Title: From Anthrax to ZIP Codes- The Handwriting is on the Wall
1From Anthrax to ZIP Codes-The Handwriting
is on the Wall
- Venu Govindaraju
- Dept. of Computer Science Engineering
- University at Buffalo
- venu_at_cedar.buffalo.edu
2Outline
- Success in Postal Application
- Role of Handwriting Recognition
- Recognition Models
- Interactive Cognitive Models
- New Research Areas
- Other Applications
3USPS HWAI Background
- Postal Sponsorship Started 1984
- 370 Academic Articles Published
- Millions of Letters Examined
- Many Experimental Systems Built and Tested
- Migrated from Hardware to Software System
- Only Postal Research Continuously Funded
4Pattern Recognition Tasks
- Items to be Recognized, Read, and Evaluated
(Machine printed and Script) - Delivery address, senders address, endorsements
- Linear Codes, Mail Class
- Indicia (2D-Codes, Meter Marks)
5Deployed..
- USA
- 250 PDC sites
- 27 Remote Encoding Centers
- 25 Billion Images Processed Annually
- 89 Automated Bar-coding
- UK
- 67 Processing Centers
- 27 Million Pieces Per Day,
- 9.7 Million Pieces Per Hour Peak
- Australia
6RCR Overview
7At the Right Price
- Processing Type Cost/1000 Pieces
- Manual 47.78
-
- Mechanized 27.46
-
- Automated 5.30
880 encode rate and counting!
9Impact
- Applications of CEDAR research helping to
automate tasks at IRS and USPS - 1st year that USPS used CEDAR-developed software
to read handwritten addresses on envelopes, saved
100 million - 1997-1999 USPS deployment of CEDAR-developed
RCRs, USPS saved 12 million work hours and over
340 million - 500 scientific publications and 10 patents
10Outline
- Success in Postal Application
- Role of Handwriting Recognition
- Recognition Models
- Interactive Cognitive Models
- New Research Areas
- Other Applications
11Role Handwriting Recognition in Address
Interpretation
12Context Provided by Postal Directories
- ltZIP Code, Primary Numbergt
- Create street name lexiconlt06478, 110gt
- DPF yields 8 street names
- ZIP4 yields 31 street names (on average about 5
times more) - HAWLEY RD 1034NEWGATE RD 1533BEE
MOUNTAIN RD 1615DORMAN RD 1642BOWERS
HILL RD 1757FREEMAN RD 1781PUNKUP RD 1784
PARK RD 6124
13CEDAR
Context
- One record per delivery point in USA
- Provided weekly by USPS, San Mateo
- Raw DPF
- 138 million records
- 15 GB (114 bytes per record)
- 41,889 ZIP Code files
- Fields of interest to HWAI
- ZIP Code, street name, primary number, secondary
number, add-on
14CEDAR
Power of Context
- ZIP Code
- 30 of ZIP Codes contain a single street name
- 5 of ZIP Codes contain a single primary number
- 2 of ZIP Codes contain a single add-on
- ltZIP Code, primary numbergt
- Maximum number of records returned is 3,071
- ltZIP Code, add-ongt
- Maximum number of records returned is 3,070
15Outline
- Success in Postal Application
- Role of Handwriting Recognition
- Recognition Models
- Interactive Cognitive Models
- New Research Areas
- Other Applications
16Handwriting Recognition
Ranked Lexicon
Context
17Multiple Choice Question
Ranked Lexicon
Context
18Lexicon Driven Model
Distance between lexicon entry word first
character w and the image between - segments 1
and 4 is 5.0 - segments 1 and 3 is 7.2 - segments
1 and 2 is 7.6
Find the best way of accounting for characters
w, o, r, d buy consuming all segments 1
to 8 in the process
19Lexicon Free Model
- Image from 1 to 3 is a in with 0.5 confidence
- Image from segment 1 to 4 is a w with 0.7
confidence - Image from segment 1 to 5 is a w with 0.6
confidence and an m with 0.3 confidence
w.6, m.3
w.7
d.8
o.5
u.5, v.2
i.8, l.8
i.7
r.4
u.3
m.2
m.1
Find the best path in graph from segment 1 to 8 w
o r d
20Holistic Features
Reference Lines
Slant Norm
Turn Points
Ascender
Position Grid and gaps
Descender
21Lexicon Reduction and Verification
22Outline
- Success in Postal Application
- Role of Handwriting Recognition
- Recognition Models
- Interactive Cognitive Models
- New Research Areas
- Other Applications
23Grapheme Models
24Structural FeaturesBAG
Loops
End
Junction
End
Loop
Turns
25Feature Extraction and Ordering
Critical node removal disconnects a connected
component.
