Dual-path deep neural network architecture with explicit features for offline signature recognition

Abstract

In this work, with the aim of separating the genuine and forgery samples of the signature, we developed a new dual-path architecture using deep neural network and a traditional descriptor for feature extraction toward an automatic offline signature recognition. The proposed approach is an extended version of VGG-16, which is enhanced using our two paths architecture. In the first path, we explore features using a deep convolutional neural network, and in the second path, we discover global features using a traditional heuristic approach. For classical feature extraction, an innovative idea is presented, in which the descriptor is stable for some common changes, such as magnification and epoch, in the signature samples. Our traditional approach extracts global features that are stable with rotation and scaling. The proposed method was analyzed and compared with three well-known databases of CEDAR, UTsig, and GPDS signature images. A dual-patched model architecture is significantly more accurate than the basic model when compared to the basic model. In agreement with the proposed method, the best signature recognition accuracy on the CEDAR database is in the range of 98.04-99.96%, while the best recognition accuracy on the GPDS and UTsig databases is 98.04% and 99.56%, respectively. Furthermore, this technique has been compared with four popular methods such as VGG-S, VGG-M, VGG-16, and LS2Net. The presented approach achieved a recognition rate of 99.96% using a diverse signature database. Experimental results demonstrate that the proposed VGG-16 based signature recognition system is superior over texture-based and deep-learning methods and also outperforms the existing state-of-the-art results in this regard. It is expected that the proposed system will provide fresh acumen to the researchers in developing offline signature verification and recognition systems in other scripts.