Académique Documents
Professionnel Documents
Culture Documents
Shou-Shiung Chou
Tunghai University/Department of Hospitality Management, Taichung, Taiwan
Email: shchou@thu.edu.tw
Abstract—This work proposes a reversible data hiding Fridrich et al. [3,4] compressed the least significant
algorithm that is based on the interleaving max-min bit (LSB) plane to obtain additional space for embedding
difference histogram. The method divides the cover secret data. Celik et al. [5-7] improved the method of
image into non-overlapping identical blocks. In each Fridrich et al. and proposed the Generalized-LSB (G-
block, the maximum or minimum pixel is selected to LSB) scheme by compressing the quantization residuals
calculate absolute differences between the gray level of pixels to yield additional space to embed a message.
value of the selected pixel and that of the other pixels. Awrangjeb and Kankanhalli [8,9] presented a scheme that
Then, these differences are used to generate a detects the textured blocks; extracts the LSBs of the
histogram and the histogram shifting method is pixel-values from these textured blocks based on the
adopted to embed data. This method, to propose an Human Visual System (HVS), and concatenates the
offset distortion method, exploits the feature that the authentication information with the compressed bit-
change of the pixel gray level upon shifting and string.
embedding obtained using the max difference is 2) Difference expansion
contrary to that obtained using the min difference. Tian [10] presented a method that expands the
The method of interleaving uses the max difference difference between two neighboring pixels to obtain
and the min difference to counteract the change in the redundant space for embedding a message. Alattar [11]
gray levels of some pixels. The total of the changed used the difference expansion of vectors of adjacent
gray level values is thus reduced. Therefore, the PSNR pixels to obtain more embedding space. Using Tian′s
value can be increased and the quality of the stego scheme of difference expansion, Chang and Lu [12]
image improved. Experimental results reveal that the calculated the difference between a pixel and the mean
proposed method yields a higher embedding capacity value of its neighboring pixels to embed a message.
than that of the median difference histogram. Weng et al. [13,14] used the correlations among four
pixels in a quad, and embedded data by expanding the
Index Terms—reversible data hiding, histogram differences between one pixel and each of its three
shifting, difference, max difference, min difference neighboring pixels.
3) Histogram shifting
Ni et al. [15] used the histogram of the pixel values
I. INTRODUCTION in the cover image to embed secret data into the
Reversible data hiding [1,2], also known as lossless maximum frequency pixels. Fallahpour and Sedaaghi
data hiding, enables marked media to be restored to their [16] divided a cover image into several blocks, and
original form without any distortion. This technique is embedded secret data into the histogram of each block.
applied in such fields as content authentication of Lin and Hsueh [17] applied the histogram shifting of Ni
multimedia data, law enforcement, medical imagery and et al. to the pixel differences, which were obtained from
astronomical research. Various reversible data hiding the absolute differences between pairs of neighboring
methods have been proposed for grayscale images, and pixels.
these can be divided into the following categories: 4) Integer wavelet transform
1) Data compression
*
Corresponding author. Tel:+886 4 27016855 ext. 1381; Fax:+886 4 27026986
Xuan et al. [18-20] proposed a lossless data hiding Step 2. Divide cover image into blocks
technique based on integer wavelet transform, which The CI is divided into non- overlapping identical w×h
embeds high capacity data into the most insensitive bit- blocks, w and h are the width and height of the block,
planes of wavelet coefficients. Yang et al. [21] presented respectively. Record w and h as OI.
a symmetrical histogram expansion scheme in the Step 3. Calculate block difference
transform domain of Piecewise-Linear Haar (PLHaar). In a block, select a pixel as the base pixel and each other
Data are embedded into the PLHaar coefficients of the pixel can get an absolute difference of gray level value
images from the pivotal bin of a histogram of PLHaar (called difference value) between the base pixel and it.
coefficients symmetrically toward both sides of the The calculated difference values are called block
histogram. difference.
