Geometric DistortionResilient Image Hashing Scheme and Its Applications on Copy Detection and Authen

1
Geometric Distortion-Resilient Image Hashing
Scheme and Its Applications on Copy Detection and
Authentication

• Source Multimedia Systems, Vol.11, No.2,
  December 2005, pp.159 -173
• SCI 69(54/78, Impact factor 0.528)
• Authors Chun-Shien Lu and Chao-Yong Hsu
• Speaker Shu-Wei Guo(???)
• Date 2006/02/17

2
Outline

• Introduction
• Problem Statement
• Proposed Method
• Analysis
• Experiments
• Conclusions

3
Introduction

• Image hashing (vs. traditional watermarking)

Image Hashing
0110010001
4
Introduction

• Image hashing (vs. traditional watermarking)
• Non-invasive embedding
• Condensed representation
• Work together with a feature database
• Resist either malicious or incidental attacks

5
Problem Statement
I original image X the set of images modified
from I and perceptually similar to I Y the set
of image modified from I but dissimilar to I Z
the set of images that are irrelevant to I
6
Problem Statement
Traditional cryptographic hashing
although x may be very similar to I
Image hashing function
, where d(.) is Hamming distance
Two conditions must be satisfied
1. If two images are similar, their corresponding
hashes much be highly correlated 2.
Collision-free
7
Proposed Method
8
Proposed Method (cont.)
Step 1 Triangulation
Harris Delaunay Algorithms
9
Proposed Method (cont.)
Step 2 Normalization
10
Proposed Method (cont.)
Step 3 Hash code generation
11
Proposed Method (cont.)
Step 3 Hash code generation
AC0(1), AC1(1), AC2(1), AC3(1), AC4(1), AC5(1)
12 32 4 51 22
9
X
X
X
51, 32, 22
Hash code
0 1 0 1 1
0
Collision free
12
Analysis
Samples 20,000
Robustness and Discrimination
13
Similarity Measurement
Im and In are considered to be similar if at
least N mesh-pairs are matched
N3 in this paper
M1 hash code 00100101
M2 hash code 11000001
T0.25
unmatched
BER(M1, M2)4/8 gt T
(Bit Error Rate)
14
Experiments - Copy Detection
Baboon
SPA(Signal processing attack), GLGT(general
linear geometric transform), CAR(change of the
aspect ratio), LR(line removal),
RC(rotationcropping), RRS(rotationre-scaling),
RB(random bending)
15
Copy Detection
16
Failure Image Query
Rotatedcropped
noisy
cropped
Convolution filtered
17
Content Authentication
Tampering
18
Content Authentication
TamperingRotation
19
Content Authentication
Tampering
20
Content Authentication
TamperingJPEG Compression(QF 30)
21
Conclusions

• Mesh-based robust hash generation
• Hash database construction for error resilient
  and fast searching

22
Harris Detector Mathematics
Harris Method
23
Harris Detector Mathematics
flat regionno change in all directions
edgeno change along the edge direction
cornersignificant change in all directions
24
Harris Detector Mathematics
Change of intensity for the shift u,v
25
Harris Detector Mathematics
For small shifts u,v we have a bilinear
approximation
where M is a 2?2 matrix computed from image
derivatives
26
Harris Detector Mathematics
Intensity change in shifting window eigenvalue
analysis
?1, ?2 eigenvalues of M
direction of the fastest change
Ellipse E(u,v) const
direction of the slowest change
(?max)-1/2
(?min)-1/2
27
Harris Detector Mathematics
?2
Edge ?2 gtgt ?1
Classification of image points using eigenvalues
of M
Corner?1 and ?2 are large, ?1 ?2E
increases in all directions
?1 and ?2 are smallE is almost constant in all
directions
Edge ?1 gtgt ?2
Flat region
?1
28
Delaunay method
Delaunay Method
29
Delaunay method
Delaunay Method
X
30
Appendix