Catálogo de publicaciones - libros

This paper introduces Transductive Support Vector Machine (TSVM) into fingerprint verification. An improved fingerprint matching approach using TSVM is presented. In the proposed approach, the traditional minutiae-based fingerprint matching task is transformed to a classification task using TSVM. The paper presents an analysis of why TSVM are well suited for fingerprint matching, especially for small training sets. The approach is supported by experiments on five test collections, including both international and domestic fingerprint verification competition databases. Experimental results show that our approach is insensitive to noise as well as with effective performance.

It is a difficult and challenge task to extract the accurate orientation field for the low quality fingerprints. This paper proposed a robust orientation estimation with feedback algorithm to get the accurate fingerprint orientation. First, the fingerprint image is segmented into recoverable and unrecoverable regions. The following orientation field estimation and orientation correction algorithms were only processed in the recoverable regions. Second, the coarse orientation image is estimated from the input fingerprint image using the gradient-based approach. Then we computed the reliability of orientation from the gradient image. If the reliability of the estimated orientation is less than pre-specified threshold, the orientation will be corrected by the proposed mixed orientation model. The proposed algorithm has been evaluated on the databases of FVC2004. Experimental results confirm that the proposed algorithm is a reliable and effective algorithm for the extraction orientation field of the low quality fingerprints.

Unlike traditional minutiae-based fingerprint matching approach, which establishes the minutiae correspondence between two fingerprints, in this paper a novel ridge-based fingerprint matching algorithm, which establishes the exact ridge correspondence between two fingerprints is proposed. The ridge correspondence is called ‘exact’ because correspondences between points on ridges are also established. The matching score is computed as the ratio of the length of matched ridges to that of all ridges. Experimental results on FVC2002 databases show that the ridge matching approach performs comparably to minutiae-based one, and an obvious improvement on matching performance can be obtained by combining the two matchers.

Fingerprint enhancement is a crucial step in automatic fingerprint recognition system, undiscriminating repeated filtering easily lead to false structure in low-quality regions of a fingerprint image. This paper presents a new adaptive enhancement algorithm that can automatically adjust the parameters of filters and the time of filtering according to the quality factors in different regions. In order to improve filtering efficiency, a template bank of 4-dimenstion scaleable array is also designed to quantize the filter and forms the basis of a new fused filter. Experimental results in eight low-quality images from FVC2004 data sets show that the proposed algorithm is higher 23.7% in Good Index (GI), and saves 54.06% time consumptions than traditional Gabor-based methods. Since the eight images are extremely bad, a little improvement is very meaningful.

Automatic fingerprint classification is an effective means to increase the matching speed of an Automatic Fingerprint Identification System with a large-scale fingerprint database. In this paper, an automatic fingerprint classification method is proposed to classify the fingerprint image into one of five classes: Arch, Left loop, Right Loop, Whorl and Tented Arch. First, the information of core points, which is detected with a two-stage method, is applied to determine the reference point in fingerprint image. Then three different features based on statistical properties of small image blocks, which are likely to degrade with image quality deterioration, are calculated from the region of interest and form a 300-dimension feature vector. The feature vector is inputted into a three-layer Back Propagation Network (BPN) classifier and a 5-dimension vector is outputted, each dimension of which corresponds to one of 5 fingerprint classes. Finally, the fingerprints are classified with integrate analysis of the BPN classifier output and singular point information. The accuracy of 93.23% with no rejection is achieved on NIST-4 database and experimental results show that the proposed method is feasible and reliable for fingerprint classification.

Reliable detection of irises is a necessary precondition for any iris-based biometric identification system. It is also important in measuring the movement of eyes in applications such as alertness detection and eye-guided user interfaces. This paper introduces a novel iterative algorithm for the accurate and fast localization of irises in eye images.

Iris-based biometric recognition outperforms other biometric methods in terms of accuracy. In this paper an iris normalization model for iris recognition is proposed, which combines linear and non-linear methods to unwrap the iris region. First, non-linearly transform all iris patterns to a reference annular zone with a predefined , which is the ratio of the radii of inner and outer boundaries of the iris. Then linearly unwrap this reference annular zone to a fix-sized rectangle block for subsequence processing. Our iris normalization model is illuminated by the ‘minimum-wear-and-tear’ meshwork of the iris and it is simplified for iris recognition. This model explicitly shows the non-linear property of iris deformation when pupil size changes. And experiments show that it does better than the over-simplified linear normalization model and will improve the iris recognition performance.

In this paper, we propose a new feature extraction method based on independent component analysis (ICA) for iris recognition, which is known as the most reliable biometric system. We extract iris features using a bank of filters which are selected from the ICA basis functions. The ICA basis functions themselves are sufficient to be used as filter kernels for extracting iris features because they are estimated by training iris signals. Using techniques of the ICA estimation, we generate many kinds of candidates ICA filters. To select the ICA filters for extracting salient features efficiently, we introduce the requirements of the ICA filter. Each ICA filter has a different filter size and a good discrimination power to identify iris pattern. Also, the correlation between bandwidths of the ICA filters is minimized. Experimental results show that the EER of proposed ICA filter bank is better than those of existing methods in both the Yonsei iris database and CASIA iris database.

In this paper, we present an iris recognition system considering counterfeit attacks. The proposed system takes multi-spectral images instead of one infrared iris image. The energy of the multi-spectral images is checked and the authentication is failed if the amount of the energy is not in the proper range. Then the images are normalized and merged into a grayscale image by using a gradient-based image fusion algorithm. In the fusion process, the images considered to be from a counterfeited iris are merged into a poor-quality image which successively generates poor matching score. We show that the proposed scheme successfully maintains the performance of real iris images preventing the counterfeit attacks with experimental results.

In this paper, we propose an iris detection method to determine iris existence. The method extracts 4 types of features, i.e., contrast feature, symmetric feature, isotropy feature and disconnectedness feature. Adaboost is adopted to combine these features to build a strong cascaded classifier. Experiments show that the performance of the method is promising in terms of high speed, accuracy and device independence.