CN103984932A - Anti-light face recognition method based on transform domain robust watermark under big data - Google Patents

Abstract

The invention discloses an anti-light face recognition method based on transform domain robust watermarks under big data. The anti-light face recognition method mainly includes a watermark embedding process and a watermark extraction process, wherein face recognition is realized in the watermark extraction process. The anti-light face recognition method mainly comprises the following steps that in the watermark embedding process, firstly, global DCT is conducted on all original faces so that feature vectors can be worked out, and secondly, the watermark of each face is associated with the feature vector of the corresponding face through a hash function in cryptology; in the watermark extraction process, thirdly, the feature vector of the face to be tested is obtained, the correlation coefficient maximum value of the feature vector of the face to be tested and the original face feature vector is worked out, face recognition is completed according to a serial number corresponding to the correlation coefficient maximum value, and the corresponding embedded watermark is obtained, and fourthly, the feature vector of the face to be tested is used for watermark extraction, and the correlation coefficient of the watermark is calculated. The anti-light face recognition method has the advantages of being applicable to the big data because sampling training is not needed, and has high resistance to light, occlusion and other attacks.

Description

Anti-illumination Attack Face Recognition Method Based on Transform Domain Robust Watermarking under Big Data

technical field

The invention relates to the field of multimedia signal processing, in particular to a face recognition method against light attack based on transform domain robust watermark under big data.

technical background

As an effective biometric identification technology, face recognition technology has been paid more and more attention by industry and academia in the past 40 years. Because face recognition technology has the advantages of being highly acceptable, natural, and not easy to be detected, it has great uses in entertainment, criminal investigation, access control systems, and military affairs.

At present, face recognition methods are mainly based on machine learning methods such as PCA, neural network, and SVM. Due to the need for training and learning, the recognition samples are relatively large. In the big data environment, the learning time is relatively long, and the current face Recognition methods are sensitive to illumination changes, expression changes, or masks. Therefore, how to solve face recognition methods that resist attacks such as illumination changes, expression changes, or masks in a big data environment is of great significance.

Digital watermarking technology was originally used for copyright protection of digital media on the Internet, and its important characteristics are robustness and invisibility; this invention can hide a person's signature or ID number as a watermark in its corresponding face image , using the robustness of watermark to realize the face recognition algorithm, especially it has good robustness to attacks such as illumination and occlusion. At present, there are few researches on face recognition methods based on big data environment and anti-illumination and occlusion attacks, and no public reports have been seen so far. Therefore, it is of great significance to study the robust watermarking technology based on transform domain to realize the face recognition method against light attack under big data.

Contents of the invention

Purpose of the invention: The technical problem to be solved by the present invention is to provide a high-speed, high-robust face recognition algorithm for the deficiencies of the prior art. Specifically, a face recognition method against light attack based on transform domain robust watermark under big data is disclosed, which is a zero watermark scheme, and the embedding of the watermark does not affect the original face image.

The basic principle of the present invention is: firstly carry out global DCT transformation to all face images; select the first 8x8 coefficients of the low-frequency part, then find a feature vector against light attack in the coefficients, and combine the watermark sequence with the feature vector Correlation realizes the embedding of the watermark; then for the image to be tested, first calculate its feature vector, then calculate the correlation coefficient of the feature vector of the image to be tested and the original image, and use the maximum value of the correlation coefficient to realize the detection of the face; and realize Watermark extraction.

Now the method of the present invention is described in detail as follows:

First, select a meaningful binary sequence as the watermark to be embedded in the face image, recorded as W={w(j)|w(j)=0,1; 1≤i≤L}; at the same time, select F as Original face image, expressing: F={f(i,j)|f(i,j)∈R; 1≤i≤M1, 1≤j≤N1}. Among them, f(i,j) represents the pixel value of the original face.

