as follows:

Given a face image I(x,y), define its face image pyramid as P(x,y), which is as follows:

among them,

I(x,y;0)=I(x,y)

I(x,y;s)=[I(x,y;s-1)*Φ(x,y)]↓2,s=1,2,...S

In equation (2), * represents the convolution operation, s is the index of the scale, ↓ 2 represents the downsampling, and Φ(x, y) represents the Gaussian kernel function of the same size as the image I(x, y), which can be expressed as

The image I(x, y) at the scale s-1 is convolved with the Gaussian kernel function to obtain a smoothed image, and then the image under the next scale s is obtained by the downsampling operation, and the image size is at the previous scale. A quarter of the image, the image resolution is also gradually reduced, and the layer-by-layer convolution downsampling, the pyramid of the face image is obtained.

In most cases, features that are not easily perceived by a face image at a single scale are easily captured at multiple scales. Therefore, in the present invention, a Gaussian kernel function convolution smoothing and a downsampling operation are preferably used to obtain a face image pyramid. Structure, since the face image pyramid structure is a multi-scale, multi-resolution representation of the face image, the method can more fully describe the face image feature and enhance the accuracy of the identity authentication in the present invention.

As an improvement of the present invention, step S104 may further be: calculating a feature vector for the GOP feature of the obtained face image sample pair, and using SVM binary classification processing to calculate whether the face image sample pair belongs to the same person.

In this step, the SVM classification algorithm is preferably used to first calculate the probability that the output label corresponding to the input feature is 1, and compare with a preset threshold. The threshold is greater than the threshold, and the two faces are from the same person. Conversely, assigning to -1 means that it is from a different person. For two types of linear learning tasks, the SVM needs to find a hyperplane with the largest interval to separate the two types of samples. As shown in Figure 3, the maximum spacing ensures that the hyperplane has the best generalization ability. For the non-linear learning task in the identity authentication process of the present invention, the SVM feature classification algorithm used in the present invention converts the actual problem into a high-dimensional feature space through a nonlinear transformation, and constructs a linear discriminant function in a high-dimensional space. The nonlinear discriminant function in the original space is realized, and the SVM algorithm can obtain the optimal solution for a limited sample without multi-sample acquisition, which is suitable for the single photo case of the ID card in the invention, further improving the identity authentication of the present invention. The accuracy. The specific algorithm of SVM classification can refer to the following process:

(1) Support Vector Classifier (SVC)

For two types of linear separable tasks, SVC needs to find a hyperplane with the largest interval to separate the two types of samples. The large interval can ensure the best generalization ability of the hyperplane. From the visual point of view, the interval is two classes. The size of the space between them or the degree of separation determined by the hyperplane. Geometrically, the interval corresponds to the shortest distance from the data point to the hyperplane, and zeros w and b represent the weight vector and the optimum, respectively. Hyperplane offset, the corresponding hyperplane can be defined as

w ^T x+b=0 (5)

The geometric distance from the sample data x to the hyperplane is

Where g(x)=w ^T +b is the discriminant function determined by the hyperplane and also becomes the functional interval of x for a given w and b.

有SVC is to find the optimal hyperplane parameter values w and b to maximize the classification interval between the two classes. Therefore, SVC is also called the maximum interval classifier. Fix the interval to 1, then for a given training set

Have

There are some data points that satisfy the formula (7), called the support vector, and the distance from the support vector to the optimal hyperplane is

The classification interval can be expressed as

In order to find the hyperplane with the largest interval, SVC should maximize ρ with w and b:

Equivalent to

In general, the Lagrangian dual solution (11) is optimized.

For partial deviations of w and b of L(w, b, α), and let it be zero, get the optimization condition

Bring in (12) to get the dual problem

After determining the optimal Lagrangian multiplier, the optimal weight vector w ^{* in} (13) can be calculated.

