Forgery detection using machine learning and image processing

Abstract

In an era where digital transactions and document authentication play a central role, the need for strong and secure methods to verify handwritten signatures is paramount. This thesis explores the intersection of traditional signature analysis and modern machine learning techniques to develop a reliable signature verification system. The survey begins with a comprehensive review of existing signature verification methods, including both conventional image processing techniques and advanced machine learning approaches. Challenges posed by the inherent variability of human signatures are identified, including differences in writing style and the presence of noise. The proposed solution leverages advanced machine learning algorithms, specifically adapted for signature recognition such as SVM, Decision Trees, kNN, and others. As in all cases of machine learning, in this one too, we use datasets. The datasets used for training and evaluation include a diverse collection of handwritten signatures, capturing variations in styles, angles, and levels of complexity. In addition, a key feature of this work is that feature extraction techniques are used to transform the raw signature data into meaningful representations for the machine learning models. For this purpose, four different feature sets of increasing complexity are proposed to improve the overall system performance.