Spatial Pyramid Pooling of Selective Convolutional Features for Vein Recognition

Abstract

Deep neural network (DNN) has demonstrated astounding performance in large-scale image recognition task, and pre-trained DNN models trained for one task have also been applied to domains different from their original purposes. Following such an idea, a novel hand-dorsa vein recognition model is constructed by adopting DNN pre-trained on a large-scale database as a universal feature descriptor. Unlike most of these studies which adopt activations of the fully connected layer of DNN as the image representation, we adopt convolutional activations as the region representation. However, not local features of all regions are equally important for final classification. Thus, to solve this issue, a novel selective convolutional feature model based on spatial weighting is proposed to acquire more robust and discriminative feature representation. In specific, a spatial weighting scheme is applied on the convolutional activations to weigh the importance of local features at different regions for classification, which further enhances the discriminability of feature representation. Besides, to take full advantage of the spatial information of the convolutional activations, spatial pyramid pooling is introduced to obtain feature representation with rich spatial information. The final image representation is formed by concatenating the local features of different level of the spatial pyramid. Series rigorous experiments on the lab-made database are conducted to evidence the effectiveness and feasibility of the proposed model. What is more, an additional experiment with subset of PolyU database illustrates its generalization ability and robustness.