Optik 125 (2014) 3972–3976

Contents lists available at ScienceDirect

Optik
journal homepage:www.elsevier.de/ijleo

A high capacity image steganography using multi-layer embedding

a,∗ b a
Mingwei Tang , Jie Hu , Wen Song
aSchool of Mathematics and Computer Science Technology, Xihua University, Chengdu 610039, China
b School of Information Science and Technology, Southwest Jiaotong University, Chengdu 610031, China

article info abstract

Article history: Until today, high-capacity is still one of important research parts in information hiding. After the secret information
Received 29 July 2013 embedded is extracted, the demand for the image reversibility for the total recovery of the original object without any
Accepted 26 January 2014 distortion goes high. This paper proposes a high capacity steganography using multilayer embedding (CRS), which can
enhance the performance of information hiding system. The experimental results show the proposed CRS scheme has better
Keywords: performance than others. Moreover, the proposed CRS method can display the advantages of good quality image and low
Steganography
complexity of computation.
Image interpolation
High capacity
© 2014 Elsevier GmbH. All rights reserved.
Multi-layer embedding

1. Introduction embed several bits in every vector. Ni et al. [5] presented a novel reversible
data hiding method which uses zero or the minimum points in the histogram
Steganography is not only a secret information communica-tion of an image and slightly changes the pixel grayscale values to embed
technology and but also an important branch of information hiding [1,2]. It is messages. It cannot only embed more data, but also recover the original image
usually used to protect confidential and sensi-tive information in the without any distortion from the stego image. Weng et al. [6] brought the
transmission process of a public network. In an information hiding system, companding technique into traditional difference expansion (DE) to present
there are three main research targets which are capacity, security and another novel reversible information hiding method. Wang et al.
robustness. The high secu-rity and capacity are studied significantly for the
steganographic methods [3]. Because of digital multimedia (such as images) [7] proposed a novel hierarchical information embedding scheme by
redun-dancy, they have been widely used for information hiding as the study attenuating the unbalance level by level. Peng et al. [8] presented an adaptive
objects. The capability (or embedding rate) is the number (or length) of reversible data hiding method using integer transform. The method payload is
information bits which can be embedded in each pixel of cover image (bpp). up to 2.17 bpp. By using reference pixel and multi-layer embedding, Zeng et
The least significant bits (LSB) embedding is the most popular and earliest in al. [9] proposed a reversible data hiding scheme (RHS). The RHS scheme is
the steganography. It is a widely used technique by which the LSB of image based on the pixel difference histogram shifting to expend capacity for
pixels may be changed pseudo randomly on spatial domain. It is provided information embedding.
with many advantages, such as higher payload, image quality and low
computational com-plexity. So far, it is still the hot research in this field. Image interpolation is widely applied to medical imaging field
[12]. By the technique, a small digital image is amplified larger. In addition,
the medical image needs a large amount of data and very high precision. So, it
When the secret messages embedded are extracted, the demand of the is significant for a better interpolation techniques to improve the processing
method reversibility may fully recover the original object without any speed and image capacity.
distention. So far, the researchers have proposed many reversible data hiding Based on the above analysis, the paper proposed a high capacity reversible
methods, such as [4–10] and [11]. Alattar [4] improved the hiding capability steganography using multilayer embedding (CRS). The rest of paper is
and the computation efficiency using difference expansion of vectors. This organized as follows. Section 2 describes simply the image interpolation
method may technique and information hiding scheme proposed by Jung and Yoo [10] and
by Lee and Huang [11]. Section 3 presents the proposed information hiding
scheme in detail how an improvement in the image interpolating method can
∗
Corresponding author. Tel.: +86 028 87720101. E-mail
increase the embedding capacity while preserving good image quality.
address: [email protected] (M. Tang).

http://dx.doi.org/10.1016/j.ijleo.2014.01.149
0030-4026/© 2014 Elsevier GmbH. All rights reserved.
 M. Tang et al. / Optik 125 (2014) 3972–3976

Fig. 2. 3 × 3 same blocks of interpolating image for embedding and extracting pro-cedure.

