Professional Documents
Culture Documents
https://doi.org/10.1007/s11554-018-0822-8
Received: 19 June 2018 / Accepted: 13 September 2018 / Published online: 21 September 2018
© Springer-Verlag GmbH Germany, part of Springer Nature 2018
Abstract
Halftone images are usually used in facsimile and halftone image steganography can be used for facsimile channel. In recent
years, real-time image processing becomes more and more important. In this paper, an efficient block-based steganographic
method for halftone images is proposed. This method is based on optimal dispersion degree (DD), which can measure the
complexity of the region texture. To reduce the visual distortion, the blocks with complex texture can be selected as carriers
according to the dispersion degree. Finally, the secret messages are embedded by flipping the pixels that can minimize the
changes of texture structure. The experiments demonstrate that the proposed scheme maintains a good image visual quality
and realizes acceptable statistical security with high capacity.
13
Vol.:(0123456789)
602 Journal of Real-Time Image Processing (2019) 16:601–609
13
Journal of Real-Time Image Processing (2019) 16:601–609 603
Considering some examples in Fig. 2, from left to right, the 2.3 Pixel flipping strategy
dispersion degree of blocks increases. We can see that the
pixel distribution in blocks with high dispersion degree is After selecting the suitable blocks according to the disper-
more decentralized, which means that the texture is more sion degree, we can embed the message by forcing the parity
complex. On the other hand, the blocks with uniform texture of the number of black pixels in blocks. Parity embedding is
have lower dispersion degree. a commonly used method of data hiding [2, 5, 6]. When the
parity of the block pixel is inconsistent with the correspond-
ing message bit, the parity needs to be modified by flipping
a pixel within the block.
Flipping different pixels in the same block will lead to
different visual quality and statistical security of anti-stega-
nalysis. For the best performance of steganography, we need
to design an optimal pixel flipping strategy.
The scatter structure of the halftone image makes it more
difficult to embed message. Flipping one pixel at a random
position will bring disturbing noise, and it is easy to destroy
Fig. 2 Dispersion degree of sample patterns from “Baboom”. The
the structural correlation between the blocks and hence
patterns in the red box are the selected patterns in the five example
patterns. The numbers under the blocks are dispersion degrees (DD) reduce the security. To maintain the original texture and
which are calculated by Eq. (3) structure of the halftone image, we need to minimize the
13
604 Journal of Real-Time Image Processing (2019) 16:601–609
destruction of the association of the image blocks caused variation of cluster area after flipping changes is the small-
by the pixel flipping. est. The formula is as follows:
Image texture consists of many small and large clusters y = arg mini |G0 − Gi |,
which are formed by the aggregation of white or black pixels (5)
and even a pixel can also be considered a special cluster. where G0 and Gi represent the sum of weighted cluster areas
Each image block is a composition of different clusters in of the cover block and the stego block flipping i-th pixel,
a specific way. So we can control the change of the clusters respectively.
composition after flipping pixels to control the change of the
image texture structure indirectly.
The sum of weighted cluster areas is a measure of how
3 Embedding and extraction procedure
clusters form an image block. It can be represented as
∑
G= g(i) × g(i), In this section, the embedding and extraction procedure
(4) of the proposed steganographic scheme is presented. The
i∈I
proposed dispersion degree can filter out suitable blocks in
where I is the set of all clusters with area larger than 1 in
all blocks for message embedding. Then flipping the pixel
the image block and g(i) denotes the area of the i-th cluster
in the inconsistent block by minimizing the variation of
which is the number of directly connected pixels with the
clusters area. The embedding and extraction procedures are
same value (as shown in Fig. 3). The first g(i) is equivalent
described in detail in the following two subsections.
to the weight of the i-th cluster so that the larger the cluster
area is, the greater the weight will be set because larger clus-
ters have a greater influence on the image texture structure. 3.1 Embedding procedure
So we can calculate the change in the sum of weighted
cluster area after flipping each pixel, and finally choose the For a given halftone image X, we first divide X into m × n
pixel with the smallest change. It should be noted that to non-overlapped blocks. The length of message should be no
ensure that all the stego blocks can be selected in the process more than the total number of the cover image blocks. The
of extracting message, the dispersion degree of the image dispersion threshold is set automatically according to the
block after flipping pixel needs to be maintained above the actual embedding capacity. The larger the embedding capac-
threshold. ity is, the larger the threshold will be set, and vice versa.