Part 1: Embedding of Watermark

1) By performing global DCT transformation on all original face images F(n), the feature vector set V(n) of the original image is obtained;

First, the global DCT transformation is performed on the original face image F(n), and the first 8×8 coefficients FD ₈ (i,j) are selected in the low-frequency coefficient matrix FD(i,j) in the frequency domain, and then the selected The coefficient matrix FD ₈ (i, j) is binarized. When the coefficient is greater than or equal to zero, it takes 1, and if it is less than 0, it takes zero to obtain the feature vector V. The main process is described as follows:

FD ₈ (i,j)=DCT2(F(i,j))

V(n)=BINARY(FD ₈ (i,j))

2) Using the cryptography HASH function to generate a binary key sequence Key(n) containing watermark information to realize the embedding of zero watermark;

Key(n)＝V(n)⊕W(n)

Key(n) is generated by the eigenvector V(n) of all original images and the corresponding n digital watermarks W(n) through the Hash function commonly used in cryptography; here W(n) consists of a random sequence with a length of 64 bits ;Save Key(n), which will be used when extracting the watermark below; Apply to a third party by using Key(n) as a key to obtain the right to use and ownership of the face image;

Part II: Face recognition and watermark extraction

3) Find the feature vector V' of the human face F' to be tested;

Assuming the face to be tested is F', after the global two-dimensional DCT transformation, the low-frequency coefficient matrix is obtained as FD'(i,j), and the low-frequency coefficient matrix is binarized according to step 1) to obtain the face to be tested eigenvector V';

FD ₈ '(i,j)=DCT2(F'(i,j))

V'=BINARY(FID'(i,j))

4) Calculate the correlation coefficient NC(n) between the eigenvector V' of the human face to be tested and the V(n) of the original human face, and carry out recognition of the human face;

Calculate the n value corresponding to the maximum value of the correlation coefficient between V' and V(n), and set n=k; according to the value of k, the key Key(k) can be obtained, and the original face image can be identified as F(k) and embedded in For the watermark value W(k) of F(k), the formula for calculating the normalized correlation coefficient of the feature vector is as follows:

$\mathrm{<span\; class="notranslate">\; NC</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; V</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}$

5) Utilize the binary logic key sequence Key(k) existing in the third party and the feature vector V' of the face to be tested to extract the watermark W'(k) in the image to be tested;

W'(k)=Key(k)⊕V'

6) Calculate the correlation coefficient between W' and W(k); calculate the correlation coefficient between the extracted watermark and the embedded watermark, and judge the content of the image to be tested according to the watermark.

Compared with the existing face recognition technology, the present invention has the following advantages:

First of all: better organically combine the robustness and invisibility of the watermark with face recognition; use the robustness algorithm of the watermark to realize the anti-lighting, masking, face distortion, etc. of the face algorithm attack, and the embedding of the watermark does not affect the pixel value of the original image; secondly: the watermark can also protect the privacy of the individual, and only authorized users can perform face recognition; finally, and because of the face recognition technology, there is no need for samples Learning, so it is suitable for a large number of face recognition, suitable for face recognition under big data.

The following is explained from the theoretical basis and experimental data:

1) Discrete cosine transform

The use of DCT in image coding is currently widely used in JPEG compression and MPEG-1/2 standards. DCT is the second best orthogonal transformation after the K-L transformation obtained under the minimum mean square error condition, and it is a lossless chief transformation. It has fast operation speed and high precision, and is famous for the best balance between the ability to extract feature components and operation speed.

The two-dimensional discrete cosine transform (DCT) formula is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>\underset{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; cos</span>}\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; m</span>}}\mathrm{<span\; class="notranslate">\; cos</span>}\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; v</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}$

u=0,1,...,M-1; v=0,1,...,N-1;

In the formula

$\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\sqrt{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}}& \mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; m</span>}}& \mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right.$ $\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\sqrt{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; N</span>}}& \mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}\\ \sqrt{\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; N</span>}}& \mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right.$

The two-dimensional inverse discrete cosine transform (IDCT) formula is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; cos</span>}\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; m</span>}}\mathrm{<span\; class="notranslate">\; cos</span>}\frac{\mathrm{<span\; class="notranslate">\; \&pi;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; v</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; N</span>}}$

x=0,1,...,M-1; y=0,1,...,N-1

Among them, x, y are sampling values in the space domain; u, v are sampling values in the frequency domain, and usually digital images are represented by a square matrix of pixels, that is, M=N

It can be seen from the above formula that the coefficient sign of DCT is related to the phase of the component.

