Transcript 下載/瀏覽Download

IEEE Transactions on Circuits and Systems for Video Technology, vol. 19, no. 6,
June 2009
Adviser： Chih-Hung Lin and
Ten-Chuan Hsiao
Speaker：Chia-Wei Chang
Date:2010/06/11
1
Authors
• Wei-Liang Tai is with the Department of Computer Science
and Information, National Chung Cheng University, Chiayi 621,
Taiwan
• Chia-Ming Yeh and Chin-Chen Chang, Fellow, IEEE are with
Department of Information Engineering and Computer
Science, Feng Chia University, Taichung 407, Taiwan
•
•
•
•
Manuscript received May 14, 2008.
Revised August 25, 2008.
First version published March 16, 2009.
Current version published June 19, 2009.
2
Outline
• Introductions
• Proposed scheme
– Related Work
– Our proposed
• Experimental Results
• Conclusions
3
Introductions
=
embedding
4
Introductions
Distortion
Reversible
image
such as military images, medical
images, or artwork preservation.
5
Introductions
• [7]M. U. Celik, et. Al. proposed “Lossless
generalized-LSB data embedding”
• [8]J. Tian, proposed “Reversible data
embedding using a difference expansion”
• [9]A. M. Alattar, proposed “Reversible
watermark using the difference expansion of a
generalized integer transform,”
6
Introductions
• [18]Z. Ni, et al. proposed “Reversible data
hiding”
• [19]M. Fallahpour and M. H. Sedaaghi,
proposed “High capacity lossless data hiding
based on histogram modification”
• [20]S. K. Lee et. al., proposed “Reversible image
authentication based on watermarking”
7
Related Work
Fig. 1. Distribution of differences.
8
Related Work
• Histogram Modification on Pixel Differences
0  i  N  1, xi  Z , xi  0, 255
(1)di:pixel difference xi-1 and xi are pixels
if i  0,
 xi ,
di  
| xi 1  xi |, otherwise.
(2)Determine the peak point P from the pixel differences.
(3) Scan the whole image in the same inverse s-order as in Step 1. If di > P, shift xi by 1 unit
if i  0 or di  P,
 xi ,

yi  xi  1, if di  P and xi  xi 1 ,
 x  1, if d  P and x  x ,
i
i
i 1
 i
yi is the watermarked value of pixel i .
9
Related Work
(4)if di = P, modify xi according to the message bit
 xi  b, if di  P and xi  xi 1
yi 
 xi  b, if di  P and xi  xi 1
where b is a message bit to be embedded.
10
Related Work
Extract
The message bit b can be extracted by
0, if | yi  xi 1 | P
b
1, if | yi  xi 1 | P  1
where xi−1 denotes the restored value of yi−1. The original pixel value of xi can be restored
by
 yi  1, if | yi  xi 1 | P and yi  xi 1

xi   yi  1, if | yi  xi 1 | P and yi  xi 1
y ,
otherwise.
 i
11
Our proposed
B. Binary Tree Structure
Fig. 2. Auxiliary binary tree for the proposed scheme.
12
Our proposed
Fig. 3. Histogram shifting. (a) Original histogram. (b) Histogram shifting.
13
Our proposed
• Embedding Process
0  i  N  1, xi  Z , xi  0, 255
(1)Determine the level L of the binary tree.
(2)Shift the histogram from both sides by 2L units.
(3)Calculate the pixel difference di between pixels xi−1 and xi .
(4)Scan the whole image in the same inverse s-order. If di ≥ 2L, shift xi by 2L units
if i  0
 xi ,

yi   xi  2 L , if di  2 L and xi  xi 1 where yi is the watermarked value of pixel i .
 x  2 L , if d  2 L and x  x
i
i
i 1
 i
(5)If di < 2L , modify xi according to the message bit
 xi  (di  b), if xi  xi 1
yi  
 xi  (di  b), if xi  xi 1
14
Our proposed
• Extraction Process
0  i  N  1, yi  Z , yi  0, 255
(1)Scan the watermarked image W in an inverse s-order.
(2)If |yi − xi−1| < 2L+1, extract message bit b by
0, if | yi  xi 1 | is even
b
 1, if | yi  xi 1 | is odd
where xi−1 denotes the restored value of yi−1.
(3)Restore the original value of host pixel xi by

 | yi  xi 1 | 
y

, if | yi  xi 1 | 2 L 1 and yi  xi 1
 i 

2




 | yi  xi 1 | 
L 1
,
if
|
y

x
|

2
and yi  xi 1
 yi  
i
i

1

xi  
2


 y  2L ,
if | yi  xi 1 | 2L 1 and yi  xi 1
i

L
if | yi  xi 1 | 2L 1 and yi  xi 1
 yi  2 ,
y ,
otherwise
 i
15
Experimental Results
• six commonly used grayscale images sized 512×512, “Lena,” “Mandrill,”
“Boat,” “Jet,” “Pepper,” and “GoldHill.”
Table 1:Hiding capacity and distortion for test images with L=0
16
Experimental Results
Table 2:Pure payload for test images with tree level L
17
Experimental Results
Fig. 4. PSNR versus pure payload size for test images.
18
Experimental Results
Fig. 5. Watermarked “Lena” and “Mandrill”
images.
(a) 48.32 dB embedded with 0.0854 b/pixel.
(b) 48.21 dB embedded with 0.0375 b/pixel.
(c) 30.75 dB embedded with 0.9049 b/pixel.
(d) 23.49 dB embedded with 0.8092 b/pixel.
19
Experimental Results
• Comparison With Other Recent Schemes
Fig. 6. Performance comparison for the “Lena” image with existing reversible schemes
[7]–[9], [18]–[20].
20
Conclusions
• We have presented an efficient extension of the
histogram modification technique by considering
the differences between adjacent pixels rather than
simple pixel value.
• We introduced a binary tree that predetermines
the multiple peak points used to embed messages.
• This enables us to achieve large hiding capacity
while keeping embedding distortion low.
21