Objective

1
Objective
Handwritten mail rejected by domestic HWAI-X
Foreign Mail Classifier
X USA Canada
Australia UK
2
Algorithm
3
1
2
Steps 1. Consider only the last three lines of
the mail address for processing 2. Consider the
one-word hypothesis as a possible location of
the country name 3. Use the Postal Code and City
and State name/abbreviations later in the
recognition process to validate/enforce the
country name recognition
3
Obtain word hypotheses
Address
Word hypotheses
Line separation
Parsing
...
4
Algorithm
5
Naïve Bayes Classifier

• Want to classify a new instance (a1, a2,, an)
  into finite number of categories (targets) from a
  set V.
• Bayesian approach
• assign the most probable category vMAP given (a1,
  a2,, an).
• Can we estimate the probabilities from the
  training data?

6
Naïve Bayes Classifier

• Second probability easy to estimate
• estimated by the frequency of each target value
  in the training data
• The first probability difficult to estimate
• cannot use frequency for argmax P(a1,a2,,anv)
  unless we have a very large set of training data
  to get a reliable estimate
• Assume independence (this is the naïve bit)

• P(aiv) can be estimated by frequency or by
  m-estimate

7
Foreign Mail Recognition

• Question How to use Naïve Bayes Classifier for
  Handwritten Foreign Country Name Recognition?
• Set of targets V Albania, Afghanistan,,Zambia
  .
• Which attributes to choose?
• Address features (of lines,of words/line,etc)
• Postal Code estimated location in the address
  block
• Postal code format obtained from parsing
  (ddddd,ddttdd,etc)
• First/Last letter in the Country name(e.g France
  F or P
• Bigrams,Trigrams frequency
• Structural features (average character height,
  characters aspect ratio)
• Low-level features( of holes,sphericity,etc -
  helpful for language identification)
• Relative order of the address elements in the
  mail address

8
Learn Probabilities

• For all countries vj in V
• mailsV total number of mail pieces in the
  Training set
• mailsvj subset of mail pieces from the Training
  set with the destination country vj
• mailsFRANCE mail pieces from the Training set
  with the destination country France
• Calculate P(vj)
• P(FRANCE) mailsFRANCE/mailsV
• Calculate P(dddddvj) using the m-estimate
  method
• P(dddddFRANCE) (ncmp)/(nm), where
• p 1/k, where k of different Postal Code
  formats
• n mailsFRANCE
• n of samples from the mailsFRANCE set that
  have the PC format ddddd
• m total of country destinations

9
Classification
10
Naïve Bayes Classifier for Foreign Mail
Classification

• Training Set 1450 mail pieces
• 119 different country destinations
• we have 20 mail pieces/country for most countries
• some countries have less than 20 entries
• Test Set 500 mail pieces
• Preliminary Set of Features used
• Number of Lines in the Mail Piece
• Number of components - First Line
• Number of Components - Second Line
• Number of Components - Third Line
• First Letter on the Last Line
• Last Letter on the Last Line
• Length of the word on the Last Line

11
Naïve Bayes Classifier for Foreign Mail
Classification

• Sample Vectors
• For South Africa
• POR02174 REPUBLICOFSOUTHAFRICA 7 9 4 10 S A
  11
• POR02533 REPUBLICOFSOUTHAFRICA 5 8 15 12 S A
  11
• POR02958 REPUBLICOFSOUTHAFRICA 7 4 6 11 R A
  21
• POR02965 REPUBLICOFSOUTHAFRICA 7 4 9 11 S A
  11
• For Australia
• POR16191 AUSTRALIA 6 2 2 2 E I 5
• POR16240 AUSTRALIA 5 7 8 8 T A 7
• POR16257 AUSTRALIA 5 3 4 4 N S 11
• POR16417 AUSTRALIA 6 4 3 4 A A 9
• POR16534 AUSTRALIA 5 5 8 8 A A 9

12
Naïve Bayes Classifier for Foreign Mail
Classification

• Compute P(vj) and P(aivj) for each country
  destination from the Training Set
• Example
• Australia
• P(Australia) 20/1450
• P(NoLines 5Australia) (101)/(208)
• P(NoComps on Line 1 13Australia) (51)/(20
  31)
• P(FirstLetterAAustralia) (111)/(2025)
• P(LastLetter ZAustralia) (01)/(2020)
• P(LastLineLength 9Australia) (131)/(2018)

13
Naïve Bayes Classifier for Foreign Mail
Classification

• Classify the vectors from the Test Set
• Example Vector for unknown country
• PRT26371 UNKNOWN DESTINATION 5 1 6 3 T A 7
• P(Destination Australia) P(NrLines5Australia
  ) P(No.Components on Line 1 13Australia) .
  P(LastLineLength 7Australia)
• P(Destination Zimbabwe) .