3. The proposed CRS scheme

3.1. High capacity reversible steganography using multilayer

Fig. 1. The flowchart of data hiding scheme in [10,11].
embedding (CRS)

Section 4 shows the experimental results and analysis. Conclusions are finally The CRS algorithm takes lager advantages of the similar prop-erties of
drawn in Section 5. neighboring pixels difference value. Suppose there is an original image I sized
(R/2) × (L/2). The CRS proposed is applied to produce a R × L cover image C
2. Related work which is used as secret information embedded. CRS can be described in Eq.
(1).
In 2009, Jung and Yoo [10] first proposed to integrate image interpolation
with data hiding in the spatial domain. They pre-sented an information hiding Imin = min{C (i, j), C (i + 2, j), C (i, j + 2), C (i + 2, j + 2)} Imax
method using neighbor mean interpolation (NMI). The NMI can embed more
= max{C (i, j), C (i + 2, j), C (i, j + 2), C (i + 2, j + 2)}
secret data. And then, Lee and Huang [11] proposed a higher capacity image
hid-ing scheme by using an interpolating method with neighboring pixels
(INP). In order to improve the performance of data hiding scheme, INP AD 3 × Imin + Imax
method takes advantage of maximum difference val-ues. The experimental = 4
results show that the scheme INP has good performance, whose payload is up
to 2.28 bpp. C (i, j) = I(i, j)
(1)
As shown in Fig. 1, the scheme proposed in [10,11] changes a R × L size (C (i, j) + C (i, j + 2))/2
input image down to 1/4 of its earliest size, which then turns into the original C (i, j 1) AD
image I. That is, the length and width of the original image are R/2 and L/2,
+ = + 2
respectively. Next, an original image of size R/2 × L/2 is scaled up into a
cover image C of size R × L. After secret information is embedded in the C (i 1, j) AD (C (i, j) + C (i + 2, j))/2
cover image, A stego image S comes into being. + = + 2

The receiver can recover embedded secret messages from the stego image. C (i 1, j 1) (C (i, j) + C (i + 1, j) + C (i, j + 1) ,
The detailed process of NMI and INP is described in [10,11]. + + = 3
i = 2h, j = 2l, h, l = 0, 1, 2, . . ., k

The pixel C(i, j) = I(i, j), i = 2h, j = 2l is used as a unchanged pixel which
will remain unchanged in the process of information embedding. The
procedure for CRS is sketched as follows.

Algorithm 1: Embedding Algorithm

Input:
An original imageI sized (R/ 2) × (L/ 2), whereI(i, j) is a pixel at (i, j) in
the original imageI, i ← 0, 1, . . . , (R/ 2) − 1; j ← 0, 1, . . . , (L / 2) − 1
Output:
A stego image S sized R × L produced
1 begin
2 Step 1. Initially set index variables i ← 0, j ← 0
3 Step 2. A cover image C sized R × L is produced based on Equ.1
4 Step 3. If cover image C has been used completely then
5 Proceed with Step 7
6 Else
7 Select a 3 × 3 block. Compute di based on Equ .2
8 End If
9 Step 4. Compute ni based on Equ .3
10 Step 5. Secret information is embedded based on Equ .4
11 Step 6. If all the secret information is embedded completely then
12 Output a stego image S sized R × L. Proceed with Step 7
13 Else
14 Proceed with Step 3
15 End If
16 Step 7. End (Algorithm)
17 end
3974 M. Tang et al. / Optik 125 (2014) 3972–3976

