J. Software Engineering & Applications, 2009, 2: 383-387
doi:10.4236/jsea.2009.25051 Published Online December 2009 (http://www.SciRP.org/journal/jsea)
Copyright © 2009 SciRes JSEA
383
Information Hiding Method Based on Block DWT
Sub-Band Feature Encoding
Qiudong SUN, Wenxin MA, Wenying YAN, Hong DAI
School of Electronic and Electrical Engineering, Shanghai Second Polytechnic University, Shanghai, China.
Email: {qdsun, wxma, wyyan, daihong}@ee.sspu.cn
Received August 12th, 2009; revised September 17th, 2009; accepted October 10th, 2009.
ABSTRACT
For realizing of long text information hiding and covert communication, a binary watermark sequence was obtained
firstly from a text file and encoded by a redundant encoding method. Then, two neighboring blocks were selected at
each time from the Hilbert scanning sequence of carrier image blocks, and transformed by 1-level discrete wavelet
transformation (DWT). And then the double block based JNDs (just noticeable difference) were calculated with a visual
model. According to the different codes of each two watermark bits, the average values of two corresponding detail
sub-bands were modified by using one of JNDs to hide information into carrier image. The experimental results show
that the hidden information is invisible to human eyes, and the algorithm is robust to some common image processing
operations. The conclusion is that the algorithm is effective and practical.
Keywords: Sub-Band Feature Encoding, Redundant Encoding, Visual Model, Discrete Wavelet Transformation,
Information Hiding
1. Introduction
With the development of information technology, people
have paid more and more attention to the information
security. Information hiding in a digital image has be-
come the focus of the information security research. For
an effective information hiding scheme, three basic re-
quirements should be satisfied: transparency, robustness
and security. The former two are in conflict with each
other. To dissolve this conflict availably, we can consider
using the masking characteristic of human visual system
(HVS) [1]. Duo to its good time-frequency localization
function is similar to the visual masking of HVS, the
DWT has been used widely in the field of information
hiding [2]. A good hiding technique should also extract
the hidden information from stego-image blindly.
In recent years, many algorithms based on HVS and
DWT had been proposed for information hiding [1–6].
And some [1–3] of them also had implemented the blind
extraction of hidden information. But those algorithms
are mostly armed at binary iconic watermark. So they are
unsuitable for hiding the text information and covert
communication. The reference [5] proposed a robust en-
cryption technique for text information. Though the
transparency, robustness and security of that algorithm
were all good, the embedded capacity was restricted due
to only one bit watermark can be hidden in two blocks,
whose sizes were settled as 8×8.
Use for reference [5] in watermark embedding, we
propose an adaptive information hiding method based on
average value relation of corresponding DWT sub-bands
of two neighboring blocks with double JND thresholds
and adjustable block size. As mentioned previously, in
order to adjust the input image for transparent water-
marks, we employ a visual model [2,5] to calculate the
different double block based JND thresholds for deter-
mining the intensity of watermarking at the different lo-
cation of image. We also give a redundant encoding
method for robustness.
This paper is organized as follows. In Section 2, we
give the JND threshold calculation equation for control-
ling the embedding intensity. Section 3 presents the in-
formation hiding algorithm and its extraction in detail.
Section 4 examines the performance of proposed algo-
rithm, and shows that the proposed scheme yields more
effective and better performance, both in terms of trans-
parency and robustness through simulation. Section 5
gives the conclusion of this paper.
2. JND Threshold Calculation
2.1 Visual Model
Under the background gray f, the human eyes relative
sensitivity to gray change

(f)=f/f, which is a non-linear
function of f, can be approximated by the equation as
follows [5]:
Information Hiding Method Based on Block DWT Sub-Band Feature Encoding
384

