Catálogo de publicaciones - libros

This paper describes characteristics of adaptive vector quantization according to the equinumber principle. Three methods of adaptive vector quantization are presented with the objective of avoiding the initial dependency of reference vectors. The present approaches which have output units without neighboring relations equalize the numbers of inputs in a partition space. The first approach is a creation method which sequentially creates output units to reach a predetermined number of neurons founded on the equinumber principle in the learning process. The second is a reduction method which sequentially deletes output units to reach a prespecified number. The third is an unification method of the creation and reduction methods, which deletes units after creating under the predetermined number. Experimental results show the properties of the present techniques.

This paper deals with the design of radar target tracking and GPS (Global Positioning System) navigation based on the ANFIS (adaptive network-based fuzzy inference system) aided adaptive Kalman filtering approach. To achieve good filtering solutions, the Kalman filter designers are required to have good knowledge on both dynamic process and measurement models, in addition to the assumption that both the process and measurement are corrupted by zero-mean Gaussian white sequences. To prevent divergence problem when the Kalman assumptions are violated, the ANFIS is employed as the dynamic model corrector. The performance improvement will be demonstrated and discussed based on the proposed method.

Discrimination between the moving foreground objects and the complex dynamic background is a challenging task. In this paper, we have proposed a probabilistic graphical model – a recurrent stochastic network, which is able to learn the temporal and the spatial correlation from the video input data and make inference with a generalized belief propagation algorithm. Experiments have shown that the proposed recurrent network can model the dynamic backgrounds containing swaying trees, bushes and moving ocean waves. Very promising segmentation results have been obtained.

Considering the fast respond and high generalization accuracy of the min-max modular support vector machine (M^3-SVM), we apply M^3-SVM to solving the gender recognition problem and propose a novel task decomposition method in this paper. Firstly, we extract features from the face images by using a facial point detection and Gabor wavelet transform method. Then we divide the training data set into several subsets with the ‘part-versus-part’ task decomposition method. The most important advantage of the proposed task decomposition method over existing random method is that the explicit prior knowledge about ages contained in the face images is used in task decomposition. We perform simulations on a real-world gender data set and compare the performance of the traditional SVMs and that of M^3-SVM with the proposed task decomposition method. The experimental results indicate that M^3-SVM with our new method have better performance than traditional SVMs and M^3-SVM with random task decomposition method.

The conventional approaches to suppress the narrow-band interference of spread spectrum systems mostly use the adaptive LMS filter to predict the narrow-band interference and subtract the predicted interfering signal from the polluted received signal before de-spreading. However, since these approaches take no account of complexity control and have no guarantee of global minimum, they often suffer from unsteady performance. In this paper, a novel approach to narrow-band interference suppression is proposed, in which – support vector regression method is used to predict the narrow-band interference instead of adaptive LMS filter. With the help of practical parameter selection rules, it is not only effective but also easy to handle. Computer simulations show that it outperforms the conventional approaches in most cases and thus is a desirable choice for narrow-band interference suppression in spread spectrum systems.

Video watermarking hides information (e.g. ownership, recipient information, etc) into video contents. Video watermarking research is classified into (1) extension of still image watermarking, (2) use of the temporal domain features, and (3) use of video compression formats. In this paper, we propose a watermarking scheme to resist geometric attack (rotation, scaling, translation, and mixed) for H.264 (MPEG-4 Part 10 Advanced Video Coding) compressed video contents. Our scheme is based on auto-correlation method for geometric attack, a video perceptual model for maximal watermark capacity, and watermark detection based on natural image statistics. We experimented with the standard images and video sequences and the result shows that our video watermarking scheme is robust against H.264 video compression (average PSNR = 31 dB) and geometric attacks (rotation with 0-90 degree, scaling with 75-200%, and 50%~75% cropping).

This paper analyzes the network structure and algorithm model of Multi-Input and Multi-Output (MIMO) blind deconvolution, proposes a novel blind deconvolution algorithm based on output signals’ context information, and puts forward a new optimum method using Quantum Inspired Genetic Algorithm (QIGA). The simulation results demonstrate the effectiveness of the algorithm to the separation of communication signals.

In this paper, a novel differential demodulator for OFDM combining traditional differential demodulation with neural computation has been introduced for differential detection. Simulations using a two-path channel model and M-ary differential phase shift keying (MDPSK) modulation have been run to investigate the performance characteristics of the proposed scheme. The results show that it adapts very well to channel conditions with both strong delay and Doppler spread. The new structures are superior when compared to the traditional differential demodulation in frequency selective fast fading channels without any extra computing complexity.

A hierarchical matching of time series(HMTS) algorithm is proposed in this paper. The trend information of the time series is extracted using EMD(empirical mode decomposition) at first, subsequently piecewise linear segmentation is used to represent the trend of the series and the segmental line information is translated into 0-1 character, which substantially reduces the computational amount when comparing to the raw data. Finally the reduced series along with the series’ details are matched. As a result, the algorithm significantly improves the efficiency and accuracy of the similarity search, and overcome the difficulties of the direct linear segmentation representation of the raw data. The experimental results illustrate the effectiveness of this algorithm.

A novel moment, called 3D polar-radius-invariant-moment, is proposed for the 3D object recognition and classification. Some properties of these new moments including the invariance on translation, scale and rotation transforms are studied and proved. Then structure moment invariants are given to distinguish complicated similar shapes. Examples are presented to illustrate the performance and invariance of these moments. With the help of these moment invariants, the 3D models are distinguished accurately.