Fig. 2 shows an example of the procedure of CRS. Let I be a orig- The extraction process is a reversible process of the embedding
inal image is as shown in Fig. 2(a). The unchanged pixels in image process described in CRS.
C (Fig. 2(b)) can be from the proposed CRS. In the extracting procedure, every nine stego pixels within a
same block are S(i, j), S(i + 1, j), S(i + 2, j), S(i, j + 1), S(i + 1, j + 1),
Imax − C (i, j + 1), C (i, j + 1) < (Imin + Imax )/2 S(i + 2, j + 1), S(i, j + 2), S(i + 1, j + 2) and S(i + 2, j + 2). The I max and
d= I ,
1 C (i, j 1) min otherwise Imin are computed by determining the maximum value of the
+ − four unchanged pixels S(i, j), S(i + 2, j), S(i, j + 2) and S(i + 2, j + 2),
d
2 Imax − C (i + 1, j), C (i + 1, j) < (Imin + Imax )/2 (2) Imax = max{S(i, j), S(i + 2, j), S(i, j + 2), S(i + 2, j + 2)}, Imin = min{S(i, j),
= C (i + 1, j) − Imin, otherwise S(i + 2, j), S(i, j + 2), S(i + 2, j + 2)} and AD = (3 × Imin + Imax)/4.
The next step is to compute the difference value between
Imax − C (i + 1, j + 1), C (i + 1, j + 1) < (Imin + Imax )/2 each selected pixel, Imax and Imin to find the number of embed-
d3 =
C (i + 1, j + 1) − Imin, otherwise ded secret information. In Fig. 2(b), the original image recovered
can be kept unchanged. This is because the four unchanged pix-
After obtaining the maximum and minimum values Imax and
els of cover image C are equal to original image as shown in
Imin, respectively, we may calculate the difference values d1, d2 and Fig. 2. Afterwards, the receiver can compute the other interpo-
d3 between pixels C(i, j + 1), C(i + 1, j) and C(i + 1, j + 1), respectively, lating pixels C(0, 1) = 167, C(1, 0) = 164, C(1, 1) = 172 by CRS. So,
Imax and Imin the values Imax and Imin are obtained from a 3 × 3 same the restoration of original image is completed. There will be the
block in a cover image, but the difference values are only calculated
maximum value Imax = 188 and minimum value Imin = 152 from the
from three neighboring pixels of the unchanged pixel C(i, j). That is,
four unchanged pixels of C(0, 0), C(0, 2), C(2, 0) and C(2, 2) in
d1, d2 and d3 are calculated as follows in Eq. (2). When the difference the same block. And then, we can calculate the difference val-
values are obtained, dk can determine how many bits or length of a ues between Imax and Imin and C(0, 1), C(1, 0) and C(1, 1) to find
secret sub-message can be hidden into a selected pixel. The amount
the number of secret information embedded. The number bits of
of information in bits can be conformed as nk which are as shown secret sub-message n1, n2 and n3, are all four. The difference val-
in Eq. (3) [11].
ues b1, b2 and b3 are 7, 2 and 12, respectively. So, the sub-message
01112, 00102 and 11002 are calculated. At last, the secret message
nk = log2dk , k = 1, 2, 3 (3)
SI = “01112”||“00102”||“11002” = “0111001011002”.
Let bk be on behalf of a bit string of nk secret bits embedded.
Then, for each unchanged pixel C(i, j), its neighboring selected pixels
C(i, j + 1), C(i + 1, j) and C(i + 1, j + 1) can embed the secret informa- 4. Experimental results and performance analysis
tion bk, for k = 1, 2 and 3, respectively. Thus, the stego pixels can be
calculated as Eq. (4). This section describes the experimental results and compares
the algorithm performance of the proposed CRS scheme with oth-
S(i, j) = C (i, j) ers. Twenty 8-bit gray scale test images, each of whose sizes are
S(i, j + 1) = C (i, j + 1) − b1 all 512 × 384, were used in the experiment as shown in Fig. 3. The
peak signal-to-noise ratio (PSNR) was applied to appraise the image

