Using Perturbed Handwriting to Support Writer ID in the Presence of Severe Data Constraints

Usually we assume large amount of data for writer identification, what if only little data is available, e.g., 1 page, 1 paragraph, or even 1 line? Several reasonable assumptions in forensics: • The writers are no longer available. • The writers might not be collaborative. The question is then: what can we do in this situation? • Use of synthetic data has been studies in fields like: signature verification, handwriting recognition. • In general, there are two methodologies involved: 1. Model Generated Handwriting (MGH)-build models to simulate people' s handwriting behavior [1,2]. 2. Model Perturbed Handwriting (MPG)-build models to manipulate people' s real handwriting, also known as deformation [3]. In this work, we only study model perturbed handwriting. Issues Techniques for handwriting recognition and writer ID share a lot in common, but: • Handwriting recognition engines strive for inter-writer commonalities, i.e., the characters/words people write. • Writer ID engines strive for inter-writer variances, i.e., the idiosyncratic styles people write in. Varga and Bunke propose a MPH model for handwriting recognition where four deformation methods are presented [1] :