CN102129549A - Image Hash method based on thumbnail and singular value decomposition - Google Patents

Abstract

The invention relates to an image Hash method based on thumbnails and singular value decomposition, which first preprocesses the image, and then divides it into blocks; applies the singular value decomposition to the image block and takes the first and second singular values as features; uses double The linear interpolation method generates the thumbnail of the original image and performs singular value decomposition, taking the first and second singular values as features; normalizes the features of the image block and thumbnail, and rearranges the position of the feature sequence of the image block ;Based on the first and second singular values of the thumbnail, calculate the Euclidean distance between each pair of singular values and it, and concatenate all the distance values to be the image Hash. When judging the similarity, calculate the L1 norm of the two image Hash , if it is less than the set threshold, the corresponding images are considered similar, otherwise they are considered different images. The invention is robust to common digital processing such as JPEG compression, moderate noise interference, brightness adjustment and contrast enhancement, and has good uniqueness.

Description

Image Hash Method Based on Thumbnail and Singular Value Decomposition

technical field

The invention relates to the field of signal processing and computer technology, in particular to an image Hash method based on thumbnails and singular value decomposition.

Background technique

With the popularity of image acquisition devices such as digital cameras, digital images are increasing exponentially, how to effectively manage and retrieve digital images has become a problem that people are facing. At the same time, the characteristics of digital images that are easy to copy and modify make the problems of infringement, tampering and forgery of image content more and more serious, and how to better protect image copyright has become more and more important. Image Hash (hashing) is an emerging technology of multimedia information security. It uses short sequence extracted from the image to identify the image, and can be widely used in image authentication, copy detection, digital watermarking, image retrieval and other fields.

Because cryptographic Hash functions such as SHA-1 and MD5 are very sensitive to changes in input data, any 1-bit change will completely change the output Hash value, so they are not suitable for images. In practical applications, it is often necessary to perform normal digital processing on the image, such as enhancement, JPEG compression, etc., and its content has not changed substantially. It is hoped that the image Hash will remain unchanged. Usually, the image hash should meet two conditions: 1) perceptual robustness, that is, for two images with similar perception, regardless of whether their internal data are consistent, their hashes have a high probability of being the same or very close; 2) uniqueness , that is, different images have different image Hash.

According to different construction techniques, the existing image hashing methods can be roughly divided into the following five categories: (1) methods based on image statistics; (2) methods based on invariant relations; (3) methods based on rough representation of images; (4) Utilize the low-level semantic features of the image; (5) Use matrix decomposition technology. Please refer to the following documents for details:

1.R.Venkatesan, S.-M.Koon, M.H.Jakubowski, et al., Robust image hashing[C], in Proc.of the IEEE International Conference on Image Processing, 2000, 3:664-666, Vancouver, BC, Canada, September 10-13, 2000.

2. C.Y.Lin and S.F.Chang, A robust image authentication system distinguishing JPEGcompression from malicious manipulation[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2001, 11(2): 153-168.

3. A. Swaminathan, Y. Mao and M. Wu, Robust and secure image hashing [J]. IEEE Transactions on Information Forensics and Security, 2006, 1(2): 215-230.

4. V.Monga and B.L.Evans, Perceptual image hashing via feature points: performance evaluation and trade-offs [J]. IEEE Transactions on Image Processing, 2006, 15(11): 3453-3466.

5.V.Monga and M.K.Mihcak, Robust and secure image hashing via non-negative matrixfactorizations[J].IEEE Transactions on Information Forensics and Security, 2007, 2(3): 376-390.

Most of the above-mentioned existing technologies are robust to certain digital processing, such as JPEG compression, digital filtering, and geometric transformation, but generally have the disadvantage of poor uniqueness.

Contents of the invention

The object of the present invention is to provide an image Hash method based on thumbnail and singular value decomposition, which can extract robust Hash from digital images, is robust to common image processing, and has good uniqueness.

An image Hash method based on thumbnails and singular value decomposition. First, the input image is preprocessed, and then non-overlapping blocks are performed; then the singular value decomposition is applied to each image block, and the first and second singular values are taken as Block features, use bilinear interpolation to convert the input image into an image with the same size as the image block and take the brightness component to represent it, generate a thumbnail of the original image, and apply the singular value decomposition to the thumbnail, taking the first and second Singular values as features, normalize the eigenvalues of the image blocks and thumbnails, and rearrange the position of the feature sequence of the image blocks, based on the first and second singular values of the thumbnails, calculate the remaining pairs of singular values Value and its Euclidean distance, concatenating all the distance values is the image Hash, when judging the similarity of two image Hash, calculate their L1 norm, if it is less than the set threshold, the corresponding image is considered to be the same, otherwise it is considered to be different images.

