Handwriting Analysis

CSAFE built a prototype model that uses glyph frequencies derived from a questioned document using the handwriter package to estimate the posterior probabilities of authorship from among a set of writers. Expanding and enhancing this model is a high priority. We are continuing to collect standard writing samples from a large set of writers to use in this research project. The data will be publicly available to support the efforts of other researchers.

Document examiners are often asked to determine the source of a handwritten document. The type of document to which we refer includes ransom notes, faked legal documents and other such documents, where the information about the source is contained in the handwriting itself rather than in the content of the document.

At present, document examiners rely on visual comparisons and subjective assessments of the similarity between two handwriting samples. They focus on attributes such as the width and length of loops, the crossing of t’s, the overall slant of the characters and such to determine whether a specific person might have been the author of the document.

Several years ago, a proprietary software called FLASH ID® [1] was introduced in the market. FLASH ID decomposes writing into graphical structures called graphemes, which are then clustered into a large number of deterministic groups depending on their morphology and other characteristics. From a large but closed set of reference documents, FLASH ID® selects those that are most like a question document using a set of quantitative criteria.

Researchers in CSAFE have produced a different program for the analysis of handwriting that is in the public domain. Handwriter, like FLASH ID®, decomposes writing into graphical structures called glyphs but uses a different set of rules to do so. Furthermore, glyphs in a document are allocated to 40 groups using a stochastic clustering algorithm called k-means, which appears to be more robust than a deterministic grouping approach. Also with a closed set of writers (that can be very large), the frequency with which a writer contributes glyphs to each group, and the attributes of the glyphs in the clusters are then used to compute the probability that anyone of the writers in the reference set may have authored the questioned document.

Initial results are promising; using a set of 90 writers recruited nationwide to participate in a study organized by CSAFE, we were able to correctly establish authorship of a questioned document with about 98% accuracy. At present, authorship is established if the posterior probability of writership is 0.9 or higher.

While there are many cases in which comparing a questioned document to samples from a closed set of writers is adequate (e.g., a bomb threat note found in a school that must have been written by an unknown student in the school), extending this approach to the more general case is an objective of the work we propose here. We also aim to explore the effect of image quality on the results of the writership determinations based on the data provided by handwriter.