Bidirectional Aggregated Features Fusion From Cnn For Palmprint Recognition

In this paper, we present a novel bidirectional aggregated features representation from convolutional neural networks (CNNs) with score-level fusion for palmprint recognition. Our method adopts the vector of locally aggregated descriptors (VLAD) to encode the convolutional features from two directions, i.e., vertical and horizontal directions, to mine both the local and global descriptions of palmprint image. Then, three score-level fusion rules are respectively employed to integrate the matching scores of the bidirectional features. We extensively evaluate the performance of convolutional features, vertical and horizontal encoding together with the score-level fusion rules through recent deep network VGG-F on the PolyU palmprint and multispectral palmprint databases. Experiments demonstrate that horizontal encoding significantly outperforms vertical encoding on red, green, blue and near-infrared (NIR) palmprint image subsets while it is slightly worse on PolyU palmprint database, moreover, the effective performance improvement can be achieved after the fusions.