Catálogo de publicaciones - libros

We propose a new classifier combination scheme for the ensemble of classifiers. The Pairwise Fusion Matrix (PFM) constructs confusion matrices based on classifier pairs and thus offers the estimated probability of each class based on each classifier pair. These probability outputs can then be combined and the final outputs of the ensemble of classifiers is reached using various fusion functions. The advantage of this approach is the flexibility of the choice of the fusion functions, and the experiments suggest that the PFM combined with the majority voting outperforms the simple majority voting scheme on most of problems.

Because of the lack of a clear guideline or technique for selecting classifiers which maximise diversity and accuracy, the development of techniques for analysing classifier relationships and methods for generating good constituent classifiers remains an important research direction. In this paper we propose a framework based on the Bayesian Belief Networks (BBN) approach to classification. In the proposed approach the multiple-classifier system is conceived at a meta-level and the relationships between individual classifiers are abstracted using Bayesian structural learning methods. We show that relationships revealed by the BBN structures are supported by standard correlation and diversity measures. We use the dependency properties obtained by the learned Bayesian structure to illustrate that BBNs can be used to explore classifier relationships, and for classifier selection.

Classifier combining rules are designed for the fusion of the results from the component classifiers in a multiple classifier system. In this paper, we firstly propose a theoretical explanation of one important classifier combining rule, the sum rule, adopting the Bayes viewpoint under some independence assumptions. Our explanation is more general than what did in the existed previous by Kittler [1]. Then, we present a new combining rule, named SumPro rule, which combines the sum rule with the product rule in a weighted average way. The weights for combining the two rules are tuned according to the development data using a genetic algorithm. The experimental evaluation and comparison among some combining rules are reported, which are done on a biometric authentication set. The results show that the SumPro rule takes a distinct advantage over both the sum rule and the product rule. Moreover, this new rule gradually outperforms the other popular trained combining rules when the classifier number increases.

Error Correcting Output Coding (ECOC) is an established technique to face a classification problem with many possible classes decomposing it into a set of two class subproblems. In this paper, we propose an ECOC system with a reject option that is performed by taking into account the confidence degree of the dichotomizers. Such a scheme makes use of a coding matrix based on Low Density Parity Check (LDPC) codes that can also be usefully employed to implement an iterative recovery strategy for the binary rejects. The experimental results have confirmed the effectiveness of the proposed approach.

Face as a biometric is known to be sensitive to different factors, e.g., illumination condition and pose. The resultant degradation in face image quality affects the system performance. To counteract this problem, we investigate the merit of combining a set of face verification systems incorporating image-related quality measures. We propose a fusion paradigm where the quality measures are quantised into a finite set of discrete , e.g., “good illumination vs. “bad illumination”. For each quality state, we design a fusion classifier. The outputs of these fusion classifiers are then combined by a weighted averaging controlled by the probability of a quality state given the observed quality measures. The use of quality states in fusion is compared to the direct use of quality measures where the density of scores and quality are jointly estimated. There are two advantages of using quality states. Firstly, much less training data is needed in the former since the relationship between base classifier output scores and quality measures is not learnt jointly but separately via the conditioning quality states. Secondly, the number of quality states provides an over the complexity of the resulting fusion classifier. In all our experiments involving XM2VTS well illuminated and dark face data sets, there is a improvement in performance over the baseline method (without using quality information) and the direct use of quality in two types of applications: as a quality-dependent score normalisation procedure and as a quality-dependent fusion method (involving several systems).

A biometric system produces a matching score representing the degree of similarity of the input with the set of templates for that user. If the score is greater than a prefixed threshold, then the user is accepted, otherwise the user is rejected. Typically, the performance is evaluated in terms of the Receiver Operating Characteristic (ROC) curve, where the correct acceptance rate is plotted against the false authentication rate. A measure used to characterise a ROC curve is the Area Under the Curve (AUC), the larger the AUC, the better the ROC. In order to increase the reliability of authentication through biometrics, the combination of different biometric systems is currently investigated by researchers. In this paper two open problems are addressed: the selection of the experts to be combined and their related performance improvements. To this end we propose an index to be used for the experts selection to be combined, with the aim of the AUC maximisation. Reported results on FVC2004 dataset show the effectiveness of the proposed index.

The use of quality measures in pattern classification has recently received a lot of attention in the areas where the deterioration of signal quality is one of the primary causes of classification errors. An example of such domain is biometric authentication. In this paper we provide a novel theoretical paradigm of using quality measures to improve both uni- and multimodal classification. We introduce  − , a classifier stacking method in which feature similarity scores obtained from the first classification step are used in ensemble with the quality measures as features for the second classifier. Using two-class, synthetically generated data, we demonstrate how  −  helps significantly improve both uni- and multimodal classification in the presence of signal quality degradation.

We present three new voting schemes for multi-classifier biometric authentication using a reliability model to influence the importance of each base classifier’s vote. The reliability model is a meta-classifier computing the probability of a correct decision for the base classifiers. It uses two features which do not depend directly on the underlying physical signal properties, verification score and difference between user-specific and user-independent decision threshold. It is shown on two signature databases and two speaker databases that this reliability classification can systematically reduce the number of errors compared to the base classifier. Fusion experiments on the signature databases show that all three voting methods (rigged majority voting, weighted rigged majority voting, and selective rigged majority voting) perform significantly better than majority voting, and that given sufficient training data, they also perform significantly better than the best classifier in the ensemble.

Matching systems can be used in different operation tasks such as verification task and identification task. Different optimization criteria exist for these tasks - reducing cost of acceptance decisions for verification systems and minimizing misclassification rate for identification systems. In this paper we show that the optimal combination rules satisfying these criteria are also different. The difference is caused by the dependence of matching scores produced by a single matcher and assigned to different classes. We illustrate the theory by experiments with biometric matchers and handwritten word recognizers.

Ensembles are often capable of greater predictive performance than any of their individual classifiers. Despite the need for classifiers to make different kinds of errors, the majority voting scheme, typically used, treats each classifier as though it contributed equally to the group‘s performance. This can be particularly limiting on unbalanced datasets, as one is more interested in complementing classifiers that can assist in improving the true positive rate without signicantly increasing the false positive rate. Therefore, we implement a genetic algorithm based framework to weight the contribution of each classifier by an appropriate fitness function, such that the classifiers that complement each other on the unbalanced dataset are preferred, resulting in significantly improved performances. The proposed framework can be built on top of any collection of classifiers with different fitness functions.