128
1
)256(
1
1
281
ee02.0)( ff
f
f
f
(1)
where e is the base of natural logarithm. In experiment,
we can use the gray mean of K×K image block Buv lo-
cated at (u, v) as the background gray f, i.e. f=mean (Buv).
2.2 JND Calculation
To ensure the watermark has good transparency and ro-
bustness, we can use JND to adjust the intensity of wa-
termark-embedding [2,4,5]. The image block Buv is
DWT-transformed into an approximate image and three
detail sub-band images ( sHH, HL, LH, represent
the three detail sub-bands of diagonal, horizontal and
vertical directions respectively). The JNDs of three detail
sub-bands are represented as follows:
s
uv
D
suv
s
uv FTJ (2)
where is the normalized value of
uv
T
uvuv ff ET
at
the range [a,b], while Euv is the normalized entropy of
Buv. When sHH, Fs equals 2, otherwise it is 1 [4].
3. Adaptive Information Hiding Scheme
3.1 Watermark Embedding
Let W represents a watermark sequence, Buv1 and Buv2 are
two neighboring image blocks and their DWT-trans-
formed detail sub-bands are and (simply
marked by D1 and D2, or by a universal symbol Dt, t{1,
2}). Now we can define the admissible distortion factor
of sub-band coefficients of DWT as follows:
s
uv1
Ds
uv2
D
2} {1,
) mean(t
t
t
t
D
D
λ (3)
where
is a positive number, which is an effect factor of
absolute values of detail sub-band coefficients to embed-
ding intensity. When the block size is 2×2, whatever the
value of

is, the Equation (3) is constant and can be
simplified to .
1
t
λ
We assume that

is the mean value of JNDs of two
neighboring blocks. It is represented by equation as fol-
lows:
s
uv
s
uv 21
2
1JJ 
(4)
If the crytic normalized range of [a,b] in Equation (2)
is set by two different un-overlapped ascend ranges [a0,b0]
and [a1,b1], such as [1,2] and [6,7], then we can get two
different
from Equations (2) and (4). They can be rep-
resented by
0and
1, or by a universal symbol
r,
r{0,1}.
We also assume that d is the corresponding DWT
detail sub-band coefficients difference of two neighbor-
ing blocks at same direction, and

is the adjustment
intensity matrix of detail sub-bands coefficients. They
are represented by equations as follows respectively:
)mean()mean(Sign12 DDW
k
d (5)

dt k
tkt 1
Sign
2
1
W
λWε
(6)
where Sign(·) is a sign function, it is defined as follows:

odd 1
even 1
Sign

x
x
x
Wk is the k-th element of binary watermark sequence W.
If the DWT detail sub-band features of two neighbor-
ing blocks after embedded should satisfy the relationship
with the consecutive two bits Wk and Wk+1 of watermark
sequence as shown in Table 1. We can prove that the
watermark embedding rule is as follows:



otherwise
0or and 1 if
11
t
kktt
t
d
D
WWεD
D
(7)
From the embedding rule, we know that each couple
corresponding detail sub-bands of two neighboring
blocks can be embedded 2 watermark bits. And the two
neighboring blocks have three couple corresponding de-
tail sub-bands. So, if the size of carrier image is M×N,
the information hiding capacity of this algorithm can
reach to the value of (3MN)/(K2) bits. It is double than
that of [5]. For example, if the size of carrier image is
512×512 and the block size is 2×2, then the full informa-
tion hiding capacity is 196608 bits or 24576 bytes. It is
large enough to hide information. If applied to hide short
text information into a carrier image, this method can
bring an enough redundancy to ensure its robustness.
3.2 Watermark Extraction
Being the same with watermark embedding, we should
select two neighboring blocks Buv1 and Buv2 each time
from Hilbert scanning sequence of stego-image blocks,
and a couple of their DWT detail sub-bands 1
D
ˆ and 2
D
ˆ.
d set th=(b0+a1)/2. Then, we can prove that the wa-
termark extraction rule is as follows:
An
Table 1. The relationship between watermark codes and the
DWT detail sub-band features of two neighboring blocks
Wk and Wk+1 The size relationship of corresponding
DWT detail sub-bands
00 012 )mean()mean(
DD
01 112 )mean()mean(
DD
10 012 )mean()mean(