To sum up, our pixel flipping strategy is that, within the The embedding procedure contains following steps:
dispersion threshold range, flipping the pixel where the
1. Divide halftone image X into non-overlapped blocks of
size m × n;
2. Calculate the dispersion degree for each block;
3. Set the dispersion threshold according to the embedded
capacity and the specific dispersion distribution of the
blocks;
4. Scan all blocks and select them whose dispersion degree
is above the threshold, then construct a specific block
sequence;
5. If the parity of the number of black pixels is inconsistent
with the corresponding message bit, flip the pixel which
minimizes the change of cluster area in block;
6. Repeat Step 5 until all message bits are embedded;
7. Successively replace each selected block in cover image
with the corresponding stego block and the stego image
is obtained.
3.2 Extraction procedure
13
Journal of Real-Time Image Processing (2019) 16:601–609 605
1. Divide stego image Y into non-overlapped blocks of size clusters of pixels. What is worth mentioning, people usually
m × n; tend to measure the amount and the size of the salt-and-
2. Calculate the dispersion degree for each block and select pepper clusters due to disturbing feature of larger clusters.
the blocks with dispersion degree above the same thresh- As in [13], the authors defined the following five scores:
old;
3. The message bit is ‘0’ if the parity of the number of ∑
4
4.1 Experiment condition ∑
4
S5 = iNi = S2 − S1 , (10)
The images we used in our experiments come from Boss- i=0
Based-1.01 [1]. The size of these halftone images is resized
into 256 × 256 pixels. Then, we converted these images to where Ni is the total number of the locations of the pixels
halftone images by error diffusion [12, 17]. Since some of which are modified by steganography and the flipped pixels
halftone images have lots of uniform blocks, we drop out having i neighbors with same pixel values in the 4-neighbor-
1000 halftone images with lots uniform blocks and we have hood. The N0 corresponds to the number of visually pleasing
9000 halftone images. isolated elements in the locations of the flipped pixels. S1
To evaluate the proposed scheme, some experiments have gives the total number of two types of element that black
been conducted, including visual quality comparison and the pixel in bright region (class 1) and white pixel in dark region
statistical security comparison. For the only two status in (class 4) in the locations of the flipped pixels. S2 gives the
halftone images, flipping pixels in unsuitable positions may total area covered by the clusters with class 1 and class 4 in
cause significant visual distortion. Unlike grayscale images A. The bigger S2 is, the more clusters in class 1 and class 4
steganographic methods, the previous halftone stegano- are in the halftone images. Previous works [13] demonstrate
graphic methods focused more on the improvement of the that pixels in class 1 and class 4 causes remarkable “salt
visual quality. “Pixel slave” strategy is a popular strategy and pepper” noise which degrades visual quality of half-
to improve visual quality after embedding messages. By tone images. Therefore, the smaller S1 and S2 are, the better
employing “pixel slave” strategy, the stego images gener- visual quality of halftone image is. S3 gives the average area
ated by DHSPT and GIM have less cluster and salt and pep- per cluster. S3 means that a pixel is flipped in the region of
per noises. In the experiments, we first show the “Baboom” opposite brightness will generate S3 clusters which are black
halftone image to compare the subject visual quality. in bright region or white in dark region and cause remark-
To further evaluate the visual quality, some objective able black or white region. S4 is the number of elements of
visual quality assessments are employed in this section. A associated with clusters of size 3 or more, which is useful
Even though objective distortion measures such as Mean because clusters of size 1 or 2 are not very visually disturb-
Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), ing. S5 is a perceptual measure with a linear penalty model.
and Structural Similarity Index Measure (SSIM) do not It gives a zero penalty score to isolated black or white pixels
match well with the evaluation for halftone images, [13] which look visually pleasing.
