Author
emma-spencer
View
214
Download
0
Tags:
Embed Size (px)
1
A Compression Based Distance Measure for Texture
Bilson J. L. CampanaEamonn J. Keogh
University of California – [email protected]
2
Outline of the Talk• What makes texture important?• Why is texture hard to mine?• The CK method and CK-1 measure.• Rival methods.• Datasets and experimenting.• The results.
3
exture is Everywhere!!
4
But what IS texture?The old forget,
the young don’t know!
• Global scalars• Entropy• Standard Deviation• Energy• …
• Global vectors• Wavelet coefficients• Fourier Transforms• …
• Local Features• SIFT descriptors• Textons• …
5
Mining Textures
• Textures are ubiquitous in images• Proper analysis of an image should take into
account many details– Texture– Color– Shape– Geospatial data– Etc.
• Current approaches for texture analysis require far to much tuning– Cannot simply use texture algorithms correctly for
many datasets
There seems to be texture, but I don’t want to spend the time setting up and tuning if it doesn’t work!
We’ve formed a simple solution to your problems!
6
The CK MethodEverything should be made as simple as
possible, but not simpler.-Albert Einstein
• Simple things are easily understood, accepted and used.
Measure image similarity by exploiting video
compression
7
Kolmogorov Complexity
ababababababababababababababababb4w1x8nb2y39abgk5q85s7arjqj0cvab
We have images!
The Kolmogorov complexity K(x) of a string x is a measure of the resources needed to specify x
Incomputable!!
Consider this example…And now, conditional
complexity K(x|y)…
8
HowMPEG1Works
• Three types of frames• I, B, P
• Encoder settings are intuitively set and empirically tested
I B P
In this example, the P frame has 1 reference to the I frame.
9
x
• Query images are used to create a two frame movie.
yC(x|y)
yxC(y|x)
xxC(x|x)
yyC(y|y)
You can’t control what you can’t measure.-Tom DeMarco
The CK-1 Measure
10Apply Invariance to RotationAs you’ll see. CK-1
is very FAST!So you can just measure two
images several times while
rotating them?
PRECISELY!
11
Rival Algorithms• Gabor Filter Banks*
– Widely used for its ability to be tuned to many applications– Six orientations and four scales– Filters are convoluted through the image and responses
are gathered into a response vector• Textons**
– Classification from clustered filter responses– Extended from the previous filter bank implementation
*P. Wu, B. S. Manjunath, S. Newsam, H. D. Shin, A texture descriptor for browsing and similarity retrieval, Signal Processing: Image Communication, Volume 16, Issues 1-2, Pages 33-43, September 2000.
**M. Varma, A. Zisserman, A Statistical Approach to Texture Classification from Single Images, Int. J. Comput. Vision 62, 1-2, 61-81, Apr. 2005.
12
A World to Be Measured
• 15 experimental datasets• Many demonstrations
– Arachnology– Forensic Science– Biology– Archeology– Biometrics– Historical Texts– Texture benchmarks– And more!
A LOT of datasets!
13
• One nearest neighbor, leave-one-out cross validation
• Texton measure is trained on the entire dataset• All experiments, demonstrations, and figures
are completely reproducible• All datasets and source are available online
Experiments and Reproducing!
Hey Doc! Start reproducing at www.cs.ucr.edu/~bcampana/texture.html
14
Speed Check
Because it’s simple! Go with blue!!
Why is CK-1 so fast?
15
Perception is Key!
Filter Bank
CK-1
16
Performance at a Glance
CK-1 is DEFINITELY a contender!
17
In Summary
• Presented a compression based framework and measure for texture.
• Simple.• Empirically tested• Freely and easily available.• Fast.• Accurate.
Simplicity, carried to an extreme, is elegance.-Jon Franklin