The concrete steps of this method are as follows:

(1) Image preprocessing: Preprocessing the input image I, including image size normalization and color space conversion. First use the bilinear interpolation method to normalize the image into M×M size. If the input is a color image, convert it to the YCbCr space, and take the Y component to represent the image. Denote the preprocessed image as J;

(2) Image block: Divide J into non-overlapping blocks with a size of t×t, and there are N=M ² /t ² image blocks in total (when the image is normalized, take M as an integer multiple of t), press Number the blocks in order from top to bottom and from left to right, and record the i-th image block as B _i (1≤i≤N);

(3) Calculate image block features: Let the singular value decomposition of Bi be recorded as [ _{U i S i} V _i ]=SVD(B _i ), where U _i and V _i are unitary matrices, S _i is a diagonal matrix, S The elements on the diagonal of _i are the singular values of _Bi , and the first and second singular values of _Bi are taken as the image block features, which are recorded as p _i and q _i respectively, that is, p _i =S _i (1,1) and q _i =S _i (2,2). Connecting the first and second features of the image block respectively, the feature vector p=[p ₁ ,p ₂ ,...,p _N ] and q=[q ₁ ,q ₂ ,.. ., _qN ];

(4) Generate thumbnails: Use bilinear interpolation to normalize the image I into thumbnails with a size of t×t. For color images, further convert them to YCbCr space, take the Y component to represent the image, and record the final thumbnail thumbnail for R;

(5) extract thumbnail feature: carry out singular value decomposition to R, get its 1st, 2nd singular value as feature, denote as s ₁ and s ₂ respectively;

(6) Feature normalization: Use vectors u and v to represent the first and second singular values of image blocks and thumbnails respectively, namely u _i =p _i (1≤i≤N), u _N+1 =s ₁ ;v _i =q _i (1≤i≤N), v _N+1 =s ₂ , normalize the elements of u and v respectively, that is, x _i =(u _i -μ _u )/δ _u , y _i =(v _i -μ _v )/δ _v , wherein μ _u and μ _v are the mean values of u and v respectively, and δ _u and δ _v are their standard deviations respectively;

(7) Feature scrambling: Under the control of the key, use a pseudo-random generator to rearrange the positions of the first N elements of the normalization vectors x and y, and keep the N+1th element unchanged, and get the set Random back vectors x' and y';

(1≤i≤N)，即可得到图像Hash h＝[h₁，h₂，...，h_N]；(8) Hash extraction, calculation

(1≤i≤N), the image Hash h=[h ₁ , h ₂ ,..., h _N ] can be obtained;

(9) Similarity judgment, let h ⁽¹⁾ and h ⁽²⁾ be two image Hash sequences respectively, h _i ⁽¹⁾ and h _i ⁽²⁾ represent their i-th elements respectively, and calculate the L1 norm If d is smaller than the set threshold T, the images corresponding to h ⁽¹⁾ and h ⁽²⁾ are considered to be the same, otherwise they are considered to be different images.

Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant progress: the present invention uses singular value decomposition to extract image information, takes the two most important singular values of image blocks as features; introduces image thumbnails as reference , using the distance between the singular value of the image block and the singular value of the thumbnail as the Hash value; the extracted image Hash is robust to common digital processing such as JPEG compression, moderate noise interference, brightness adjustment, and contrast enhancement, and has good uniqueness.

Description of drawings

Fig. 1 is a test image used in the embodiment of the present invention, size is 600 * 400;

Fig. 2 is another image used in the embodiment of the present invention;

Figure 3 is the result of normalizing the size of Figure 1;

Fig. 4 is the result of taking the Y component (brightness component) of Fig. 3, that is, the final result of preprocessing Fig. 1;

Fig. 5 is a block diagram of Fig. 4;

Figure 6 is a thumbnail of Figure 1;

Fig. 7 is a statistical distribution diagram of the L1 norm in the embodiment.

Detailed ways

A preferred embodiment of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following embodiment.

