Offline Signature Verification Using Convolutional Neural Network

Abstract

Offline handwritten signature verification is widely used important form of biometrics. It is a challenging task due to time-variant nature of signature. To address the above difficulty, a new approach is proposed in this paper to compute the features of signatures. The proposed approach is divided into two parts: 1) writer-independent approach, 2) writer-dependent approach. Writer-independent approach is utilized for fine-tuning of VGG16 convolutional neural network (CNN). In writer-dependent approach, this fine-tuned CNN is utilized to extract the features from the signature. The signature is passed through this fine-tuned CNN and the vector obtained at first fully connected layer (after last convolutional layer) is used as feature vector. To obtain the accurate features for classification of signatures, the pixels of thinned signature image are replaced by their Gaussian Weighting Based Tangent Angle (GWBTA) in both writer-independent and writer-dependent approach. The computed features which are obtained in writer-dependent approach are fed to Support Vector Machine (SVM) classifier to classify the signature into genuine or forgery class. The performance results obtained on different databases of offline handwritten signatures confirms the validity of proposed methodology for offline signature verification.