Digital watermarking technology plays a powerful role in the effective protection of digital media copyright, image authentication, image sharing, image information transmission and other fields. Driven by strong demand, digital image watermarking technology has aroused widespread research interest and has gradually developed into one of the most active research directions in information science. In this paper, we present a novel robust digital watermarking algorithm based on discrete radon transform tight frame in finite-set (FDRT). FDRT of the zero mean image is a tight frame, the frame boundary A = B = 1, the dual of the frame is itself. The decomposition and reconstruction of the FDRT tight frame will not cause the phenomenon of image distortion. The embedding of hidden watermark is to add a weak signal to the strong background of the original image. Watermark extraction is to effectively identify the embedded weak signal. The feasibility of the watermarking algorithm is analyzed from two aspects of information hiding and robustness. We select the independent Gaussian random vector as the watermark series, and the peak signal-to-noise ratio (PSNR) as the visual degradation criterion of the watermark image. Basing the FDRT compact stand dual operator, we derived the relationship among the strength parameter, square sum of watermark series, the PSNR. Using Checkmark system, the simulation results show that the algorithm is robust enough to some very important image processing attacks such as lossy compression, MAP, filtering, segmentation, edge enhancement, jitter, quadratic modulation and general geometric attack (scaling, rotation, shearing), etc.
In recent years, the research activities in digital watermarking have increased rapidly, mainly because the Internet and digital technology have brought us rich online digital images, which need effective copyright protection, image authentication, image sharing, and image information transmission [
The embedding of image implicit watermark is to superimpose a weak signal on the strong background of the original image. Watermark extraction is to effectively identify embedded weak signals [
The embedded watermark for protection should not degrade image quality. In addition, it must be robust to distortions caused by image processing algorithms. Image processing does not modify only the image but also the embedded watermark signal as well. Thus, the watermark may become undetectable after intentional or unintentional image processing attacks. A general watermarking framework for copyright protection has been described in [
Most watermarking techniques in the literature involve watermark signal detection based on similarity. However, the similarity detector is optimal only if the watermark series follows a Gaussian distribution and minimizes the error probability [
The direct discrete Radon transform in Euclidean geometry will introduce information redundancy when describing digital images [
As defining continuous Radon transformations in Euclidean geometry, finite Radon transformations are defined in finite geometry set. The FDRT of a real function f defined on a finite set S is the real function on subset G (G Ì S), the values of which are obtained by summing f over the subset G. When set S is considered a special algebraic structure like image, subset G is the lattice points on a line, the FDRT of the image can be attained [
The FDRT of an image f [ i , j ] ( i , j ∈ Z p , Z p = { 1 , 2 , … , p − 1 } , p is a prime number) is defined as:
ℜ k [ l ] = FDRT( f [ k , l ] ) = 1 p ∑ ( i , j ) ∈ L k , l f [ i , j ] (1)
where, the factor 1 / p is introduced to normalize the l2-norm between the input and output of the FDRT, and ensure the energy matching between the original image and the reconstructed image. Lk,l denotes the point set that makes up a line on the lattice Z p 2 .
L k , l = {( i , j ): j = k i + l ⋅ mod ( p ) , i ∈ Z p } , 0 ≤ k ≤ p − 1 L p , l = {( l , j ): j ∈ Z p } k = p , l ∈ Z p (2)
where, due to existing of the module operation, the line Lk,l exhibits a wrap-around effect. In other words, the FDRT treats the input image as a periodic image.
The line Lk,l on the finite set Z p 2 is uniquely represented by its slope or direction k ∈ Z p * ( Z p 2 = { 0 , 1 , 2 , … , p − 1 , p } ) and its intercept l ∈ Z p , where k = p corresponds to infinite slope or close to vertical lines. There are total lines p × ( p + 1) defined in this way and every line contains p points. Two lines of different slopes intersect at exactly one point, and any two distinct points on Z p 2 belong to just one line. For any given slope, there are p parallel lines that provide a complete cover on the lattice Z p 2 . Traversing all image Z p 2 , we have:
∑ l ∈ Z p , k ∈ Z p * ℜ k [ l ] = 1 p ∑ l ∈ Z p , k ∈ Z p * ∑ ( i , j ) ∈ Z p 2 f [ i , j ] = E p (3)
where the constant E is the total energy of the original image f.
