Research On Finger Vein Image Retrieval And Fusion Methods

The amount of finger vein images increases dramatically over recent years as the applications of finger vein recognition system are in wide use.The rapid retrieval of large-scale finger vein images has become a key problem to be solved in the field of finger vein research.So far there are few works aiming at large-scale finger vein image indexing and retrieval,therefore efficient finger vein image indexing and retireval is still a challenging problemThis thesis aims at the problem of large-scale finger vein image retrieval and carries out intensive exploration and research.We also make the first attempt to integrate finger vein image with Electrocardio(ECG)signal.The main work and innovations are as follows1.A finger vein image retrieval method based on non-negative locality-constrained vocabulary tree is proposed.To decrease the large quantization error in traditional vocabulary tree model,the vector quantization coding method of vocabulary tree is improved.By introducing non-negative constraint,only the nearest neighbor words located in the same subspace as the local feature to be encoded are selected,which improves the coding accuracy.Experiments have been conducted on the self-established large-scale finger vein fusion database Merged-FV2040.The results show that,compared with the traditional vocabulary tree model and other existing finger vein image retrieval methods,the proposed method achieves higher retrieval accuracy and efficiency2.The finger vein image feature based on bag of subspaces model and nonlinear subspace coding method is proposed.In traditional bag of words model,the generation of codebook does not consider geometric structure characteristics,therefore the codebook of visual words is extended to the codebook of subspaces Each local feature is mapped to top k neighbor subspaces rather than top k visual words,then the finger vein image features are extracted by nonlinear subspace coding method.Experiments are conducted on the fusion database Merged-FV2040,and the results show that compared with the feature based on traditional bag of words model,the proposed feature obtains stronger discrimination power and characterization capability.3.A finger vein image retrieval method based on affinity-preserving k-means hashing algorithm is proposed.The algorithm constructs codebook and encodes feature vectors simultaneously,and encodes the encoding feature vector with binary index of visual words which are closest to this vector,thereby the feature vectors in high-dimensional Euclidean space are effectively mapped into compact binary codes.Since the similarity of two feature vectors is not determined by the Euclidean distance of binary indexes of corresponding visual words,but by the hamming distance of these binary indexes,the matching speed is greatly accelerated.The experimental results on the fusion database Merged-FV2040 show that the proposed method achieves high retrieval accuracy while maintaining low computational complexity.4.A finger vein image retrieval method based on binary hash codes learning is proposed.For the finger vein image retrieval method based on affinity-preserving k-means hashing algorithm,class label information is not leveraged,therefore the proposed method combines binary hash codes learning and class label information and adds the discirminitivity and the stability constraints to the binary coding learning model to improve the recognition ability and compactness of the learned binary codes.Experimental results on the finger vein public benchmark database FVPolyU and the fusion database Merged-FV2040 verify the high retrieval accuracy and efficiency of the proposed method.5.A multimodal biometric identification system based on the fusion of finger vein and ECG signal is proposed.So far there have been no multimodal fusion research based on finger vein image and ECG signal,thus the first attempt is made to study the fusion of finger vein and ECG signal,and a variety of score-level and feature-level fusion strategies are introduced in the multimodal system.Experimental results on a virtual multimodal dataset consists of finger vein and ECG unimodal datasets show that,compared the two unimodal system,multimodal recognition system has achieved a higher recognition accuracy and security,and validate the strong complementarity between the two unimodal biometrics and the effectiveness of multimodal fusion.