DD
11 112 )mean()mean(

DD
Copyright © 2009 SciRes JSEA
Information Hiding Method Based on Block DWT Sub-Band Feature Encoding385
else 1,
)mean()mean( if , 012
DD
W
ˆˆ
ˆk
(8)

else 1,
)mean()mean(Sign if , 012
1

th
ˆˆ
ˆk
k
DDW
W
(9)
where is the k-th element of watermark sequence
, which is extracted from stego-image blindly. From
Equation (9), we know that the anti-interference ability
of this algorithm lies on the interval value between b0 and
a1. The larger interval value is, and the better anti-inter-
ference ability is.
k
ˆ
W
W
ˆ
3.3 Information Hiding Algorithm
Step 1: Read a text file and convert it into a bit stream W,
which is called the original watermark.
Step 2: In order to enable that the length of original
watermark W is just equal to 3 times of total blocks
number of carrier image, some zeros can be appended to
the end of it.
Step 3: For improving the robustness of watermarking,
the redundancy of embedded watermarks should be en-
sured. So the original watermark W should be extended
periodically as follows:
1,0,1,1;,0,1, ;
ex LlCrnlLnm
l
mWW
(9)
where Wex is the extended watermark, represents its
m-th element, Cr is the extended factor.
m
ex
W
Step 4: In order to improve the security of water-
marking, Wex should be scrambled randomly.
Step 5: In order to keep the relativity of two neighbor-
ing image blocks, we can scan the original carrier image
by Hilbert scanning to obtain a Hilbert scanning se-
quence.
Step 6: Select two neighboring image blocks Buv1 and
Buv2 each time from the Hilbert scanning sequence, and
embed the watermark according to the method as men-
tioned in Section 3 until all DWT detail sub-bands of all
blocks have been processed.
Step 7: After applied the inverse DWT for all water-
mark embedded blocks, we can get a stego-image I.
3.4 Information Recovering Algorithm
Step 1: Scan the stego-image I by Hilbert scanning with
the same order as that in information hiding.
Step 2: Select two neighboring image blocks Buv1 and
Buv2 each time from the Hilbert scanning sequence, and
extract the watermark according to the method as men-
tioned in Section 3 until all DWT detail sub-bands of all
blocks have been processed.
Step 3: After that, we can get a watermark sequence
ex
W
ˆ
, which involves the Cr copies of original watermark.
Step 4: If there was a scrambling when watermark was
embedded, here we should do unscrambling to
ex
W
ˆ
.
Step 5: The final watermark can be obtained from
ex
W
ˆ
as follows:
else ,0
2
if , 1
1
0
ex

Cr
n
lLn
l
Cr
ˆ
ˆW
W (10)
Step 6: The binary watermark sequence should
be converted back into a text file.
W
ˆ
4. Experimental Results
The peak signal-to-noise ratio (PSNR) is employed to
evaluate the quality of stego-image, meanwhile the bit
error rate (BER) is employed to evaluate the quality of
recovered secret information.
In the experiment, the proposed algorithm was evalu-
ated on the gray image “Lena” (512×512×8). The block
size K can be 8, 4 or 2. In order to ensure the algorithm’s
anti-interference ability is good, [a0,b0] should be set
smaller and [a1,b1] should be set bigger. So we set
[a0,b0]=[1,2]. Figure 1 is the relationship between PSNR
of stego-image and the effect factor
at different block
sizes. Figure 2 is the relationship between BER of the
recovered secret bits and the effect factor

Figure 3 is
the relationship between PSNR and the JND normalized
range [a1,b1] at full capacity embedding.
Figure 1. Relationship between PSNR and

Figure 2. Relationship between BER and
Copyright © 2009 SciRes JSEA
Information Hiding Method Based on Block DWT Sub-Band Feature Encoding
386
Figure 3. Relationship between PSNR and a1 ([a0,b0]=[1,2],
b1=a1+1)
As shown in Figure 1, the PSNRs of stego-image are
constant when K =2, but when K >2, the bigger
is, the
higher PSNRs are. But as known in Section 3, the bigger
will bring down on the embedding intensity, it leads
that the performance of recovering algorithm will be
worse, especially when K =8, the BERs will be higher.
As shown in Figure 2, we know that BERs are always
zeros when