(4)
S(i + 1, j) = C (i + 1, j) − b2
S(i + 1, j + 1) = C (i + 1, j + 1) − b3, i = 2h, j = 2l,

h, l = 0, 1, 2, . . ., 127
3.2. Example

We list the example of the original, the cover and the stego image by CRS
in Fig. 2.
Let SI be a secret message. Suppose SI = “0111001011002” which is to be
embedded into a cover image C. Let i and j be 0 and the max-imum value
Imax = max{C(0, 0), C(0, 2), C(2, 0), C(2, 2)} = max{185, 161, 152, 188} =
188, the minimum value Imin = min{C(0, 0), C(0, 2), C(2, 0), C(2, 2)} =
min{185, 161, 152, 188} = 152 and AD = 161 of the four pixels C(i, j), C(i, j +
2), C(i + 2, j) and C(i + 2, j + 2) in a 3 × 3 same block. Calculate the
difference values d1, d2 and d3 to determine the amount of secret information
to be embedded. The difference values d1, d2 and d3 are calculated as follows:
d1 = 188 − 167 = 21, d2 = 188 − 164 = 24 and d3 = 172 − 152 = 20.
 Calculate the number of bits, nk, of secret information the CRS based on
Eqs. (2) and (3).

n1 = log2d1 = log221 = 4
n
2 = log2d2 = log224 = 4 (5)
n
3 = log2d3 = log220 = 4
The length of secret information b1 is 4 bits, 4 bits for b2 and
b3, too. With the difference values d1 = 21, d2 = 24 and d3 = 20 as
well as the number bits of data, their values are b1 = “01112” = 710,
b2 = “00102” = 210 and b3 = “11002” = 1210, respectively. Finally, the
secret sub-information can be embedded using Eq. (4) and the

stego-pixels are as shown in Fig. 2(c). Fig. 3. Twenty test grayscale images.
 M. Tang et al. / Optik 125 (2014) 3972–3976

Table 1
Comparison of image quality (PNSR) and capacity (bpp) with CRS, INP and NMI.

Test image CRS INP NMI

PSNR (dB) Capacity (bpp) PSNR (dB) Capacity (bpp) PSNR (dB) Capacity (bpp)
Image01 32.35 2.20 32.96 1.81 32.82 1.12
Image02 33.05 1.82 33.60 1.45 33.10 0.96
Image03 34.33 1.60 35.15 1.24 34.92 0.61
Image04 30.70 2.18 30.57 2.16 30.27 1.52
Image05 34.15 2.27 33.76 1.82 32.99 1.11
Image06 32.42 2.41 32.24 1.97 31.65 1.28
Image07 33.27 2.12 33.19 1.67 32.78 0.93
Image08 32.38 2.20 32.27 2.04 31.56 1.34
Image09 32.19 2.45 32.09 1.96 31.72 1.21
Image10 35.65 1.49 35.66 1.15 35.49 0.49
Image11 32.57 2.25 32.54 1.79 32.15 0.75
Image12 34.68 1.75 34.81 1.44 34.17 0.95
Image13 34.50 1.82 34.03 1.15 33.51 1.31
Image14 31.86 2.14 31.81 2.01 31.25 1.38
Image15 36.98 1.20 36.76 0.92 36.17 0.58
Image16 34.73 1.79 34.33 1.39 33.68 1.40
Image17 30.66 2.02 30.22 2.13 29.70 1.91
Image18 34.65 1.75 34.31 1.42 33.70 0.94
Image19 34.75 1.72 34.41 1.26 33.80 0.82
Image20 31.08 2.05 30.73 2.03 30.35 1.63