can be used to define some objective distortion measures In steganographic performance evaluation, the statisti-
for halftone images. Using the similar standard, we measure cal security is another important criteria. Herein, some
the visual quality of our algorithms again in this paper. The experiments are conducted to evaluate the security of
distortion resulting from the proposed algorithm is always our proposed steganographic scheme. RLCM-100D fea-
exhibited in the form of salt-and-pepper artifacts due to local tures calculate the high-order difference images using
13
606 Journal of Real-Time Image Processing (2019) 16:601–609
the run-length and co-occurrence matrices. The number 4.2 Results and comparisons
following each feature name is the dimension of corre-
sponded feature. We used these features and SVM with an As shown in Fig. 4, the visual quality of these stego images
optimized Gaussian kernel to construct the steganalyzers. is acceptable in vision. DHSPT (shown in Fig. 4b) and GIM
The decision error rate PE is used to measure the decision (shown in Fig. 4c) are embedded in all blocks, therefore the
performance, which is defined as: flipped pixels are uniformly distributed in the whole images.
Using “pixel slave” strategy, the region around flipped pixels
1
PE = (P + PFn ), (11) has no remarkable clusters. The proposed methods (shown
2 Fp
in Fig. 4d), the embedding positions are mainly in the blocks
where PFp and PFn stand for the probabilities of false positive with complex texture. The stego image has no remarkable
(detecting cover as stego) and false negative (detecting stego noise, which demonstrates that the proposed methods can
as cover), respectively. maintain the visual quality.
The experimental results are conducted with the help The results of the objective visual quality comparison
of MATLAB R2016a. Intel(R) Core(TM) i5-7500 CPU can better evaluate the visual quality of the performance
and speed 3.40GHZ and its capacity is 8GB of RAM. of the halftone steganographic schemes. From Table 1, the
five scores of the proposed schemes are smaller than that
of DHSPT and GIM. As Fu and Au proposed in [13], the
smaller of the five scores means that good visual quality
13
Journal of Real-Time Image Processing (2019) 16:601–609 607
Table 1 Average scores (S1–S5) of various schemes on the halftone RLCM-100D mainly captures the lower order statistical
images dataset with 1024 bits embedded features while PMMTM-320D captures more structural fea-
S1 S2 S3 S4 S5 tures, the proposed scheme has better performance under
RLCM-320D. It is noted that in low payload, the advantage
Proposed 402 1209 3.04 295 807
of filtering out carriers is not obvious, because the flipped
DHSPT 792 2446 3.12 477 1654
pixels by all schemes are few and the changes of texture
GIM 695 2289 3.27 402 1594
structures are not obvious.
Generally, minimizing the changes of texture struc-
tures can reduce changes of co-occurrence and Pixel mesh
0.5 Markov transition matrix, so that the performance under
proposed scheme
0.45 DHSPT RLCM-100D and PMMTM-320D can be improved. GIM
0.4
GIM and DHSPT employ “pixel slave” strategy to flip pixels in
pairs rather than individually in each block, which inevitably
0.35
increase destructions between texture. To retain the depend-
0.3 ence between texture structures, the proposed scheme flips
one pixel by minimizing the change of the cluster composi-
E
0.25
P
0 4.3 Complexity analysis
0 200 400 600 800 1000 1200
payload
(a) In this section, we analyze the time complexity of differ-
ent halftone steganographic schemes. For a given halftone
0.3
proposed scheme image X with NX pixels, critical message consisting of r
DHSPT bits needs to be embedded in it. In DHSPT [13], the master
GIM
pixels are set in advance by the random seed, so the execu-
tion time is mainly spent on selecting slave pixels. To select
0.2
the suitable slave pixel, the connection weights of neighbor
pixels need to be calculated in a 3 × 3 neighborhood, which
E
13
608 Journal of Real-Time Image Processing (2019) 16:601–609
is significantly short, because our scheme can be efficiently International Workshop on Digital Watermarking, pp. 514–528
realized through the operation of lower time complexity, (2013)
6. Feng, B., Lu, W., Sun, W.: Secure binary image steganography
mainly involving the operation of filtrating blocks and based on minimizing the distortion on the texture. IEEE Trans.