The goal of these methods is to realize a high Step 4. Generate difference histogram
embedding capacity while yet maintaining high image Let f(x) denotes the frequency of a difference value x,
quality. The proposed scheme uses the block difference to where “frequency” is the number of instances of a
generate histogram to embed data, and propose an offset difference value. Then the difference histogram is { f(x) |
distortion method to reduce the total of the changed gray 0≦x≦255 }.
level value. Experimental results reveal that the proposed Step 5. Shifting and embedding
method has a greater embedding capacity and produces Find the difference value of maximum and minimum
less image distortion than previous schemes. frequency in the difference histogram, a and b. The area
The rest of this paper is organized as follows. between a+1 and b-1 are shifted to right by one, and the
Section 2 describes the proposed scheme. Section 3 frequency of a+1 is leaved unoccupied. Then the
summarizes the experimental results. Section 4 draws frequency of a is used to embed data, a pixel has
conclusions. difference value a can embed one bit. If the embedded
data is bit 1, then the difference value is increased by one,
II. PROPOSED SCHEME i.e. become a+1. If the embedded data is bit 0, then the
The proposed interleaving max-min difference difference value is remained. When the difference value
histogram method is based on block difference histogram of pixel is increased by one, the gray level value of the
and an offset distortion method (called interleaving max- pixel needs to change according to the relationship of
min difference), and they are described below. sorted order between the pixel and the base pixel. Record
a and b as OI.
A. Block Difference Histogram Let Pij is the j-th row and i-th column pixel in CI, Pb
The embedding process (called EP) and extraction is the selected base pixel, g(Pij) denotes the gray level
process (called E′P) are described below. value of Pij, and d(Pij) denotes the difference value of Pij
1) Embedding Process and is calculated as d(Pij)=|g(Pij)-g(Pb)|, Pij ≠ Pb. Let
The block difference histogram method is shown as o(Pij) is the sorted order of Pij in a block, c(Pij) is the
Figure 1. The step 1-5 are described below. change of g(Pij) when shifting and embedding, e is the
Cover Image 2. Divide cover image embedded bit. As shown in Figure 2, the sorted pixels in
1. Handle into blocks
LBP and
a block are divided into several pixel sets. The
3. Calculate
UBP block definitions, process and c(Pij) of the pixel sets are listed
difference
in Table 1. After EP, the changed CI is saved as stego
image (called SI).
Stego Image Difference Histogram of Baboon Stego Image Difference Histogram of Baboon Cover Image
9000 9000
8000 8000
7000 7000
frequency frequency
6000 5. Shifting 6000 4. Generate
5000
4000
and 5000
4000
difference
3000 embedding 3000 histogram
2000 2000
1000 1000
0
0 50 100 150
difference value
200 250
0
0 50 100 150
difference value
200 250 Figure 2. Pixel sets in a block and how to change their gray level value
when shifting and embedding
Figure 1. Block difference histogram method
Figure 4. Pixel sets in a block and their intersection in EP2k+1 and EP2k+2
TABLE 3.
Figure 3. Pixel sets in a block and how to change their gray level value C′′(PIJ) OF PIXELS IN THE PIXEL SET INTERSECTIONS IN EP2K+1 AND EP2K+2
when extracting and shifting back
Pixel Set c′′(Pij)
B∩Pm′ if C={} -1
TABLE 2. B∩A′ -1
DEFINITIONS, PROCESS AND C′(PIJ) OF THE PIXEL SETS IN A BLOCK IN E′P
B∩B′ 0
o(Pij)<o(Pb) o(Pij)>o(Pb) A∩B′ +1
Definition Process
Pixel Set c′(Pij) Pixel Set c′(Pij) B′∩Pm if C′={} +1
{ Pij | 0≦d(Pij)≦a-1 } nothing Sa 0 Sa ′ 0 2) Embedding Process
A A′
{ Pij | d(Pij)=a } extract 0 Ea0 0 Ea0′ 0 Figure 5 displays EP in the proposed scheme. Secret
{ Pij | d(Pij)=a } extract 1 Data (called SD) are embedded into CI by repeatedly
Ea1 +1 Ea1′ -1
shift back B B′ performing the embedding process until SD have all been
{ Pij | a+1≦d(Pij)≦b-1 } shift back LaSb +1 LaSb′ -1 fully embedded; finally, SI and key are output. The key
{ Pij | b≦d(Pij)≦255 } nothing C Lb 0 C′ Lb′ 0 contains the information that must be used by the receiver
to extract SD. EP has three steps, which are as follows.
