Handwritten Signature Recognition: A Convolutional Neural Network Approach

Abstract

Handwritten Signature Recognition is an important behavioral biometric which is used for numerous identification and authentication applications. There are two fundamental methods of signature recognition, on-line or off-line. On-line recognition is a dynamic form, which uses parameters like writing pace, change in stylus direction and number of pen ups and pen downs during the writing of the signature. Off-line signature recognition is a static form where a signature is handled as an image and the author of the signature is predicted based on the features of the signature. The current method of Off-line Signature Recognition predominantly employs template matching, where a test image is compared with multiple specimen images to speculate the author of the signature. This takes up a lot of memory and has a higher time complexity. This paper proposes a method of off-line signature recognition using Convolution Neural Network. The purpose of this paper is to obtain high accuracy multi-class classification with a few training signature samples. Images are preprocessed to isolate the signature pixels from the background/noise pixels using a series of Image processing techniques. Initially, the system is trained with 27 genuine signatures of 10 different authors each. A Convolution Neural Network is used to predict a test signature belongs to which of the 10 given authors. Different public datasets are used to demonstrate effectiveness of the proposed solution.