240 likes | 261 Views
Explore the differences between authentic handwriting and forgery, measure wrinkle patterns, and devise an automatic forgery detection model. Delve into the legal motivation and individuality of handwriting. Visit csis.pace.edu/~scha/handwriting.html for more details.
E N D
Automatic Detection of Handwriting forgery Dr. Sung-Hyuk Cha & Dr. Charlies C. Tappert School of Computer Science & Information Systems
Recognition Examination Personality identification (Graphology) On-line Off-line Writer Identification Writer Verification Natural Writing Forgery Disguised Writing Handwriting Analysis Taxonomy Analysis of Handwriting
Overview • Background • Differences b/w authentic handwriting & forgery • Measure of Wrinkliness • Automatic Forgery Detection Model • Conclusion
Legal Motivation To determine the Validity of Individuality in Handwriting Daubert vs. Merrell Dow (1993) testing, peer review, error rates Frye vs. US (1923) scientific community U.S. vs. Starzecpyzel (1995) “skilled” testimony GE vs. Joiner (1997) weight of evidence Kumho vs.Carmichael (1999) reliability standard
Individuality of Handwriting Each person writes differently.
Authentic vs. Forgeries (b) Forgeries of (a) (a) Authentic handwriting samples from one writer
3 Differences b/w authentic & forgery 1. Shape 2. Pressure 3. Speed
Angular and Magnitude Type Element String Image Stroke Direction Stroke Width Angular Magnitude
Stroke Width Extraction w7 w8 w9 w10 w1 w2 w3 w4 w5 w6 w7 w6 w5 w4 w3 w2 w1 2.83 2.83 2.83 5 3 6 5 4 5 5 4.24 4.24 4.24 2.83 4.24 4.24 4.24 min(wi) = 3 min(wi) = 2.83 (a) Vertical & horizontal stroke width (b) Diagonal stroke width
Fractal: How wrinkly is Handwriting? (b) (a) (a) Number of in the boundary = 69 (b) Number of in the boundary = 32
Computational Features (d-e) ascender & descender
Computational Features (f) stroke width (g-i) projected histogram and gradient histogram
Automatic Forgery Detection Model sample1 by x sample2 by x sample1 by x Forgery of x by y Feature Extractor Distance computing d-dimensional within-authentic- handwriting distance set d-dimensional between-authentic- handwriting & forgery distance set
Feature distances Truth Inputs & Truth cent slant wid zone side-h bot-h grad A A A A A A F F F F F F .49 .70 .71 .13 .47 .32 .21 .49 .75 .70 .26 .54 .35 .18 .49 .67 .74 .23 .48 .32 .22 .72 .33 .47 .66 .60 .42 .10 .74 .33 .48 .60 .59 .45 .10 .79 .36 .54 .60 .59 .52 .09 .30 .61 .66 .70 .71 .57 .10 .42 .72 .64 .67 .74 .53 .10 .40 .75 .67 .75 .70 .54 .11 .30 .60 .59 .66 .60 .36 .10 .32 .60 .59 .60 .59 .39 .10 .30 .66 .60 .60 .59 .34 .09
Artificial Neural Network Authentic sample from a known source Feature extraction Distance compu- tation Original/ Forgery? Handwriting sample in question
d ( , ) Decision boundary d ( , ) forgery identified authentic authentic identified as forgery Distributions and Errors within authentic distance between authentic & forgery distance
Design of Experiment between class within class Random selection 180 60 dichotomizer dichotomizer d-error d’-error s-error s’-error estimate
Conclusion • Authentic handwriting and forgery handwritten word images were collected. • Differences b/w authentic handwriting and forgery • Measure of Wrinkliness • Automatic Forgery Detection Model using the dichotomy approach. • Further quantitative study with more samples is necessary.
The End Thank you. http://www.csis.pace.edu/~scha/handwriting.html