Thus determining the optimal offset b ^* =1-w ^T x _s when y _s = +1

(2) Soft interval optimization of support vectors

For real-world data, there are often some anomalies. These points are training errors for the maximum interval hyperplane. The soft interval is to extend the SVC algorithm so that the hyperplane allows a small amount of noise to exist. By introducing the slack variable ζ _i Quantifier classifier error

Parameter C is used to balance the complexity and fault tolerance of the classifier. It is a regularization parameter selected by the user. Using the same Lagrangian multiplier method as before, the optimal weight w ^* and the optimal offset b ^* can be determined.

(3) Kernel function

The kernel function transforms data from a low-dimensional nonlinear space to a high-dimensional space, thereby making the data The high dimensional feature space is linearly separable, and the kernel function is defined as follows:

K(x,x')=φ(x) ^T φ(x') (17)

表示从低维输入空间到高维特征空间的映射，带入公式(15)中的最优权重向量表达式中，可得最优分类器among them

The mapping from the low-dimensional input space to the high-dimensional feature space is introduced into the optimal weight vector expression in equation (15), and the optimal classifier can be obtained.

In general, the RBF core is preferred.

(4) Probability prediction

Given the sample x to be tested, we want to get the probability of categorizing it as a class to compare with a preset threshold to determine the classification label. Fit with sigmod function, conditional probability:

通过极大似然估计确定模型参数A、B。对待测样例x，首先由训练好的SVM模型得到f，再通过此概率模型便可得到类别标签为1的概率，将此概率值与预先设定好的阈值相比较，最终确定类别标签。阈值的选取需要通过验证集确定，在具体的做法中，可将验证集中每对验证样本的类别概率从小到大排序，依次作为阈值，得到每个阈值对应的误判率和认证率，根据事先要求的误判率确定最佳阈值。Training sample set

The model parameters A, B are determined by maximum likelihood estimation. To test sample x, first obtain f from the trained SVM model, and then obtain the probability that the category label is 1 by using the probability model, compare the probability value with a preset threshold, and finally determine the category label. The selection of the threshold needs to be determined by the verification set. In a specific method, the class probability of each pair of verification samples in the verification set can be sorted from small to large, and sequentially used as a threshold, and the false positive rate and the authentication rate corresponding to each threshold are obtained, according to the prior The required false positive rate determines the optimal threshold.

若为1，则判为同一个人，为-1则判为不同人。After determining the classifier model through the training sample, for a ID image to be tested and a face image collected in the field, after normalization, first extract the respective GOP features, and calculate the classification feature x, the trained SVM classifier, Forecast label

If it is 1, it is judged to be the same person, and if it is -1, it is judged as a different person.

In the feature classification process of the present invention, in addition to the SVM classification, a random forest method can also be adopted. The so-called random forest is a classifier that contains multiple decision trees. The so-called randomness means that the generation of each decision tree is random (the training samples of each tree are randomly selected from the original samples). The set of attributes used in the decision tree construction process is also selected by random selection of equal probability, that is, random selection of training samples and random selection of splitting attributes. When there is one When a new input sample enters, each decision tree in the forest judges the input samples separately, obtains the respective decision results, and then summarizes the decision results to the forest output. For classification problems, simply count all tree-to-category votes, select the one with the highest number of votes, and assign the input samples to that category.

As a further improvement of the present invention, step S104 may further be: performing a point multiplication operation on the GOP features of the obtained face image sample pairs, obtaining a pair of cosine similarity vectors as feature vectors, and using SVM two-class processing to calculate and judge Whether the face image sample pair belongs to the same person.

In this step, the specific calculation process of the cosine similarity vector is as follows:

Face authentication can be regarded as a two-category problem. Given a pair of face images I ₁ (x, y) and I ₂ (x, y), which represent ID photos and live video photos, our purpose is It is judged whether I ₁ and I ₂ are from the same person, and if so, the judgment result is 1, and if not, it is judged as -1, therefore, it is necessary to map I ₁ and I ₂ to the feature space.

x=F(I ₁ ,I ₂ ) (20)

Where x∈R ^d represents the feature vector extracted from a pair of faces, and the feature extraction function F: I×I→R ^{d is} defined as follows:

In formula (21), .× denotes the point multiplication operation of the matrix, and the cosine similarity between the GOP features of the pair of face images calculated by the equation is point by point, and the result is a d-dimensional column vector as the feature classification process. input of.