An inverse converter f ˜ is the sum of the finite Radon coefficients of all lines passing through a given point, that is,
f ˜ [ i , j ] = IFDRT( ℜ k [ l ]) = 1 p ∑ ( k , l ) ∈ P i , j ℜ k [ l ] , ( i , j ) ∈ Z p 2 (4)
where Pi,j denotes the set of indices of all the lines that go through a point, as
P i , j = {( k , l ) : l = j − k i ⋅ mod( p ) , k ∈ Z p } ∪ {( p , i )} (5)
so as
f ˜ [ i , j ] = IFDRT( ℜ k [ l ]) = 1 p ⋅ ∑ ( k , l ) ∈ P i , j ∑ ( i , j ) ∈ L k , l f [ i , j ] = 1 p ( p ⋅ f [ i , j ] + ∑ l ∈ Z p , k ∈ Z p * ∑ ( i , j ) ∈ L k , l f [ i , j ] ) = f [ i , j ] + E / p (6)
The constant E is the total energy of image f [ i , j ] . Equation (6) shows that the result of reconstructing f ˜ [ i , j ] can be completely reconstructed except for adding a constant E / p to each pixel.
If the total energy E = 0, let the matrix ψ [ i , j ] studied is a zero-mean image, then Equation (6) shows that there is no redundancy or loss of information/energy in the FDRT and IFDRT of ψ [ i , j ] .
Let the discrete Radon transform of ψ [ i , j ] be expressed as FDRT(ψ) = Hψ. The elements of matrix H(p+1)×p is
h k , l = p − 1 / 2 δ L k , l ( i , j ) ,
where δ L k , l ( i , j ) = { 1 , ( i , j ) ∈ L k , l 0 , o t h e r w i s e . If ψ [ i , j ] the zero-mean image as a set of column vectors, then the matrix H reduced to H = p − 1 / 2 R , where R is the correlation matrix between the set H and its subset ( R L k , l ) ( i , j ) = { 1 , ( i , j ) ∈ L k , l 0 , o t h e r w i s e .
The FDRT of the zero-mean image ψ [ i , j ] satisfies the frame condition, and the frame { h k , l : k ∈ Z p * , l ∈ Z p } is a tight frame, the frame boundary A = B = 1. i.e. the dual frame { h ^ k , l : k ∈ Z p * , l ∈ Z p } is itself [
ψ ˜ = ∑ l ∈ Z p ∑ k ∈ Z p * 〈 ψ ˜ , h k , l 〉 h k , l , ∀ f ⊂ Z p * (7)
Therefore, the decomposition and reconstruction of the FDRT tight frame for zero mean images ψ [ i , j ] is a reversible operation. That is, the FDRT and IFDRT of the frame have the same algorithm structure. The decomposition and reconstruction of the FDRT tight frame will not cause the phenomenon of image distortion.
Watermark embedding can be seen as a weak signal superimposed on the strong background of the original image. According to the masking characteristics of the vision system, only when the superimposed signal exceeds certain intensity can it be detected by the vision system. In the image domain, the strength of the watermark can’t cause changes in the objective quality of the original image (such as PSNR) and visual effects. The watermark can be better hidden in the original image, making it lossless information [
In the FDRT domain, the energy of image is best compacted on the projections with different slops. That is to say, the projection matrix Q is composed of p + 1 projection vectors. The watermark which is unchangeable and decipherable consists of a real series W = { ω i , i = 1 , … , N } , where each value ωi is chosen independently according to N ( 0 , 1 ) [
We select l = 0 on the k'-th ( k ′ < p − int ( N / p ) ) projection embedding the watermark W. By Zigzag scanning, we have the watermarked matrix Wmk is
W ℜ [ k , l ] = ℜ k [ l ] + α ⋅ W m k [ l ] k ∈ Z p 2 (8)
where, α is strength factor of the watermark, so
W m k [ l ] = { W { k ′ ≤ k ≤ int( N / p ) , l ∈ Z p } ∪ { k = int( N / p ) + 1 ， 0 < l ≤ ( N − p ⋅ int( N / p ))} 0 o t h e r w i s e (9)
Considering the equality (6), the watermarked image f ′ [ i , j ] is:
f ′ [ i , j ] = IFDRT( W ℜ [ k , l ]) = f [ i , j ] + α p ∑ ( k , l ) ∈ P i , j W m k [ l ] = f [ i , j ] + α ⋅ W ′ [ i , j ] (10)
The first item is reconstructed image f [ i , j ] , and the second item is the inverse projection of the embedded watermark in FDRT domain. The pixels W ′ [ i , j ] which satisfies the wrap around stack path Lk,l in Equation (4) are defined over the entire image domain, viz, the Equation (10) shows that the energy of the embedded watermark uniformity distributes on all pixels of the image.
