Research On Hand-metrics Recognition For Single Sample Biometrics

Biometrics is one of the most secure and convenient way to satisfy the requirements for identity digitalization and virtualization in the coming network society, which refers to the automatic identification of an individual by using certain physiological or behavioral traits associated with the person. In some real-world tasks, for example, in the law enforcement scenarios, smart card, passport, only one image per person may be available for training, which may lead to bad recognition result. How to improve the performance of single sample biometrics has become the focus of recent studies. Multimodal biometrics, which utilizes multiple biometric traits, is considered to be efficient alternative for single sample biometrics.This dissertation studies the hand-metrics, which refers to palmprint, kunckleprint, finger features and fusion strategy, to improve the performance of system with single sample for training. First, we present several algorithms for palmprint and finger-based recognition to explore more discriminant features for single sample biometrics. Then, matching score level fusion can greatly improve the performance of single sample biometrics. Finally, a hierarchical recognition strategy based on multiple hand-metrics fusion is proposed, and the evaluation results validate its effectiveness. The main content of this dissertation includes the following aspects:1. Analyze the performances of subspace methods used in palmprint identification, and propose a feature weighted principal component analysis (FWPCA). For PCA based palmprint identification, the system could achieve better performance when removing some principal components. We find the chief reason for such situation is the magnitudes with larger eigenvalues are extraordinarily greater than others, which restrains the abilities of principal components with other eigenvalues. The proposed FWPCA can achieve better performance with lesser features by weighting principal components. In addition, we give a feature weighted locality preserving projections (FWLPP) by adding a graph incorporating neighborhood structure of the palmprint space. The experimental results demonstrate the effectiveness of FWPCA and FWLPP. Finally, analyze the generalization capability of PCA for palmprint recognition, and find it can be improved by using enough samples for training. 2. A wavelet sub-bands fusion scheme for single sample palmprint recognition is proposed. Wavelet decomposition has been applied in palmprint recognition successfully. However, only the low frequency sub-band was used for further feature extraction, while the high frequency sub-bands were considered to be unsuitable for palmprint recognition due to their sensitivity to noise and shape distortion. We utilize mean filtering to enhance the robustness of the high frequency sub-bands. Experimental results show that the performances of the horizontal and vertical high frequency sub-bands can be promoted up to a competitive level, and the fusion scheme, which combines the matching scores of high frequency sub-bands with that of low frequency sub-band, improves the performance of single sample palmprint recognition greatly, and is superior to the conventional recognition methods.3. Personal identification algorithms based on finger features are explored. A new location algorithm of inner knuckleprint is proposed. The rigid deformation property of finger is used to keep rotation invariance. High-pass filter with horizontally is used to enhance the line features in finger image. Radon projection verifies the translational invariance, and the location of inner knuckleprint in original finger image is accurately located. Subspace and texture analysis are utilized for features extraction, and matching score fusion can further extend the recognition performance. Furthermore, finger-based biometrics, which includes inner knuckleprint and finger shape features, would give more discriminant information. We extract PCA and LPP features based on index finger, middle finger, ring finger and little finger. The experimental results demonstrate the effectiveness of finger-based biometrics in recognition accuracy even with single sample for training. They provide the basis for hand-metrics fusion.4. An evaluation method of fusion operator for matching score fusion is proposed. It can provide a theoretical support to evaluate the performance of fusion operators. Fusion operator, which contains sum rule, product rule, max rule and min rule, is common and simple scheme for information fusion and is considered as special case of compound classification based on the Bayesian framework. They are usually used and validated in most multimodal biometric systems, but the optimal operator is obtained experimentally in real-world systems. The proposed method for optimal fusion operator selection by first estimating the probability density function (PDF) of each feature score and then calculating the PDFs of fusion operators on the assumption that the representations used are conditionally statistically independent. The distance between the class of genuine and impostor, which is based on the theory of probability density distribution, can be used to evaluate the capability of fusion operators. The results of40experiments based on Hand database validate the proposed method.5. A hierarchical hand-metrics identification scheme is proposed for single sample biometrics. At the first level, the matching scores of middle finger and little finger features are fused for recognition. Most of individuals are recognized accurately, and few of them will fall into following levels. At the second level, the matching scores of index finger and ring finger features are fused for recognition. At the third level, the matching scores of palmprint wavelet sub-bands features are fused for recognition. If some individuals can not be recognized, they will fall into the fourth level, and all above scores will be fused for final identification. The experimental results demonstrate the proposed hierarchical identification scheme could perform almost perfectly for single sample biometrics under laboratory conditions.