In this step, the cosine similarity vector of the gradient direction pyramid is used as the feature vector, which can comprehensively realize the feature extraction of the face image, and the cosine similarity vector is robust to external factors such as illumination, expression and age, and further strengthens The accuracy and anti-interference of the present invention.

In the present invention, as shown in FIG. 4, before step S102, the method may include:

Step S1021: performing face detection processing on the face image sample pair;

Step S1022: Perform feature point setting on the face image sample pair after the face detection processing Bit processing

Step S1023: Normalize the face image sample pairs subjected to the feature point positioning process.

In this step, face detection is the primary part of face analysis, and the problem is to confirm whether there is a face image in the image. In the present invention, Adaboost face detection algorithm is preferably adopted, and Adaboost algorithm is a classifier algorithm. The basic idea is to use a large number of simple classifiers with general classification ability to superimpose them by a certain method to form a strong classifier with strong classification ability, and then connect several strong classifiers into a classifier to complete image search and detection. The detector has a fast detection speed and a short development cycle, which is feasible. After the face detection is completed, if there is a face image, the feature image is positioned on the face image. In the present invention, the feature point location may adopt various methods known to those skilled in the art, such as a method based on geometric features, based on template matching. Methods, model-based methods, etc. In the present invention, a method based on gradation characteristics and geometric features is preferably adopted, which can realize feature point localization of a face image simply and effectively. After the feature point positioning is completed, for each face image, according to the binocular positioning coordinates, the eyes are aligned to the same position set in advance by image registration, and normalization is achieved. In the present invention, the face image is normalized. The processing may also employ various methods known to those skilled in the art, such as maximum-minimum normalization, Fourier transform-based face image normalization, and the like.

On the other hand, corresponding to the above method, the present invention further provides an identity authentication device, as shown in FIG. 5, including:

The obtaining module 11 is configured to obtain a pair of face image samples, wherein one face image is an ID card photo, and another face image is an image collected in the field;

The sampling module 12 is configured to process the acquired face image sample pair by using a face image pyramid algorithm to obtain a face image pyramid structure;

The calculating module 13 is configured to calculate a grayscale gradient direction for the face image of each scale in the pyramid image structure of the obtained face image sample pair, to obtain a gradient direction pyramid GOP feature;

The determining module 14 is configured to calculate, according to the GOP feature of the obtained face image sample pair, whether the pair of face image samples belongs to the same person.

The identity authentication apparatus of the present invention first acquires a pair of face image samples by the module 11; The post-sampling module 12 processes the acquired face image sample pairs by using the face image pyramid algorithm to obtain a pyramid structure of the face image. The method uses a multi-resolution method to represent the face image, and can comprehensively describe the face image. The feature effectively improves the accuracy of the identity authentication of the present invention; then the calculation module 13 calculates the grayscale gradient direction of the face image of each scale in the pyramid image structure of the obtained face image sample pair to obtain the gradient direction. The pyramid GOP feature enhances the anti-interference of the present invention, and further improves the accuracy of the face authentication result. Finally, the determining module 14 calculates, according to the obtained GOP feature of the face image sample pair, whether the face image sample pair is the same. personal. The identity authentication method of the invention effectively avoids interference of factors such as illumination, age and expression, and also significantly improves the accuracy of identity authentication.

As an improvement of the present invention, the sampling module 12 can be further configured to obtain a face image pyramid structure by using Gaussian kernel convolution smoothing and downsampling on the acquired face image sample pairs.

