CN104766063B - A kind of living body faces recognition methods - Google Patents

Abstract

The invention discloses a living body face recognition method, which is divided into a training stage and a recognition stage. In the training stage, multiple live face images and photo face images are obtained, and then the grayscale image of each live face image is extracted. The eigenvectors of each photo face image are used as positive samples, and the eigenvectors of the grayscale images of each photo face image are extracted as negative samples, and then all positive samples and negative samples are input into the SVM classifier for training to obtain the SVM classifier training model; In the recognition stage, a frame of face image is obtained, and the face recognition technology is used for recognition first. When the recognition result is a legitimate user, the feature vector is extracted by the living body detection technology, and then the feature vector is input into the SVM classifier training model for live body detection; The advantage is to use face recognition technology to determine whether the face is a legitimate user, and then use live body detection technology to determine whether the source of the face is a live face or a fake photo face when the user is a legitimate user, thus effectively eliminating the security risks caused by photo faces .

Description

A live face recognition method

technical field

The invention relates to a face recognition technology, in particular to a living face recognition method.

Background technique

Face recognition technology is a kind of biometric identification technology, which has achieved rapid development in recent years due to its advantages of convenience, speed and accuracy. The input of the face recognition system is generally a face image containing the identity to be detected, and several face images of known identities in the face database, and its output is a series of face similarity scores, In this way, the identity of the recognized face is indicated. At present, face recognition technology has been widely used in criminal investigations, banking systems, customs inspections, civil affairs departments, work and attendance, and other fields. However, with the continuous expansion of the application range of face recognition technology, some security problems have also occurred. Criminals use forged face photos to deceive the face recognition system, thus causing major economic losses to legitimate users. Therefore, it is particularly important to judge the authenticity of the source of the face image, which is liveness detection.

After Google released Android 4.0, it brought the function of unlocking mobile phones through face recognition to the majority of mobile phone users, but there have been reports of using personal photos instead of real people to unlock mobile phones, so Google has been using face recognition cautiously and conservatively. recognition technology. In order to make the face recognition system mature, the potential safety hazards of replacing real living human faces with such photos must be solved, so it is necessary to study a technology for recognizing living human faces.

Contents of the invention

The technical problem to be solved by the present invention is to provide a living face recognition method, which can determine whether the face is a legitimate user, and can also determine whether the source of the face is a living face, effectively eliminating the security risks brought by the photo face. Hidden danger.

The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a living body face recognition method, which is characterized in that it comprises the following steps:

① Obtain M pieces of live face images containing different face objects with a size of 256×256, and then obtain photo face images of each live face image, and the size of each photo face image is 256×256; then convert M live face images and M photo face images into grayscale images, and form a training image set with 2M grayscale images; then calculate each grayscale image in the training image set The feature vector of the grayscale image of each piece of live face image is used as a positive sample, and is marked with +1, and the feature vector of the grayscale image of each photo face image is used as a negative sample, and Marked by -1; finally, all positive samples and all negative samples are input into the SVM classifier for training, and the SVM classifier training model is obtained;

② When live face recognition is required, obtain a frame of face image containing the face object to be recognized, and then intercept the smallest rectangular area where the face object is located in the face image, and then calculate the size of the rectangular area Regularly, obtain the image of the face area to be recognized with a size of 256×256, and then convert the image of the face area to be recognized into a grayscale image;

③Recognize the grayscale image of the face area image using face recognition technology, if the recognition result is a legitimate user, then perform step ④; if the recognition result is an illegal user, reject the face verification, and the face verification fails;

④Using the living body detection technology, first calculate the feature vector of the grayscale image of the face area image, and then input the feature vector of the grayscale image of the face area image into the SVM classifier training model. If the SVM classifier training model outputs + 1, it means that the source of the face area image is a live face, and the face verification is successful; if the SVM classifier training model outputs -1, it means that the source of the face area image is a photo face, and the face verification is rejected. verification failed.

