Multiresolution SceneBased Video Watermarking Using Perceptual Models

1
Multi-resolution Scene-Based Video Watermarking
Using Perceptual Models

• by Mitchell D. Swanson, Bin Zhu, and Ahmed H.
  Tewfik
• from IEEE Journal on Selected Areas in
  Communications

Presenter Wei-Cheng Lin Project leader
B.H. Advisor Prof. Ja-Ling Wu
2
Outline

• Introduction
• Author Representation v.s deadlock
• Visual Masking
• Temporal Wavelet Transform
• Watermark Design
• Watermark Detection
• Visual and Robustness Results
• Conclusion

3
Introduction (1/2)

• Digital watermarking has been proposed as a means
  to identify the owner and distribution path of
  digital data.
• Some issues when applying watermark
• data redundancy between frames
• identical watermark v.s statistical invisibility

4
Introduction (2/2)

• Major contributions of this paper
• A Perceptual-Based Video Watermarking Procedure
• A Scene-Based Multi-scale Watermark
  Representation
• An Author Representation Which Solves the
  Deadlock Problem
• A Dual Watermark

5
Author Representation and Deadlock (1/2)

• The Deadlock Problem and Rightful Ownership ( See
  Figure )
• Dual watermarks
• One watermarking procedure requires the original
  data set for watermark detection while the other
  doesnt.
• The second watermark is meant to protect the
  video that the pirate claims to be his original.

6
Author Representation and Deadlock (2/2)

• Using a pseudo random sequence to represent the
  author
• use two keys and a suitable generator, say RSA,
  Rabin, Blum/Micali , etc.
• one key is author dependent the other is
  computed from the video signal using a one-way
  hash function, such as RSA, MD4.
• due to the computationally infeasible and signal
  dependent key, the pirate is unable to fabricate
  a counterfeit original that generate the desired
  watermark!!

7
Visual Masking (1/4)

• Frequency Masking
• compute the contrast threshold at certain
  frequency.

8
Visual Making (2/4)

• to find the contrast threshold at a frequency,
  first use DCT and then sum rule below.
• if the contrast error at f is less than c(f), the
  model predict the error is invisible.

9
Visual Masking (3/4)

• Spatial Masking
• based on the threshold vision model proposed by
  Girod, it accurately predicts the making effects
  near edges and in uniform background.
• first compute the contrast saturation

10
Visual Masking (4/4)

• compute the luminance on the retina
• then obtain the tolerable error level for the
  pixel (x,y) by following formula
• w1 and w2 are based on Girods model. The change
  to pixel less than ds(x,y) introduce no
  perceptible distortion.

11
Temporal Wavelet Transform (1/2)

• We employ the wavelet transform along the
  temporal axis of the video sequence.
• After the wavelet transform, we can get the
  static and dynamic components (i.e. lowpass
  frames and highpass frames) of the original
  signal.
• See the figure.

12
Temporal Wavelet Transform (2/2)

• The watermark embedded in the static components (
  lowpass frames ) exists throughout the entire
  video scene.
• The watermark embedded in the dynamic components
  ( highpass frames ) are highly localized in time
  and change rapidly from frame to frame.

13
Watermark Design

• First break the video sequence into scenes.
• Diagram of watermarking procedure

14
Watermark Detection (1/3)

• Detection I - Watermark Detection with Index
  knowledge

15
Watermark Detection (2/3)

• scalar similarity
• the overall similarity is computed as the mean
  of Sk for all k. and compared with the threshold
  to determine to presence.
• If the length of test video is the same as the
  original, perform the test in wavelet domain.

16
Watermark Detection (3/3)

• Detection II - Watermark Detection without Index
  Knowledge
• The hypothesis test is formed by removing the low
  temporal wavelet frame from the test frame and
  the computing the similarity with the watermark
  for it.

17
Visual and Robustness Results (1/2)

• Visual Results ( See the figure and table )
• Robustness Results
• detect the watermark when one exists and reject a
  video when none exists.
• for a given distortion, the overall performance
  may be ascertained by the relative difference
  between the similarity when one present and none
  present.
• use the first 32 frames for test of both
  detection methods.

18
Visual and Robustness Results (2/2)

• Attacks
• Colored Noise
• Coding ( MPEG CR 1001 )
• Multiple Watermarks
• Frame Averaging
• Printing and Scanning

19
Conclusion (1/2)

• The watermarking technique directly exploits the
  masking phenomena of HVS to guarantee the
  invisibility.
• The pseudo random sequence is generated by two
  keys one is author dependent and the other is
  signal dependent.

20
Conclusion (2/2)

• Wavelet-based watermark exists at multiple scales
  in the video.
• The watermark can be detected with and without
  the index knowledge in the distortions.