selecting flipped pixels, which only needs to calculate the Inf. Forensics Secur. 10(2), 243–255 (2014)
dispersion degree of each block and the variation of the clus- 7. Feng, B., Lu, W., Sun, W.: Binary image steganalysis based on
ter area after each candidate pixel is flipped, respectively. pixel mesh markov transition matrix. J. Vis. Commun. Image
Represent. 26(C), 284–295 (2015)
Therefore, the proposed steganography scheme can be effi- 8. Feng, B., Lu, W., Sun, W.: Novel steganographic method based
ciently applied in real-time image processing. on generalized k-distance n-dimensional pixel matching. Mul-
timed. Tools Appl. 74(21), 9623–9646 (2015)
9. Feng, B., Lu, W., Sun, W., Liang, Z., Liu, J.: Blind watermark-
ing based on adaptive lattice quantization index modulation. In:
5 Conclusions International Workshop on Digital Watermarking, pp. 239–249.
Springer International Publishing, New York (2015)
In this paper, we have proposed an efficient halftone image 10. Feng, B., Weng, J., Lu, W., Pei, B.: Multiple watermarking
using multilevel quantization index modulation. In: Interna-
steganographic scheme based on optimal dispersion degree. tional Workshop on Digital Watermarking, pp. 312–326 (2016)
By employing the proposed concept of dispersion degree, 11. Filler, T., Fridrich, J.: Design of adaptive steganographic
we can quantify the region texture complexity. Thus 1the schemes for digital images. Proc SPIE 7880(1), 181–197 (2011)
regions with large dispersion degree can be selected as car- 12. Floyd, R.W., Steinberg, L.: Adaptive algorithm for spatial grey-
scale. In: Proceedings of SID, pp. 75–77 (1976)
riers since embedding messages in complex texture regions 13. Fu, M.S., Au, O.C.: Halftone image data hiding with intensity
are visually imperceptible. The secrete messages are embed- selection and connection selection. Signal Process. Image Com-
ded by minimizing the changes of texture structure. Experi- mun. 16(10), 909–930 (2001)
ments show that the visual quality of the proposed scheme 14. Guo, J.M., Tsai, J.J.: Data-hiding in halftone images using
adaptive noise-balanced error diffusion. IEEE Multimed. 18(2),
is better than existing schemes and the statistical security is 48–59 (2010)
acceptable. For practicability, the proposed scheme can hide 15. Guo, M., Zhang, H.: High capacity data hiding for halftone
information with very low processing time. image authentication. In: International conference on digital
For halftone steganography, the visual quality and the forensics and watermarking, pp. 156–168 (2013)
16. Huang, F., Huang, J., Shi, Y.Q.: New channel selection rule
statistical security have a closer relationship with each other. for jpeg steganography. IEEE Trans. Inf. Forensics Secur. 7(4),
They are two in one and they can help each other forward. 1181–1191 (2012)
Herein, in the future works, we may focus more on the tex- 17. Jarvis, J.F., Judice, C.N., Ninke, W.H.: A survey of techniques
ture correlation to design a more secure steganographic for the display of continuous tone pictures on bilevel displays.
Comput. Graphics Image Process. 5(1), 13–40 (1976)
method with good visual quality. 18. Kang, L.C., Pieprzyk, J.: Blind steganalysis: A countermeasure
for binary image steganography. In: International Conference
Acknowledgements This work is supported by the National Natu- on Availability, Reliability, and Security, pp. 653–658 (2010)
ral Science Foundation of China (No. U1736118), the National Key 19. Lin, X., Feng, B., Lu, W., Sun, W.: Content-adaptive residual for
R&D Program of China (No. 2017YFB0802500), the Natural Sci- steganalysis. In: International Workshop on Digital Watermark-
ence Foundation of Guangdong (No. 2016A030313350), the Special ing, pp. 389–398 (2014)
Funds for Science and Technology Development of Guangdong (No. 20. Lu, W., Sun, W., Chung, F.L., Lu, H.: Revealing digital fakery
2016KZ010103), the Key Project of Scientific Research Plan of Guang- using multiresolution decomposition and higher order statistics.
