Handwritten signature retrieval and identification

1
Handwritten signature retrieval and identification

• ? ? ?

2
? ?

• Introduction
• Signature representation,indexing and matching
• Detailed system description
• Performance results
• Future research

3
Introduction
query
Image database
image
Input image
Requirement (within some specified similarity
measure)
Requirementinformation is added
??1. Flow of Similarity retrieval of images
4

• Similarity retrieval of images has two
  components(1) What should be the measure of
  similarity?(2) How should similar images be
  identified?
• Two components are the issue of image
  representation,indexing and matching so that
  images similar to query image can be retrieved
  without excessively searching the image database.
• The aim of this paper is to address these issues
  in the context of database of handwritten
  signature images

5
2.Signature representation,indexing and matching
Three component of the signature identification
system
Signature representation Search
organization(indexing) Match scheme
6

2-1 Signature representation

• Requires the transformation of a signature image
  into a compact and meaningful description through
  the extraction of certain features.
• In our approach, the signature description is
  obtained by first extracting a set of geometric
  and topologic features from a signature image.

7

• The geometric features horizontal bars,
  vertical bars, and loops
• The topologic features end points, branch
  points, crossing points, convex points, and
  concave points.

8
2-2 search organization (the indexing scheme)

• Based on the string hashing version
• for example a group of k, k lt n,.
• Through overlapping, n-k1 local associations
  for a string of length n
• Each local association is then hashed through a
  suitable perfect hashing function to generate a
  hash address in a hash index table which stores a
  pointer to a linked list of signature images

9
2-3 the matching

• 2D string is used to take into account the
  2-dimensional spatial relationships of the
  different features present in a signature.
• The matching criterion used is the longest common
  subsequence(LCS) shared by the two 2D strings.
• After matching the most similar signature found
  in the DB is sent to a final decision process to
  decide whether the identity of the query
  signature is accepted or rejected.
• By checking the match score against a predefined
  threshold value

10
?? 2 A SAMPLE HASH INDEX TABLE
11
3.Detailed system description
Enrollment identification
Input Signature Image
Preprocessing
Image Storage / Retrieval
Matching and Identification
Feature Extraction
Image Database
Signature identity
Normalization and Feature Mapping
12
(a) A query image
This image will be used in our descriptionto
illustrate different processing stages.
13
(b) Preprocessing

• Threshold binary image
  Thinned signature

Background subtraction Thresholding Noise
cleaning Gap filling Thinning
14
(c) Geometric features

• Include horizontal bars, vertical bars and loops
• Horizontal and vertical bars are extracted from
  the binary image of the signature using the
  morphological hi-or-miss operation
• The loops are extracted from the skeleton image
  of the signature using the 4-connected component
  labeling algorithm

15
(d)Topological features

• The set of topologic features consists of end
  points,branch points, crossing points, convex
  points, and concave points.

End points
16
Branch points and Crossing points
Suggested by Simon and Zerhoumi(1991),Simon and
Baret(1992),Simon(1992)
17
Concave and Convex
Use the off-line tracing algorithm suggested by
Lee and Pan(1992) And the sailent curve-point
selection algorithm by Fischler and Wolf (1994)
18
(e) All the features

• This map shows the feature map of the example
  signature image by indicating all the detected
  features in it.
• All the detected features are represented as
  points
• Ex) a loop feature is represented by a point at a
  position that corresponds to the center of the
  loop.
• Ex) a detected horizontal/vertical bar feature is
  represented by its two end point.

19
(f) Feature Normalization Feature mapping
mapping
20
(g) Enrollment phase identification
phase

• In the enrollment phase,The local associations of
  the representation of the input signature are
  hashed to generate addresses to set up the
  reference signature image database.
• In the identification phase, the local
  associations of the representation of the query
  signature are mapped into the hash index tables
  to vote for candidate signatures to be retrieved.

21

• To perform indexing, a perfect hashing function
  is used to map the X string and Y string of the
  representation of a signature into two hash index
  tables,
• h(P) Si2 62 Si1 61 Si 60
• ( 1 lt i ltn-2 )
• where h(P) is the hashing address,
• P is the X string or the Y string,
• Si2, Si1, Si are the three consecutive
  coding numerals in P according to table.
• n is the length of P.

22
(h) Signature retrieval stage

• An accumulator array is used to record the voting
  results generated by the query signature.
• For each hash address, the corresponding linked
  list votes for the corresponding signature.
• After voting, the accumulator array records the
  voting numbers of all the signatures in the DB.
• The top-three vote-getting signatures in the two
  hash tables are selected to form the candidate
  set of signatures which are retrieved from the DB
  for matching and identification.

23
(i) In the matching and identification

• the 2D string LCS is used as a measurement of
  similarity between the query signature and a
  retrieved candidate signature.
• The candidate signature yielding the longest LCS
  is considered most similar to the query image

24
4. Performance results

• Based on two sets of 120 signatures the
  reference set of signatures the test set of
  signatures
• No constraints, Each signer was asked for two
  signatures written on the paper, Digitized by a
  CCD camera, OS unix ,Implementation is C and
  C
• These experiments included testing the system
  with partial query signatures. The partial query
  signatures were generated by masking the already
  collected test signatures.

25
(No Transcript)
26
The performance of the system

• The average number of candidate signature
  retrieved for every query image
• indicates the quality of the features as
  well as the local associative indexing scheme.
• for discriminatory features and good
  indexing,
• the average number of candidate signatures
  retrieved should be small.
• Hit ratio
• how many times the signature of correct
  identity was retrieved as a member of the
  candidate set.

27

• 3. Recognition rate
• capability to identify a signature correctly.
• this results are for different amounts of
  query signature masking.
• in each case, the results are presented for
  three threshold settings.

28
5. Future research

• The use of multiple signatures per signer in the
  database to improve recognition and rejection
  rates
• The use of the signature slant and size
  information to group signatures for a
  hierarchical organization of the signature
  database

29
Threshold 20 mask