The acquisition process of the feature vector of each grayscale image in the training image collection in the described step 1. is the same as the acquisition process of the feature vector of the grayscale image of the human face area image in the described step 4., the training image collection Each grayscale image of and the grayscale image of the face region image are taken as an image to be processed, and the process of obtaining the feature vector of the image to be processed is:

a. Divide the image to be processed into A non-overlapping image block with a size of 64×64;

b. Define the i-th image block currently to be processed in the image to be processed as the current image block, wherein, 1≤i≤16, and the initial value of i is 1;

c. Use a sliding window with a size of 3×3 to slide pixel by pixel in the current image block, and divide the current image block into (64-2)×(64-2) overlapping sizes of 3×3 sub-block;

d. Convolve Sobel operators in eight different directions with each sub-block in the current image block to obtain the gradient values of each sub-block in the current image block in eight different directions, and use the current image block in the The gradient value of the jth sub-block in the kth direction is recorded as Among them, eight Sobel operators in different directions are 0° Sobel operator, 45° Sobel operator, 90° Sobel operator, 135° Sobel operator, 180° Sobel operator, 225° Sobel operator Sobel operator, 270° Sobel operator, 315° Sobel operator, 1≤j≤(64-2)×(64-2), 1≤k≤8;

e. According to the order of all sub-blocks in the current image block, arrange the gradient values of all sub-blocks in the current image block in each direction to form the dimension of the current image block in each direction as (64-2)×( 64-2) gradient value vector, the dimension of the current image block formed by arranging the gradient values of all sub-blocks in the k-th direction in the current image block in the k-th direction is (64-2)×(64- 2) The gradient value vector is denoted as Among them, the symbol “[]” is a vector representation symbol, Indicates the gradient value of the first sub-block in the current image block in the k-th direction, Indicates the gradient value of the second sub-block in the current image block in the k-th direction, Indicates the gradient value of the (64-2)×(64-2)-1th sub-block in the current image block in the k-th direction, Indicates the gradient value of the (64-2)×(64-2)th sub-block in the current image block in the k-th direction;

f, make i=i+1, use the next image block to be processed in the image to be processed as the current image block, and then return to step c to continue until all image blocks in the image to be processed are processed, and the image to be processed is obtained The dimension of each image block in eight different directions is a gradient value vector of (64-2)×(64-2), wherein, "=" in i=i+1 is an assignment symbol;

g, according to the order of all image blocks in the image to be processed, the gradient value vectors of all image blocks in the image to be processed are arranged in eight different directions respectively to form the feature vector of the image to be processed, denoted as T, T=[TV ₁ ¹ ,TV ₂ ¹ ,…,TV ₈ ¹ ,TV ₁ ² ,TV ₂ ² ,…,TV ₈ ² ,…,TV ₁ ¹⁶ ,TV ₂ ¹⁶ ,…,TV ₈ ¹⁶ ], where the symbol “ []” is a vector representation symbol, TV ₁ ¹ represents the gradient value vector whose dimension of the first image block in the first direction is (64-2)×(64-2), TV ₂ ¹ represents the first image The dimension of the block in the second direction is (64-2)×(64-2) gradient value vector, TV ₈ ¹ means that the dimension of the first image block in the eighth direction is (64-2)× The gradient value vector of (64-2), TV ₁ ² represents the gradient value vector whose dimension of the second image block in the first direction is (64-2)×(64-2), TV ₂ ² represents the second The dimension of the second image block in the second direction is the gradient value vector of (64-2)×(64-2), TV ₈ ² means that the dimension of the second image block in the eighth direction is (64-2 )×(64-2) gradient value vector, TV ₁ ¹⁶ represents the gradient value vector whose dimension of the 16th image block in the first direction is (64-2)×(64-2), TV ₂ ¹⁶ represents The dimension of the 16th image block in the second direction is the gradient value vector of (64-2)×(64-2), TV ₈ ¹⁶ means that the dimension of the 16th image block in the 8th direction is (64 -2)×(64-2) vector of gradient values.

Compared with the prior art, the present invention has the advantages of:

1) The inventive method has added living body detection technology on the basis of existing face recognition technology, and the inventive method obtains the SVM classifier training model to a plurality of living body face images and a plurality of photo face images in the training stage , in the identity authentication process, use face recognition technology to determine whether the face is a legitimate user, and use live detection technology to first calculate the feature vector of the face area image, and then input the feature vector of the face area image into the SVM classifier training In the model, it is determined whether the source of the face is a live face or a fake photo face, thereby effectively eliminating the security risks caused by the photo face and realizing the double protection of private information security.

