Catálogo de publicaciones - libros

In this paper, a class of GMM-based discriminative kernels is proposed for speaker identification. We map an utterance vector into a matrix by finding the sequence of components, which have the maximum likelihood in the GMM for the all frame vectors. And the weights matrix was used, which were got by the GMM’s parameters. Then the SVMs are used for classification. A one-versus-rest fashion is used for the c class problem. Results on YOHO in text-independent case show that the method can improve the performance greatly compared with the basic GMM.

In this paper, Individual Dimension Gaussian Mixture Model (IDGMM) is proposed for speaker identification. As to the training-purpose feature vector series of a certain register, its joint probability distribution function (PDF) of is modeled by the product of the PDF of each dimension (marginal PDF), the scalar-based Gaussian Mixture Model (GMM) serving as the marginal PDF. For a good discriminative capability, the decorrelation by Schmidt orthogonalization and the Mixture Component Number (MCN) decision are adopted during the train. A close-set text-independent speaker identification experiment is also given. The simulation result shows that the IDGMM accelerates the training process remarkably and maintains the discriminative capability in testing process.

Several works related to information fusion for signature verification have been presented. However, few works have focused on sensor fusion and sensor interoperability. In this paper, these two topics are evaluated for signature verification using two different commercial Tablet PCs. An enrolment strategy using signatures from the two Tablet PCs is also proposed. Authentication performance experiments are reported by using a database with over 3000 signatures.

Function-based methods for on-line signature verification are studied. These methods are classified into local and regional depending on the features used for matching. One representative method of each class is selected from the literature. The selected local and regional methods are based on Dynamic Time Warping and Hidden Markov Models, respectively. Some improvements are presented for the local method aimed at strengthening the performance against skilled forgeries. The two methods are compared following the protocol defined in the Signature Verification Competition 2004. Fusion results are also provided demonstrating the complementary nature of these two approaches.

Handwriting recognition is a traditional and natural approach for personal authentication. Compared to signature verification, text-independent writer identification has gained more attention for its advantage of denying imposters in recent years. Dynamic features and static features of the handwriting are usually adopted for writer identification separately. For text-independent writer identification, by using a single classifier with the dynamic or the static feature, the accuracy is low, and many characters are required (more than 150 characters on average). In this paper, we developed a writer identification method to combine the matching results of two classifiers which employs the static feature (texture) and dynamic features individually. Sum-Rule, Common Weighted Sum-Rule and User-specific Sum-Rule are applied as the fusion strategy. Especially, we gave an improvement for the user-specific Sum-Rule algorithm by using an error-score. Experiments were conducted on the NLPR handwriting database involving 55 persons. The results show that the combination methods can improve the identification accuracy and reduce the number of characters required.

Gait is a biometric feature and gait recognition has become a challenging problem in computer vision. New wavelet velocity moments have been developed to describe and recognize gait. Wavelet moments are translation, scale and rotation invariant. Wavelet analysis has the trait of multi-resolution analysis, which strengthens the analysis ability to image subtle feature. According with the psychological studies, reflective symmetry features are introduced to help recognition. Combination of wavelet velocity moments and reflective symmetry not only has the characteristic of wavelet moments, but also reflects the person’s walking habit of symmetry. Experiments on two databases show the proposed combined features of wavelet velocity moments and reflective symmetry are efficient to describe gait.

Interest in automated biometrics continues to increase, but has little consideration of time. This paper deals with a problem of recognition by gait when time-dependent and time-invariant covariates are added. We have shown previously how recognition rates fall significantly for data captured over lengthy time intervals. We suggest predictive models of changes in gait due both to time and now to time-invariant covariates. A considerable improvement in recognition capability is demonstrated, with potential generic biometric application.

Application and research of ear recognition technology is a new subject in the field of biometrics recognition. Earlier research showed that human ear is one of the representative human biometrics with uniqueness and stability. Feasibility and characteristics of ear recognition was discussed and recent advances in 2D and 3D domain was presented. Furthermore, a proposal for future research topics was given, such as ear database generation, ear detection, ear occluding problem and multimodal biometrics with face etc.

In this paper, we present a new biometric identification system based on dental features. First, we collected the plaster figures of teeth which were done at dental hospital, department of oral medicine. Second, we developed a representation of dental images based on principal component analysis(PCA) representation included the 100 principle components as the features for individual identification. The PCA basis vectors had reflected well the features for individual identification from the parts of teeth. Finally, the classification for individual identification based on the nearest neighbor(NN) algorithm from the part of teeth was created. The identification performance is 98% for the part of teeth excluded the right-molar and back teeth, 95% for the part of teeth excluded the front teeth and 80% for the part of teeth excluded the left-molar and back-teeth.

A multimodal biometric scheme using watermarking technique to provide more secure and reliable personal recognition is proposed in this paper. Two distinct biometric traits have been under consideration: palmprint and knuckleprint. The palmprint image is chosen to be the host image. Knuckleprint biometric feature is selected to use as watermark hidden in the host image. Such that knuckleprint watermark not only protects palmprint biometric data, but also can be used as a covert recognition. Meanwhile, the bimodal biometrics recognition provides the improvement in the accuracy performance of the system. The experiment results demonstrate the effectiveness of the proposed method.