Catálogo de publicaciones - libros

This paper presents a novel framework of error-correcting output coding (ECOC) addressing the problem of multi-class classification. By weighting the output space of each base classifier which is trained independently, the distance function of decoding is adapted so that the samples are more discriminative. A criterion generated over the Extended Pair Samples (EPS) is proposed to train the weights of output space. Some properties still hold in the new framework: any classifier, as well as distance function, is still applicable. We first conduct empirical studies on UCI datasets to verify the presented framework with four frequently used coding matrixes and then apply it in RoboCup domain to enhance the performance of agent control. Experimental results show that our supervised learned decoding scheme improves the accuracy of classification significantly and betters the ball control of agents in a soccer game after learning from experience.

Pairwise coupling is a popular multi-class classification approach that prepares binary classifiers separating each pair of classes, and then combines the binary classifiers together. This paper proposes a pairwise coupling combination strategy using individual logistic regressions (ILR-PWC). We show analytically and experimentally that the ILR-PWC approach is more accurate than the individual logistic regressions.

We show that the novelty detection approach is a viable solution to the class imbalance and examine which approach is suitable for different degrees of imbalance. In experiments using SVM-based classifiers, when the imbalance is extreme, novelty detectors are more accurate than balanced and unbalanced binary classifiers. However, with a relatively moderate imbalance, balanced binary classifiers should be employed. In addition, novelty detectors are more effective when the classes have a non-symmetrical class relationship.

In this paper we present a novel multistage classification strategy for handwriting Chinese character recognition. In training phase, we search for the most representative prototypes and divide the whole class set into several groups using prototype-based clustering. These groups are extended by nearest-neighbor rule and their centroids are used for coarse classification. In each group, we extract the most discriminative feature by local linear discriminant analysis and design the local classifier. The above-mentioned prototypes and centroids are optimized by a hierarchical learning vector quantization. In recognition phase, we first find the nearest group of the unknown sample, and then get the desired class label through the local classifier. Experiments have been implemented on CASIA database and the results show that the proposed method reaches a reasonable tradeoff between efficiency and accuracy.

In this article we extend the (recently published) unsupervised information theoretic vector quantization approach based on the Cauchy–Schwarz-divergence for matching data and prototype densities to supervised learning and classification. In particular, first we generalize the unsupervised method to more general metrics instead of the Euclidean, as it was used in the original algorithm. Thereafter, we extend the model to a supervised learning method resulting in a fuzzy classification algorithm. Thereby, we allow fuzzy labels for both, data and prototypes. Finally, we transfer the idea of relevance learning for metric adaptation known from learning vector quantization to the new approach.

A modal symbolic classifier for interval data is presented. The proposed method needs a previous pre-processing step to transform interval symbolic data into modal symbolic data. The presented classifier has then as input a set of vectors of weights. In the learning step, each group is also described by a vector of weight distributions obtained through a generalization tool. The allocation step uses the squared Euclidean distance to compare two modal descriptions. To show the usefulness of this method, examples with synthetic symbolic data sets are considered.

Hough transform neural network is adopted to detect line pattern of direct wave and hyperbola pattern of reflection wave in a seismogram. The distance calculation from point to hyperbola is calculated from the time difference. This calculation makes the parameter learning feasible. The neural network can calculate the total error for distance from point to patterns. The parameter learning rule is derived by gradient descent method to minimize the total error. Experimental results show that line and hyperbola can be detected in both simulated and real seismic data. The network can get a fast convergence. The detection results can improve the seismic interpretation.

This paper describes an evidence-theoretic classifier which employs global k-means algorithm as the clustering method. The classifier is based on the Dempster-Shafer rule of evidence in the form of Basic Belief Assignment (BBA). This theory combines the evidence obtained from the reference patterns to yield a new BBA. Global k-means is selected as the clustering algorithm as it can overcomes the limitation on k-means clustering algorithm whose performance depends heavily on initial starting conditions selected randomly and requires the number of clusters to be specified before using the algorithm. By testing the classifier on the medical diagnosis benchmark data, iris data and Westland vibration data, one can conclude classifier that uses global k-means clustering algorithm has higher accuracy when compared to the classifier that uses k-means clustering algorithm.

This paper proposed a new method of extracting texture features based on Gabor wavelet. In addition, the application of these features for bark classification applying radial basis probabilistic network (RBPNN) has been introduced. In this method, the bark texture feature is firstly extracted by filtering the image with different orientations and scales filters, then the mean and standard deviation of the image output are computed, the image which have been filtered in the frequency domain. Finally, the obtained Gabor feature vectors are fed up into RBPNN for classification. Experimental results show that, first, features extracted using the proposed approach can be used for bark texture classification. Second, compared with radial basis function neural network (RBFNN), the RBPNN achieves higher recognition rate and better classification efficiency when the feature vectors have low-dimensions.

Handwritten Chinese Address Recognition describes a difficult yet important pattern recognition task. There are three difficulties in this problem: (1) Handwritten address is often of free styles and of high variations, resulting in inevitable segmentation errors. (2) The number of Chinese characters is large, leading low recognition rate for single Chinese characters. (3) Chinese address is usually irregular, i.e., different persons may write the same address in different formats. In this paper, we propose a comprehensive and hybrid approach for solving all these three difficulties. Aiming to solve (1) and (2), we adopt an enhanced holistic scheme to recognize the whole image of words (defined as a place name) instead of that of single characters. This facilitates the usage of address knowledge and avoids the difficult single character segmentation problem as well. In order to attack (3), we propose a hybrid approach that combines the word-based language model and the holistic word matching scheme. Therefore, it can deal with various irregular address. We provide theoretical justifications, outline the detailed steps, and perform a series of experiments. The experimental results on various real address demonstrate the advantages of our novel approach.