Double Biologically Inspired Transform Network for Robust Palmprint Recognition
Kaijun Zhou,Xiancheng Zhou,Lingli Yu,Lizhi Shen,Shaoqian Yu
DOI: https://doi.org/10.1016/j.neucom.2018.07.083
IF: 6
2019-01-01
Neurocomputing
Abstract:Most state-of-the-art methods in recognizing rotated, scaled, and translated (RST) palmprint image have difficulty to achieve the tradeoff between accuracy and robustness. Inspired by vision processing mechanic, we propose a double biologically inspired transform (DBIT) network. DBIT involves two stages, and each stage has several layers. In the first stage, we first build a convolution layer to highlight the edges at all direction and interval. Then, we construct another two convolution layers to measure local spatial frequency (LSF), which can improve the selectivity of feature maps. Since there is no negative spike in biological neurons, a rectified linear unit and sum pooling are employed to produce a neuron spike. Then, each spike is normalized by their original orientation edge maps, and finally, all spikes are combined to a feature vector. In the second stage, two convolution layers including orientation edge detection and single LSF followed by several subsequent operations are performed once again, which transforms image shift to an invariant spike on final feature maps. Experiments on PolyU multispectral, PolyU II, CASIA, IITD, and COEP database illustrate the effectiveness and robustness of the proposed approach. Remark progresses in feature invariance and selectivity are achieved by the proposed method. Comparing with several existing methods, our approach can achieve better recognition performance under rotation, scaling, translation, and even noise interference. (c) 2019 Elsevier B.V. All rights reserved.