2) A robust visual feature vector selection method for face images

At present, the reason why most face recognition algorithms have poor anti-attack ability to lighting, masking, and expression changes is that when these attacks are implemented, in the spatial domain, the pixel value changes greatly, especially the influence of indoor and outdoor lighting; It is of great significance to find a feature vector that is resistant to attacks such as lighting and masking. If it is possible to find an eigenvector that reflects the geometric characteristics of the face image, when the face image undergoes small geometric transformations and illumination changes, the eigenvector will not undergo significant mutations, then according to the eigenvector, use the maximum value of the correlation coefficient , we obtain the corresponding original image, and extract the corresponding watermark.

To this end, the experimental data of some conventional attacks such as light attack, mask attack and filtering are selected, as shown in Table 1. The original image used for testing in Table 1 is Figure 1, which is the first face in the ORL face database, created by the AT&T laboratory of the University of Cambridge; "Column 1" in Table 1 shows the face recognition algorithm subjected to Attack type, the face image after being attacked by different intensities of light is shown in Figure 2 to Figure 4, and the light intensity is from small to large; occlusion attack is shown in Figure 5 to Figure 7; including, research common glasses occlusion, mask occlusion and hat Occlusion; Figure 8 is a squeeze attack, which is similar to an expression change attack; conventional attacks such as Gaussian noise, JPEG compression, and median filtering are shown in Figures 9 to 12. "Column 2" of Table 1 indicates the peak signal-to-noise ratio (PSNR) of the face image after being attacked; "Column 3" to "Column 10" of Table 1 are low-frequency coefficients after DCT transformation, where Select eight low-frequency variation domain coefficients such as "D(1,1), D(1,2), ... D(2,4)". "Column 11" of Table 1 is the sequence of symbols used for feature extraction. Through the data in this table, we found that for attacks such as illumination and partial distortion of masks, the low-frequency values D(1,1), D(2,4) in the DCT transform domain may undergo some transformations, but the coefficient signs are still different. change, we record the coefficients greater than or equal to 0 as 1; those less than 0 as 0, then for the original face data, the coefficient values D(1,1), D(1,2)...D (2,4) and so on correspond to the symbol binary sequence: "1100 0011", see the 11th column of Table 1, observe this column, we can find that no matter the light attack, mask attack, distortion attack or conventional attack, the symbol sequence It is similar to that of the original data, and the normalized correlation coefficient of the symbol sequence of the original data is relatively large, as shown in column 12 of the table. The results show that the correlation coefficients are all greater than 0.5, which is the correlation coefficient value obtained by taking the first 64-bit symbol sequence of the low-frequency coefficients.

Table 1 The change value of DCT coefficients and some coefficients of face images after different attacks

*DCT number unit 1.0e+002

In order to further illustrate that the symbol sequence is a feature vector of a face image, we obtain the visual feature vectors of different face images, and then calculate the correlation coefficient between them; here we take the first 10 normal expressions of the ORL face database Faces are tested; Figure 13-Figure 22, from a statistical point of view, here are the first 8×8 64 DCT coefficients of the low-frequency part of the DCT domain. And calculate the correlation coefficient between each eigenvector, and the calculation results are shown in Table 2.

Table 2. Correlation coefficients between 10 different face image feature vectors (vector length 64bit)

V1 V2 V3 V4 V5 V6 V7 V8 V9 V10

V1 1.00 0.226 0.452 0.30 0.436 0.254 0.007 0.373 0.07 0.134 V2 0.226 1.00 0.029 0.09 0.095 -0.03 0.21 0.283 0.029 0.029 V3 0.452 0.029 1.00 0.254 0.38 0.25 0.058 0.192 0.247 0.059 V4 0.30 0.09 0.254 1.00 0.309 0.251 0.004 0.435 0.192 0.317 V5 0.436 0.095 0.38 0.309 1.00 0.188 0.129 0.184 0.066 0.192 V6 0.254 -0.03 0.25 0.251 0.188 1.00 0.375 0.125 0.50 0.187 V7 0.007 0.21 0.058 0.004 0.129 0.375 1.00 0.004 0.310 0.059 V8 0.373 0.283 0.192 0.435 0.184 0.125 0.004 1.00 0.310 0.129 V9 0.07 0.029 0.247 0.192 0.066 0.50 0.310 0.310 1.00 0.059 V10 0.134 0.029 0.059 0.317 0.192 0.187 0.059 0.129 0.059 1.00

It can be seen from Table 2 that, firstly, the correlation coefficient between the face images themselves is the largest, which is 1.00; secondly, the correlation coefficients of the image features of different faces are not greater than 0.5; this is consistent with what we actually observe with the human eye , which shows that the feature vectors extracted by the method of the invention reflect the main appearance features and main contours of the human face, and the more similar the appearance of the faces, the higher the similarity of the feature vectors.

