System that Identifies Writers

Since the writer identification plays an important investigative and forensic role in many types of crime, various automatic techniques, feature extraction, comparison and performance evaluation methods have been studied (see [1] for the extensive survey). It has been practiced based on the hypothesis that people’s handwritings are as distinctly different from one another as their individual natures, as their own finger prints [2,3]. However, relatively little study has been carried out to demonstrate its scientific and statistical validity and reliability as forensic evidence [4] or to answer the question whether one can build a machine that can identify writers. For this reason, we present a simple model to establish the individuality of handwriting based on Hilton’s model. Hilton calculated the odds by taking the likelihood ratio statistic that is the ratio of the probability calculated on the basis of the similarities, under the assumption of identity, to the probability calculated on the basis of dissimilarities, under the assumption of non-identity [3,5]. There exist various parametric and non-parametric techniques to solve the multiple category classification problem or simply called polychotomizer where the number of classes is finite and small [6]. As the number of classes is too large to observe all (U.S. population), these techniques are of no use and the problem is seemingly insurmountable. For this reason, we suggest to transform a large and intractable polychotomizer to a simple dichotomizer, a classifier that places a pattern in one of only two categories: distance data between two writings of the same author and those of two different authors. In this model, one need not observe all classes and still allows the inferential classification. We state the problem as follows; given two randomly selected handwritten documents, the writer identification problem is to determine whether the two documents were written by the same person with two types of confusion error probabilities. To illustrate, suppose there are three writers, . Each writer provides three documents and two scalar value features extracted per document. Fig. 1 (a) shows the plot of two features from documents for every writer and Fig. 1 (b) represents the transformed plot in the two dimensional feature distance domain. Using eleven feature distance values, we trained an artificial neural network and obtained overall correctness.