Example:

This embodiment includes two parts: robustness verification and uniqueness verification. Robustness verification is realized by judging whether the Hash images corresponding to Figure 1 and Figure 2 are similar, where Figure 2 is obtained through continuous digital processing of Figure 1, including JPEG compression (quality factor is 60), brightness adjustment (adjustment amplitude of 20), contrast enhancement (adjustment amplitude of 20), and white Gaussian noise (mean of 0, variance of 0.01). In the following steps, (1) to (8) are the steps of extracting the Hash of Figure 1, and the steps of extracting the Hash of Figure 2 are the same as those of Figure 1, and will not be repeated. (9) is the similarity of the two image Hash Sex judgment, the specific steps are as follows:

(1) Image preprocessing: use bilinear interpolation to normalize Figure 1 to a size of 256×256, and the result is shown in Figure 3; convert Figure 3 to YCbCr space representation, take the Y component to represent the image, and obtain the preprocessed image The result, as shown in Figure 4;

(2) Image segmentation: Figure 4 is divided into non-overlapping blocks with a size of 64 × 64, and a total of 16 image blocks are obtained, and Figure 5 is a block diagram;

(3) Calculate image block features: Singular value decomposition is performed on the above 16 image blocks in turn, and the first and second singular values of each image block are taken as features, so the feature vectors p and q of the image can be obtained:

p＝[13618.85，12438.3281，6800.47315，5791.30954，13182.7201，12360.2039，7554.88336，7234.64，12772.461，11346.9951，6586.04545，7960.01579，8250.24471，5453.8304，5120.26638，5752.40614]，

q＝[130.76731，623.10121，484.34539，381.7512，112.65023，516.03795，665.47482，606.93448，193.00764，971.94397，844.79315，288.13595，1315.9135，163.94719，247.47046，221.32931]

(4) Generate a thumbnail: Utilize the bilinear interpolation method to normalize Fig. 1 into a size of 64 × 64, convert to YCbCr space at the same time, represent it with the Y component, and obtain the thumbnail as shown in Fig. 6;

(5) Extract the features of the thumbnail: perform singular value decomposition on Figure 6 to obtain the first and second singular values, namely s ₁ =9323.15671, s ₂ =130.76731;

(6) Feature normalization: use vectors u and v to represent the first and second singular values of image blocks and thumbnails, namely

u＝[13618.85，12438.3281，6800.47315，5791.30954，13182.7201，12360.2039，7554.88336，7234.64，12772.461，11346.9951，6586.04545，7960.01579，8250.24471，5453.8304，5120.26638，5752.40614，9323.15671]，

v＝[130.76731，623.10121，484.34539，381.7512，112.65023，516.03795，665.47482，606.93448，193.00764，971.94397，844.79315，288.13595，1315.9135，163.94719，247.47046，221.32931，130.76731]

Normalize u and v to get vectors x and y:

x=[1.5486, 1.16, -0.69593, -1.0281, 1.4051, 1.1343, -0.44758, -0.553, 1.27, 0.80075, -0.76651, -0.31421, -0.21867, -1.1392, -1.249, -1.0409, 0.0]

y=[-0.97634, 0.46352, 0.057717, -0.24232, -1.0293, 0.1504, 0.58744, 0.41624, -0.79432, 1.4837, 1.1119, -0.51611, 2.4897, -0.8793, -0.635704] 1, 9

(7) Feature scrambling: rearrange the positions of the first 16 elements of the normalized vectors x and y, and keep the 17th element unchanged to obtain the scrambled vectors x' and y':

x'=[-0.69593, 1.16, -0.31421, -1.249, -1.0409, -0.76651, 1.5486, 0.80075, 1.4051, -0.21867, -1.0281, -0.44758, -1.1392, -0.553, 1.13403, 1.347, 2]

y'=[0.71149, -1.0293, 1.1119, -0.24232, -0.51611, 0.58744, -0.79432, 2.4897, 1.4837, -0.63504, -0.97634, 0.41624, 0.1504, -0.8793, 0.057406]3, 5

(8) Hash extraction, calculation (1≤i≤16), the Hash of the image in Figure 1 can be obtained: h ⁽¹⁾ = [0.87166, 1.0269, 2.1359, 1.5662, 1.2623, 1.8048, 1.4258, 3.5295, 2.7688, 0.49116, 1.1627, 1.5093, 1.7006, 0.69434 , 1.4383, 1.8337].

