Blind image data hiding based on self reference

Source : Pattern Recognition Letters,

Vol. 25, Aug. 2004, pp. 1681-1689

Authors: Yulin Wang and Alan Pearmain

Speaker: Nan-I Wu (吳男益 )

Date: 2004/11/25(四 )

Introduction• Blind watermark technique: recover the

watermark without using original host data.

• The proposed method base on relative modulation of the pixel value/DCT coefficient value by referring to its estimated one.

• One is based on the estimation of pixel luminance in spatial domain

• Another based on the estimation of block DCT AC coefficients.

The proposed scheme 1 (Spatial Domain)

• For a nature image, the luminance value of one pixel normally has a relation with its neighbor.

100 100 100

100 95 100

100 100 100

LReal

The mean luminance value: Lmean

The embedding algorithm

• Embed an one-bit watermark into 3 x 3 sub-block

• Embedded bit is ‘1’

LReal ≥ Lmean + ∆1

• Embedded bit is ‘0’

LReal < Lmean - ∆2

As experienced values, ∆1 and ∆2 are selected as 5-10% of

LReal

The extracting algorithm

If LReal ≥ Lmean

extract bit ‘1’

If LReal < Lmean

extract bit ‘0’

The proposed scheme 2(Frequency Domain)

RGB to Y Cb Cr

Y = 0.299 R+ 0.587 G+ 0.114 B

R=100 G=80 B=120

90.54 = 0.299*100+ 0.587*80+ 0.114*120

DCT

The proposed scheme 2(Frequency Domain)

The embedding algorithm

Block1

DC1

Block2

DC2

Block3

DC3

Block4

DC4

Block5

DC5

Block6

DC6

Block7

DC7

Block8

DC8

Block9

DC9

0 1 2

0

1 2

AC(0,1)=1.13884 x (DC4-DC6)/8

Select every nine 8 x 8

blocks as one group,

in which 5 watermark bits

can be embedded

The embedding algorithm

Block1

DC1

Block2

DC2

Block3

DC3

Block4

DC4

Block5

DC5

Block6

DC6

Block7

DC7

Block8

DC8

Block9

DC9

0 1 2

0

1 2

AC(0,1)=1.13884 x (DC4-DC6)/8

Set ACi ≥ AC’i + ∆ to embed bit ‘1’

Set ACi ≤ AC’i - ∆ to embed bit ‘0’

∆ can be chosen as 5-15% of

the original ACi value

The embedding algorithm

Block1

DC1

Block2

DC2

Block3

DC3

Block4

DC4

Block5

DC5

Block6

DC6

Block7

DC7

Block8

DC8

Block9

DC9

0 1 2

0

1 2

AC(1,0)=1.13884 x (DC2-DC8)/8

Set ACi ≥ AC’i + ∆ to embed bit ‘1’

Set ACi ≤ AC’i - ∆ to embed bit ‘0’

The embedding algorithm

Block1

DC1

Block2

DC2

Block3

DC3

Block4

DC4

Block5

DC5

Block6

DC6

Block7

DC7

Block8

DC8

Block9

DC9

0 1 2

0

1 2

AC(0,2)=0.27881 x (DC4+DC6 – 2 x DC5)/8

Set ACi ≥ AC’i + ∆ to embed bit ‘1’

Set ACi ≤ AC’i - ∆ to embed bit ‘0’

The embedding algorithm

Block1

DC1

Block2

DC2

Block3

DC3

Block4

DC4

Block5

DC5

Block6

DC6

Block7

DC7

Block8

DC8

Block9

DC9

0 1 2

0

1 2

AC(2,0)=0.27881 x (DC2+DC8 – 2 x DC5)/8

Set ACi ≥ AC’i + ∆ to embed bit ‘1’

Set ACi ≤ AC’i - ∆ to embed bit ‘0’

The embedding algorithm

Block1

DC1

Block2

DC2

Block3

DC3

Block4

DC4

Block5

DC5

Block6

DC6

Block7

DC7

Block8

DC8

Block9

DC9

0 1 2

0

1 2

AC(1,1)=0.16213 x (DC1+DC9 – DC3 - DC7)/8

Set ACi ≥ AC’i + ∆ to embed bit ‘1’

Set ACi ≤ AC’i - ∆ to embed bit ‘0’

The extracting algorithm

• The original image is not require for the watermark bit detection, only the comparison of the relative value between ACi and its estimated value AC’i is needed.

• If ACi > AC’i, then the extracted bit is ‘1’,

• If ACi < AC’i, then the extracted bit is ‘0’.

• Of course, if ACi=AC’i, there is uncertainly about whether the watermark bit is a ‘1’ or ‘0’.

Experimental Results

Experimental Results

Experimental Results

Conclusion• This paper presents a kind of estimation

based blind image watermarking technique.

• Our DCT technique achieves the optimal trade-off among imperceptibility capacity and robustness.