1, and whatever the K is. So we set the
=1 in the following experiments. As shown in Figure 3,
the PSNRs are almost in inverse proportion to [a1,b1].
The smaller [a1,b1] is, the higher PSNRs are, and the bet-
ter imperceptibility of hidden information is. So that an
appropriate [a1,b1] is better. Here, we set [a1,b1]=[6,7].
As shown in Table 2, the algorithm’s full hiding capaci-
ties in various block sizes and their performances are
good enough to hide information.
Figure 4(a) is the original images of “Lena”. Figure
4(b) and 4(c) are the stego-image hidden with 768 bytes
text when K=8 and the difference images between the
original image and stego-image on the condition of mag-
nifying 30 times. Figure 4(d) and 4(e) are the results
when K=4. Figure 4(f) and 4(g) are the results when K=2.
As shown in Table 2 and Figure 4, considering the visual
perception in the algorithm, the hidden information in
“Lena” gray image is invisible though the PSNR is a
little low. The algorithm can fully recover the hidden
information from the stego-image.
(a) (b) (c)
(d) (e) (f) (g)
Figure 4. The information hidden results of “Lena”
Table 2. The full capacities in differenct block sizes and
their PSNRs and BERs
Block size
Capacity and Per-
formances 8×8(K=8) 4×4(K=4) 2×2(K=2)
Hiding capacity (bit) 12288 49152 196608
Hiding capacity (byte)1536 6144 24576
PSNR (dB) 37.21 36.44 34.95
BER (%) 0 0 0
Table 3. The zero BER attack defense tests of short text
information hiding
Attack items Performance under zero BER
Cropping in central region 258×258
Brightness enhancement 69%
Contrast enhancement 33%
In the experiment, we also did some attacking tests to
the algorithm on the condition of K=4 and 768 bytes text
with 8 redundant copies. As shown in Table 3, we found
that the algorithm was robust to central region cropping,
brightness enhancement and contrast enhancement.
5. Conclusions
This paper presented a new scheme of information hiding
in gray images based on DWT for long text information
hiding or covert communication. In our approach, the
comparability of corresponding DWT detail sub-bands of
two neighboring image blocks was considered, and in
order to improve the transparency of information hiding,
the visual model was also used for calculating the double
block based JNDs to determine the embedding intensity
at different locations of image. The block DWT sub-band
feature encoding technique increased the embedding ca-
pacity double than that of [5]. The adjustable block size
gave facilities for various applications. If you request the
algorithm to have a better transparency, you should select
bigger block size. Or if you request that it has a larger
embedding capacity, you should select smaller block size.
In addition, in order to improve the algorithm’s robust-
ness and ability of defense some general image process-
ing attacks, such as cropping, brightness enhancement
and contrast enhancement, the redundant encoding was
given for increasing the embedded copies of watermark.
The experiment results demonstrate that the proposed
algorithm yields the acceptable performance for trans-
parency and robustness, and also increases the embed-
ding capacity for information hiding.
6. Acknowledgements
This research project was supported by the Technological
Innovation Foundation of Shanghai Municipal Education
Commission under Grant No. 09YZ456 and the Key
Copyright © 2009 SciRes JSEA
Information Hiding Method Based on Block DWT Sub-Band Feature Encoding
Copyright © 2009 SciRes JSEA
387
Disciplines of Shanghai Municipal Education Commis-
sion under Grant No. J51801.
REFERENCES
[1] M. Barni, F. Bartolini, and A. Piva, “Improved wavelet-
based watermarking through pixel-wise masking,” IEEE
Transactions on Image Processing, Vol. 10, No. 5, pp.
783–791, 2001.
[2] H. F. Yang and X. W. Chen, “A robust image-adaptive
public watermarking technique in wavelet domain,”
Journal of Software, Vol. 14, No. 9, pp. 1652–1660,
2003.
[3] J. G. Cao, J. E. Fowler, and N. H. Younan, “An image-
adaptive watermark based on a redundant wavelet trans-
form,” In: Pitas I, ed. Proceedings of the IEEE Interna-
tional Conference on Image Processing, Thessaloniki, pp.
277–280, 2001.
[4] Z. M. Wang, Y. J. Zhang, and J. H. Wu, “A wavelet do-
main watermarking technique based on human visual
system,” Journal of Nanchang University (Natural Sci-
ence), Vol. 29, No. 4, pp. 400–403, 2005.
[5] Q. D. Sun, W. X. Ma, W. Y. Yan, and H. Dai, “Text en-
cryption technique based on robust image watermarking,”
Journal of Image and Graphics, Vol. 13, No. 10, pp.
1942–1946, 2008.
[6] L. Y. Wu and F. Yang, “An improved digital watermark-
ing algorithm based on DWT,” Control and Automation,
Vol. 23, No. 6(3), pp. 46–47, 59, 2007.