quality of the input images and the stego images. If PSNR is greater than 30 have been converted into 8-bit grayscale images before image pre-processing
dB, image distortion is generally imperceptible [11]. and multi-layer data embedding is applied to the gray scale images. A
Matlab (version: R2009a) is used a tool for this experiment. There are comparison of same capacity values occurrences for the three methods using
1500 images from [13,14] and the Internet to make up image database in the similar interpolation is shown in Fig. 5 from UCID-Image Database (v2) [14].
experiments. The randomly selected 20 images are listed in Fig. 3. The size of
all experimental images is 512 × 384 or 384 × 512. It could embed more amount of information compared with other
reversible methods that can contain higher capacity as shown
Firstly, each input 512 × 384 grayscale image was reduced to 256 × 192
as an original image. And then, it was extended to 512 × 384 by the method
CRS as a cover image. (a)6 x 10 5
The performances of CRS are better than NMI and INP, observ-ably. In CRS

addition, the image quality of the CRS is superior to the NMI and INP for INP

twenty test images in Table 1. Moreover, Table 1 shows the results of image NMI
capacity using three different algorithms for twenty test images. Based on
5
experiment results, the CRS shows better performance in relevant
(bits)

interpolating pixels.
Additionally, the CRS gains better capacity than other methods, such as
Pure capacity

NMI and INP. Although the visual image quality of CRS in eyes is almost 3
similar to that of NMI as shown in Table 1, the steganography CRS is actually
better in columns 2 and 3 than INP in columns 4 and 5 and NMI in columns 6 2

and 7 according to the results from Table 1. Results demonstrate that the
1
proposed CRS method is better than other two methods INP [11] and NMI
[10]. 0

The experimental results of pure capacities of 100 images of randomly 0 10 20 30 40 50 60 70 80 90 100

selection are listed in Fig. 4(a) for comparing with other reversible methods. Test image
The sizes of experimental images are 512 × 384 or 384 × 512. The pure (b) 3
capacity is the number of maximal pure capacity.
CRS

INP
The experimental results of capacities (bpp) of 100 images of randomly
selection are listed in Fig. 4(b) for comparing with other reversible methods. NMI

The sizes of experimental images are 512 × 384 or 384 × 512. The capacity is 2. 5

the number of maximal capacity. 2

capacity (bpp)

The experimental results showed that the proposed method is better in

comparison with others. It is shown that our method is easy to be applied to
reversible data hiding. Especially, it can embed secret data more when 1. 5

producing larger difference between neigh-boring pixels like as image01 in

Table 1. 1
The occurrence number of images shown in Fig. 5 means the number of
0. 5
images.
The proposed scheme can offer average payload 1.79 bpp and average
0
PSNR value 33.85 dB, in the sense that the proposed scheme can offer very
high payload when keeping low distortion. It should be noted that all of the 0 10 20 30 40 50 60 70 80 90 1

1338 images in UCID-Image Database (v2) Test image

(c) 50
 CRS/INP

(100%)
45

40
of capacity
35

30
rate

25
Improving

20

15

0 10 20 30 40 50 60 70 80 90

Test image

Fig. 4. Comparisons of pure capacity, capacity and the improving rate of capacity.
3976 M. Tang et al. / Optik 125 (2014) 3972–3976

maintains low computational complexity and has higher perfor-mance. In

700
addition, the CRS gains good PSNRs. Compared to the scheme INP, the
CRS
experimental results of proposed information hid-ing scheme improved the
INP
capacity. Therefore, this shows fully the better performance of the proposed
scheme.
of Image

600 NMI

500
Acknowledgments
Occurrenc Number

400
This work is supported by the key of Scientific Research Funds Project of
Educational Department of Sichuan Province (No. 13ZA0019), the key of
300 Scientific Research Funds Project of Xihua University (Nos. Z1222623 and
Z1322622) and the Scientific Research Funds Project of Science and
Technology Department of Sichuan Province (No. 2013JY0089).
200
The

100
References
0
[1] R.J. Anderson, F.A.P. Petitcolas, On the limits of steganography, IEEE J. Select. Areas
[0,0.4) [0.4,0.6) [0.6,0.8) [0.8,1) [1,1.2) [1.2,1.4) [1.4,1.6) [1.6,1.8) [1.8,2) >=2,
Commun. 16 (4) (1998) 474–481, http://dx.doi.org/10.1109/49.668971, ISSN: 0733-8716.
Payload (bpp)