2) The method of the present invention is applicable to platforms with low processing capabilities, such as the Android system.

3) The method of the present invention only needs one camera, and it can judge whether it is a living human face by extracting a frame of human face image, which greatly reduces the proportion of system resources occupied, and does not need to add additional image auxiliary equipment, and does not require the active cooperation of users , very natural.

4) The method of the present invention recognizes the images in the live face database of Nanjing University of Aeronautics and Astronautics (NUAA) in the experimental verification stage, and the accuracy rate of identifying two types of face images reaches 98.718%.

Description of drawings

Fig. 1 is the overall realization block diagram of the inventive method;

Fig. 2a is the gradient value vector distribution figure of all image blocks in the grayscale image of a living human face image respectively in eight different directions;

Fig. 2 b is the gradient value vector distribution diagram of all image blocks in the grayscale image of the photo face image corresponding to Fig. 2 a in eight different directions;

Fig. 3 is an ROC curve diagram of the working characteristics of the living body detection technology in the method of the present invention.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

A kind of live face recognition method that the present invention proposes, and it is applicable to Android system, and its overall realization block diagram is as shown in Figure 1, and it comprises the following steps:

① Obtain M pieces of live face images containing different face objects with a size of 256×256 through the mobile phone camera, and then obtain the photo face images of each live face image through the mobile phone camera. The size of the face image is 256×256; then M pieces of live face images and M pieces of photo face images are converted into grayscale images, and 2M grayscale images are formed into a training image set; then the training image set is calculated The feature vector of each gray-scale image of each piece of face image; then the feature vector of the gray-scale image of each live face image is used as a positive sample, and marked with +1, the feature vector of the gray-scale image of each photo face image As a negative sample, it is marked with -1; finally, all positive samples and all negative samples are input into the SVM classifier for training, and the SVM classifier training model is obtained; among them, M≥50, usually positive samples and negative samples in the training stage The more samples there are, the more accurate the classification result of the SVM classifier training model will be. Therefore, M can be properly set to a larger value during specific operations. For example, M=261 can be used during specific operations.

Here, the face objects in the M live face images are different, that is, M different faces are photographed to obtain different live face images; The image obtained by taking another shot.

② When live face recognition is required, a frame of face image containing the face object to be recognized is obtained through the camera of the mobile phone, and then the smallest rectangular area where the face object is located is intercepted in the face image, and then the rectangular area The size is regularized to obtain the image of the face region to be recognized with a size of 256×256, and then the image of the face region to be recognized is converted into a grayscale image.

③Use the existing face recognition technology to recognize the grayscale image of the face area image, if the recognition result is a legitimate user, then perform step ④; if the recognition result is an illegal user, then reject the face verification, and the face verification fails . Here, when the recognition result is an illegal user, the source of the face area image may be a live face, or a photo face, that is, people other than the user cannot pass the face verification.

④Using the living body detection technology, first calculate the feature vector of the grayscale image of the face area image, and then input the feature vector of the grayscale image of the face area image into the SVM classifier training model. If the SVM classifier training model outputs + 1, it means that the source of the face area image is a live face, and the face verification is successful; if the SVM classifier training model outputs -1, it means that the source of the face area image is a photo face, and the face verification is rejected. verification failed.

In this specific embodiment, the acquisition process of the feature vector of each grayscale image in the training image set in step 1. is the same as the acquisition process of the feature vector of the grayscale image of the face area image in step 4. The training image set Each gray-scale image and the gray-scale image of the face area image in the image are regarded as an image to be processed. The biggest difference between the live face and the fake photo face is that the specular reflection of the latter is much larger than that of the former. The light intensity of the diffuse reflection light generated on the surface of the face is always proportional to the light intensity of the incident light in all directions on the face surface and the cosine value of the incident angle, while the face of the fake photo does not conform to the facial unevenness, and the degree of smoothness is high. The amount of reflection is high, and the face object of the photo will be a linear combination of diffuse reflection and specular reflection components, and the specular reflection and diffuse reflection share weights, so the inventive method divides the image to be processed into 16 non-overlapping image blocks, Then divide each image block into (64-2)×(64-2) overlapping sub-blocks with a size of 3×3, and then use each sub-block to convolve with eight Sobel operators in different directions The result of the operation to obtain the feature vector of the image to be processed, that is, the process of obtaining the feature vector of the image to be processed is:

a. Divide the image to be processed into non-overlapping image blocks with a size of 64×64.

