Catálogo de publicaciones - libros

In this paper, we use three focus measures for ranking digital images, namely, L norm of image gradient, absolute central moment, and spatial frequency. The ranking scores from the focus measures are combined by the sum and product rules to result in the final decision. Experimental results on a photo album demonstrate the proposed method is very useful to users.

In the territory of text categorization, the distribution and quality of sample set is highly influential to categorization result. Associated rule categorization ARC-BC is effective under common circumstances. The accuracy of categorization obviously falls as distribution of feature words of training samples is uneven. In this paper, a Chinese text classification approach was proposed based on sample weighting associated rules (SW-ARC). The approach improved substantial classification efficiency by performing self-adapting sample weights adjustment. Experiment result shows SW-ARC can solve the quality fall caused by uneven distribution of feature words. Macro-average recall of open test increases from 50% of ARC-BC to 70% of SW-ARC, Macro-average precision increases from 28% of ARC-BC to 70% of SW-ARC.

The on-line update of classifiers is an important concern for categorizing the time-varying neurophysiological signals used in brain computer interfaces, e.g. classification of electroencephalographic (EEG) signals. However, up to the present there is not much work dealing with this issue. In this paper, we propose to use the idea of gradient decorrelation to develop the existent basic Least Mean Square (LMS) algorithm for the on-line learning of Bayesian classifiers employed in brain computer interfaces. Under the framework of Gaussian mixture model, we give the detailed representation of Decorrelated Least Mean Square (DLMS) algorithm for updating Bayesian classifiers. Experimental results of off-line analysis for classification of real EEG signals show the superiority of the on-line Bayesian classifier using DLMS algorithm to that using LMS algorithm.

The purpose of document structure analysis is to get the document structure of the source text. Document structure is defined as 3 layers in the paper. A new model of document structure analysis — DLM is proposed. The model is composed of three layers: physical structure layer, logical structure layer and semantic structure layer, which are corresponding to the definition of the document structure. The input, output and operation of each layer are illustrated in details in the paper. The model has the feature of flexible, systematic and extendible. DLM is implemented on the Automatic Summarization System. It shows that the model is feasible and good result can be achieved.

This paper presents an investigation into the use of the delay coordinate embedding technique with multi-input multi-output (MIMO) adaptive-network-based-fuzzy-inference system (ANFIS) to learn and predict the continuation of chaotic signals ahead in time. Based on the average mutual information and global false nearest neighbors techniques, the optimal values of the embedding dimension and the time delay are selected to construct the trajectory on the phase space. The MANFIS technique is trained by gradient descent algorithm. First, the parameter set of the membership functions is generated with the embedded phase space vectors using the back-propagation algorithm. Second, fine-tuned membership functions that make the prediction error as small as possible are built. The model is tested with both periodic and the Mackey-Glass chaotic time series. Moving root-mean-square error is used to monitor the error along the prediction horizon.

Video abstraction can be useful in multimedia database indexing and querying and can illustrate the important content of a longer video to quick browsing. Further, in sports video, replay scene often demonstrates the highlight of the video. The detection of replay scene in the sports video is a key clue to sports video summarizing. In this paper, we present a framework of replay scene based video abstraction in MPEG sports video. Moreover, we detect identical events using color and camera information after detecting replay scene using MPEG feature. At last, we propose a three-layer replay scene based sports video abstraction. It can achieve real time performance in the MPEG compressed domain, which is validated by experimental results.

A wavelet-based fuzzy multiscale edge detection scheme (WFMED) is presented in this paper. The dyadic wavelet transform is employed to produce the multiscale representation of the image, fuzzy logic is applied in wavelet domain and it can synthesize the information of image across scales effectively, an optimal result of edge detection can be acquired. WFMED method is used to extract the edge of pulp fibre image; the paper compares the performance of WFMED to the Canny edge detector and to Mallat’s algorithm. The results show the superiority of WFMED to these other methods.

The summary of video content provides an effective way to speed up video browsing and comprehension. In this paper, we propose a novel automatic video summarization approach. Video structure is first analyzed by combining spatial-temporal analysis and statistical learning. Video scenes are then detected based on unsupervised statistical learning. The video summary is created by selecting the most informative shots from the video scenes that are modeled as a directed graph. Experiments show that the proposed approach can generate the most concise and informative video summary.

In this paper we present a new feature of texture images which can scale the uniformity degree of image texture directions. The feature value is obtained by examining the statistic characteristic of the gradient information of the image pixels. Simulation results illustrate that this feature can exactly coincide with the uniformity degree of image texture directions according to the perception of human eyes.

Recently, there are remarkable progress in similarity computing for 3D geometric models. Few focus is put on the research of the similarity between volumetric models. This paper proposes a novel approach for performing similarity computation between two volumetric data sets. For each data set, it is performed by four stages. First, the volume data set is resampled into a unified resolution. Second, the data set is band-pass filtered and quantized to reveal its physical attributes. The resulting voxels are then normalized into a canonical coordinate system concerning the center of mass and scale. Subsequently, a series of uniformly spaced concentric shells around the center of mass are constructed, based on which spherical harmonics analysis (SHA) is applied. The coefficients of SHA constitute a rotation invariant spectrum descriptor which are used to measure the similarity between two data sets. The algorithm has been performed on a set of clinical CT and MRI data sets and the preliminary results are fairly inspiring.