3) The relationship between the length of the feature vector and the robustness of the face algorithm

According to the characteristics of human vision (HVS), low-intermediate frequency signals have a greater impact on human vision, and for two-dimensional images, it is the image contour. Therefore, when we select the appropriate transformation coefficient for the face image, we select the low intermediate frequency coefficient of the face image, the number of low intermediate frequency coefficients is selected and the size of the original face image for global DCT transformation, and the amount of information of the one-time embedded watermark It is related to the requirements, the smaller the length L of the selected feature vector, the less the amount of information embedded at one time, but the higher the robustness of the watermark. Considering the following experiments comprehensively, we choose the length of L to be 64 in the specific experiment.

Description of drawings

Figure 1 is the original face image.

Figure 2 is the face image after the light attack, the light intensity is S.

Figure 3 is the face image after the light attack, and the light intensity is M.

Figure 4 is the face image after the light attack, and the light intensity is L.

Figure 5 is an image of a face covered by glasses.

Figure 6 is a face image covered by a mask.

Figure 7 is a face image covered by a hat.

Figure 8 is a squeezed face image.

Figure 9 is a face image after Gaussian interference, and the Gaussian intensity is 3%.

Figure 10 is a face image after Gaussian interference, and the Gaussian intensity is 5%.

Figure 11 is a JPEG compressed face image with a compression quality of 5%.

Figure 12 is a face image after median filtering, the filtering parameters are [3x3], and the filtering times are 20.

Figure 13 The face image of the first person in the ORL face database.

Figure 14 The face image of the second person in the ORL face database.

Figure 15 The face image of the third person in the ORL face database.

Figure 16 The face image of the fourth person in the ORL face database.

Figure 17 The face image of the fifth person in the ORL face database.

Figure 18 The face image of the sixth person in the ORL face database.

Figure 19 The face image of the seventh person in the ORL face database.

Figure 20 The face image of the eighth person in the ORL face database.

Figure 21 The face image of the ninth person in the ORL face database.

Figure 22 The face image of the 10th person in the ORL face database.

Figure 23 The face image to be tested, without interference attack.

Figure 24 The original image detected when there is no attack.

Figure 25 is the extracted watermark when there is no attack.

Fig. 26 is a face image with an illumination intensity of -100%.

Fig. 27 When the light intensity is -100%, the detected original face image. Figure 28. Light intensity is -100% extracted watermark.

Figure 29 Face image occluded by glasses, the occlusion size is M.

Figure 30 The detected face image when the glasses are blocked.

Figure 31 The extracted watermark when the glasses are covered.

Figure 32 Face image covered by a mask, the size of the block is M.

Figure 33 Mask occlusion is the detected face image.

Figure 34 The digital watermark extracted when the mask is occluded.

Fig. 35 Face image occluded by a hat, the occlusion size is M.

Figure 36 The detected face image when the hat is occluded.

Figure 37 The extracted digital watermark when the hat is covered.

Figure 38 Face image with face extrusion, extrusion intensity is 60%.

Figure 39 is the recognized face image when the squeeze strength is 60%.

Figure 40 is the extracted digital watermark when the extrusion intensity is 60%.

Figure 41 Spherically distorted face image with a distortion amount of 40%.

Figure 42 is the detected face image when the amount of spherical distortion is 40%.

Figure 43 is the extracted watermark when the amount of spherical distortion is 40%.

Figure 44 Face image attacked by Gaussian noise, Gaussian noise intensity is 5%.

Figure 45 is the recognized face image when the Gaussian noise intensity is 5%.

Figure 46. Extracted watermark at 5% Gaussian noise intensity.

Figure 47 JPEG compressed face image, the compression quality is 5%.

Figure 48 is the original image identified at a JPEG compression of 5%.

Figure 49 extracts watermark information when JPEG compression is 5%.

Figure 50 is the face image of the median filter, the filter parameters are: [3x3], and the number of filters is 10.

Figure 51 After median filtering for [3x3], the detected original image.

Figure 52. Extracted watermark with median filter parameters [3x3].

