Writer identification by means of loop and leadin features

1
Writer identification by means of loop and
lead-in features

• Vivian Blankers
• Radboud Universiteit Nijmegen
• Nijmegen Institute for Cognition and Information
• June 2007

2
Outline

• Introduction
• Why?
• What?
• How?
• Method
• Data collection
• Features
• Analysis
• Results
• Discussion
• Future work

3
Why?

• Forensic handwriting investigations
• Now humans
• Future computers

4
What?

• well-established method compare handwritings
  through allographs
• Equal letters ( k vs k )
• Similar letters ( b vs k )

5
How?

• By means of loops and lead-in strokes

6
Data Collection

• The Plucoll set, Schomaker, Segers Vuurpijl
  97

• 41 Dutch writers
• lowercase letters
• online data (tablet)

7
Groups of similar letters, loops
ascenders ( b d f h l k )
descenders ( g j p q y )
8
Groups of similar letters, lead-in strokes
a c d q
b h k l t
e f
i j m n p s u v w y z
9
Loop features
10
Length

• Length between coordinate before intersection
  point and coordinate after intersection point

loopfeatures
11
Writing speed in mm/s

• 100 Hz
• one coordinate registered every 0.01 seconds

loopfeatures
12
Surface
loopfeatures
13
Width/height-ratio
loopfeatures
14
Relative height
loopfeatures
15
Direction
loopfeatures
16
Curvature

• Average angle between each couple of succeeding
  vectors

loopfeatures
17
Lead-in features
18
Length and writing speed

• Same as in loop
• Length of lead-in from begin of sample to end of
  lead-in stroke
• Average writing speed of lead-in stroke

lead-in features
19
Direction

• The angle between the x-axis and the vector
  between the first and the last coordinate of the
  stroke

lead-in features
20
Average direction
lead-in features
21
Analysis

• Make feature vectors and calculate Euclidean
  distance
• Training
• kNN
• Testing
• Can a writer be identified by comparing equal
  letters?
• Can a writer be identified by comparing similar
  letters?
• 2 types of tests penup and pendown

22
Penup results equal
23
Pendown results equal
24
Similar letters results

• Extensive and different results
• In graph
• mean scores per group
• maximum scores per group
• minimum scores per group

• chance level 2.44

25
(No Transcript)
26
Discussion

• Equal letters
• Similar letters
• No loops or lead-in strokes no results

27
Further research

• Similar letters
• More writers
• More features
• PCA best features
• Useful in practice
• Extract penup information from scanned offline
  handwritings

28

• ?