Exploring Competitive Features Using Deep Convolutional Neural Network for Finger Vein Recognition

Abstract

Deep convolutional neural networks (DCNN) have been applied successfully for finger vein recognition and have achieved promising performance in the past three years. However, public finger vein datasets are scarce and tend to be relatively small, and thus unable to provide a substantial number of well-labeled images needed to train an effective convolutional neural network (CNN). In this paper, a CNN model pre-trained on ImageNet is adopted to develop a CNN-based local descriptor named CNN competitive order (CNN-CO) that can exploit discriminative features for finger vein recognition. The CNN filters from the first layer of the AlexNet network and the corresponding CNN filtered images are visualized and compared with Gabor filters. According to the appearances and outputs of the CNN filters in three color spaces, we select the ones that most resemble the Gabor filters. The selected CNN filters are employed to generate a competitive order (CO) image using the winner-take-all rule. Then, the pyramidal histograms calculated from the CO image in different levels are concatenated to build the final histogram. The extensive experimental results on two public finger vein datasets demonstrate the effectiveness of the proposed method in selecting CNN filters. The results show that the proposed CNN-CO scheme using the selected CNN filters outperforms the well-known local descriptors.