Transcript Introducing TRIGRAPH trimodal writer identification

Introducing TRIGRAPH
trimodal writer identification
Ralph Niels*, Louis Vuurpijl*
and Lambert Schomaker♦
* Nijmegen Institute
for Cognition and
Information
Radboud University
Nijmegen
Dutch Forensic
Institute
♦ Artificial
Intelligence Institute
University of
Groningen
ENFHEX conference - November 2005 – Budapest, Hungary
Overview
●
●
Computer assisted document examination
TRIGRAPH combines 3 methods:
I
Automatic features from image
II
Manually measured properties
III Allographic features
●
Recent achievement: “intuitive” matching
●
Summary
●
Next steps
Computer assisted document
examination
Computer assisted document
examination
Improving on current systems
●
●
Systems do not benefit from recent advances in
pattern recognition and image processing
New insights in:
–
automatically derived
handwriting features
user interface development
–
innovations in forensic writer identification systems
–
●
Aim: Suspected document in top-100 hit list from
database of > 20,000 writers
Design requirements
●
Improve on currently available performance
●
Minimize amount of manual labor
●
Exploit human cognition and expertise
●
Correspond to expectations of human experts
WANDA
●
Integrate techniques in WANDA Workbench
(Franke et al., ENFHEX News 2004; Van Erp et al., JFDE (16) 2004)
Three approaches
I
Automatic features from images
II Manually measured properties
III Allographic features
Automatic features from images (1)
I
●
Layout and spacing
●
Ink morphology
(Franke)
Automatic features from images (2)
I
●
Local shape (Bulacu)
Automatic features from images (3)
I
●
Grapheme-fraglet tables (Schomaker)
II
Manually measured properties
Fish
●Script
●Wanda
●
III
●
Allographic properties (1)
(Vuurpijl, Niels) Matching characters by:
–
Considering global shape characteristics
–
Reconstructing and comparing production process
–
Zooming in on particular features
III “Intuitive”
matching (1)
10
●
●
●
Given: 2 dynamic trajectories
(one questioned, one from a
set of prototypes)
Technique: Dynamic Time
Warping (point-to-point
comparison)
Result: similarity measure that
can be used to find prototype
that is most similar to
questioned sample
1
7
1
●
●
●
III “Intuitive” matching (2)
Experiment: compare various techniques
Result: Dynamic Time Warping yields visually
convincing (or “intuitive”) results
Our work on DTW was previously presented at:
9th International Workshop on Frontiers in Handwriting Recognition
(IWFHR-2004), Japan.
th Conference of the International Graphonomics Society
● 12
(IGS-2005), Italy.
th International Conference on Document Analysis and Recognition
● 8
(ICDAR-2005), South-Korea.
●
III
●
●
Allographic properties (2)
(Semi-)automatic extraction of dynamic information:
–
Automatically extract traces from scanned document
–
Verify resulting trajectories with allograph prototypes
–
Start user-interaction in case of confusion
Advantages:
–
More reliable measurements
–
Online character recognition techniques
–
Search for particular allographs in documents
–
Visually convincing matching techniques
Summary
●
●
●
Computers can help forensic experts in measuring
handwriting and searching databases
In TRIGRAPH, new insights from different scientific
areas will be used
In TRIGRAPH, new UI methods will be combined with
techniques developed in three modalities:
I Automatic features from images
II
III
Manually measured properties
Allographic features
Next steps
●
Automatic extraction of dynamical information
from scanned images
●
Supervised character segmentation
●
Allograph based verification of results
Introducing TRIGRAPH
trimodal writer identification
Ralph Niels*, Louis Vuurpijl*
and Lambert Schomaker♦
* Nijmegen Institute
for Cognition and
Information
Radboud University
Nijmegen
Dutch Forensic
Institute
♦ Artificial
Intelligence Institute
University of
Groningen
ENFHEX conference - November 2005 – Budapest, Hungary