Catálogo de publicaciones - libros

In this paper, an approach based on lacunarity to locate address blocks in postal envelopes is proposed. After computing the lacunarity of a postal envelope image, a non-linear transformation is applied on it. A thresholding technique is then used to generate evidences. Finally, a region growing is applied to reconstruct semantic objects like stamps, postmarks, and address blocks. Very little knowledge of the envelope images is required. By using the lacunarity for several ranges of neighbor window sizes onto 200 postal envelope images, the proposed approach reached a success rate over than 97% on average.

The coarse to fine search methodology is frequently applied to a wide variety of problems in computer vision. In this paper it is shown that this strategy can be used to enhance the recognition of on-line handwritten characters. Some explicit knowledge about the structure of a handwritten character can be obtained through a structural parameterization. The Frame Deformation Energy matching (FDE) method is a method optimized to include such knowledge in the discrimination process. This paper presents a novel parameterization strategy, the Djikstra Curve Maximization (DCM) method, for the segments of the structural frame. Since this method distributes points unevenly on each segment, point-to-point matching strategies are not suitable. A new distance measure for these segment-to-segment comparisons have been developed. Experiments have been conducted with various settings for the new FDE on a large data set both with a single model matching scheme and with a NN type template matching scheme. The results reveal that the FDE even in an ad hoc implementation is a robust matching method with recognition results well comparing to the existing state-of-the-art methods.

Civil structures could undergo hysteresis cycles due to cracking or yielding when subjected to severe earthquake motions or even high wind. System identification techniques have been used in the past years to assess civil structures under lateral loads. The present research makes use of a polynomial artificial neural network to identify and predict, on-line, the behavior of such nonlinear structures. Simulations are carried out using the Loma Prieta and the Mexico City seismic records on two hysteretic models. Afterwards, two real seismic records acquired on a 24-story concrete building in Mexico City are used to test the proposed algorithm. Encouraging results are obtained: fast identification of the weights and fair prediction of the output acceleration.

techniques are increasing the use of the speaker’s voice to control access to personalized telephonic services. This paper describes the use of vector quantization as a feature matching method, in an automatic speaker recognition system, evaluated with speech samples from a SALA Spanish Venezuelan database for fixed telephone network. Results obtained reflect a good performance of the method in a text independent job in the context of sequences of digits.

We propose a novel classification method to identify boar spermatozoid heads which present an intracellular intensity distribution similar to a model. From semen sample images, head images are isolated and normalized. We define a model intensity distribution averaging a set of head images assumed as normal by veterinary experts. Two training sets are also formed: one with images that are similar to the model and another with non-normal head images according to experts. Deviations from the model are computed for each set, obtaining low values for normal heads and higher values for assumed as non-normal heads. There is also an overlapped area. The decision criterion is determined to minimize the sum of the obtained false rejected and false acceptance errors. Experiments with a test set of normal and non-normal head images give a global error of 20.40%. The false rejection and the false acceptance rates are 13.68% and 6.72% respectively.

This paper presents an approach for the automatic speech re-cognition using syllabic units. Its segmentation is based on using the Short-Term Total Energy Function (STTEF) and the Energy Function of the High Frequency (ERO parameter) higher than 3,5 KHz of the speech signal. Training for the classification of the syllables is based on ten related Spanish language rules for syllable splitting. Recognition is based on a Continuous Density Hidden Markov Models and the bigram model language. The approach was tested using two voice corpus of natural speech, one constructed for researching in our laboratory (experimental) and the other one, the corpus Latino40 commonly used in speech researches. The use of ERO parameter increases speech recognition by 5% when compared with recognition using STTEF in discontinuous speech and improved more than 1.5% in continuous speech with three states. When the number of states is incremented to five, the recognition rate is improved proportionally to 97.5% for the discontinuous speech and to 80.5% for the continuous one.

The watersheds method is a powerful segmentation tool developed in mathematical morphology, which has the drawback of producing over-segmentation. In this paper, in order to prevent its over-segmentation, we present a strategy to obtain robust markers for atherosclerotic lesions segmentation of the thoracic aorta. In such sense, we introduced an algorithm, which was very useful in order to obtain the markers of atherosclerotic lesions. The obtained results by using our strategy were validated calculating the false negatives (FN) and false positives (FP) according to criterion of pathologists, where 0% for FN and less than 11% for FP were obtained. Extensive experimentation showed that, using real image data, the proposed strategy was very suitable for our application.

In this paper, we show how the heat-kernel can be used to construct a scale-space for image smoothing and edge detection. We commence from an affinity weight matrix computed by exponentiating the difference in pixel grey-scale and distance. From the weight matrix, we compute the graph Laplacian. Information flow across this weighted graph-structure with time is captured by the heat-equation, and the solution, i.e. the heat kernel, is found by exponentiating the Laplacian eigen-system with time. Our scale-space is constructed by varying the time parameter of the heat-kernel. The larger the time the greater the the amount of information flow across the graph. The method has the effect of smoothing within regions, but does not blur region boundaries. Moreover, the boundaries do not move with time and this overcomes one of the problems with Gaussian scale-space. We illustrate the effectiveness of the method for image smoothing and edge detection.

In this work, the problem of estimating high density regions from univariate or multivariate data samples is studied. To be more precise, we estimate minimum volume sets whose probability is specified in advance. This problem arises in outlier detection and cluster analysis, and is strongly related to One-Class Support Vector Machines (SVM). In this paper we propose a new simpler method to solve this problem. We show its properties and introduce a new class of kernels, relating the proposed method to One-Class SVMs.

The function of the Cerebral Blood Flow Autoregulation (CBFA) system is to maintain a relatively constant flow of blood to the brain, in spite of changes in arterial blood pressure. A model that characterizes this system is of great use in understanding cerebral hemodynamics and would provide a pattern for evaluating different cerebrovascular diseases and complications. This work posits a non-linear model of the CBFA system through the evaluation of various types of neural networks that have been used in the field of systems identification. Four different architectures, combined with four learning methods were evaluated. The results were compared with the linear model that has often been used as a standard reference. The results show that the best results are obtained with the FeedForward Time Delay neural network, using the Levenberg-Marquardt learning algorithm, with an improvement of 24% over the linear model (p<0.05).