Analysis of the Off-line Signature Verification Techniques

Abstract

A signature is an integral part of a persons individual identity and a mark of his true self. It is the term used to mean the depiction of someone's signing, name or even a single letter which can be any letter or an "A" which is written by a person on a document. While signatures have a widespread acceptance by the public and a huge importance in niche applications like validating documents, validating papers, banking applications and many more makes the signatures an interesting tool of veri cation. But at the same time it imparts a huge amount of risk in case the signatures of a person are forged so as to access his bank accounts and trespass his property and wealth by the e orts of a skilled forger or intruder. Thus, signature veri cation has gained recognition from the past three decades so that no frauds could be accomplished by the use of a persons signature. Even with the presence of skilled signature veri cation algorithms some forgeries are out of scope owing to the skill of the forger or to the change a persons signature witnesses to even a slight extent every time the person signs on a piece of paper. Hence, nowadays signature veri cation is often combined with the biometric ngerprint veri cation because two people can produce same signatures but cannot have same ngerprints at all. Still, improvements in the eld of signature veri cation are required and similar kind of work has been proposed in this study where the dataset consisting of both forged and genuine signatures has been used to train various models of machine learning so that the predictability of the fact that which signature is forged and which is genuine is improved to a greater extent. Comparisons have been done on the basis of the situation as to where data preprocessing and dimensionality reduction have been applied and where only the simple data has been used without any changes in the values of the attributes used. The signatures have been converted into vectors on the grounds of the strength of the pixel value in the image of the signature and this study has been applied only on the assessment of o -line signatures. It has shown the maximum e ciency in the terms of accuracy in the case of subspace ensemble of kNN classi er.