The Research Of The Feature Fusion Algorithms In Multi-spectral Finger-vein

With the rapid development of electronic information technology, single-spectral finger-vein recognition technology has been unable to meet the demand of the social security and the individual information security. However, the optical characteristics of the biological tissue vary with the spectral of near infrared light, so the multi-spectral finger-vein images can present some different tissue information compared with the single-spectral finger-vein images. Therefore, the fused image will have more abundant information, which can achieve the purpose of perfecting finger-vein information and improving image quality.Due to the advantages of feature-level fusion in multispectral biometric-based technology, the fusion of multi-spectral finger-vein images are implemented at feature-level.Before the multi-spectral finger-vein images are fused, some relevant preprocessing(image restoration and image registration) will be done to better extract feature information. With respect to the fusion method at feature-level, we mainly put forward two kinds of different fusion strategies according to the research purpose of this subject, namely the fusion strategy based on image transform and the fusion strategy based on feature vector decomposition.In order to get the fused image and to establish the fused image gallery, Non-subsampled Contourlet Transform(NSCT) here is employed to accomplish feature-level fusion of multi-spectral finger-vein images. And a novel feature extraction algorithm based on NSCT coefficients classification is proposed. However, because the fusion speed is slow in transform domain and cannot meet the real-time requirement, another fusion strategy is put forward at feature-level. Firstly, finger-vein code is produced based on a unified Gabor filter framework.Then, in order to fuse two different spectral images at feature-level reasonably and effectively,the fused feature vectors from multi-spectral finger-vein images are generated utilizing sparsity preserving projections(SPP) and a novel weighted fusion strategy. Experimental results demonstrate that the proposed methods have a good performance in feature-level and multi-spectral finger-vein recognition.