(9) Similarity judgment, extract the image Hash in Figure 2 to get h ⁽²⁾ = [0.9612, 1.1092, 2.3191, 1.7667, 1.4140, 2.0047, 1.1778, 3.5129, 2.8024, 0.7206, 1.1775, 1.7879, 1.8365, 0.7194, 1.6644, 1.9184]; set the threshold T=1.0, calculate the L1 norm of h ⁽¹⁾ and h ⁽²⁾ , and obtain d=0.64987; since d is smaller than T, it can be considered that Fig. 1 and Fig. 2 are similar images.

The present invention uses 100 different images (with a size of 256×256～1994×2592) as test data, extracts their image Hash (taking M=256, t=64) and calculates the L1 norm between pairs of Hash, there are 4950 The results are shown in Figure 7. Among them, the minimum distance is 1.6458, the maximum distance is 7.5218, the average distance is 4.3755, and the standard deviation is 0.8587. It is found from the embodiment that the distances between the hashes of any two different images are greater than the set threshold T=1.0, which shows that the present invention has better uniqueness.

Claims (2)

1. A kind of image Hash method based on thumbnail and singular value decomposition, it is characterized in that: earlier to input image preprocessing, carry out non-overlapping block again, then apply singular value decomposition to each image block, get the first and The second singular value is used as the block feature, and the input image is converted into an image with the same size as the image block by bilinear interpolation method and represented by the brightness component to generate a thumbnail of the original image, and the singular value decomposition is applied to the thumbnail, Take the first and second singular values as features, normalize the feature values of the image block and thumbnail, and rearrange the position of the feature sequence of the image block, based on the first and second singular values of the thumbnail , calculate the Euclidean distance between each pair of singular values and it, concatenate all the distance values to be the image Hash, when judging the similarity of two image Hash, calculate their L1 norm, if it is less than the set threshold, consider the corresponding image similar, otherwise they are considered different images. 2. the image Hash method based on thumbnail and singular value decomposition according to claim 1, is characterized in that: concrete steps are as follows: (1) Image preprocessing: Preprocessing the input image I, including image size normalization and color space conversion. First use the bilinear interpolation method to normalize the image into M×M size; if the input is a color image, convert it to the YCbCr space, use the Y component to represent the image, and record the preprocessed image as J; (2) Image block: Divide J into non-overlapping blocks with a size of t×t. There are N=M ² /t ² image blocks in total. Number the blocks in order from top to bottom from left to right, and record the i-th image block as Bi (1≤i≤N); (3) Calculate image block features: Let the singular value decomposition of Bi be recorded as [ _{U i S i} V _i ]=SVD(B _i ), where U _i and V _i are unitary matrices, S _i is a diagonal matrix, S The elements on the diagonal of _i are the singular values of _Bi , and the first and second singular values of _Bi are taken as the image block features, which are recorded as p _i and q _i respectively, that is, p _i =S _i (1,1) and q _i =S _i (2, 2), connect the first and second features of the image block respectively, and obtain the feature vector p=[p ₁ , p ₂ ,..., p _N ] describing the entire image and q=[q ₁ , q ₂ , . . . , q _N ]; (4) Generate thumbnails: Use bilinear interpolation to normalize the image I into a thumbnail with a size of t×t. For a color image, convert it to the YCbCr space, take the Y component to represent the image, and record the final thumbnail for R; (5) extract thumbnail feature: carry out singular value decomposition to R, get its 1st, 2nd singular value as feature, denote as s ₁ and s ₂ respectively; (6) Feature normalization: use vectors u and v to represent the features of the image block and image thumbnail respectively, namely u _i =p _i (1≤i≤N), u _N+1 =s ₁ ; v _i =q _i (1≤i≤N), v _N+1 = s ₂ , normalize the elements of u and v respectively, that is, x _i = (u _i -μ _u )/δ _u , y _i = (v _i -μ _v )/δ _v , wherein, μ _u and μ _v are the average value of u and v respectively, and δ _u and δ _v are their standard deviations respectively; (7) Feature scrambling: Under the control of the key, use a pseudo-random generator to rearrange the positions of the first N elements of the normalization vectors x and y, and keep the N+1th element unchanged, and get the set Random back vectors x' and y'; (8) Hash extraction: calculation (1≤i≤N), the image Hash h=[h ₁ , h ₂ ,..., h _N ] can be obtained; (9) Similarity judgment: Let h ⁽¹⁾ and h ⁽²⁾ be two image Hash sequences respectively,

and

represent their i-th elements respectively, and calculate the distance

If d is smaller than the set threshold T, the images corresponding to h ⁽¹⁾ and h ⁽²⁾ are considered to be similar, otherwise they are considered to be different images.

Images