Bayesian Structural Content Abstraction (BSCA) for Image Authentication

1
Bayesian Structural Content Abstraction (BSCA)
for Image Authentication

• Wei FENG (w.feng_at_student.cityu.edu.hk)
• Zhi-Qiang LIU
• School of Creative Media
• City University of Hong Kong
• Aug 27, 2004

2
Outline

• Contributions
• Backgrounds
• BSCA Scheme
• BSCA Modeling
• Spurious Region merging
• Extending BSCA
• Experimental Results
• Conclusion Future Work

3
Contributions

• A new model for parameterized image
  authentication which satisfies
• robustness to NCOs
• sensitivity to COs
• the user-defined NCO/CO division
• An extensible image authentication scheme

4
Backgrounds

• Image authentication (IA) is to protect the
  integrity of the content. An ideal IA approach
  should be able to tolerate content preserving
  operations (i.e., NCO, e.g. compression
  rotation etc.) robustly, while detecting content
  altering" (i.e., CO, e.g. object removing
  replacement) modifications sensitively.

A fundamental requirement of IA
5
Backgrounds constraints
A good IA scheme
NCO/CO division is very application dependent
6
Backgrounds former works

• Fragile/semi-fragile watermarking
• M. Wu B. Liu, ICIP, 1998
• W. Zeng B. Liu, IEEE Trans. IP, 1999
• Wong, ICIP 1998, IEEE Trans. IP, 2003
• Digital signature
• M. Schneider S.F. Chang, ICIP, 1996
• C.Y. Lin S.F. Chang, 1998
• C.S. Lu H.Y. Liao, IEEE Trans. MM, 2003

7
Backgrounds our solution

• Compared with DS, F SF water-marking do not
  need an added secure channel. But the tamper
  detection capability is limited.
• Both DS and watermarking are based on the
  selection of content represen-tation.

8
Backgrounds a general IA scheme
9
BSCA Scheme modeling

• Generally, any feature extracted from the image
  (color, texture, edge, shape and structure etc.)
  can be viewed as a representation of its content.
  
• The diverse features depict very different
  aspects of the image, thus have non-uniform
  representability to the content. If we use
  granularity to describe a feature's descriptive
  fineness as a content representation, we can see
  that color and texture are fine and local content
  representations, while structure is a global one.

10
BSCA Scheme modeling

• Lu Liao found that large granularity features,
  such as structure etc., are more reliable than
  small granularity features in IA. They designed
  an incidental distortion resistant scheme called
  SDS, based on the inter-scale relations of
  wavelet coefficients. Although they considered
  carefully the tradeoff between robustness and
  fragility, SDS still cannot perform robustly to a
  wide range of NCOs, such as low quality
  compression, global color adjustment and
  geometric distortions.
• This is because that the structure in SDS was
  derived directly from pixels.
• We define our content representation based on
  regions.

11
BSCA Scheme modeling

• We model the image I as the sum of its underlying
  content C and an observation noise process N. We
  think the content should include at least two
  kinds of information homogeneous region index
  map L that represents the structure, and
  principle color p(L) of each region that
  corresponds to the images small-granularity
  content.

12
BSCA Scheme modeling

• Note that p(L) is a deterministic function which
  maps each homogeneous region to its dominant
  observation property univocally.
• We can also assume that N is a Gaussian noise
  with zero-mean.
• From (1) and (2), we can draw a statistical
  optimal estimation of the content within the
  Bayesian framework.

13
BSCA Scheme modeling

• Furthermore, we can firstly get the optimal
  region map.
• We model the region map L as a MRF with a
  second-order neighborhood system. Thus, the
  optimal L can be gotten by DElias tree MRF
  segmentation algorithm.

14
BSCA Scheme modeling

• Actually, the estimation of an statistical
  optimal region map L is a Bayesian segmentation
  of the image.
• This is rational because
• The spatial distribution of homogeneous regions
  reflects humans global perception of the image.
• It will not change dramatically under a wide
  range of NCOs, but changes apparently under COs,
  such as object adding and removing etc.

15
BSCA Scheme modeling

• The formulation can be further improved by
  introducing an explicit combination of NCOs.

16
BSCA Scheme modeling
The concrete configuration of clique we used.
17
BSCA Scheme spurious region merging

• We find many small fragments may exist in L.
  This is due to the existence of spurious regions
  in N and errors in the estimation to the noise
  process which should be removed.

18
BSCA Scheme spurious region merging

• Regions reordering by their sizes
• Region size thresholding
• Spurious regions merging according to local
  continuity and scale

19
BSCA Scheme extending

• Till now, a BSCA is a non-directed graph.
• The graph can be naturally extended by
  integrating other favorable local features
  (regarding them as attributes of the associated
  vertex).

20
Experimental Results

• We already know that reference generation and
  transmission is the most important part in the
  IA. And, a desirable reference should be
  consistent with the content of the protected
  image. Thus, as an image content representation
  scheme, BSCA can be used to generate the standard
  reference feature.
• Here, we only present a simple but efficient
  image authentication scheme which achieves both
  robustness to NCOs and sensitivity to COs.

21
Experimental Results

• Experiments design

22
Experimental Results
NCOs COs
BSCA-Watermark 85 90
23
Experimental Results
24
Conclusion Future Work

• We proposes a hierarchical image content
  representation scheme BSCA for IA. Because it is
  based on the fundamental criterion, it is robust
  against NCOs with little sacrifice of sensitivity
  to COs.
• Moreover, as a general content representation
  scheme, BSCA can also be used in many other image
  processing and computer vision problems, such as
  compression and CBR etc.
• We may explore the possibility of using BSCA to
  represent individual objects in a scene and their
  transitions in a video sequence.
• How to construct a rotation invariant distance
  metric of BSCA?

25

• Thank you very much!