Detailed ways

Below in conjunction with accompanying drawing, the present invention is further described, simulation platform is Matlab2010a, produces 1000 groups of independent binary pseudo-random sequences (value+1 or 0), each group of sequence length is 64bit, these 1000 groups of data are corresponding 1000 numbers Watermark, W(n). Each watermark corresponds to an original image. Here, the 500th group of watermarks corresponds to the first image of ORL, as shown in Figure 1. Its size is 92x112. This image is used here as an original image and a test image subject to zero attack.

In order to test the convenience of anti-illumination and mask attack experiments, another 1000 sets of binary random sequences (value +1, or 0), each set of sequence length is 64bit, and these 1000 sets are used as the features of 1000 original images Vector, V (n); The 500th unit is stored in the original feature vector of the first face figure in the ORL face storehouse;

The basic idea is, let F' be a face image to be tested without interference, see Figure 23, find its feature vector V', and then, by finding the correlation coefficient between V' and V(n) of all original images , find its maximum value; then determine the corresponding serial number k, so that when n=k, the correlation coefficient of V' and V(k) is the largest, according to the serial number k value, obtain the original image F(k) of recognition, and The embedded watermark W(k) and key Key(k) corresponding to the original image F(k), where Key(k)=W(k)⊕V(k); according to V' and Key(k), find The watermark information W' in the image F' to be tested, W'=V'⊕Key(k), and the correlation coefficient between W and W' is calculated, and the degree of correlation between the watermark and the image to be tested is determined according to the value of NC. owner.

Figure 23 is the face image to be tested when no interference is added;

Fig. 24 is the detected original face image;

Figure 25 is the extracted watermark, it can be seen that NC=1.00, the watermark can be extracted accurately.

Next, we use specific experiments to judge the ability of the robust watermark-based face recognition algorithm to resist light, occlusion and conventional attacks.

First test the face recognition algorithm against illumination, occlusion and distortion attacks.

(1) Light attack experiment

Table 3 is the experimental data of anti-light attack. It can be seen that when the light intensity of the image to be tested is reduced and the light intensity is -100%, the signal-to-noise ratio of the image to be tested is 7.21dB. By finding the maximum value of the correlation coefficient of the eigenvector, the corresponding original image is correctly detected. And the digital watermark can be extracted accurately, the correlation coefficient NC=0.81; while using the image search function of Google, when the light intensity of the image to be tested is -100%, the correct original image cannot be searched. It can be seen from Table 3 that when the light intensity is 100%, the signal-to-noise ratio of the image to be tested is 9.34dB, and the correct original image and watermark can still be detected, and the watermark correlation coefficient is 0.94; The face recognition algorithm of the watermark has a good ability to resist light attack.

Figure 26 is an image to be tested with an illumination intensity of -100%; the image is visually darker, and the signal-to-noise ratio (PSNR) is 7.21dB;

Fig. 27 is the recognized face image, and the correlation coefficient between the eigenvector of this figure and the eigenvector of the image to be tested is the maximum value;

Figure 28 is the extracted watermark, NC=0.81, it is easy to extract the watermark.

Table 3 Anti-light attack experimental data

(2) Blocking attack experiment

Occlusion is a difficult problem in the process of face recognition, but face recognition based on robust watermarking, using a robust watermarking algorithm, can better solve this problem. Table 4 is the experimental data of anti-occlusion attack. Including glasses occlusion, mask occlusion and hat occlusion; in the table, S, M, and L represent the relative size of the occluded area; as can be seen from Table 4, when the glasses occlude more, the occlusion size is L, and the PSNR of the occluded image = 13.30dB, the signal-to-noise ratio is low, but the original image can still be correctly identified and the watermark extracted, NC = 0.62; when the mask occludes a lot and the occlusion size is L, the PSNR of the occluded image is 16.20dB, and the watermark can still be extracted Out of the watermark, NC=0.59; when the hat is covered and the occlusion area is large, L, then the signal-to-noise ratio is PSNR=10.38dB, and the watermark can still be extracted, NC=0.56

Figure 29 is glasses occlusion, the image to be tested with an occlusion size of M; at this time, there is a certain occlusion effect, and the signal-to-noise ratio PSNR is 14.78dB;

Figure 30 is the detected original face image; the corresponding original image can be correctly identified;

Figure 31 is the extracted watermark, NC=0.75, the watermark can be extracted correctly.

