Convolutional neural networks for face recognition and finger-vein biometric identification

Abstract

The Convolutional Neural Network (CNN), a variant of the Multilayer Perceptron (MLP), has shown promise in solving complex recognition problems, particularly in visual pattern recognition. However, the classical LeNet-5 CNN model, which most solutions are based on, is highly compute-intensive. This CNN also suffers from long training time, due to the large number of layers that ranges from six to eight. In this research, a CNN model with a reduced complexity is proposed for application in face recognition and finger-vein biometric identification. A simpler architecture is obtained by fusing convolutional and subsampling layers into one layer, in conjunction with a partial connection scheme applied between the first two layers in the network. As a result, the total number of layers is reduced to four. The number of feature maps at each layer is optimized according to the type of image database being processed. Consequently, the numbers of network parameters (including neurons, trainable parameters and connections) are significantly reduced, essentially increasing the generalization ability of the network. The Stochastic Diagonal Levenberg-Marquadt (SDLM) backpropagation algorithm is modified and applied in the training of the proposed network. With this learning algorithm, the convergence rate is accelerated such that the proposed CNN converges within 15 epochs. For face recognition, the proposed CNN achieves recognition rates of 100.00% and 99.50% for AT&T and AR Purdue face databases respectively. Recognition time on the AT&T database is less than 0.003 seconds. These results outperform previous existing works. In addition, when compared with the other CNN-based face recognizer, the proposed CNN model has the least number of network parameters, hence better generalization ability. A training scheme is also proposed to recognize new categories without full CNN training. In this research, a novel CNN solution for the finger-vein biometric identification problem is also proposed. To the best of knowledge, there is no previous work reported in literature that applied CNN for finger-vein recognition. The proposed method is efficient in that simple preprocessing algorithms are deployed. The CNN design is adapted on a finger-vein database, which is developed in-house and contains 81 subjects. A recognition accuracy of 99.38% is achieved, which is similar to the results of state-of-the-art work. In conclusion, the success of the research in solving face recognition and finger-vein biometric identification problems proves the feasibility of the proposed CNN model in any pattern recognition system.