zhou (No. 201804020068), the Fundamental Research Funds for the Eng. Appl. Artif. Intell. 24(4), 666–672 (2011)
Central Universities (No. 16lgjc83 and No. 17lgjc45). 21. Mei, Q.G., Wong, E.K., Memon, N.D.: Data hiding in binary
text documents. Proc. SPIE 4314, 369–375 (2001)
22. Pei, S.C., Guo, J.M.: High-capacity data hiding in halftone
images using minimal-error bit searching and least-mean square
References filter. IEEE Trans. Image Process. 15(6), 1665–1679 (2006)
23. Qin, C., Chang, C.C., Chiu, Y.P.: A novel joint data-hiding and
1. Bas, P., Filler, T., Pevn, Y.T.: Break our steganographic system: compression scheme based on smvq and image inpainting. IEEE
the ins and outs of organizing boss. J. Am. Stat. Assoc. 96(454), Trans. Image Process. 23(3), 969–78 (2014)
488–499 (2011) 24. Qin, C., Chang, C.C., Huang, Y.H., Liao, L.T.: An inpainting-
2. Byun, J.Y., Jung, K.H., Yoo, K.Y.: Improved data hiding method assisted reversible steganographic scheme using a histogram
by block parity for binary images. In: International Conference on shifting mechanism. IEEE Trans. Circuits Syst. Video Technol.
Computer Science and Software Engineering, pp. 931–934 (2008) 23(7), 1109–1118 (2013)
3. Chen, J., Lu, W., Fang, Y., Liu, X., Yeung, Y., Xue, Y.: Binary 25. Shapiro, L.: Digital halftoning: Robert ulichney. Comput. Vis.
image steganalysis based on local texture pattern. J. Vis. Commun. Graphics Image Process. 41(3), 382–382 (1988)
Image Represent. 55, 149–156 (2018) 26. Yadav, G.S., Ojha, A.: A fast and efficient data hiding scheme in
4. Chen, J., Lu, W., Yeung, Y., Xue, Y., Liu, X., Lin, C., Zhang, Y.: binary images. In: Eighth international conference on intelligent
Binary image steganalysis based on distortion level co-occurrence information hiding and multimedia signal processing, pp. 79–84
matrix. Comput. Mater. Contin. 55(2), 201–211 (2018) (2012)
5. Feng, B., Lu, W., Sun, W.: High capacity data hiding scheme
for binary images based on minimizing flipping distortion. In:
13
Journal of Real-Time Image Processing (2019) 16:601–609 609
27. Zhang, Y., Qin, C., Zhang, W., Liu, F., Luo, X.: On the fault- Yuileong Yeung received the B.S. degree in Optical Information Sci-
tolerant performance for a class of robust image steganography. ence and Technology from Sun Yat-sen University, China, in 2016. He
Signal Process. 146, 99–111 (2018) is currently a Master’s student with the School of Data and Computer
Science, Sun Yat-sen University, China. His research interests include
multimedia security and multimedia signal processing.
Yingjie Xue received the B.S. degree in School of Mathematics and
Statistics from Central South University, China in 2016. He is currently Xianjin Liu received the B.S. degree from Hainan University, China in
a Master’s student with School of Electronics and Information Tech- 2015. He is currently a Master’s student with the School of Data and
nology, Sun Yat-sen University, China. His research interests include Computer Science, Sun Yat-sen University, Guangzhou, China. His
multimedia security and multimedia signal processing. research interests include multimedia security and multimedia signal
processing.
Wanteng Liu received the B.S. degree in School of Data and Computer
Science, Sun Yat-sen University, Guangzhou, China in 2018. He is cur- Hongmei Liu received the M.S. degree in Computer Science from Tsin-
rently a Master student in School of Data and Computer Science, Sun ghua University, China, and the Ph.D. degree in Radiophysics from
Yat-sen University, Guangzhou, China. His research interests include Sun Yat-sen University, China, in 1996 and 2001, respectively. From
multimedia security and multimedia signal processing. November 2002 to December 2003, she was a postdoctoral fellow at
ENIC, Lille 1, France. She is now an Associate Professor with the
Wei Lu received the B.S. degree in Automation from Northeast Univer- School of Data and Computer Science, Sun Yat-sen University, Guang-
sity, China in 2002, the M.S. degree and the Ph.D. degree in Computer zhou, China. Her current research interests include information hiding,
Science from Shanghai Jiao Tong University, China in 2005 and 2007, image/video watermarking, and video compression.
respectively. He is currently an Associate Professor with the School
of Data and Computer Science, Sun Yat-sen University, Guangzhou,
China. His research interests include multimedia forensics and security,
multimedia signal processing, image/video intelligent analysis.
13