Digital Watermarking

1
Digital Watermarking

• Annual Progress Seminar III
• Jayalakshmi. M
• Roll NO. 03407003
• Guided by
• Prof. S. N. Merchant

2
Previous work (Annual Progress Seminar 12)

• Significant pixels in wavelet domain for robust
  watermarking
• Quantization of significant pixels with respect
  to HVS model
• Watermarking in Contourlet Domain - Non-blind
  methods
• Found suitable for images like maps which contain
  lot of lines, texts and curves
• Error Concealment using Watermarking (DCT and DWT
  domain)

3
Overview

• Wavelet Domain Watermarking
• - Significant pixels with quantization given
  by HVS model with a single copy of the watermark
• - Chair -Varshney rule for improved
  detection when multiple copies are inserted
• Watermarking in Contourlet Domain
• -Non-blind techniques
• -Application mainly in geographical maps
• -Blind watermarking in contourlet domain -
  improved detection after filtering
• Error Concealment by Watermarking
  - Improvement in DCT domain
• - Wavelet domain

4
1. Significant Pixel Watermarking in Wavelet
Domain

• Significance factor for each pixel
• Quantization of highest significant pixels with
  respect to HVS model -gives maximum allowable
  quantization for every pixel at any resolution
• No pixel in the original image carries more than
  one watermark bit

5
Result 16x16 binary watermark

Watermarked
Original
6
Results of wavelet domain significant pixel
watermarking

• PSNR for invisibility
• Correlation coefficient after attacks like
• -mean filtering, Gaussian noise addition,
• salt pepper noise addition with median
  filtering, quantization ,JPEG compression
  cropping
• Cases considered- 2,3 4 level decomposed image
  with and without significant pixels.
• Performance evaluated with 16x16, 32x32 binary
  watermark and pseudorandom watermark

7
Results of wavelet domain significant pixel
watermarking

• Retrieval with significant pixels better than
  the highest absolute pixels
• Significance factor becomes meaningful as moved
  towards higher decomposition levels
• The distortion to the image should be kept to
  minimum- 3rd decomposed level was found a good
  choice for embedding
• Since a single copy of watermark is embedded,
  cropping any part of visual importance will not
  result in detection

8
Chair- Varshney rule with multiple copies of
watermark

• Multiple copies of watermark (K)

9
Pd and Pm for each bit

• Both Pd and Pm are fixed for all the bits in
  every copy of the watermark
• Then, these values are increased or decreased
  depending on whether a correct detection has
  taken place or not
• This could give rise to false positives

10
Results with CV rule

• Correlation coefficients after JPEG compression

11
Results with CV rule

• JPEG compressed image (Q3) and retrieved
  watermarks

CV rule
Majority rule
12
Results with CV rule

• Correlation coefficients after mean filter and
  mean filter with JPEG compression (Q10)

13
Results with CV rule

• Mean filtered and compressed image (Q10) and
  retrieved watermarks

CV rule
Majority rule
14
Results with CV rule

• Correlation coefficients after histogram
  equalization

15
Results with CV rule

• Histogram equalized and retrieved watermarks

CV rule
Majority rule
16
Results with CV rule

• Correlation coefficients after Gaussian blur

17
Results with CV rule

• Gaussian blurred image and retrieved watermarks

CV rule
Majority rule
18
Results with CV rule

• Correlation coefficients after pixelization

19
Results with CV rule

• Pixelized image and retrieved watermarks

CV rule
Majority rule
20
False Positives with Random Watermarks

• 200 random samples of watermark

21
2. Contourlet Domain Methods

• Sparse representation of 2-D piecewise smooth
  signals
• 2-D wavelets formed by tensor product of 1-D
  wavelets are good at catching discontinuities at
  edge points
• Wavelets do not see smoothness along the contours
• Contourlets make use of Pyramidal Directional
  Filter Bank (PDFB)
• PDFBLPDFB

22
Directional decomposition used in proposed methods
23
Method 1- High absolute coefficients

• Directional decomposition doubles at every scale
• Highest absolute coefficients in D is watermarked
• DDa m

24
Method 2 (Significance factor)
25
Generalized Neighborhood
26
Watermarked Images
Wavelet based
27
Watermarked Images
Highest absolute pixel (Multiple6)
Significant pixel (Multiple3)
28
Watermarked Images
Gen. neighborhood
Highest abs. (curve scale)
29
Watermarked Images