[2] F. Petitcolas, R. Anderson, M. Kuhn, Information hiding – a survey, Proc. IEEE 87

Fig. 5. Test results on all of the 1338 images in UCID-Image Database (v2). (7) (1999) 1062–1078, http://dx.doi.org/10.1109/5.771065, ISSN: 0018-9219.
[3] W. Lin, S. Horng, T. Kao, R. Chen, Y. Chen, C. Lee, T. Terano, Image copy -right
Table 2 protection with forward error correction, Expert Syst. Appl. 36 (2009) 11888–11894.
Test results on all of the 1338 images in UCID-Image Database (v2).
[4] A. Alattar, Reversible watermark using the difference expansion of a general-ized integer
Algorithm PSNR (dB) Capacity (bpp) transform, IEEE Trans. Image Process. 13 (8) (2004) 1147–1156,
http://dx.doi.org/10.1109/TIP.2004.828418, ISSN: 1057-7149.
Max Min Average Max Min Average [5] Z. Ni, Y. Shi, N. Ansari, W. Su, Reversible data hiding, IEEE Trans. Circuits Syst. Video
Technol. 16 (2006) 354–362, ISSN: 1051-8215.
RHS 36.89 28.41 30.88 1.81 0.21 1.11 [6] S. Weng, Y. Zhao, J. Pan, A novel reversible data hiding scheme, Int. J. Innov. Comput.
NMI [10] 39.83 29.45 33.24 2.04 0.05 0.96 Inf. Control 4 (2008) 351–358.
INP [11] 40.36 30.20 33.79 2.28 0.23 1.59 [7] K. Wang, Q. Liu, L. Chen, Hierarchical reversible data hiding based on statisti-cal
CRS 40.95 30.01 33.85 2.68 0.49 1.79 information: preventing embedding unbalance, Signal Process. 92 (2012) 2888–2900.

[8] F. Peng, X. Li, B. Yang, Adaptive reversible data hiding scheme based on integer
in Fig. 4(c), which the PSNR was guaranteed to be higher than 30 dB as transform, Signal Process. 92 (2012) 54–62.
[9] X. Zeng, Z. Li, L. Ping, Reversible data hiding scheme using reference pixel and multi-
shown in Table 1. layer embedding, AEU – Int. J. Electr. Commun. 66 (2012) 532–539.
The proposed scheme has been applied to all of the 1338 images in UCID-
Image Database (v2) [14], which includes many different types of images and [10] K. Jung, K. Yoo, Data hiding method using image interpolation, Comput. Stand. Interfaces
31 (2009) 465–470.
the experimental results are displayed in Fig. 5 and Table 2. [11] C. Lee, Y. Huang, An efficient image interpolation increasing payload in reversible data
hiding, Expert Syst. Appl. 39 (2012) 6712–6719.
[12] P. Thevenaz, T. Blu, M. Unser, Interpolation revisited, IEEE Trans. Med. Imaging 19 (7)
(2000) 739–758, http://dx.doi.org/10.1109/42.875199, ISSN: 0278-0062.
5. Conclusion
[13] T.U. of Southern California, USC-SIPI Image Database, http://sipi.usc.edu/
services/database/database.html (18.05.13).
We proposed a high capacity reversible steganography using multilayer [14] L. University, UCID Image Dataset, http://homepages.lboro.ac.uk/cogs/datasets/
embedding (CRS). And then, the data hiding method, based on the proposed ucid/data/ucid.v2.tar.gz (20.05.13).

maximizing the difference values between neighboring pixels, was proposed,

which applicable to information hiding and reversible information hiding
method.
This paper modified the Lee and Huang approach to develop an improved
method for CRS and information hiding scheme. Like the approach INP,
interpolating images are designed by the properties between neighboring
pixels. Therefore, the proposed scheme still