Step 1.
B. Interleaving Max-Min Difference Handle LBP and UBP in CI. Use the method
The proposed offset distortion method, embedding described in Section 2.1.1, and save LUBPI as OI.
process, and extraction process are described below. Step 2.
1) Offset Distortion Method Let EPk be the k-th EP. If k is odd, divide CI into
The calculated difference when the maximum pixel blocks and in each block select the maximum pixel
in every block is designated the base pixel is called the (defined in Section 2.2.1) as base pixel to calculate max
max difference. The calculated difference when the difference. Find a and b (defined in Section 2.1.1), and
minimum pixel in each block is chosen as the base pixel save w,h,a,b as OI.
is called the min difference. The change of the pixel gray If k is even, use the block size in EPk-1 to divide CI
level upon shifting and embedding obtained using the and in each block select the minimum pixel (defined in
max difference is contrary to that obtained using the min
Section 2.2.1) as base pixel to calculate min difference.
difference. The method of interleaving uses the max
Find a and b, and save w,h,a,b as OI.
difference and the min difference to counteract the
Step 3.
change in the gray levels of some pixels.
Use the block difference to embed SD and OI. Use
After dividing CI into blocks, let P is a pixel and the method described in Section 2.1.1.
o(P) denotes the sorted order of P in a block. Let
Pm = arg max(
P
o ( P ))
, then Pm is called the maximum pixel in
the block; let Pm' = arg min(
P
o ( P ))
, then Pm′ is called the
N.3
1.3
2.3
Embed
Embed
Embed
Embed
N-1.2
N-1.1
(w,h,a
N-2.2
Embed
Embed
N.1
Embed
Embed
2.1
1.1
2.2
1.2
(w,h,a
LUBI2
LUBIN
LUBIN-1
LUBI1
(w,h,a
(w,h,a
(w,h,a
,b) N+
N.2
N-1.3
N.3
1.3
2.3
,b) N-1
,b
,b) N-2
,b
N+C N
)2
)1
N-1.2
N.2
2.2
1.2
Extract
Extract
Extract
Extract
Cover Stego
Image E′PN E′PN-1 ... E′P2 E′P1 Image
2.1
(w,h,a
Extract
2.3
N-1.3
N-1.1
Extract
3.1
Extract
Extract
N.3
1.3
N.1
(w,h
LUBIN-1
(w,h,a
LUBIN
LUBI1
(w,h,a
LUBI2
(w,h,a
,b) N+
1.1
,a,b) N
N-1.2
N.2
1.2
2.2
,b) N-2
,b
N+ C N
,b) 2
-1
)1
1.4
1.2
1.0
Proposed(3X4)
Payload (bpp)
L&H2 [52]
0.8
F&S16 [51]
Ni [50]
Figure 11. 512×512 grayscale cover images and secret data image tested 0.4
in experiments
The performance of the proposed method was 0.2
Cover Image
approached 30dB. Table 4 presents the results of this
process, where the unit of payload is bpp (bits per pixel). Figure 12. Comparison of performance of proposed methods with that
The 3×4 block yielded the largest average payload, of histogram shifting category methods under different cover images
1.120bpp.
The performance was measured and compared with Figures 13, 14, 15 and 16 compare the average
those of Ni [15], F&S4 [16], F&S16 [16] and L&H [17]. performance (payload and PSNR value) of the proposed
The embedding process was repeated until the PSNR methods with those of the histogram shifting, data
approached 30dB. Table 5 presents the results of this compression, difference expansion and integer wavelet
process, where the unit of payload is bpp. For each image transform category methods, respectively. The figures
and on average, the proposed method outperformed the reveal that the proposed method outperforms other
other methods. methods.
