Professional Documents
Culture Documents
www.uk.mitsubishielectric-rce.eu
WORKSHOP ON MOBILE VISUAL SEARCH Fingerprinting the Web with Image Signatures
Miroslaw Bober Paul Brasnett Mitsubishi Electric R&D Centre Europe December 2009
VIS09-PPT168
www.uk.mitsubishielectric-rce.eu
Who are her friends? Who holds the copyright? Who is she? What web-pages does she appear at? show me all! Is that a part of a larger image? Does she know Leonardo?
In which museum can I see this painting? Do I already have a copy of that image on my laptop?
What is her telephone number? Somebody is taking this painting across the border is that OK?
VIS09-PPT168
Overview
Need for standardised Visual Signature Tools MPEG7 Image and Video signature tools Image Signature Requirements Evaluation & Development Process Design of Image Signature Performance Use Scenarios Conclusions
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
www.uk.mitsubishielectric-rce.eu
There is a need to identify multimedia content Identifier standards exist e.g. International Standard Audio-visual Number However:
99.999% digital content exists without any identifier even if image content is registered, no automated search/identification is possible
www.uk.mitsubishielectric-rce.eu
Work started in January 2006, a part of ISO/IEC 15938-part 3 (MPEG-7 Visual) Tough requirements formulated July 2007 Call for proposals issued April 2009 Image Signature Standard completed Work on Video Signature close to completion
Region Shape
Colour Layout
Camera Motion
Boom up Track right Dolly forward Boom down Dolly backward Track left
Dominant Colour
Edge Histogram
Homogeneous Texture
Colour Structure
Contour Shape
VIS09-PPT168
www.uk.mitsubishielectric-rce.eu
How many images on the Web? 100 billion?+ What does this mean for the requirements? False alarm rate 1 in 20,000,000 Matching speed 100 million+ per second on a typical PC (global)
General Requirements: Robustness Independence Fast Matching Fast Extraction Compactness Non-alteration of content Self-contained Coding Independence
2009 R&D Centre Europe B.V. Visual & Sensing Division
Image Signature Requirements Partial matching (cropping) Heavy compression Scaling & aspect ratio change
VIS09-PPT168
Performance Evaluation
Two Datasets
www.uk.mitsubishielectric-rce.eu
Colour/grayscale photographs, logos & graphics of varying sizes Independence: 135,000 images, ~10 billion pairs Robustness: 10,000 images x 44 modifications
VIS09-PPT168
www.uk.mitsubishielectric-rce.eu
Technologies Received
Standard Published
VIS09-PPT168
Trace Transform
www.uk.mitsubishielectric-rce.eu
The trace transform is a generalization of the Radon transform The image is traced by straight lines along which a functional of the image is computed. Each line is characterised by two parameters; Trace transform is effectively the map of the values of the functional as a function of these parameters By choosing the functionals , one can construct features that are invariant to various image distortions, like affine transforms or illumination changes. A second functional (also known as diametrical functional) is applied to the columns of the trace transform producing 1D function called a circus function Problems of occlusion can be tackled with the help of local image kernels, a localized version of the Trace transform
VIS09-PPT168
www.uk.mitsubishielectric-rce.eu
3.85
x 10
x 10
3.8
2.5
3.75
2
3.7
3.65
3.6
3.55
FFT magnitude
1.5
3.5
0.5
3.45
0 0 10 20 30 40 50 60 70 80 90 100
3.4
100
200
300
400
500
600
x 10
0.5
-0.5
-1
-1.5
10
20
30
40
50
60
70
80
90
100
VIS09-PPT168
10
www.uk.mitsubishielectric-rce.eu
3.85 3.9 3.85 3.85 3.8 3.8 3.8 3.75 3.75 3.75 3.7 3.7 150 150 3.7 3.65 3.65 3.6 3.6 250 250 3.55 3.55 3.55 3.5 3.5 3.5 3.45 3.45 3.45
x 104 x 10
4 4
50 50
100 100
200 200
300 300
350 350
0 0
100 100
200 200
300 300
400 400
500 500
600 600
3.4 3.4 0 0
100 100
200 200
300 300
400 400
500 500
600 600
x 10
2.5
1.5
0.5
10
20
30
40
50
60
70
80
90
100
VIS09-PPT168
11
www.uk.mitsubishielectric-rce.eu
We need to increase robustness Idea: create a multiresolution image representation by building a pyramid of trace transforms Fast can be done directly in trace domain Linear and angular dimension
VIS09-PPT168
12
Trace Cube
www.uk.mitsubishielectric-rce.eu
A three-dimensional (3D) representation of the image is generated in terms of the parameters d, and t This new representation, which is a 3D volume, is referred to as the trace-cube. The trace-cube carries significantly more information then the conventional trace transform, including information related to non-linear regions in the image, such as circles or ellipses.
