Carlos Vivaracho-Pascual

An efficient low cost approach for on-line signature recognition based on length normalization and fractional distances

This work presents a new proposal for an efficient on-line signature recognition system with very low computational load and storage requirements, suitable to be used in resource-limited systems like smart-cards. The novelty of the proposal is in both the feature extraction and classification stages, since it is based on the use of size normalized signatures, which allows for similarity estimation, usually based on dynamic time warping (DTW) or hidden Markov models (HMMs), to be performed by an easy distance calculation between vectors, which is computed using fractional distance, instead of the more typical Euclidean one, so as to overcome the concentration phenomenon that appears when data are high dimensional. Verification and identification tasks have been carried out using the MCYT database, achieving an EER (common threshold) of 6.6% and 1.8% with skilled and random forgeries, respectively, in the first task and 3.6% of error in the second. The proposed system outperforms DTW-based and HMM-based ones, even though these have proved to be very efficient in on-line signature recognition, with storage requirements between 9 and 90 times lesser and a processing speed between 181 and 713 times greater than the DTW-based systems.

Practical On-Line Signature Verification

A new DTW-based on-line signature verification system is presented and evaluated. The system is specially designed to operate under realistic conditions, it needs only a small number of genuine signatures to operate and it can be deployed in almost any signature capable capture device. Optimal features sets have been obtained experimentally, in order to adapt the system to environments with different levels of security. The system has been evaluated using four on-line signature databases (MCYT, SVC2004, BIOMET and MyIDEA) and its performance is among the best systems reported in the state of the art. Average EERs over these databases lay between 0.41% and 2.16% for random and skilled forgeries respectively.

On the Use of Mobile Phones and Biometrics for Accessing Restricted Web Services

In this study, an application that allows a mobile phone to be used as a biometric-capture device is shown. The main contribution of our proposal is that this capture, and later recognition, can be performed during a standard web session, using the same architecture that is used in a personal computer (PC), thus allowing a multiplatform (PC, personal digital assistant (PDA), mobile phone, etc.) biometric web access. The review, which is from both an academic and commercial point of view, of the biometry and mobile device state of the art shows that in other related works, the biometric capture and recognition is either performed locally in the mobile or remotely but using special communication protocols and/or connection ports with the server. The second main contribution of this study is an in-depth analysis of the present mobile web-browser limitations; thus, it is concluded that, in general, it is impossible to use the same technologies that can be used to capture biometrics in PC platforms (i.e., Applet Java, ActiveX Control, JavaScript, or Flash); therefore, new solutions, as shown here, are needed.

Improving Automatic Classification of Prosodic Events by Pairwise Coupling

This paper presents a system that automatically labels tones and break indices (ToBI) events. The detection (binary classification) of prosodic events has received significantly more attention from researchers than its classification because of the intrinsic difficulty of classification. We focus on the classification problem, identifying eight types of pitch accent tones, nine types of boundary tones and five types of break indices. The complex multi-class classification problem is divided into several simpler problems, by means of pairwise coupling. We propose to combine two-class classifiers to achieve the multi-class classification because two-class problems provide high accuracy results. Furthermore, complementarity between artificial neural networks and decision trees classifiers has been exploited to improve the final system, combining their outputs using a fusion method. This proposal, together with the adequate feature extraction that includes the use of features such as the Tilt and Bézier parameters, allows us to achieve a total classification accuracy of 70.8% for pitch accents, 84.2% for boundary tones and 74.6% for break indices, on the Boston University Radio News Corpus. The analysis of the misclassified samples shows that the types of mistakes that the system makes do not differ significantly from the common confusions that are observed in manual ToBI inter-transcriber tests.

A fuzzy classifier to deal with similarity between labels on automatic prosodic labeling

This paper presents an original approach to automatic prosodic labeling. Fuzzy logic techniques are used for representing situations of high uncertainty with respect to the category to be assigned to a given prosodic unit. The Fuzzy Integer technique is used to combine the output of different base classifiers. The resulting fuzzy classifier benefits from the different capabilities of the base classifiers for identifying different types of prosodic events. At the same time, the fuzzy classifier identifies the events that are potentially more difficult to be labeled. The classifier has been applied to the identification of ToBI pitch accents. The state of the art on pitch accent multiclass classification reports around 70% accuracy rate. In this paper we describe a fuzzy classifier which assigns more than one label in confusing situations. We show that the pairs of labels that appear in these uncertain situations are consistent with the most confused pairs of labels reported in manual prosodic labeling experiments. Our fuzzy classifier obtains a soft classification rate of 81.8%, which supports the potential of the proposed system for computer assisted prosodic labeling.