If the strength factor α is chosen suitably, the watermark signal W ′ [ i , j ] in space domain is very weak, so that the degradation of the watermark image is very small. The embedded watermark signal is invisibility, i.e. the visual quality of the watermarked image is not debased.
In the image domain, the watermark embedding process shown in Equation (10) actually introduces noise pollution. The PSNR of the watermarked image is
PSNR = 10 ⋅ lg 255 ⋅ p ∑ ( i , j ) ∈ Z p ( α ⋅ W ′ ( i , j )) 2 (11)
Due to the quadratic sum of watermark signal W ′ [ i , j ] in image domain equal to the quadratic sum of the watermark series { α ω i , i = 1 , 2 , … , N } embedded in FDRT domain, viz. ∑ ( i , j ) ∈ Z p ( α ⋅ W ′ ( i , j )) 2 = ∑ i = 1 N ( α ω i ) 2 . The watermark series elements ωi are independent random variables, let E ω = ∑ i = 1 N ( ω i ) 2 . We have
PSNR = 10 ⋅ lg ( p ) − 10 ⋅ lg α − 5 ⋅ lg ( E ω ) (12)
The PSNR is only related to the strength parameter α, square sum Eω of watermark series. For a given watermark series W = { ω i , i = 1 , … , N } , the smaller the strength parameter α of the watermark, the lower the visibility of the embedded watermark, the smaller the quality of the watermark image is better; otherwise, the quality of the watermark image is more reduced. This conforms to the masking characteristics of the visual system.
The watermark extraction is essentially the reverse of the insertion process. The FDRT of the original and watermarked image is first performed. The watermark
matrix C* is then obtained by subtracting the original image projections from the watermarked image projections. We have
C * = FRAT f ′ [ k , l ] − FRAT f [ k , l ] = { W ′ m k [ l ] { k ′ ≤ k ≤ int( N / p ) , l ∈ Z p } ∪ { k = int( N / p ) + 1,0 < l ≤ ( N − p ⋅ int( N / p ))} 0 k ≠ k ′ , l ∈ Z p (13)
After extracting the sub-matrix W*mk from the matrix C*, the one-dimensional series W* is restored by Zigzag scanning.
The presence of the watermark is then evaluated based on the similarity between the extracted W* and original watermark W. The similarity detector is given by
K ( W * , W ) = ∑ i = 1 N ( W * ( i ) ⋅ W ( i )) / ∑ i = 1 N ( W * ( i )) 2 (14)
A threshold T is selected previously. If K ( W * , W ) > T one can determine that the image f ′ [ i , j ] has been inserted watermark W. The similarity measure depends on N, the length of the watermark series. Larger values of N tend to cause larger similarity values when W* and W are genuinely related, without causing larger similarity values when W* and W are independent. For the smaller number N, setting the threshold at bigger value will not cause spurious matching. We should select the T basing on the length of the original watermark series. For example, N = 1024, then T = 6.
Using the FDRT, the energy or information of the watermarked signal which is uniformly incumbent on the image is congregated renewed on the correspondence projections in FDRT domain. Even if the watermarked image is exposed to attack, this energy or information can be yet collected on the exactitude projections, and can be detected by the similarity detector. The proposed watermarking algorithm has stronger robustness.