The comparison of performance (the payloads when
the PSNR approached 30dB) of proposed methods with
that of histogram shifting category methods under
different cover images is shown as Figure 12. The figure
shows that our proposed methods perform better than
other methods.
1.2 1.2
Proposed(3X4) Proposed(3X4)
1.0 1.0
L&H2 [52] MD-optimal
Payload (bpp)
Payload (bpp)
F&S4 [51]
Xuan et al.[54]
0.6 0.6
Ni [50]
0.4
0.4
0.2
0.2
0.0
0.0
30 35 40 45 50 55 PSNR (dB)
30 35 40 45 50 55
PSNR (dB)
Figure 13. Comparison of average performance of proposed method Figure 16. Comparison of average performance of proposed method
with those of histogram shifting category methods with those of integer wavelet transform category methods
1.2
IV. CONCLUSIONS
Proposed(3X4)
1.0 This paper proposed a block difference histogram
Celik et al.[41] method and an offset distortion method to embed data
0.8 with reversible. In the experiment, when the PSNR
Awrangjeb and Kankanhalli[44]
approached 30dB, the optimal payloads of the
Payload (bpp)
0.0
[1] M. Awrangjeb, “An overview of reversible data
30 35 40 45 50 55 hiding,” International Conference on Computer and
PSNR (dB) Information Technology (ICCIT), 2003, pp. 75-79.
[2] Y.Q. Shi, Z. Ni, D. Zou, C. Liang, and G. Xuan,
Figure 14. Comparison of average performance of proposed method “Lossless data hiding: fundamentals, algorithms and
with those of data compression category methods applications,” IEEE International Symposium on
Circuits and Systems (ISCAS), Vol. II, 2004, pp. 33-
1.2
36.
[3] J. Fridrich, M. Goljan, and R. Du, “Invertible
Proposed(3X4)
authentication,” Proc. SPIE Security and
1.0
Weng et al.[48] Watermarking of Multimedia Contents, 2001, pp.
Chang and Lu[47] 197–208.
0.8
[4] J. Fridrich, M. Goljan, R. Du, “Lossless data
Payload (bpp)
Alattar[14]
embedding-new paradigm in digital watermarking,”
0.6 Tian[46] EURASIP Journal on Applied Signal Processing,
Vol. 2, 2002, pp. 185–196.
0.4 [5] M.U. Celik, G. Sharma, A.M. Tekalp, and E. Saber,
”Reversible data hiding,” Proc. IEEE Int. Conf.
0.2
Image Processing, Vol. 2, 2002, pp. 157-160.
[6] M.U. Celik, G. Sharma, A.M. Tekalp, and E. Saber,
“Lossless generalized-LSB data embedding,” IEEE
0.0
30 35 40 45 50 55
Trans. Image Processing, Vol. 14, No.2, 2005, pp.
PSNR (dB) 253–266.
[7] M.U. Celik, G. Sharma, A.M. Tekalp, “Lossless
Figure 15. Comparison of average performance of proposed method watermarking for image authentication: a new
with those of difference expansion category methods framework and an Implementation,” IEEE Trans.
Image Processing, Vol. 15, No.4, 2006, pp. 1042–
1049.
[8] M. Awrangjeb and M.S. Kankanhalli, “Lossless based on integer wavelet transform,” in Pro. Int.
Watermarking Considering the Human Visual Sym. Circuits and Systems, Vol. 2, 2004, pp. 29-32.
System,” Lecture Notes in Computer Science, Vol. [20] G. Xuan, Y.Q. Shi, C. Yang, Y. Zheng, D. Zou, snd
2939, 2004, pp. 329-336. P. Chai, “Lossless Data Hiding Using Integer
[9] M. Awrangjeb and M.S. Kankanhalli, “Reversible Wavelet Transform and Threshold Embedding
watermarking using a perceptual model,” Journal of Technique,” IEEE Int. Conf. Multimedia and Expo,
Electronic Imaging, Vol. 14, No. 1, 013014, 2005, 2005, pp. 1520-1523.
