Generalizing convolutional neural networks for pattern recognition tasks

Abstract

Convolutional Neural Network (CNN) promises automatic learning and less effort for hand-design heuristics in building an efficient pattern recognition system. It requires simple and minimal preprocessing stages for data preparation. These features enable CNN architecture to be applied to various pattern recognition applications. This paper proposes a fourlayered CNN architecture that caters to face recognition and finger-vein biometric identification case studies. The methodology applied in designing the network is discussed in detail. For face recognition, the design is evaluated on three facial image databases with different levels of complexities. These databases are AT&T, AR Purdue, and FERET. The same four-layered CNN architecture is also tuned for finger-vein biometric identification problems. The design performance is evaluated on finger-vein biometric database developed in-house, consisting of 81 subjects. The results obtained from these case studies are promising. For face recognition applications, 100%, 99.5%, and 85.16% accuracies were obtained for AT&T, AR Purdue, and FERET, respectively. On the other hand, the result obtained from the finger-vein biometric identification case study has 99.38% accuracy. The results have shown that the proposed design is feasible for any pattern recognition problem