b. Define the i-th image block currently to be processed in the image to be processed as the current image block, where 1≤i≤16, and the initial value of i is 1.

c. Use a sliding window with a size of 3×3 to slide pixel by pixel in the current image block, and divide the current image block into (64-2)×(64-2) overlapping sizes of 3×3 subblock.

d. Convolve Sobel operators in eight different directions with each sub-block in the current image block to obtain the gradient values of each sub-block in the current image block in eight different directions, and use the current image block in the The gradient value of the jth sub-block in the kth direction is recorded as It is obtained by convolving the Sobel operator in the k-th direction with the j-th sub-block in the current image block; where, as listed in Table 1, the Sobel operators in eight different directions are 0° Sobel operators , 45° Sobel operator, 90° Sobel operator, 135° Sobel operator, 180° Sobel operator, 225° Sobel operator, 270° Sobel operator, 315° Sobel operator, 1≤j≤(64-2)×(64-2), 1≤k≤8.

Table 1 Sobel operators in eight different directions

e. According to the order of all sub-blocks in the current image block, arrange the gradient values of all sub-blocks in the current image block in each direction to form the dimension of the current image block in each direction as (64-2)×( 64-2) gradient value vector, the dimension of the current image block formed by arranging the gradient values of all sub-blocks in the k-th direction in the current image block in the k-th direction is (64-2)×(64- 2) The gradient value vector is denoted as Among them, the symbol “[]” is a vector representation symbol, Indicates the gradient value of the first sub-block in the current image block in the k-th direction, Indicates the gradient value of the second sub-block in the current image block in the k-th direction, Indicates the gradient value of the (64-2)×(64-2)-1th sub-block in the current image block in the k-th direction, Indicates the gradient value of the (64-2)×(64-2)th sub-block in the current image block in the k-th direction.

f, make i=i+1, use the next image block to be processed in the image to be processed as the current image block, and then return to step c to continue until all image blocks in the image to be processed are processed, and the image to be processed is obtained The dimensions of each image block in eight different directions are (64-2)×(64-2) gradient value vectors, where "=" in i=i+1 is an assignment symbol.

g, according to the order of all image blocks in the image to be processed, the gradient value vectors of all image blocks in the image to be processed are arranged in eight different directions respectively to form the feature vector of the image to be processed, denoted as T, T=[TV ₁ ¹ ,TV ₂ ¹ ,…,TV ₈ ¹ ,TV ₁ ² ,TV ₂ ² ,…,TV ₈ ² ,…,TV ₁ ¹⁶ ,TV ₂ ¹⁶ ,…,TV ₈ ¹⁶ ], where the symbol “ []” is a vector representation symbol, TV ₁ ¹ represents the gradient value vector whose dimension of the first image block in the first direction is (64-2)×(64-2), TV ₂ ¹ represents the first image The dimension of the block in the second direction is (64-2)×(64-2) gradient value vector, TV ₈ ¹ means that the dimension of the first image block in the eighth direction is (64-2)× The gradient value vector of (64-2), TV ₁ ² represents the gradient value vector whose dimension of the second image block in the first direction is (64-2)×(64-2), TV ₂ ² represents the second The dimension of the second image block in the second direction is the gradient value vector of (64-2)×(64-2), TV ₈ ² means that the dimension of the second image block in the eighth direction is (64-2 )×(64-2) gradient value vector, TV ₁ ¹⁶ represents the gradient value vector whose dimension of the 16th image block in the first direction is (64-2)×(64-2), TV ₂ ¹⁶ represents The dimension of the 16th image block in the second direction is the gradient value vector of (64-2)×(64-2), TV ₈ ¹⁶ means that the dimension of the 16th image block in the 8th direction is (64 -2)×(64-2) vector of gradient values.

