近似最近傍探索の 多段階化に よる物体の高速認識

Copyright Protection of Images Based
on Large-Scale Image Recognition
Koichi Kise, Satoshi Yokota, Akira Shiozaki
Osaka Prefecture University
Outline
Demo
 Image Recognition with Local Descriptors
 Our Research Topics
 Copyright Protection and Digital Watermarking
 Proposed Method
 Experiments
 Conclusion

2
DEMO
Large-Scale Image Recognition
Outline
Demo
 Image Recognition with Local Descriptors
 Our Research Topics
 Copyright Protection and Digital Watermarking
 Proposed Method
 Experiments
 Conclusion

4
Image Recognition
with Local Descriptors
Image 1
Image 2
Image 4
Typically a few hundreds to
a few thousands are extracted
Database
Image 3
Images are represented by
local descriptors
(feature vectors)
Query
5
Database
Image 2
1 2 3 4
Image 4
Nearest vector
in the DB
Image 1
max. votes
Image 3
no. of votes
Image Recognition
with Local DescriptorsNearest Neighbor Search
Query
6
Pros and Cons
Locality and stability
of descriptors
Large number of
expensive descriptors
CONS
PROS

Robust to occlusion
and variations


7
Extraction of descriptors
a few seconds / query
Matching
By brute-force NN search
with the DB of 1,000 images
315 seconds / query
Outline
Demo
 Image Recognition with Local Descriptors
 Our Research Topics
 Copyright Protection and Digital Watermarking
 Proposed Method
 Experiments
 Conclusion

8
Our Research Topics
Extraction of
Descriptors
Baseline
Software
Matching
(with 10,000 images)
2.3 sec. Brute-force
Existing
Technologies
3,000 sec.
ANN (Approximate
Nearest Neighbor)
50 ms
60,000 times faster
Our Methods GPU & CUDA
0.57 sec. Hash & Cascade
4 times faster
Application Copyright Protection
9
0.78 ms
60 times faster
Outline
Demo
 Image Recognition with Local Descriptors
 Our Research Topics
 Copyright Protection and Digital Watermarking
 Proposed Method
 Experiments
 Conclusion

10
Copyright Protection and
Digital Watermarking

Digital Watermarking is a technology to embed
information (copyright notice) to original digital data.


Embedded information (watermark) is often unnoticeable.
We focus on digital images as original data.
Copyright: A
Purchaser: B
original

Methods:


11
Non-blind
Blind
watermarked
Non-Blind and Blind Watermarking

Non-blind watermarking
watermarked
image

&
original
image
Blind watermarking
watermarked
image

Copyright: A
Purchaser: B
Copyright: A
Purchaser: B
History

From non-blind to blind watermarking


12
Non-blind watermarking needs matching of images
 spoils the scalability
Blind watermarking has been the focus of interest
Limitations of Blind Watermarking

StirMark


distorts images imperceptibly
severely damages the wartermark
Original
13
StirMark
StirMark: Some Image Transformations
cropping
fraction of pixel displacement
original
punch / pinch
14
corner shift
Outline
Demo
 Image Recognition with Local Descriptors
 Our Research Topics
 Copyright Protection and Digital Watermarking
 Proposed Method
 Experiments
 Conclusion

15
Proposed Method: Ideas

How do we solve the problem of blind watermarking?

Cancel the effect of StirMark,
especially geometric transformations
normalization
StirMark’ed

Normalization could be done by comparing with the original
image
StirMark’ed

16
normalized StirMark’ed
Blind to non-blind
Original
normalized StirMark’ed
Proposed Method: Ideas

How to solve the problem of non-blind watermarking?

Matching problem


solved by the efficient and robust image recognition
How about the normalization?

