Research On Finger Biometric Transformation And Hypersphere Granulation

Traditional unimodal biometrics cannot meet the requirement for accurate authentication because of having poor universality.Therefore,to improve the effectiveness and reliability of identification,multimodal based biometrics has attracted increasing attentions.However,effectively fusing multimodal biometric is difficult due to the varieties in image scale and feature dimension.To deal with this problem,this thesis proposes a feature transformation method and construct a hypersphere granule fusion model.The proposed method is verified with finger vein,finger print and finger-knuckle-print as testing data.The main works of this thesis are as follows:(1)Two methods of feature transformation are implemented to deal with the problem of scale variety.1)In spatial domain,the images of different modalities are divided into image blocks with the same number by uniform division or super-pixel division.Then,features with identical dimension are extracted from each image block.2)A neighborhood feature matrix(NFM)generation method is proposed.The method can transform Gabor maximum response images with different scales into their corresponding NFMs with same size.Thus,the data dimension and the computational cost can be reduced effectively.(2)A hypersphere granule clustering fusion model is constructed to fuse the multimodal finger features.In this model,the union operator and threshold operation are utilized to fuse the three unimodal atomics to a hypersphere granule.The fusion granules are used for identification.Moreover,the triangular fusion model is used to further prove the effectiveness of hypersphere granule in multimodal feature fusion.Extensive experiments are conducted to verify the proposed methods.The experimental results show that combining NFM and the hypersphere granule fusion model can achieve higher recognition accuracy than traditional methods.