In most cases, features that are not easily perceived by a face image at a single scale are easily captured at multiple scales. Therefore, in the present invention, it is preferable to use a Gaussian kernel function convolution smoothing and downsampling operation to the face image pyramid. The structure performs multi-scale and multi-resolution representation of the face image, and the method can describe the face image feature more comprehensively, and enhances the accuracy of the identity authentication in the present invention. It is preferably employed in the present invention.

Preferably, the determining module 14 is further configured to calculate a feature vector for the GOP feature of the obtained face image sample pair, and use the SVM binary classification process to calculate whether the face image sample pair belongs to the same person.

For the non-linear learning task in the identity authentication process of the present invention, the SVM feature classification algorithm adopted by the judging module 14 converts the actual problem into a high-dimensional feature space by nonlinear transformation, and constructs a linear discriminant function in a high-dimensional space to realize The nonlinear discriminant function in the original space, and the SVM algorithm can obtain the optimal solution for a limited sample, without multi-sample acquisition, and is suitable for the single photo case of the ID card in the present invention, further improving the identity authentication of the present invention. accuracy.

As an improvement of the present invention, the determining module 14 may be further configured to perform a point multiplication operation on the GOP features of the obtained face image sample pairs, and obtain a pair of cosine similarity vectors as feature vectors, and use SVM two-class processing to calculate and judge Whether the face image sample pair is the same personal.

In the present invention, the cosine similarity vector of the gradient direction pyramid is used as the feature vector, which can comprehensively realize the feature extraction of the face image, and the cosine similarity vector is robust to external factors such as illumination, expression and age, and further strengthens. The accuracy and anti-interference of the present invention.

As a further improvement of the present invention, as shown in FIG. 6, the acquisition module 11 and the sampling module 12 are further connected with:

The detecting module 121 is configured to perform face detection processing on the face image sample pair;

The positioning module 122 is configured to perform feature point positioning processing on the face image sample pair subjected to the face detection processing;

The normalization module 123 is configured to perform normalization processing on the face image sample pairs subjected to the feature point positioning processing.

In the present invention, the problem that the detection module 121 processes is to confirm whether there is a face image in the image, so as to prevent the collected non-face image from entering the device, resulting in a waste of device computing resources. The positioning module 122 adopts feature point location processing on the acquired face image sample pair, thereby avoiding the influence of posture and location information on identity authentication. The normalization module 123 normalizes the face image sample pairs processed by the feature point positioning to enhance the consistency of the face image sample pairs, thereby avoiding the influence of factors such as illumination, direction and noise, and improving the invention. Anti-interference.

In the present invention, as shown in FIG. 7, in the specific work, the following steps can be performed:

Step 1: Obtain a pair of face image samples, wherein one face image sample is taken from the ID database, and another face image sample is taken from the live collection video library;

Step 2: adopting feature point location processing on the acquired face image sample pair;

Step 3: normalizing the face image sample pairs after the feature point positioning processing;

Step 4: obtaining a face image pyramid structure by normalizing the face image sample pair by Gaussian kernel convolution smoothing and downsampling;

Step 5: Calculate the gradient direction of the face image at each scale in the pyramid structure of the face image, and obtain the gradient direction pyramid GOP feature of the face image.

Step 6: Calculate the cosine similarity vector of the gradient direction pyramid for the face image sample pair to obtain the face feature expression.

Step 7: Using the SVM classification algorithm to perform class probability prediction on the obtained face image representation;

Step 8: The obtained class probability prediction is compared with a given threshold (Th);

Step 9: Obtain the authentication result, wherein the probability value is greater than Th, then the value is 1, indicating that the face graph sample pair is from the same person, and the probability value is less than Th, then the value is -1, indicating that the face graph sample pair is different. People.

In the above embodiment, the face gradient direction pyramid is used to obtain the feature expression of the face, and the gradient information of the face at different scales is utilized, and only the gradient direction is discarded, and the gradient amplitude is discarded. Experiments show that the method can catch people more. The essential information of the face is robust to factors such as illumination, expression and age. At the same time, experiments show that the algorithm is simple and efficient, and the time complexity and space complexity are relatively low. The algorithm directly solves the ID card authentication problem from the perspective of face authentication, avoids the problem caused by the single sample of the ID card photo, and the test of the second generation ID card photo of 229 volunteers and a video photo taken on site. In the library test, when the system's error acceptance rate is 1%, 10%, the corresponding correct acceptance rates are 82.97% and 99.13%, respectively.