In order to further illustrate the feasibility and effectiveness of the method of the present invention, the method of the present invention is verified experimentally.

The sample library selected on the Matlab experimental platform uses the public NUAA live detection face database, and 522 images are randomly selected from the NUAA live detection face database, including 261 live face images and photos of each live face image. Face images (that is, a total of 261 photo face images), the size of each image is 256×256. The type of each image is an RGB three-channel color image, which is converted into a grayscale image to facilitate image processing. Divide each image into Non-overlapping image blocks with a size of 64×64, and then use a sliding window with a size of 3×3 to slide pixel by pixel in each image block, and divide each image block into (64-2)× (64-2) overlapping sub-blocks with a size of 3×3, and then perform convolution operations on eight Sobel operators in different directions with each sub-block in each image block to obtain each image block The gradient value of each sub-block in eight different directions, and then obtain the gradient value vector of each image block in each direction. Figure 2a shows the gradient value vector distribution of all image blocks in the grayscale image of a living face image in eight different directions. It can be seen from Figure 2a that the gradient value of all image blocks in each direction The percentages of vectors vary. Figure 2b shows the gradient value vector distributions of all the image blocks in the grayscale image of the photographic face image corresponding to the live face image in Figure 2a in eight different directions. As can be seen from Figure 2b, all image blocks The percentage of the vector of gradient values in each direction varies. Comparing Figure 2a and Figure 2b, it can be seen that the level of distribution shown in Figure 2b is quite different.

According to the gradient value vectors of all the image blocks in the grayscale images of all living human face images obtained above in eight different directions, the feature vectors of the grayscale images of all living human face images are obtained, and randomly selected M eigenvectors are used as positive samples; and according to the gradient value vectors of all image blocks in the grayscale image of the photo face image of all live face images obtained above in eight different directions, obtain all live face images. The feature vector of the grayscale image of the photo face image, and the feature vector corresponding to the positive sample is used as a negative sample; all positive samples and all negative samples are input into the SVM classifier for training, and the SVM classifier training model is obtained. the remaining The eigenvectors of the grayscale images of M live face images and the remaining The eigenvectors of the gray-scale image of the photos of M live face images are input in the SVM classifier training model as the eigenvectors of the test respectively, and Fig. 3 has provided the operating characteristics ROC of the living body detection technique in the method of the present invention Curve, in Fig. 3, abscissa represents false positive probability (False Positive Rate), and ordinate represents true probability (True Positive Rate), from the area below the ROC curve that provides in Fig. 3 as can be known, the living body detection technology in the inventive method The detection rate reaches 98.718%, and the detection takes 167 seconds in total, and the average processing time of each face is less than one second, which fully demonstrates that the method of the present invention has lower complexity, and provides a better solution for the application of the Android platform. Theoretical basis, this makes the living body detection technology in the method of the present invention greatly improve the feasibility of perfecting the face recognition technology, and more effectively guarantees the use safety of the face recognition system.

Claims (1)