Loops
End
End
Turns
Junction
Turns
Loop
2-degree critical nodes keep feature ordering
from left to right.
Right Component
Left Component
26Continuous Attributes
grapheme pos orientation angle
Down cusp 3.0 -90o
Up loop
Down arc
27Stochastic Model
28Observations
29Results
Lex size Top WMR SM CA
10 1 96.86 96.56
2 98.80 98.77
100 1 91.36 89.12
2 95.30 94.06
1000 1 79.58 75.38
2 88.29 86.29
20000 1 62.43 58.14
2 71.07 66.49
30Interactive Models McClelland and Rumelhart,
Psychological Review, 1981
ABLE
TRIP
TRAP
Words
A
T
N
Letters
Features
31Interactive Recognition
Lexicon 1 Lexicon 2 Lexicon 3
West Central StreetWest Main StreetSunset
Avenue
West Central StreetEast Central StreetSunset
Avenue
West Central StreetWest Central AvenueSunset
Avenue
Interactive Model
features
T-crossings, loops, ascenders, descenders, length
image
32Adaptive Character Recognition Park and
Govindaraju, IEEE CVPR 2000
- Adaptive selection of features
- Adaptive number of features
- Adaptive resolutions
- Adaptive sequencing of features
- Adaptive termination conditions
33Features
4 gradient features
5 moment features
Vector code book
34Feature Space
- V x Nc x Ixy
- 29 x 10 x 85 (quad tree, 4 levels)
- Recognition rate and feature V
- GSC V 2512
- Tradeoffs space vs accuracy
- Hierarchical space with additional resolution and
features as needed
35Active Recognition Using Quad Trees
36Experimental Results
37(No Transcript)
38Results
Classifier Active Model Neural Net KNN
Top 1 95.7 96.4 95.7
Templates 612 976 3,777
Msec/char 1.45 11.5 384
Training hrs 1 24 1
25656 training and 12242 test (Postal NIST)
39Outline
- Success in Postal Application
- Role of Handwriting Recognition
- Recognition Models
- Interactive Cognitive Models
- New Research Areas
- Other Applications
40Fast Recognition
-Reuse matched characters -Reuse matched
sub-strings -Parallel processing
41Combination and Dynamic Selection Govindaraju
and Ianakiev, MCS 2000
image
WR 1
WR 3
1
Top 50
Lexicon
lt55
WR 2
Top 5
- Optimization problem
- Combinatorial explosion in
- arrangement of recognizers
- lexicon reduction levels
42Lexicon Density Govindaraju, Slavik, and Xue,
IEEE PAMI 2002
Lexicon 1 Lexicon 2 Me MeHe MemoSo MemoryTo
MemoirsIn Mellon
43Classifier Performance Prediction Xue and
Govindaraju, IEEE PAMI 2002
q probability that recognizer make a unit
distance errors D average distance between any
two words in the lexicons n lexicon size p
performance a, k, model parameters ln (-ln p)
(ln q) D a ln ln n ln k
44Outline
- Success in Postal Application
- Role of Handwriting Recognition
- Recognition Models
- Interactive Cognitive Models
- New Research Areas
- Other Applications
45Bank Check Recognition
46PCR Trend Analysis
47NYS EMS PCR Form
- NYS PCR Example
- Thousands are filed a day.
- Passed from EMS to Hospital.
- PCR Purpose
- Medical care/diagnosis
- Legal Documentation
- Quality Assurance
- EMS Abbreviations
- COPD Chronic Obstructive Pulmonary Disease
- CHF Congestive Heart Failure
- D/S Dextrose in Saline
- PID Pelvic Inflammatory Disease
- GSW Gunshot Wound
- NKA No known allergies
- KVO Keep vein open
- NaCL Sodium Chloride
48Medical Text Recognition and Data Mining
49Reading Census Forms
Lexicon Anomalies Space sales man and
salesman Morphology acct manager and
account management Abbreviation Plural
school and schools Typographical managar
and manager
50Binarization
51Historic Manuscripts
52Summary
- Handwriting recognition technology
- Pattern recognition task
- Lexicon holds domain specific knowledge
- Adaptive methods
- Classifier combination methods
- Many applications