Since the watermark is Gaussian random series with zero-mean and is a full frequency domain signal, the FDRT and IFDRT do not change the frequency features of the watermark. The watermark signal is covered in the whole image space in full frequency band, this watermark embedding and extraction detection algorithm can resist the attack of various types of convolution filtering (low pass filter, band pass filter, high pass filter, notch filter, etc. in
| No | attacks | the similarity detector | |||
|---|---|---|---|---|---|
| Lena | Baboon | Head | |||
| 1 | wavelet lossy compression | 30.121 | 30.815 | 31.984 | |
| 2 | DCT compression (8 × 8) | 31.485 | 31.04 | 30.217 | |
| 3 | MAP 9 × 9 | wiener filter | 16.558 | 10.685 | 15.456 |
| soft threshold | 10.67 | 10.095 | 12.544 | ||
| hard threshold | 12.599 | 10.748 | 12.372 | ||
| 4 | border cut | pixel = 6 | 30.683 | 30.27 | 31.592 |
| pixel = 18 | 26.916 | 22.991 | 28.555 | ||
| 5 | scale change [a] | (1/2, 2) | 14.864 | 9.884 | 14.854 |
| (1/3, 3) | 6.98 | 6.388 | 7.358 | ||
| 6 | median filter (7 × 7) | 6.183 | 6.711 | 6.553 | |
| 7 | binary segmentation [b] | 19.01 | 16.846 | 15.074 | |
| 8 | edge enhancement | Laplacian | 27.768 | 26.157 | 26.823 |
| log | 28.4 | 27.747 | 28.969 | ||
| unsharp | 30.996 | 28.691 | 28.975 | ||
| average | 10.189 | 7.368 | 6.455 | ||
| Gaussian | 29.512 | 29.747 | 30.422 | ||
| 9 | contrast change | log(f + 2) | 11.7 | 14.834 | 19.002 |
| log(f − 2) | 7.837 | 9.203 | 13.607 | ||
| 10 | jitter | 12.62 | 12.161 | 11.041 | |
| 11 | whitening (25%) | 15.733 | 15.3611 | 14.597 | |
| 12 | rotation 0.65˚ | 6.908 | 6.246 | 8.514 | |
| 13 | partial extraction (25%) | 8.589 | 9.333 | 11.577 | |
| 14 | template removal | 30.858 | 30.769 | 31.315 | |
| 15 | high-pass filter [c] | 12.886 | 12.987 | 12.778 | |
| 16 | notch filter [d] | 23.053 | 20.128 | 23.98 | |
| 17 | secondary modulation | DPR | 14.273 | 13.938 | 17.882 |
| DPRCORR | 12.096 | 11.723 | 17.698 | ||
Note: [a] The parameters (a, b), reducing the size to the original 1/a according to the nearest neighbor method, and then zooming in b times; [b] “Lana”, “Baboon”, the maximum variance method calculates the gray threshold value. “Head”, The image gray threshold is the average value of the image; [c] The lower bound frequency is 3/4 of the highest frequency; [d] The notch frequency band is the highest frequency 1/4 - 2/4.
For various types of shear, edge enhancement, contrast adjust, segmentation, geometrical attacks, FDRT can aggregate the remaining watermark information in the original embedded position of the watermark in the Radon domain, the existence of the watermark can still be detected. That is, the correlation between the watermark to be tested and the original watermark is still high (in
In order to estimate the watermarking scheme, we subjected the watermarked image to a series of image processing and collusion style attacks. In these experiments, the length of the Gaussian white noise series is N = 1024. The strength parameter is a = 0.1. The origin point is chosen at the first element of the projection with slope k = 30.
Experiment 1: The original image is 256 × 256 × 8 bit “Lena” images with High and low frequency information rich, as shown in
Experiment 2: Noise pollution and compression. The original images selected are 256 × 256 × 8 bit “Lena”, “Baboon”, “Brain” images. The “Baboon” has the complex texture areas. The medical image contains large dark or bright areas; the “Brain” is the human brain MR image.
Because of the diversity of noise sources, digital images are easily polluted. The watermarked image was polluted by Gaussian white noise.
shows the relationship between the correlation detection and the degree of pollution (PSNR) of the watermark image. As the PSNR is reduced to 26.61 dB, the visual quality of watermark image has become very poor. But the correlation degree of watermark detection is still more than the threshold T = 6.
Compression of is very important for image transmission and storage, the robustness of anti-compression is one of the important indexes of watermarking algorithm.
Experiment 3, the robustness of watermarking scheme must be tested and evaluated by various attack. This paper uses attack methods and strategies proposed in [
A novel watermark algorithm based on the finite Radon transform tight frame has been presented. We derived the relationship among the PSNR, the strength parameter, quadratic sum of watermark series. According to the assigned image quality, the strength parameter α and Square sum of watermark series can be estimated. Experimental results reveal that the proposed watermarking algorithm yields watermarked images with superior imperceptibility and robustness to common attacks, such as removal/interference (filtering, MAP, JPEG, dithering, segmentation, noise pollution) and geometrical attacks (translation, cropping, rotation).
The authors declare no conflicts of interest regarding the publication of this paper.
Zhang, J.G. and Qi, H.B. (2020) A Robust Digital Watermarking Algorithm Based on Finite-Set Discrete Radon Transform Tight Frame. Journal of Computer and Communications, 8, 123-133. https://doi.org/10.4236/jcc.2020.812012