Figure 32 is mask occlusion, the occlusion size is M, at this time the occlusion effect is obvious, the signal-to-noise ratio is low, and the PSNR is 17.31dB;

Figure 33 is the original image identified, it can be seen that the original image can be correctly identified;

Figure 34 is the extracted watermark, NC=0.78, the watermark of the image to be tested can be correctly extracted.

Figure 35 is a hat occlusion, the image to be tested with an occlusion size of M; at this time, the occlusion effect is obvious, and the signal-to-noise ratio PSNR is 10.64dB;

Figure 36 is a recognized face image; the corresponding original image can be correctly identified;

Figure 37 is the extracted watermark, NC=0.62, it is easy to detect the watermark.

Table 4 Anti-occlusion attack

(3) Face squeeze attack experiment

Table 5 is the experimental data of anti-face extrusion attack, sometimes the change of expression can be reflected in the extrusion of the face. It can be seen from the figure that when the face extrusion degree of the image to be tested is large and the extrusion amount is 100%, the signal-to-noise ratio of the image is as low as 16.60dB, but the original image can still be detected and the watermark extracted, NC =0.59;

Figure 38 is an image to be tested with a squeeze intensity of 60%; the facial expression has changed significantly, and the signal-to-noise ratio is 18.12dB;

Fig. 39 is the recognized face image, which can correctly identify the face image;

Figure 40 is the extracted watermark, NC=0.81, it is easy to detect the watermark.

Table 5 Face squeeze attack

Extrusion quantity (%) -100 -80 -60 -40 -20 0 20 40 60 80 100 PSNR(dB) 19.10 19.30 19.61 20.01 21.02 20.32 19.03 18.12 16.36 16.60 NC 0.68 0.72 0.74 0.81 0.94 1 0.96 0.90 0.77 0.62 0.59

(4) Spherical Distortion Attack

Table 6 is the experimental data of watermarking algorithm against face spherical distortion attack. It can be seen from the table that when the spherical distortion of the image to be tested is large and the amount of distortion is 60%, the signal-to-noise ratio of the image to be tested is as low as 17.61dB, but the original image can still be detected and extracted watermark, nc = 0.53;

Figure 41 is a spherical distortion attack, the image to be tested with a distortion amount of 40%; the deformation of the face is obvious, and the signal-to-noise ratio is 18.68dB;

Figure 42 is a recognized face image, which can detect the original face;

Figure 43 is the extracted watermark, NC=0.62, the watermark can be extracted correctly.

Table 6 Spherical Distortion Attacks

The above experiments mainly carried out experiments on attacks such as lighting and occlusion, and the following tests are performed on conventional attacks such as filtering;

(1) Noise attack experiment

Use the imnoise() function to add Gaussian noise to the image under test.

Table 7 is the experimental data of anti-Gaussian noise interference. It can be seen that when the Gaussian noise intensity of the image to be tested is as high as 30%, the PSNR of the image to be tested drops to 7.94dB. At this time, the quality of the image to be tested is poor; the original image can still be detected correctly, and the watermark can be extracted. NC = 0.60. However, using Google's image search function, when the Gaussian noise intensity of the image to be tested is only 5%, it is impossible to recognize the normal original image. This shows that the invention has a good anti-Gaussian noise capability.

Figure 44 is the image under test with 5% Gaussian noise intensity, which is visually blurred, PSNR=12.98dB;

Figure 45 is the original image of the recognized face, which can be correctly identified;

Figure 46 is the extracted watermark, the watermark can be extracted accurately, NC=0.81.

Table 7 Anti-Gaussian interference

Noise intensity (%) 1 2 3 5 10 20 30 PSNR 19.36 16.35 14.9 12.98 10.67 8.83 7.94 NC 1.0 0.90 0.90 0.81 0.87 0.62 0.60

(2) JPEG compression attack experiment

The image compression quality percentage is used as a parameter to perform JPEG compression on the image to be tested; Table 8 shows the experimental data of JPEG compression on the image to be tested. When the compression quality is only 2%, the compression quality is relatively low, PSNR=22.21dB, the watermark can still be extracted, NC=0.75.

Figure 47 is the image to be tested with a compression quality of 5%, and the block effect has appeared in this picture.

PSNR = 25.12dB;

Figure 48 is the recognized original image, which can be correctly recognized;

Figure 49 is the watermark extracted from the image to be tested, NC=0.90, the watermark can be extracted accurately.

