Image Analysis: Old Problems and New Challenges

molassesitalianΤεχνίτη Νοημοσύνη και Ρομποτική

6 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

74 εμφανίσεις

VIPER

SSIAI'98


Slide
1


Image Analysis:

Old Problems and New Challenges

Edward J. Delp

Purdue University

School of Electrical and Computer Engineering

Purdue Multimedia Testbed

Video and Image Processing Laboratory (
VIPER)

West Lafayette, IN 47907
-
1285

+1 765 494 1740

+1 765 494 0880 (fax)

email: ace@ecn.purdue.edu

http://www.ece.purdue.edu/~ace


VIPER

SSIAI'98


Slide
2

Outline


Image Analysis: What’s Right and Wrong


Comparison to Other Scientific Areas


Technical Societies


How to Make Money In The Next Millennium!


Conclusions

VIPER

SSIAI'98


Slide
3

Image Analysis


Image Processing


use of computer techniques to “process” a digital
image
-

result is another image


Image Analysis


use of computer techniques to extract information
from a digital image


Image Analysis a.k.a. Computer Vision, Image
Understanding … others

VIPER

SSIAI'98


Slide
4

The Typical Image Analysis System


Preprocessing


Segmentation


Feature Extraction or Measurement


Feature Processing


Decision


Various steps can and/or should be model
-
based



VIPER

SSIAI'98


Slide
5

Application Domains


Medical Imaging


Target Recognition


Stereo Imaging


Remote Sensing


Document Image Analysis


Image and Video Databases


Automatic vehicle guidance



… many more

VIPER

SSIAI'98


Slide
6

Image Analysis


Image Analysis is more than 30 years old



Most of the impact in “digital imaging” has come from
image processing



VIPER

SSIAI'98


Slide
7

The Open Problems


Segmentation: Edge Detection


Shape from X


Stereo Correspondence


Object Tracking and Identification



VIPER

SSIAI'98


Slide
8

Image Analysis: What’s Wrong?


Has image analysis earned it’s keeps?


By most standards of scientific achievement: YES and
NO!


Much has been promised and little has been delivered


Only simple problems have been solved

VIPER

SSIAI'98


Slide
9

Important Success Stories


CMU’s autonomous vehicle



Fingerprint systems




Papnet




Mathematics Experiences Through Image Processing
(METIP)





VIPER

SSIAI'98


Slide
10

Important Success Stories


Factory Automation
-

IC manufacturing and



Iris Recognition and forms of biometrics



Handwriting Recognition



Image Database
-

IBM QBIC


Commercial systems are characterized by the needs for
simple solutions with low computational complexity and
in many cases the need for human interaction


VIPER

SSIAI'98


Slide
11

Why has there been so little impact?


The problems are ill
-
posed inverse problems and are too
hard!


Problems require applications
-
specific solutions


Too much work using the “human observer” as a
metaphor for how the problem should be solved


Too much work rediscovering what has already been
done


Not enough experiments (small sample size problem)


Not enough validation experiments



VIPER

SSIAI'98


Slide
12

Comparison: Pharmaceutical Industry


Large impact on society


How are problems addressed?


Experiments with lots of data


Complete disclosure of all experimental parameters


Defined performance metrics


VIPER

SSIAI'98


Slide
13

Issues


Metrics


Data sets and data sharing


Complete disclosure of techniques


Ability to reproduce an experiment and publish it

VIPER

SSIAI'98


Slide
14

Issues: Metrics


How do we measure performance?


mean square error


statistical validation


test images


VIPER

SSIAI'98


Slide
15

Metrics

VIPER

SSIAI'98


Slide
16

Metrics

VIPER

SSIAI'98


Slide
17

Issues: Data Sets





NOT LENA!

VIPER

SSIAI'98


Slide
18

Issues: Data Sets


This problem is starting to be addressed


What is needed:


freely or cheaply available, ground
-
truthed and
copyright cleared imagery


Examples


JPEG and MPEG test data


Haralick’s document images and RADIUS images


CMU calibrated stereo images


Digital Database for Screening Mammography




VIPER

SSIAI'98


Slide
19

Data Sets






1998 ICPR will have a
contest based on
calibrated data

VIPER

SSIAI'98


Slide
20

Issues: Data Sets

Proposal


a paper will not be published unless “standard”
images are used


if images are not used, the author must make data
available and must certify the “pedigree” of the data


Academics must obey the copyright law


VIPER

SSIAI'98


Slide
21

Issues: Disclosure


We need to fully describe how our techniques work and
what parameters must be “tweaked.”


All experiments reported in a paper must be described in
enough detail so that it can it replicated


The author must be willing to share both the original
data and “processed” images

VIPER

SSIAI'98


Slide
22

New Challenges


Two new challenges:


how does one maintain, store, search, and manipulate
a large database of images and/or video (digital
library)


how does one protect privacy

VIPER

SSIAI'98


Slide
23

Image Database

Zoom in

Zoom out

Zoom in

Zoom out

VIPER

SSIAI'98


Slide
24


The Role of Technical Societies

Goals of the IEEE:


The technical objectives of the IEEE focus on
advancing the theory and practice of electrical,
electronics and computer engineering and computer
science. To realize these objectives, the IEEE sponsors
technical conferences, symposia and local meetings
worldwide: publishes nearly 25% of the world's
technical papers in electrical, electronics and
computer engineering; provides educational
programs to keep its members' knowledge and
expertise state
-
of
-
the
-
art.

VIPER

SSIAI'98


Slide
25

The Role of Technical Societies


The IEEE is today:


conference organizer


publisher


everything else



The IEEE is becoming:


conference organizer


library and less of a publisher


everything else


VIPER

SSIAI'98


Slide
26

Technical Societies as Libraries


Technical Societies and Web
-
based publishing


Advantages:


support the distribution of things other than
words and static images


papers could have supplemental material


get the information when you need it


Disadvantages:


papers could have supplemental material


must be maintained (cannot “forget” about a
paper after it is published)


VIPER

SSIAI'98


Slide
27

How to Make Money in


the Next Millennium!



Two HOT areas to invest:


Paper Mills and related industries


Data transcoding

VIPER

SSIAI'98


Slide
28

Paper:

Are you kidding?!


Paper use has been increasing by 6% for more than 20
fifteen years


recent evidence is that paper use my even be
increasing faster!


Why?


Old Model


Print

䑩獴物扵r攠

却潲S 潮o䉯潫獨敬f


New Model


Distribute

偲楮琠

周牯r 慷慹


VIPER

SSIAI'98


Slide
29

Date Transcoding


The popular wisdom is that we will own more bits and
less paper, BUT how will we maintain these bits?


Will you be able to read your bits in 10 years?


Will the hardware exists to do this?


Will you be able to read your CD
-
ROM in 10 years?
(will you want to!!)


Will the web address exist in 5 years and will the
material be readable?


A whole new industry will be needed for data
transcoding




VIPER

SSIAI'98


Slide
30

Conclusion


Image analysis is still a promising area but it must
deliver


We must do our research and publishing in a different
way


Bits are not forever!


Buy dead wood!