pp. 1-8. [21] L. Yang, P. Hao, and C. Zhang, “Progressive
[10] J. Tian, “Reversible data embedding using a reversible data hiding by symmetrical histogram
difference expansion,” IEEE Transactions on expansion with piecewise-linear haar transform,”
Circuits and Systems for Video Technology, Vol. 13, IEEE International Conference on Acoustics,
No.8, 2003, pp. 890–896. Speech, and Signal Processing (ICASSP), Vol. II,
[11] A.M. Alattar, “Reversible watermark using the 2007, pp. 265-268.
difference expansion of a generalized integer
transform,” IEEE Trans. Image Processing, Vol. 13, Hsien-Wei Yang is an Assistant Professor of Department of
No.8, 2004, pp. 1147–1156. Information Management at Overseas Chinese University,
[12] C.C. Chang, and T.C. Lu, “A difference expansion Taiwan, ROC. His research interests include Data Hiding, Web
oriented data hiding scheme for restoring the original Service, and E-Learning. From 1999 to date, he was a faculty of
the Information Management Department at Oversea Chinese
host images,” The Journal of Systems and Software, University. He received his Ph.D. in Applied Mathematics from
Vol. 79, 2006, pp. 1754–1766. ChungHsing University in 2009. The contact address of Dr.
[13] S. Weng, Y. Zhao, J.S. Pan, R. Ni, “A novel Yang is Department of Information Management, Overseas
reversible watermarking based on an integer Chinese University, No:100, Chiao Guang Rd., Taichung 407,
transform,” IEEE Int. Conf. Image Processing, Vol. Taiwan, R.O.C. E-mail: fiber@ocu.edu.tw.
3, 2007, pp. 241-244.
[14] S. Weng, Y. Zhao, J.S. Pan, and R. Ni, “Reversible Kuo-Feng Hwang is an Assistant Professor of Department
watermarking based on invariability and adjustment of Information Management at Overseas Chinese University,
on pixel pairs,” IEEE Signal Processing Letters, Vol. Taiwan, ROC. His research interests include Internet Security,
Data Mining, Distance Learning, and E-Commerce. From 1999
15, 2008, pp. 721-724. to date, he was a faculty of the Information Management
[15] Z. Ni, Y.Q. Shi, N. Ansari, and W. Su, “Reversible Department at Oversea Chinese University. He received his
data hiding,” IEEE Trans. Circuits Syst. Video Ph.D. in Applied Mathematics from ChungHsing University in
Technol., Vol. 16, No. 3, 2006, pp. 354-362. 2006. The contact address of Dr. Hwang is Department of
[16] M. Fallahpour and M.H. Sedaaghi, “High capacity Information Management, Overseas Chinese University,
lossless data hiding based on histogram No:100, Chiao Guang Rd., Taichung 407, Taiwan, R.O.C. E-
modification,” IEICE Electronics Express, Vol. 4, mail: kfhwang@ocu.edu.tw.
No. 7, 2007, pp. 205-210.
[17] C.C. Lin and N.L. Hsueh, “Hiding data reversibly in Shou-Shiung Chou is a Lecturer of Department of
Hospitality Management at Tunghai University, Taiwan, ROC.
an image via increasing differences between two His research interests include Marketing Services and Hotel
neighboring pixels,” IEICE TRANS. INF. & SYST., Management. He is studying doctor degree in Department of
Vol. E90–D, No. 12, 2007, pp. 2053-2059. Business Administration, National Taipei University. After Mr.
[18] G. Xuan, J. Chen, J. Zhu, Y.Q. Shi, Z. Ni, and W. Chou finished the Certificate in Hotel Management in
Su, “Lossless data hiding based on integer wavelet Singapore International Hotel And Tourism College, he went as
transform,” IEEE Workshop on Multimedia Signal a Hotelier in Evergreen Laurel Hotel Hong Kong and Taichung
Processing, 2002, pp. 9-11. for seven years..
[19] G. Xuan, Y.Q. Shi, Z.C. Ni, J. Chen, C. Yang, Y.
Zhen, J. Zheng; “High capacity lossless data hiding