DCT
Hadamard
30
Retrieved watermarks after mean filtering
Wavelet (wave4)
Highest absolute pixels(multiple6)
Significant pixels(multiple3)
Gen. neighborhood(gen1)
Curve scaling relation(gen2)
DCT Hadamard
31
Conclusion Non-blind techniques in contourlet
domain

• Contourlet based algorithms performed better than
  wavelet, DCT Hadamard transform domain methods
  in images like maps
• Significant pixels in contourlet domain were
  found the best choice for watermark embedding for
  images containing lot of curves and texts
• l

32
Blind Watermarking in Contourlet Domain

• Contourlet decomposition of images
• Binary watermark embedded using spread spectrum
  method
• Additive embedding is performed
• Robustness verification is done using Stir Mark
  attack
• Recovered watermark is evaluated by finding its
  correlation with original logo
• The visual similarity is improved by median
  filtering of the retrieved logo

33
Additive Embedding

• Y Set of original pixels
• Y Corresponding watermarked pixels
• PPseudorandom sequence generated using a key
• w watermark bit

Original image
Water- marked image
Pseudo random sequence
key
34
Embedding- Results

• Original image Logo
  Watermarked image

35
Embedding- Results

• PSNR with embedding

36
Retrieved logo

• Retrieved logo with authorized key
• Retrieved logo with unauthorized key
• Retrieved logo after post processing

37
Correlation coefficients with a

• Retrieval results

38
Correlation coefficients after attacks
39
Mean filtering (256x256, a0.225)

• Mean filtered image Retrieved logo
  After post-processing

g0.7077 g 0.8704
40
Quantization to multiples of 50 (256x256,
a0.225)

• Quantized image Retrieved logo
  After post-processing

g0.7914 g 0.9301
41
Quantization to multiples of 100 (256x256,
a0.225)

• Quantized image Retrieved
  logo After post-processing

g0.6808 g 0.8332
42
JPEG Compression (Q40) (256x256, a0.225)

• Quantized image Retrieved
  logo After post-processing

g0.6939 g 0.8419
43
Watermarked Image with 128x384 Coefficients
(a0.275)

• Watermarked Retrieved logo
  After post-processing

g0.8021 g0.9438
g0.8322 g0.9735
44
Mean filtering with 128x384 Coefficients
Watermarked ( a0.275)

• Mean filtered image Retrieved
  logo After post-processing

g0.7280 g 0.8984
45
Mean filtering with 128x384 Wavelet Coefficients
Watermarked (a0.275)

• Mean filtered image Retrieved logo
  After post-processing

g0.6682 g 0.8341
46
Conclusion blind watermarking in contourlet
domain

• Blind watermarking in contourlet decomposed
  images is performed using spread spectrum
  technique
• Correlation based detection is improved using
  post processing after detecting all the bits
• Shows good robustness against attacks
• Directional bands can be effectively used to
  achieve robustness against geometrical attacks

47
3. Error Concealment Using Watermarking

• Watermark derived from image itself

48
Algorithm

• Approximate band embedded in LH1 HL1
• N x N approximate band represented by 8N2 bits
• LH1 HL1 together gives 32N2 bits
• 4 copies of watermark hidden with a shift in the
  bit stream
• Sign of every pixel remains unchanged after
  embedding
• Completely blind algorithm
• Results compared with BNM technique

49
Results of EC in DWT domain -8X80 block errors

• 20 line errors

PSNR18.94 PSNR
33.4 PSNR 32.58
Error image Proposed
method BNM
50
Error Concealment in DCT Domain

• Existing algorithm embeds ROI into ROB in spatial
  domain
• ROI embedded into the compressed JPEG bit streams
  
• Good resistance to compression
• 32x32 ROI into the JPEG quantized pixels through
  LSB substitution
• Only 7th to 16th pixels in the zigzag scanned
  pattern of each 8x8 block in textured regions are
  selected for hiding the watermark bits.

51
Results with medical image

• ROI defined on the fracture

52
Results with only ROI embedded(after compression)

• ROI Spatial
  Proposed

53
Results with only ROI embedded(after compression)

• ROI Spatial
  Proposed

54
Results with only ROI embedded(after compression)

• PSNR of error concealed ROI

55
Error Concealment through watermarking

• DWT domain algorithm has two levels of security
  using a scale factor and a key for pseudorandom
  generation
• Also efficiently conceals error in any part of
  the image
• DCT based method needs lot of texture region in
  the ROB for hiding the watermark and ROI defined
  is small compared to the size of the image

56
Future work

• Improvement in error concealment and PAPR with
  redundancy in MC-CDMA systems
• Effect of super resolution on watermarking

57
Thank you