The above is a preferred embodiment of the present invention, and it should be noted that those skilled in the art can also make several improvements and retouchings without departing from the principles of the present invention. It should be considered as the scope of protection of the present invention.

Claims (10)

An identity authentication method, comprising: Obtaining a pair of face image samples, wherein one of the face images is an ID card photo, and the other face image is an image acquired in the field; The face image pyramid pair is processed by the face image pyramid algorithm to obtain a pyramid structure of the face image; Calculating a grayscale gradient direction on the face image of each scale in the pyramid image structure of the obtained face image sample pair to obtain a gradient direction pyramid GOP feature; According to the obtained GOP feature of the face image sample pair, it is calculated whether the face image sample pair belongs to the same person. The identity authentication method according to claim 1, wherein the face image pyramid algorithm is used to process the acquired face image sample pair, and the pyramid structure of the face image is further: A face image pyramid structure is obtained by using Gaussian kernel convolution smoothing and downsampling for the acquired face image sample pairs. The identity authentication method according to claim 1, wherein the calculating, according to the GOP feature of the obtained face image sample pair, determining whether the face image sample pair belongs to the same person is further: The feature vector is calculated for the GOP feature of the obtained face image sample pair, and the SVM binary classification process is used to calculate whether the face image sample pair belongs to the same person. The identity authentication method according to claim 1, wherein the calculating, according to the GOP feature of the obtained face image sample pair, determining whether the face image sample pair belongs to the same person is further: The GOP feature of the obtained face image sample pair is subjected to a point multiplication operation to obtain a pair of cosine similarity vectors as the feature vector, and the SVM binary classification process is used to calculate whether the face image sample pair belongs to the same person. An identity authentication method according to any one of claims 1 to 4, characterized in that The face image pyramid algorithm is used to process the acquired face image sample pairs, and the face image pyramid structure is obtained before: Perform face detection processing on face image sample pairs; Perform feature point location processing on the face image sample pairs after the face detection processing; The face image sample pairs processed by the feature point are normalized. An identity authentication device, comprising: The acquisition module is configured to obtain a pair of face image samples, wherein one face image is an ID card photo, and another face image is an image acquired in the field; Sampling module: used to process the acquired face image sample pairs by using the face image pyramid algorithm to obtain a pyramid structure of the face image; The calculation module is configured to calculate a gray gradient direction of the face image of each scale in the pyramid image structure of the obtained face image sample pair to obtain a gradient direction pyramid GOP feature; The judging module is configured to calculate, according to the GOP feature of the obtained face image sample pair, whether the pair of face image samples belong to the same person. The identity authentication apparatus according to claim 6, wherein the sampling module is further configured to obtain a face image pyramid structure by using Gaussian kernel convolution smoothing and downsampling on the acquired face image sample pairs. The identity authentication device according to claim 6, wherein the determining module is further configured to calculate a feature vector for the GOP feature of the obtained face image sample pair, and use SVM binary classification processing to calculate and determine the face image sample. Whether it belongs to the same person. The identity authentication apparatus according to claim 6, wherein the determining module is further configured to perform a point multiplication operation on the GOP features of the obtained face image sample pair to obtain a pair of cosine similarity vectors as feature vectors. The SVM binary classification process is used to calculate whether the face image sample pairs belong to the same person. The identity authentication device according to claim 6, wherein the connection between the acquisition module and the sampling module is: a detecting module, configured to perform face detection processing on a face image sample pair; a positioning module for performing feature points on the face image sample pairs after the face detection processing Positioning processing The normalization module is configured to normalize the face image sample pairs after the feature point positioning processing.

Images