| On many occasions of daily life, we often need to express our own identity to others in order to obtain certain permissions. For example, when logging into the operating system, we prove our identity by entering a user name and password to obtain computer use privileges; when checking-in at the airport, we prove our identity by showing our ID to airport security officers in order to obtain the permission to board on a plane; when accessing to the customs, we prove our identity through iris recognition or fingerprint identification to customs officers in order to obtain immigration permits, etc. Therefore, how to perform personal identity authentication effectively, conveniently and quickly, protect people's legitimate rights and interests, and ensure legitimacy and effectiveness of various social activities have become a key social issue that must be resolved quickly.Authentication is to prove our own identity to others in order to obtain certain rights through a variety of technical or non-technical means. Traditional authentication methods can be categorized as knowledge-based authentication (such as passwords) and device-based authentication (such as keys, ID). However, these methods mainly depend on external objects, once the password is leaked, whose identities may be pretended (or substituted) by others. Thus, the traditional authentication methods, which have brought a lot of inconveniences and potential security problems to our lives, become difficult to meet the needs of society.Biological characteristics (such as iris, retina and fingerprint), due to their advantages such as unique, invariant, unable to be forgotten and lost, difficult to forge and steal, have become a new way to perform personal identity authentication with broad application prospects. Biometrics is a technology which is used to identify an individual based on the physiological (such as face, iris and fingerprint) or behavioral (such as speech, signature and gait) characteristics. Presently, there are several different characteristics which are widely used for the personal identification, including face, iris, retina, palmprint, fingerprint, speech, signature and gait. We can construct the unimodal biometric system by using one of these characteristics (modalities). Also, we can construct the multimodal biometric system by utilizing two or more individual characteristics (modalities). Since biometrics has incomparable advantages over traditional authentication methods, it has been widely used in financial services, human–computer interaction, video surveillance, information security, forensic identification and other fields in recent years. On the other hand, there is a misconception in many people's mind that biometrics is already mature. On the contrary, there are many problems that still need to be resolved in biometrics, and biometrics is still a challenging and important research topic. For example: low-quality, small-area and distorted fingerprint matching problem in fingerprint identification; open problems in face recognition; small sample size problem and spoof attacks in unimodal biometric technology, and so on.Since there are many unresolved problems in biometrics, this thesis has studied some key issues including orientation field estimation from low-quality fingerprint image, super-resolution from low-quality face image, popularity problem of unimodal biometric algorithm, robustness of multimodal biometric technology.The major research contents and results of this thesis are as follows:(1) We proposed a fingerprint orientation field estimation based on the primary and secondary ridges within the fingerprint block. As a global feature of fingerprints, the orientation field plays an important role in automatic fingerprint identification systems. Although many algorithms have been proposed for orientation field estimation, the results are not very satisfactory and the computational cost is expensive. In this paper, a novel algorithm based on the primary and secondary ridges within the fingerprint block is proposed for the orientation field estimation. The algorithm comprises four steps, preprocessing original fingerprint image, determining the primary and secondary ridges of fingerprint foreground block using the top semi-neighbor searching algorithm, estimating block direction based on straight-line model of such a primary ridge and correcting the spurious block directions. The proposed algorithm is suitable for almost all types of fingerprints. Experimental results show that it achieves satisfying estimation accuracy with high computational efficiency. A further experiment shows that it is more accurate and robust to noise compared with the previous works and can improve the performance of the fingerprint recognition system, even on low-quality fingerprint databases.(2) We proposed a correlative two-step approach to hallucinating faces. Face hallucination is to synthesize high-resolution face image from the input low-resolution one. Although many two-step learning-based face hallucination approaches have been developed, they suffer from the expensive computational cost due to the separate calculating of the global and local models. To overcome this problem, we propose a correlative two-phase learning-based face hallucination approach which bridges a connection between global model and local model. In the first step (global phase), we build a global face hallucination framework by combining the steerable pyramid decomposition and the reconstruction. In the second step (residue compensation phase), based on the combination weights and constituent samples obtained in the global phase, a residue face image is synthesized by the neighbor reconstruction algorithm to compensate the hallucinated global face image with detailed facial features. The ultimate hallucinated face image is the composition of the global face image and the residue face image. Compared with existing approaches, in the global phase, our global face image is more similar to the original high-resolution face image. Moreover, in the residue compensation phase, we use combination weights and constituent samples obtained in the global phase to compute the residue face image, by which the computational complexity can be greatly reduced without compromising the quality of facial details. The experimental results and comparisons demonstrate that our approach can not only synthesize distinct high-resolution face images efficiently, but also has high computational efficiency. Furthermore, our proposed approach can be used to restore the damaged face images in image inpainting. The efficacy of our proposed approach is validated by recovering the seriously damaged face images with visually good results.(3) We built a unimodal biometric system based on local topology structure preserving projections. In traditional unimodal biometric systems, they usually utilize specially designed recognition algorithms for different biological characteristics. Therefore, the traditional unimodal biometric systems have popularity problem. This paper proposes a unified unimodal biometric system that is suitable for most individual modalities, e.g., face, palmprint and gait. The proposed system consists of three steps: (i) preprocessing raw biometric data, (ii) determining the intrinsic low-dimensional subspace of preprocessed data by local topology structure preserving projections (LTSPP) and (iii) performing the classification in the determined subspace using the intra-class distance sum. In the proposed system, LTSPP is a novel subspace algorithm which focuses on not only the class information but also the local topology structure. In terms of representing the separability of different classes, LTSPP projects the inter-class margin data far apart. Meanwhile, LTSPP preserves the intra-class topology structures by using linear reconstruction coefficients. In comparison with other subspace algorithms, LTSPP possesses more discriminant abilities and is more suitable for biometric recognition. In addition, both preprocessing each raw datum into unit and performing the classification using the intra-class distance sum are helpful to improve the recognition rates. We carry out various recognition experiments using the Yale and HumanID gait databases. The encouraging experimental results demonstrate the effectiveness of our unified unimodal biometric system, and the proposed LTSPP algorithm for this system can yield the best recognition rates than the other algorithms.(4) We proposed a multimodal gender recognition method using Bayesian hierarchical model. To achieve a robust and discriminative performance for gender recognition, we propose to estimate human gender from fingerprint and corresponding face information with Bayesian hierarchical model. Different from previous work of fingerprint based gender estimation that needs specially designed features, our bag-of-words model, composed of a set of visual words, is employed to structure fingerprint and face images representations respectively. In this model, we propose a novel supervised method to construct the visual words, by which the redundant feature dimensions are discarded and the important dimensions for gender classification are highlighted. In addition, such an image representation for each modality can be naturally embedded into a generative framework, Bayesian hierarchical model, for gender recognition purpose. For each modality, we train the generative models for both categories, male and female. By computing the likelihood of the two generative models, one can estimate the category label of this modality.We obtain the final recognition result by fusing different modalities at the decision level. Experiments on a large set of fingerprints and face database demonstrate the effectiveness of the proposed method of feature representation and new model. Complementary advantages from fingerprint-face fusion has benefited to our gender recognition.
|
PDF Full Text Request