14
Naïve Bayes Classifier for Foreign Mail
Classification
15
Naïve Bayes Classifier for Foreign Mail
Classification

• Errors
• Confusion
• Detected Dest. AUSTRALIA ( 9 1 0 3 7 10 7)
  Actual Dest. ARGENTINA ( 5 5 14
  11 A A 9)
• Detected Dest. CROATIA ( 6 0 1 1 9 9 0)
  Actual Dest. ITALY ( 4 4
  6 4 C A 14)
• Few samples for a country
• CROATIA ( 5 3 1 1 1 9 8)
  ETHOPIA ( 5 10 9 9 E A 7)

16
Naïve Bayes Classifier for Foreign Mail
Classification

• Total Set of 10330 mail pieces
• Divided into three sets Training,Validation,Test
• Training Set 2681 mail pieces
• 119 different country destinations, all
  represented
• one third of the samples for each country are
  included
• Test Set 2435 mail pieces
• randomly chosen

17
Cleaning the Data

• Problem
• The features extracted have to be reliable
• Some features are inherently reliable
• Number of Lines,Number of Components on different
  lines
• Other features depend on the recognizers
  performance
• First/Last Letter in the Country name,Length of
  the word,etc
• Make sure that we get accurate data

18
Cleaning the Data

• Problems
• Wrong country assigned in the truth file

• Missing country name

19
Cleaning the Data

• Country name on a different line

• Country name together with some other information

• Country name with noise

20
Cleaning the Data

• Feature Length of the country name
• important to be accurately estimated
• Use several recognizers to converge to the right
  result
• Pre-recognition parser -gt Estimate1
• If word is cursive, use a word recognizer
  -gtEstimate 2
• If word is not cursive, use a character
  recognizer -gt Estimate 3

21
Naïve Bayes Classifier for Foreign Mail
Classification
22
Structural Features

• Features to investigate
• Relative order of the address elements in the
  address
• PostCode length/format
• Gap(s) in the last line
• indicates that a country name is composed of two
  words

23
Structural Features

• Structural feature
• Is there a Gap in the country name?
• Problems

• The results from the parser are not always
  reliable

• The best segmentation is not always the one
  expected

24
Structural Features

• Word practically unbreakable

• Extra noise makes word unbreakable

• We have to be careful not to break valid words

25
Structural Features

• Feature Relative order of the address block
  components City/State name, Postal Code, Country
  name
• Problems
• We dont have all the elements present. When we
  have them, we cannot always recognize them

• For certain countries this feature is not
  reliable

26
Naïve Bayes Classifier for Foreign Mail
Classification
27
Structural Features

• Next
• Features
• PostCode length/format
• Confidences values for individual characters
• Design a module to recognize first and last
  character only

28
Features

• Noise
• Pre-recognizer classifies components in
  OTHERSTUFF and FRAGMENT when it decides that they
  are most probably not LETTERS or DIGITS
• 1 - if the level of noise is above a certain
  threshold, 0 otherwise

29
Features

• Different Alphabet
• Images of Letters/Digits in the Roman alphabet
  have an average number of horizontal/vertical
  transitions WHITE-gtBLACK
• Characters in other alphabets present a
  different average

30
Features

• PC Length and Line Position

31
Features

• Continent Id

32
Naïve Bayes Classifier for Foreign Mail
Classification

• F7 - noise
• F8 - different alphabet
• F9 - Gap in country name
• F10 - PC length
• F11 - PC line
• F12 -Continent Id

• Features
• F1 - no. of lines
• F2 - no. of components Line 1
• F3 - no. of components Line 2
• F4 - no of components Line 3
• F5 - First Letter Country Name
• F6 - Last Letter Country Name