Offline signature verification using pre-trained deep convolutional neural network Available to Purchase

Handwritten signature is the most commonly employed behavioral biometric in financial, and legal institutions. Because of its simplicity, forgers may easily counterfeit signatures and conduct fraud. Therefore, the objective of this research is to utilize the advantages of machine learning approaches to recognize forged signatures. To attain this objective, we have used the Deep Convolution Neural Network (DCNN) to classify forged and genuine signatures. Though there are extremely few signature samples available for training, we have attempted to use one of the efficient pre-trained models to improve generalization capability and decrease time complexity. VGG16 CNN has been used out of the available pre-trained models to tackle the problem of handwritten signature verification in a writer-independent approach. VGG16 model is trained and tested using the CEDAR dataset. We conducted experiments to test the efficacy of the model on three different aspects. 1. Dataset Split - To study the effect of variations in train and test dataset split. 2. Number of signature samples - To study the efficacy of the model by varying the number of signature samples per signer,3. Hyper-parameter variation - To study the dependency of the model on various hyper-parameters. Finally, the outcome of these experiments shows that the 80:20 dataset ratio, Adam optimizer, the learning rate of 0.0001,15 epochs, and the number of signature samples equal to 20 give the best results.