Catálogo de publicaciones - libros

This paper introduces a prediction method for time series that is based on the multi-resolution analysis of wavelets (MRA). The MRA is better able to decompose the non-stationary time series of nonlinear systems into different components, allowing a better separation of the general trend terms, the periodic terms and the random fluctuation terms. By applying the most suitable prediction methods(for example, the neural networks method, cosine approximation, or the ARMA model) to the components under different resolutions, this new prediction method produces more accurate prediction results. The new approach is then applied to a real example – the BRENT oil price time series – to demonstrates its usefulness and validity.

Independent component analysis (ICA) has been applied in many fields of signal processing and many ICA learning algorithms have been proposed from different perspectives. However, there is still a lack of a deep mathematical theory to describe the ICA learning algorithm or problem, especially in the cases of both super- and sub-Gaussian sources. In this paper, from the point of view of the one-bit-matching principle, we propose two adaptive matching learning algorithms for the general ICA problem. It is shown by the simulation experiments that the adaptive matching learning algorithms can efficiently solve the ICA problem with both super- and sub-Gaussian sources and outperform the typical existing ICA algorithms in certain aspects.

Computational methods detecting the transcription factor binding sites (TFBS) remain one of the most intriguing and challenging subjects in bioinformatics. Gibbs sampling is essentially a heuristic method, and it is easy to trap into a nonoptimal “local maximum”. To overcome this problem and to improve the accuracy and sensitivity of the algorithm, we present an improved Gibbs sampling strategy MPWMGMS to search for TFBS. We have tested MPWMGMS and other existing Gibbs sampling algorithms on simulated data and real biological data sets with regulatory elements. The results indicate that MPWMGMS has better performance than other methods to a great extent in accuracy and sensitivity of finding true TFBS.

In this paper, a novel Fisher criterion is introduced and shown to be equivalent to the traditional Fisher criterion. Based on this new Fisher criterion and simultaneous diagonalization technique, a -subspace Fisher discriminant (-SFD) method is developed to deal with the small sample size (S3) problem in face recognition. The proposed method overcomes some drawbacks of existing LDA based algorithms. Also, our method has good computational complexity. Two public available databases, namely ORL and FERET databases, are exploited to evaluate the proposed algorithm. Comparing with existing LDA-based methods in solving the S3 problem, the proposed -SFD method gives the best performance.

In this paper, an emotion recognition system for mandarin speech is presented. Five basic human emotions including angry, fear, happy, neutral and sad are investigated. The recognizer is based on neural network with OCON and ACON architecture. Some novel feature selection methods are also added as optional tool to enhance the efficiency and classification accuracy. The system can train speaker dependent emotion speech model through online emotional utterance recording. Experiment results show that emotion can be recognized through neural network model, the best mean accuracy is 86.7%. In addition, the feature selection module is effective to reduce the compute load and increase the generalization ability of the recognizer.

In this paper, we propose ubiFloorII, a novel floor-based user identification system to recognize humans based on their walking pattern such as stride length, dynamic range, foot angle, and stance and swing time. To obtain users walking pattern from their gait, we deployed photo interrupter sensors instead of switch sensors used in ubiFloorI. We developed a software module to extract walking pattern from users’ gait. For user identification, we employed neural network trained with users’ walking samples. We achieved about 96% recognition accuracy using this floor-based approach. The ubiFloorII system may be used to automatically and transparently identify users in home-like environments.

When a radial-basis function neural network (RBFNN) is used for pattern classification, the problem involves designing the topology of RBFNN and also its centers and widths. In this paper, we present a particle swarm optimization (PSO) learning algorithm to automate the design of RBF networks, to solve pattern classification problems. Simulation results for benchmark problems in the pattern classification area show that the PSO-RBF outperforms two other learning algorithms in terms of network size and generalization performance.

Patchouli is a common used traditional Chinese herbal medicine for flatulence or vomit with a long history in China and some other Asian countries. Patchoulis of different geographical origins usually have different curative effects. In order to evaluate their qualities, it’s very necessary to discriminate them. Our objective of this study is to develop a nondestructive and accurate identification method. The results in this paper showed that it’s difficult to use conventional infrared spectra and second derivative spectra directly to distinguish them. But it’s quite easy to use two-dimensional infrared (2D IR) correlation spectra, especially after applying wavelet transform process, to discriminate them. The resolution of the 2D IR spectrum is improved obviously after wavelet transform process, more peaks appear and the peaks become quite clear and separate. The differences of the 2D IR spectra become rather remarkable. In this way, it’s not difficult to discriminate the patchoulis of different geographical origins. This will be a nondestructive, economical and rapid way to distinguish complicated mixture like traditional Chinese medicines. Combined 2D IR and wavelet transform, Fourier transform infrared spectroscopy would become more powerful in analysis and discrimination.

This paper presents a new approach for human identification at a distance using gait recognition. Binarized silhouette of a motion object is represented by 1-D signals which are the basic image features called the distance vectors. The distance vectors are differences between the bounding box and silhouette, and extracted using four view directions to silhouette. Based on normalized correlation on the distance vectors, gait cycle estimation is first performed to extract the gait cycle. Second, eigenspace transformation based on PCA is applied to time-varying distance vectors and then Mahalanobis and normalized Euclidean distances based supervised pattern classification is finally performed in the lower-dimensional eigenspace for human identification. Experimental results on two main database demonstrate that the right person in top two matches 100% of the times for the cases where training and testing sets corresponds to the walking styles for data set of 25 people, and other data set of 22 people.

This paper presents a framework of HMM parameter adaptation technique for improving automatic speech recognition (ASR) performance in the noisy environments, which online combines the clean hidden Markov models (HMMs) with the noise model. Based on the given composite HMM corresponding to the initial recognition pass result and truncated vector Taylor series, the noise model in the cepstral domain is updated and refined using iterative Expectation-Maximization (EM) algorithm under maximum likelihood (ML) criterion. Experiments results show that the presented approach in this paper is found to greatly improve recognition performance under mismatched conditions.