Improving ANN performance for imbalanced data sets by means of the NTIL technique

This paper deals with the problem of training an Artificial Neural Network (ANN) when the data sets are very imbalanced. Most learning algorithms, including ANN, are designed for well-balanced data and do not work properly on imbalanced ones. Of the approaches proposed for dealing with this problem, we are interested in the re-sampling ones, since they are algorithm-independent. We have recently proposed a new under-sampling technique for the two-class problem, called Non-Target Incremental Learning (NTIL), which has shown a good performance with SVM, improving results and training speed. Here, the advantages of using this technique with ANN are shown. The performance with regard to other popular under-sampling techniques is compared.

Client threshold prediction in biometric signature recognition by means of Multiple Linear Regression and its use for score normalization

Abstract Biometric person authentication has become an important area of fieldwork both for research and commercial purposes in the last few years. The development of the technology, now ready for practical applications, has encouraged the scientific community to focus on practical issues. In this sense, a key question is the decision threshold estimation. Biometric authentication is a pattern recognition problem where a final decision (identity accepted/rejected) must be taken; so, to set a correct decision threshold is essential, since the best system becomes useless if an inaccurate decision threshold is fixed. This work focuses on this subject for biometric systems based on manuscript signatures. The decision threshold can be client (signatory) dependent or the same for all (common threshold). In this paper, new approaches for both problems are shown. A new solution, based on the Multiple Linear Regression model, is proposed for client dependent decision threshold estimation or prediction. The state of the art shows that only independent variables based on the Gaussian scores distribution supposition have been used. Here, new robust parameters, not based on that supposition, have been successfully included in the model. This proposal has been evaluated by means of both a statistical validation and a performance comparison with the state of the art. When a common threshold is used, the problem is to normalize the client scores. A new proposal for this task is also shown, based on the use of the predicted client threshold. Both proposals have been multi-working point, multi-corpus and multi-classifier tested. Improvements from 12% to 57% have been achieved with respect to the state of the art in threshold prediction, while these improvements are from 15% to 40% in the score normalization task.

A new approach for a priori client threshold estimation in biometric signature recognition based on multiple linear regression

This paper presents a novel approach to estimate (predict) the a priori client decision threshold for biometric recognition systems based on multiple linear regression. Biometric recognition is a complex classification problem where the goal is to classify a pattern (biometric sample) as belonging or not to a certain class (client). As in other pattern recognition problems, a correct estimation of the decision threshold is essential for optimizing the biometric systems performance. Our proposal is tested in biometric signature recognition, estimating thresholds for different system working points. A theoretical and practical performance analysis is presented, including a comparison with the state of the art, showing the advantages, in system performance, of our proposal.

A New Proposal for Score Normalization in Biometric Signature Recognition Based on Client Threshold Prediction

Score Normalization is a usual technique in pattern recognition to standardize the classifier output ranges so as to, for example, fuse these outputs. The use of score normalization in biometric recognition is a very important part of the system, specially in those based on behavioral traits, such as written signature or voice, conditioning the final system performance. Then, many works can be found that focus on the problem. A successful new approach for client threshold prediction, based on Multiple Linear Prediction, has been presented in recent works. Here, a new approach for score normalization, based on this proposal for biometric manuscript signature user verification, is shown. This proposal is compared with the state of the art methods, achieving an improvement of 19% and 16% for Equal Error Rate (EER) and 60% and 26% for Detection Cost Function (DCF) performance measures, for random and skilled forgeries, respectively.

On the Use of Score Ratio with Distance-Based Classifiers in Biometric Signature Recognition

Biometric user verification or authentication is a pattern recognition problem that can be stated as a basic hypothesis test: X is from client C (\(H_0\)) vs. X is not from client C (\(H_1\)), where X is the biometric input sample (face, fingerprint, etc.). When probabilistic classifiers are used (e.g., Hidden Markov Models), the decision is typically performed by means of the likelihood ratio: \({P(X/H_0)}/{P(X/H_1)}\). However, as far as we know, this ratio is not usually performed when distance-based classifiers (e.g., Dynamic Time Warping) are used. Following that idea, we propose, here, to perform the decision based not only on the score (“score” being the classifier output) supposing X is from the client (\(H_0\)), but also using the score supposing X is not from the client (\(H_1\)), by means of the ratio between both scores: the score ratio. A first approach to this proposal can be seen in this work, showing that to use the score ratio can be an interesting technique to improve distance-based biometric systems. This research has focused on the biometric signature, where several state of the art systems based on distance can be found. Here, the score ratio proposal is tested in three of them, achieving great improvements in the majority of the tests performed. The best verification results have been achieved with the use of the score ratio, improving the best ones without the score ratio by, on average, 24 %.