Signature Forgery Detection using Convolutional Neural Networks

Signature of a person can uniquely identify the person and it is widely used in social situations and monetary transactions with individuals and financial entities. Many fraud cases have been appearing in society where signature of a person is forged for financial and other benefits. There is need for detecting forged signatures with technology driven approaches. With the emergence of Artificial Intelligence (AI), there are unprecedented possibilities in solving problems of the real world with responsible usage of AI. Deep learning (DL) is one part of AI which extends neural networks has become very significant in computer vision applications. From the existing approaches, it is observed that there is need for a complete framework for end to end processing of signatures for efficient detection of forgeries. Towards this end, we proposed a DL based framework for automatic detection of signature forgery. The framework is designed to leverage performance of the models. We proposed an algorithm known as Learning based Signature Forgery Detection (LbSFD) which exploits pipeline of multiple DL models such as CNN, VGG16 and Siamese. All the models are CNN variants used to improve efficiency in signature forgery detection. A benchmark signature dataset is used for our empirical study. Our experiments revealed that the CNN based models are highly efficient in signature forgery detection. Highest accuracy with 98.26% is achieved when VGG16 model is employed with transfer learning.