VIS09-PPT168
13
www.uk.mitsubishielectric-rce.eu
Component binary strings are combined to create the global image signature 512 bits
+ FFT
512 bits
+ FFT
VIS09-PPT168
14
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
15
www.uk.mitsubishielectric-rce.eu
A scale-space representation of the image is constructed by low-pass filtering with Gaussian kernel. Features are detected by finding maxima in the location-scale space Two detectors are used to provide good detection of the location and scale of features modified scale corrected Laplacian of Gaussian (filter) scale-adapted Harris operator Image intensity gradients are calculated using the Scharr operator
The Laplacian of Gaussianis is applied over 12 levels in scale space The locations and scales with the highest value are chosen as potential features. Candidate features corresponding to a weak response from Harris operator are removed
2009 R&D Centre Europe B.V. Visual & Sensing Division Export Control : NLR VIS09-PPT168
16
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
17
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
18
Local
levels Modified Items Brightness change Colour to monochrome conversion JPEG compression with varying Quality factors Colour reduction Gaussian Noise Histogram equalization Auto-levels Blur Simple Rotation Geometric transformation Scaling Flip AVERAGE
2009 R&D Centre Europe B.V. Visual & Sensing Division Export Control : NLR
www.uk.mitsubishielectric-rce.eu
Heavy 99.81
Light 100.00
99.69
99.95 99.34
99.04
19
Local + Global
www.uk.mitsubishielectric-rce.eu
levels Modified Items Brightness change Colour to monochrome JPEG compression with varying Quality factors Colour reduction Gaussian Noise Histogram equalization Auto-levels Blur
Heavy 99.97
Light 100.0 0
99.97
99.99 99.84
99.76
100.00
100.0
100.0
100.00
VIS09-PPT168
20
Local + Global
www.uk.mitsubishielectric-rce.eu
Rotation Simple Rotation Scaling Translation Geometric transformation Flip Aspect Ratio Change Crop Skew Perspective Combined (C,T,S) AVERAGE
VIS09-PPT168
21
www.uk.mitsubishielectric-rce.eu
Content linking
Inferring links between content based on analysis of media content
Metadata Association
Quickly and accurately obtain content metadata information
Copyright management
Royalty collection, brand management
2009 R&D Centre Europe B.V. Visual & Sensing Division Export Control : NLR VIS09-PPT168
22
www.uk.mitsubishielectric-rce.eu
Wojciech Machowski
Index
VIS09-PPT168
23
Web Search
www.uk.mitsubishielectric-rce.eu
Approximately 1.5 million images from Flickr Several hundred random seed keywords used e.g.
oxide, rank, perplexity, mirrors, throne, walk, hens, seat etc.
24
Example Query
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
25
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
26
www.uk.mitsubishielectric-rce.eu
VIS09-PPT168
27
www.uk.mitsubishielectric-rce.eu
Local Signature
Between 3,584 and 7,424 bits per image False alarm rate of 10ppm 98.57% correct detection on cropping, translation Over 100,000 images per second searching
VIS09-PPT168
28
Conclusions
MPEG-7 Visual Signature Tools
Standardised Multimedia Content Identification Robust, Low False Alarm, Fast Searching
www.uk.mitsubishielectric-rce.eu
Suitable for usage monitoring, copyright management, content linking, content metadata association MPEG-7 Image Signature Tools
Published April 2009
Maybe this is the missing link that you need for your killer application!
VIS09-PPT168
29
www.uk.mitsubishielectric-rce.eu
MERCE R&D
VIS09-PPT168
30