1. a living body face recognition method, is characterized in that comprising the following steps: ① Obtain M pieces of live face images containing different face objects with a size of 256×256, and then obtain photo face images of each live face image, and the size of each photo face image is 256×256; then convert M live face images and M photo face images into grayscale images, and form a training image set with 2M grayscale images; then calculate each grayscale image in the training image set The feature vector of the grayscale image of each piece of live face image is used as a positive sample, and is marked with +1, and the feature vector of the grayscale image of each photo face image is used as a negative sample, and Marked by -1; finally, all positive samples and all negative samples are input into the SVM classifier for training, and the SVM classifier training model is obtained; ② When live face recognition is required, obtain a frame of face image containing the face object to be recognized, and then intercept the smallest rectangular area where the face object is located in the face image, and then calculate the size of the rectangular area Regularly, obtain the image of the face area to be recognized with a size of 256×256, and then convert the image of the face area to be recognized into a grayscale image; ③Recognize the grayscale image of the face area image using face recognition technology, if the recognition result is a legitimate user, then perform step ④; if the recognition result is an illegal user, reject the face verification, and the face verification fails; ④Using the living body detection technology, first calculate the feature vector of the grayscale image of the face area image, and then input the feature vector of the grayscale image of the face area image into the SVM classifier training model. If the SVM classifier training model outputs + 1, it means that the source of the face area image is a live face, and the face verification is successful; if the SVM classifier training model outputs -1, it means that the source of the face area image is a photo face, and the face verification is rejected. verification failed; The acquisition process of the feature vector of each grayscale image in the training image collection in the described step 1. is the same as the acquisition process of the feature vector of the grayscale image of the human face area image in the described step 4., the training image collection Each grayscale image of and the grayscale image of the face region image are taken as an image to be processed, and the process of obtaining the feature vector of the image to be processed is: a. Divide the image to be processed into A non-overlapping image block with a size of 64×64; b. Define the i-th image block currently to be processed in the image to be processed as the current image block, wherein, 1≤i≤16, and the initial value of i is 1; c. Use a sliding window with a size of 3×3 to slide pixel by pixel in the current image block, and divide the current image block into (64-2)×(64-2) overlapping sizes of 3×3 sub-block; d. Convolve Sobel operators in eight different directions with each sub-block in the current image block to obtain the gradient values of each sub-block in the current image block in eight different directions, and use the current image block in the The gradient value of the jth sub-block in the kth direction is recorded as Among them, eight Sobel operators in different directions are 0° Sobel operator, 45° Sobel operator, 90° Sobel operator, 135° Sobel operator, 180° Sobel operator, 225° Sobel operator Sobel operator, 270° Sobel operator, 315° Sobel operator, 1≤j≤(64-2)×(64-2), 1≤k≤8; e. According to the order of all sub-blocks in the current image block, arrange the gradient values of all sub-blocks in the current image block in each direction to form the dimension of the current image block in each direction as (64-2)×( 64-2) gradient value vector, the dimension of the current image block formed by arranging the gradient values of all sub-blocks in the k-th direction in the current image block in the k-th direction is (64-2)×(64- 2) The gradient value vector is denoted as Among them, the symbol “[]” is a vector representation symbol, Indicates the gradient value of the first sub-block in the current image block in the k-th direction, Indicates the gradient value of the second sub-block in the current image block in the k-th direction, Indicates the gradient value of the (64-2)×(64-2)-1th sub-block in the current image block in the k-th direction, Indicates the gradient value of the (64-2)×(64-2)th sub-block in the current image block in the k-th direction; f, make i=i+1, use the next image block to be processed in the image to be processed as the current image block, and then return to step c to continue until all image blocks in the image to be processed are processed, and the image to be processed is obtained The dimension of each image block in eight different directions is a gradient value vector of (64-2)×(64-2), wherein, "=" in i=i+1 is an assignment symbol; g, according to the order of all image blocks in the image to be processed, the gradient value vectors of all image blocks in the image to be processed are arranged in eight different directions to form the feature vector of the image to be processed, denoted as T, Wherein, the symbol “[]” is a vector representation symbol here, and TV ₁ ¹ represents the gradient value vector whose dimension of the first image block in the first direction is (64-2)×(64-2), Indicates the gradient value vector whose dimension of the first image block in the second direction is (64-2)×(64-2), TV ₈ ¹ indicates that the dimension of the first image block in the eighth direction is ( 64-2)×(64-2) gradient value vector, TV ₁ ² represents the gradient value vector whose dimension of the second image block in the first direction is (64-2)×(64-2), Indicates that the dimension of the second image block in the second direction is the gradient value vector of (64-2)×(64-2), TV ₈ ² indicates that the dimension of the second image block in the eighth direction is ( 64-2)×(64-2) gradient value vector, TV ₁ ¹⁶ represents the gradient value vector whose dimension of the 16th image block in the first direction is (64-2)×(64-2), Indicates that the dimension of the 16th image block in the second direction is the gradient value vector of (64-2)×(64-2), TV ₈ ¹⁶ indicates that the dimension of the 16th image block in the 8th direction is ( 64-2)×(64-2) vector of gradient values.