Table 8 JPEG attack experiment

compression quality 2 5 10 20 30 40 PSNR 22.21 25.12 27.7 29.86 31.15 32.08 NC 0.75 0.90 1.00 0.87 1.0 1.0

(3) Median filter attack experiment

Table 9 shows the anti-median filter attack experiment. It can be seen from the table that when the median filter parameter is [7x7] and the number of filter repetitions is 10, the signal-to-noise ratio of the image is low PSNR=20.92dB, but the watermark can still be measured The presence of , NC = 0.87.

Figure 50 is the image to be tested with the median filtering parameter [3x3] and the number of filtering repetitions being 10. The image has been blurred, PSNR=29.18dB;

Figure 51 is the detected original image in the case of the above filtering, and the original image can be correctly detected at this time;

Figure 52 is the watermark extracted from the image to be tested, NC=0.96, the watermark can be extracted accurately.

Table 9 Median filtering attack experiment

Claims (1)

1. Face recognition method based on transformation domain robust watermarking anti-light attack face recognition method under big data, which is characterized in that: firstly perform global DCT transformation on the original face, select the first 8x8 coefficients in the transformation domain as feature vectors, and then convert the watermark information Associated with the feature vector to realize the embedding of the watermark; then, for the face to be tested, its feature vector is obtained, and then the correlation coefficient between the face to be tested and all face feature vectors is calculated, and the maximum value of the correlation coefficient is obtained, according to the The value is the recognized face image and the embedded watermark; then according to the feature vector of the image to be tested, the watermark is extracted, and the watermark correlation coefficient is obtained; the embedding and extraction of the digital watermark and the recognition of the face are realized; the invention includes watermark embedding and Extract two major parts, a total of six steps: Part 1: Embedding of Watermark 1) By performing global DCT transformation on all original face images F(n), the feature vector set V(n) of the original image is obtained; First, the global DCT transformation is performed on the original face image F(n), and the first 8×8 coefficients FD ₈ (i,j) are selected in the low-frequency coefficient matrix FD(i,j) in the frequency domain, and then the selected The coefficient matrix FD ₈ (i, j) is binarized. When the coefficient is greater than or equal to zero, it takes 1, and if it is less than 0, it takes zero to obtain the feature vector V. The main process is described as follows: FD ₈ (i,j)=DCT2(F(i,j)) V(n)=BINARY(FD ₈ (i,j)) 2) Using the cryptography HASH function to generate a binary key sequence Key(n) containing watermark information to complete the embedding of the zero watermark; Key(n)＝V(n)⊕W(n) Key(n) is generated by the eigenvector V(n) of all original images and the corresponding n digital watermarks W(n) through the Hash function commonly used in cryptography; here W(n) consists of a random sequence with a length of 64 bits ;Save Key(n), which will be used when extracting the watermark below; Apply to a third party by using Key(n) as a key to obtain the right to use and ownership of the face image; Part II: Face recognition and watermark extraction 3) Find the feature vector V' of the human face to be tested; Assuming the face to be tested is F', after the global two-dimensional DCT transformation, the low-frequency coefficient matrix is obtained as FD'(i,j), and the low-frequency coefficient matrix is binarized according to step 1) to obtain the face to be tested eigenvector V'; FD ₈ '(i,j)=DCT2(F'(i,j)) V'=BINARY(FID'(i,j)) 4) Calculate the correlation coefficient NC(n) between the eigenvector V' of the human face to be tested and the V(n) of the original human face, and carry out recognition of the human face; Calculate the n value corresponding to the maximum value of the correlation coefficient between V' and V(n), and set n=k; according to the value of k, the key Key(k) can be obtained, and the original face image can be identified as F(k) and embedded in For the watermark value W(k) of F(k), the formula for calculating the normalized correlation coefficient of the feature vector is as follows: $\mathrm{<span\; class="notranslate">\; NC</span>}<span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>{\mathrm{<span\; class="notranslate">\; V</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>}$ 5) Utilize the binary logic key sequence Key(k) existing in the third party and the feature vector V' of the face to be tested to extract the watermark W' in the image to be tested; W'=Key(k)⊕V' 6) Calculate the correlation coefficient between W' and W(k); calculate the correlation coefficient between watermark W and W'(K), and obtain the content of the image to be tested.