Generating copybooks from consistent handwriting styles
230 likes | 398 Views
Ralph Niels & Louis Vuurpijl Nijmegen Institute for Cognition and Information Radboud University Nijmegen The Netherlands. Generating copybooks from consistent handwriting styles. Overview. Handwriting styles and copybooks Method Results Discussion. Handwriting styles.
Generating copybooks from consistent handwriting styles
E N D
Presentation Transcript
Ralph Niels & Louis Vuurpijl Nijmegen Institute for Cognition and Information Radboud University Nijmegen The Netherlands Generating copybooks from consistent handwriting styles
Overview • Handwriting styles and copybooks • Method • Results • Discussion
Handwriting styles • Handwriting is individual • Similar handwritings: handwriting styles • Top down ‘copybooks’ * • We defined writing styles bottom up * S.-H. Cha, S. Yoon, C.C. Tappert, 2006.
Applications • Handwriting recognition • Personalized recognizers • Handwriting synthesis • ‘Handwriting fonts’ • Forensic writer identification • Human experts use the notion of style K. Franke, 2005
Method (1) Data Data
Data • Databases: • Unipentrainset • Unipendevset • Plucoll database • Online handwritten characters(pre-segmented) 43 writers 41 writers
Method (2) Data
Clustering of prototypes • The prototype we used areaveragedshapes of actualhandwrittencharacters L. Vuurpijl & L. Schomaker, Finding Structure in Diversity, ICDAR 1997. R. Niels, L. Vuurpijl & L. Schomaker, Automatic allograph matching inforensic writer identification, IJPRAI, Feb. 2007.
Clustering of prototypes PCi Prototype clusters PCj PCk Prototypes
Method (3) Data
Create membership vectors • Relative frequency of the occurrence of each prototype cluster in a persons handwriting
Create membership vectors(example: handwriting X) 1.0 PCi Prototype clusters PCj PCk 0.2 0.8 Prototypes 0.15 0.05 0.05 0.52 0.23
Create membership vectors(example: handwriting X, Y and Z) Handwriting Y 0.47 0.41 0.00 0.09 0.03 Handwriting X 0.15 0.05 0.05 0.52 0.23 Handwriting Z 0.12 0.01 0.22 0.55 0.10
Method (4) Data
Find writing styles • Hierarchical clustering of membership vectors (handwritings) Writing styles B H Z X E D I A G B J K Y C F Handwriting
Method (5) Data
Select consistent handwriting styles Cluster parameters PCi Level selection PCj PCk B H Z X E D I A G B J K Y C F Cluster parameters Level selection
Select consistent handwriting styles • Monte Carlo simulation of combinations of parameters and levels • Large number of writing styles • Find the writing styles that occur most • By prototypes or • By writers
Results • Copybooks • Preliminary results • Visual evaluation by handwriting experts • Meaningful names • Well-known broad categories: cursive, mixed and print
Results (example handwritings) Print Cursive Mixed
Results (prototype occurrence) Print Cursive Mixed
Discussion • Applied to/with, not limited to: • Online Latin characters • Dynamic Time Warping for character comparison (human congruous) • Best of both worlds: Integrate top down and bottom up (with forensic experts) B H Z X E D I A G B J K Y C F Integrate