Document Identification using Allographic Features

Abstract

Signatures which can be captured in an electronic tablet are called online signatures. In this project a pen based tablet was used to capture data from different user, and these were stored as .mat file in a database. The x and y coordinates of the signature were of main interest.From the x and y coordinates, features were extracted which were unique to a writer.Feature extraction is a process in which useful data is collected from a huge pile of data. These reduced data set should be informative, non-redundant and it should describe the signature of an individual in a unique way. To enable signature identification data needs to be collected from the signature, these collected and processed data is called feature extraction.
The features extracted were:
• Angle with reference to the first point.
• Length with reference to the first point.
• Curvature
• Mapping
The above features were chosen because it was observed that these features can uniquely
define a signature even if the signature has been rotated and they gave an unique pattern to
every signature.
These extracted features were given to a classifier. A classifier is a model which can
group similar data together, each group of similar data is called a class, so each signature in this thesis is a class.
Types of classifiers used in this thesis are:
• Decision Tree
• Bagged Tree
• K Nearest Neighbour
The accuracy of these classifiers against the extracted features were compared. It was found
that the mapping feature surpassed the angle, length and curvature feature.