Comprehensive Guide on Drafting Directors' Report and its ROC Compliances und...
CEDARFOX-020907.pdf
1. Computational Approaches to Questioned
Handwriting Examination
Sargur (Hari) Srihari
University at Buffalo
State University of New York
2. Computational Forensics
• Forensic domains involving pattern matching
• Motivated by Importance of Quantitative
methods in the Forensic Sciences
1. Daubert Ruling
2. High Standards established by DNA
3. Computers
1. Low Cost
2. Advances in Artificial Intelligence/Pattern Recognition
4. Improved Statistical Methods for Evidence
E.g., Aitken and Taroni, Statistics and the Evaluation of
Evidence for Forensic Scientists, Wiley, 2004
3. QDE
• Bureau of Justice Statistics (2002)
– Among 50 largest publicly funded crime labs
• 57% perform QD function
• 5,231 cases requested
• 1,079 backlogged at year end
• Significantly larger case load internationally
• Handwriting is common in QD case work
4. CEDAR Research on Handwritten QDE
• Research on quantifying discriminatory
power of handwriting since 1999
– Testing on national database, twins data
• Feedback from QDE’s in developing
computational tools
– Workshops at ASQDE,
– JtMtg of MAFS,CAFS,
– SWAFDE
• Developing Statistical Evidence Theory
5. CEDARFOX software system
• Writer Verification/Identification
– Probability/Strength of Evidence Computation
• Document Properties
– Line Structure, Writer Characteristics
• Signature Verification
• Document Search
System Requirements
Pentium class processor
(P4 or higher recommended)
Windows NT, 2000 or XP
128MB of RAM
30MB available disk space
8. How is Strength of Evidence Computed?
• Handwriting characteristics are extracted
from both K and Q and their similarities
compared to the similarities in a
representative database
• Based on a data base of 1,500 writers
providing 3 pages of writing each
• Probability distributions of similarities
modeled by Gamma and Gaussian
distributions
22. Summary
• CEDAR-FOX is a system for QDE with a
focus on handwriting
• Has automated tools for writer/signature
verification/identification
• Has tools for case-work display
• Computes strength of evidence
23. Future Work
• Better Statistical Model
– Current statistical model in system uses
independence assumption
– Performance is not high as with better
theoretical models, e.g., neural networks
– Plan to incorporate a compromise model e.g.,
pairwise independence