Image recognition with local descriptors


17
provides matching of descriptors as a side effect
helpful for normalization
Proposed Method: Processing Steps
STEP1: Recognition
STEP2: Normalization
Image
Database
query
image
point matching
STEP3: Detection
Embedded Watermark
19
Proposed Method: Processing Steps
STEP1: Recognition
STEP2: Normalization
Image
Database
query
image
point matching
Local Descriptors: PCA-SIFT
Matching: ANN
20
STEP3: Detection
Embedded Watermark
Outline
Demo
 Image Recognition with Local Descriptors
 Our Research Topics
 Copyright Protection and Digital Watermarking
 Proposed Method
 Experiments
 Conclusion

21
Experiments: Overview

Experiment 1


Experiment 2


Robustness against StirMark with default setting
(imperceptible change)
Robustness against perceptible changes
Experiment 3

22
Robustness against watermarking
Experiments: Overview

Experiment 1


Experiment 2


Robustness against StirMark with default setting
(imperceptible change)
Robustness against perceptible changes
Experiment 3

23
Robustness against watermarking
Experiments: Dataset & Task
Base set
10,000 images
100
Watermark
StirMark
query
images
Input : query image
Task: find its original from 10,000 images
24
Experiment 1
10,000
with
10,000 images
100
Watermark
without
10,000 images
100
StirMark
imperceptible
100 random application
with default setting
StirMark
watermarking without with
accuracy
100%
99.49%
time (per query) 217ms
216ms
25
query
images
query
images
Experiment2

4 image transformations


cropping, fraction of pixel displacement,
punch/pinch, corner shift
perceptible
4,300
with
10,000 images
100
Watermark
without
10,000 images
26
100
StirMark
query
images
perceptible 4 trans.
43 application in total
StirMark
query
images
Cropping: Original Image
Cropping: 70; Accuracy 97%
29
Cropping: 50; Accuracy: 95%
30
Cropping: 30; Accuracy: 92%
31
Accuracy [%]
Cropping
With watermarking
Without watermarking
Cropping parameter
32
Punch / Pinch
Original Image
33
Punch / Pinch: 1; Accuracy: 99%
34
Punch / Pinch: 25; Accuracy: 99%
35
Punch / Pinch: 50; Accuracy: 97%
36
Punch / Pinch: 100; Accuracy: 50%
37
Accuracy [%]
Punch / Pinch
With watermarking
Without watermarking
Punch / Pinch Parameter
38
Conclusion

A non-blind framework for copyright protection of
images has been proposed


Fundamental experiments on image recognition


Efficient image recognition technology with local descriptors
Image recognition is robust to StirMark attacks and
watermarking
Future Work

39
Completion of overall method
Copyright Protection of Images Based
on Large-Scale Image Recognition
Thank you for your attention
Image Recognition: Local Descriptors

PCA-SIFT (PCA Scale-Invariant Feature Transform)


41
36 dimensions
Invariant to scale and rotation changes
Image Recognition: Matching


ANN(Approximate Nearest Neighbor)
Tree structure
Feature Space
Feature Vector
ANN
:Feature Vector of a query
ANN
Nearest Neighbor Search for
ANN
Approximate NN Search for
Result
This cell is not
searched
ANN
Approximate NN Search for
Correct
more approximation
with a smaller circle
Errors caused
by approximation
Wrong
Image recognition rate[%]
NNS accuracy vs. Image recognition rate
100
95
ANN
dist
E2LSH
90
85
80
0
10 20 30 40 50 60 70 80 90 100
NNS accuracy[%]
47
Corner Shift
Original Image
48
Corner Shift: 1; Accuracy: 99%
49
Corner Shift: 100; Accuracy: 98%
50
Corner Shift: 150; Accuracy: 96%
51
Corner Shift: 225; Accuracy: 74%
52
Accuracy [%]
Corner Shift
With watermarking
Without watermarking
Corner Shift Parameter
53
Fraction of Pixel Displacement
Original Image
54
Fraction of Pixel Displacement: 1;
Accuracy 99%
55
Fraction of Pixel Displacement: 3
Accuracy: 94%
56
Fraction of Pixel Displacement: 5;
Accuracy: 81%
57
Fraction of Pixel Displacement: 8
Accuracy: 58%
58
Fraction of Pixel Displacement: 10
Accuracy: 24%
59
Accuracy [%]
Fraction of Pixel Displacement
With watermarking
Without watermarking
Fraction Parameter
60