A Novel Bank Check Signature Verification Model using Concentric Circle Masking Features and its Performance Analysis over Various Neural Network Training Functions

Background: Handwritten signature is a person's unique identity. Signature verification is an economical biometric method with online and offline schemes. This paper deals with the offline verification of signatures found in bank checks. Method: Extracting feature is the most vital part of a signature verification process. An efficient feature extraction method, Concentric Circles Masking Method, is used to extract robust, scale invariant and rotation invariant features. The extracted feature values are normalized and fed to a feedforward back propagation neural network for classification of the signatures into genuine or forged ones. The feature's performance is measured with various training functions of the neural network. The system modeled is tested with the well-known CEDAR database. Findings: Experimental Analysis shows that the features extracted by this method prove to be efficient. The scanned signature is covered by concentric circles and the pixel distribution ratio in each circle is calculated and used for verification purpose. Since a circle is used, the extracted features are scale and rotation invariant which makes the feature robust. The neural network's training, validation and testing ratio are varied and the performance of various training functions is studied. It is inferred that conjugate gradient back propagation with Fletcher-Reeves updates (traincgf) training function has the maximum average accuracy of 97.89% for the CCMM features.