Writer Identification Lewis Griffin & Andrew Newell Computer Science, CoMPLEX University College London [email protected].

Writer Identification 

Lewis Griffin & Andrew Newell 

Computer Science, CoMPLEXUniversity College London

[email protected]

A

B

C

B

Example 1: same language, different text

A

B

C

C

Example 2: different language, different textExample 1: same language, different textExample 0: same language, same text

http://www.moralesforensics.com/identification.htm

Macro Features

Micro Features

• Definitive, highly-cited publication.

• Basis of the CedarFox document analysis software package.

• Analyzed the uniqueness of the macro and micro features used by experts.

Task: pick out the sample by the same writer from a set of 200 writers(same language, same text).

Srihari et al. (CedarFox?)

Using expert-style macro features, 40% correct.

Using expert-style micro features, 86% correct.

With macro & micro features,91% correct.

Newell & Griffin

Using our macro features,95% correct(on same language, different text).

+2+2 +1

+1

+1

+1

Basic Image Features (BIFs)A computational model

of Human Vision

Our Primary Research Area…

Sand grain appearance under electron microscopy reveals distinctive, geographically specific textures

Basic Image Features make the texture differences obvious to a computer

…looked sufficiently promising when used to characterize the texture of sand grain images, that…

Electron Microscopy of Sand Grains weathered at different wind speeds

Newell, A. J., Morgan, R. M., Griffin, L. D., Bull, P. A., Marshall, J. R., & Graham, G. (2012). Automated Texture Recognition of Quartz Sand Grains for Forensic Applications. Journal of Forensic Sciences.

…the Atomic Weapons Establishment funded a project to investigate the forensic potential of our texture analysis method for particulate identification

Then, by chance, we came across the problem ofwriter identification…

Not knowing the importance experts attach to comparing individual letters we tried our texture approach…

We got a perfect score, crashed the website, and won the competition.

And retained our title last year in a harder competition with more entrants.

Summary

• Computers can be more accurate than experts…

• …by using methods that do not mimic those experts.

• The Newell-Griffin method is currently world-leading.

• It is based on a method of texture analysis.

• That also works for quartz grains, biomedical image analysis,general-purpose object recognition,…

• That derives from a model of how human vision works.

• We are keen to take advice on this handwriting work…

• …and to find other forensic applications of our image analysis/recognition expertise.