A Computational Forensic Approach to the Analysis of Questioned Document Fragments

Fragments of documents are common subjects in forensic analysis of questioned documents. Forensic analysis of torn document is more challenging owing to sparse data content; for example, a document fragment might consist of only part of a word. The degree of difficulty increases when large number of such documents needs to be analyzed. A forensic expert might overlook evidences in this huge pool of data. This dissertation aims to help combat this problem by studying scientific methodologies that can narrow down the search space of a forensic expert. Automatic sorting of document fragments can be accomplished based on criteria set by the forensic expert. This demands execution of the following :(i) text/graphics segmentation;(ii) segmentation of text type (printed/handwritten);(iii) script identification of text;(iv) identification of the writer; (v) identifying the font of the printed text. Adopting various image processing and pattern recognition techniques certain methodologies are proposed for accomplishing such tasks. Rigorous experiments have been carried out to evaluate our scientific methodologies with real life torn document fragments. Feature encoding techniques have been meticulously chosen so that discriminative properties between different objects of interest are well represented, making the classification task easier. For e.g. in case of writer identification we have implemented a feature encoding scheme that reveals variations in character shape structures between different writers. The thesis consists of 10 chapters. A brief overview of every chapter is as follows: • Chapter 1 discusses the topic, and the challenges associated with it. This chapter also provides the motivation behind the research addressed in this thesis. In the beginning it briefs the problem and then provides a logical explanation about how the research aims to solve the problem of sorting torn document fragments. Later, it mentions about the contributions of this thesis followed by a brief description of all chapters in ”Thesis Outline” Section. • Chapter 2 states various background theories that have been used as the basis of solutions proposed in this thesis. Our analysis revealed that sorting of similar torn document fragments can be accomplished by exploiting information on characteristics of its content type like script and font of printed text, writer of handwritten text, etc.. Background knowledge related to these topics are discussed in this chapter. We have extensively used Support Vector Machine (SVM) classifier in all experiments, hence a theoretical discussion on with SVM is also provided. • Chapter 3 presents the existing state-of-the-art methodologies on relevant subproblems that we need to deal in order to accomplish our objective, for example we narrate here existing state-of-the-art methodologies on the following topics : (a) Text/graphics segmentation ; (b) Script identification ; (c) Writer identification ; (d) Font identification. • Chapter 4 provides a conclusion and direction towards future research. • Chapter 5 is based on an article ”Document-Zone Classification in Torn Documents” and is devoted to the problem of text/graphics segmentation and text type discrimination i.e. printed and handwritten text identification in torn document fragments.