My Master Research

My Research Interests lie in Machine Learning, Pattern Recognition, Machine Vision, Deep Learning, Transfer Learning, Multitask Learning, Metric Learning. For my master's thesis, I worked on Offline Signature Verification, specifically Offline Persian Signatures under supervision of Babak N Araabi in Machine Learning and Computational Modeling Lab (MLCM).

Handwritten signature is a well-known behavioural characteristic which is currently used as an authentication method. The verification procedure is done by Forensic Handwriting Experts (FHEs) which is a time consuming job. So that an automatic signature verification system can accelerate and even relieve mistakes from this well-accepted biometric process.

Automatic signature verification is a complicated system because signature is not a completely repetitive pattern. Moreover, although genuine ones obey some rules, training a system is hard due to sample deficiency whether genuine (positive) or forged (negative) ones.

In the literature, signature is divided into offline and online. Offline is just images but online has extra information such as temporal, speed and pressure information. Our interest is offline signature verification since it is more natural, more applicable and more challenging.

We collected an offline Persian signature dataset, University of Tehran Signature Dataset (UTSig) which is accessible from the main page and here. The paper introducing UTSig can be found in the publications. Our paper presenting a Deep Multitask Metric Learning network, specifically for signatures have been published here. Bellow you can see its graphical abstract. Moreover, our paper investigating curvature and gradient feature extraction for Persian signature has been accepted in ICCKE2016 and indexed in IEEE.

Master Thesis (Abstract): Offline Persian Signature Verification

This project investigates the problem of offline Persian signature verification. Defining the verification system as a classification problem, we address three main parts of the system, namely dataset, feature extraction, and classifier. For the dataset, we collected and published University of Tehran signature dataset (UTSig) which is a new and rich offline Persian signature dataset. In feature extraction part, Histogram of Oriented Gradients (HOG) and Histogram of Curvature (HOC) were used for the first time in the Persian signature verification and generally signature verification literature, respectively. It is the first time that the curvature information of all parts of offline signatures are considered for the verification purpose. For the classifier, Discriminate Deep Metric Learning (DDML) was employed for the first time in the signature verification literature, and a new method, Deep Multitask Metric Learning (DMML) was introduced. Results of experiments on UTSig show that HOG and HOC cause promising performance for verifying genuine, skilled, and random forged samples. It is also shown that HOC works better than HOG and two other features, fixed-point arithmetic and discrete Radon transform. Results of comparing DMML against SVM and writer-dependent and writer-independent structure of DDML on four Persian and non-Persian offline datasets, indicate better performance for DMML in terms of Equal Error Rate (EER). Moreover, using DMML on GPDSsynthetic and GPDS960GraySignatures datasets, we achieve the best results of skilled EER compared to existing results in the offline signature literature.

Download first two page of the thesis.