Catálogo de publicaciones - libros

We propose a novel subtraction-based method for visualizing segmental and subsegmental pulmonary embolism. For the registration of a pair of CT angiography, a proper geometrical transformation is found through the following steps: First, point-based rough registration is performed for correcting the gross translational mismatch. The center of inertia (COI), apex and hilar point of each unilateral lung are proposed as the reference point. Second, the initial alignment is refined by iterative surface registration. Third, thin-plate spline warping is used to accurately align inner region of lung parenchyma. Finally, enhanced vessels are visualized by subtracting registered pre-contrast images from post-contrast images. To facilitate visualization of parenchymal enhancement, color-coded mapping and image fusion is used. Our method has been successfully applied to four pairs of CT angiography.

In this work several sets of categories obtained by a statistical clustering algorithm, as well as a linguistic set, were used to design category-based language models. The language models proposed were evaluated, as usual, in terms of perplexity of the text corpus. Then they were integrated into an ASR system and also evaluated in terms of system performance. It can be seen that category-based language models can perform better, also in terms of WER, when categories are obtained through statistical models instead of using linguistic techniques. They also show that better system performance are obtained when the language model interpolates category based and word based models.

The purpose of this work is to present a comparative analysis of knowledge-based systems, artificial neural networks and statistical clustering algorithms applied to the classification of low resolution stellar spectra. These techniques were used to classify a sample of approximately 258 optical spectra from public catalogues using the standard MK system. At present, we already dispose of a hybrid system that carries out this task, applying the most appropriate classification method to each spectrum with a success rate that is similar to that of human experts.

In this paper, we propose an automatic approach to measure the minutiae quality. When image of 500 dpi is captured, immediately the enhancement, thinning and minutiae extraction processes are executed. The basic idea is to detect the spatial – Connected minutiae cluster using the Euclidean distance and quantify the number of element for each group. In general, we observe that more than five element in a group is a clue to mark all points in the cluster as bad minutiae. We divide the image in block of 20 x 20 pixels. If one block contains bad minutiae it is mark as a bad block. The goodness quality index is calculated as the proportion of bad blocks respect to the number of total blocks. The proposed index was tested on the FVC2000 fingerprint image database.

Tin Kam Ho and Ester Bernardò Mansilla in 2004 proposed to use data complexity measures to determine the domain of competition of the classifiers. They applied different classifiers over a set of problems of two classes and determined the best classifier for each one. Then for each classifier they analyzed how the values of some pairs of complexity measures were, and based on this analysis they determine the domain of competition of the classifiers. In this work, we propose a new method for selecting the best classifier for a given problem, based in the complexity measures. Some experiments were made with different classifiers and the results are presented.

Complex images contaminated by noise appear in various applications. To improve these phase images, noise effects, as loss of contrast and phase residues that deteriorate the phase unwrapping process, should be reduced. Noise reduction in complex images has been addressed by various methods, most of them dealing only with the magnitude image. Few works have been devoted to phase image de-noising, despite the existence of important applications like Interferometric Synthetic Aperture Radar (IFSAR), Current Density Imaging (CDI) and Magnetic Resonance Imaging (MRI). In this work, several de-noising algorithms in the wavelet packets domain were applied to complex images to recover the phase information. These filtering algorithms were applied to simulated images contaminated by three different noise models, including mixtures of Gaussian and Impulsive noise. Significant improvements in SNR for low initial values (SNR<5 dB) were achieved by using the proposed filters, in comparison to other methods reported in the literature.

Optical character recognition of cursive scripts present a number of challenging problems in both segmentation and recognition processes and this attracts many researches in the field of machine learning. This paper presents a novel approach based on a combination of MLP and SVM to design a trainable OCR for Persian/Arabic cursive documents. The implementation results on a comprehensive database show a high degree of accuracy which meets the requirements of commercial use.

Area based local stereo correspondence algorithms that use the simple ’winner takes all’ (WTA) method in the optimization step perform poorly near object boundaries particularly in occluded regions. In this paper, we present a new modified area based local algorithm that goes some way towards addressing this controversial issue. This approach utilizes an efficient strategy by adding the concept of a computation skip threshold (CST) to area based local algorithms in order to add the horizontal smoothness assumption to the local algorithms. It shows similar effects to Dynamic Programming(DP) and Scanline Optimization(SO) with significant improvements in occlusions from existing local algorithms. This is achieved by assigning the same disparity value of the previous neighboring point to coherent occluded points. Experiments were carried out comparing the new algorithm to existing algorithms using the standard stereo image pairs and our own images generated by a Scanning Electron Microscope (SEM). The results show that the horizontal graphical performance improves similarly to DP particularly in occlusions but the computational speed is faster than existing local algorithms, due to skipping unnecessary computations for many points in the WTA step.

The location of the texture limits in an iris image is a previous step in the person’s recognition processes. The iris localization plays a very important role because the speed and performance of an iris recognition system is limited by the results of iris localization to a great extent. It includes finding the iris boundaries (inner and outer). We present a new method for iris pupil contours delimitation and its practical application to iris texture features estimation and isolation. Two different strategies for estimating the inner and outer iris contours are used. The results obtained in the determination of internal contour is used efficiently in the search of the external contour parameters employing a differential integral operator. The proposed algorithm takes advantage of the pupil’s circular form using well-known elements of analytic geometry, in particular, the determination of the bounded circumference to a triangle. The algorithm validation experiments were developed in images taken with near infrared illumination, without the presence of specular light in their interior. Satisfactory time results were obtained (minimum 0.0310 s, middle 0.0866 s, maximum 0.1410 s) with 98% of accuracy. We will continue working in the algorithm modification for using with images taken under not controlled conditions.

Catastrophic Interference is a well known problem of Artificial Neural Networks (ANN) learning algorithms where the ANN forget useful knowledge while learning from new data. Furthermore the structure of most neural models must be chosen in advance.

In this paper we introduce a hybrid algorithm called Flexible Architecture of Self Organizing Maps () that overcomes the Catastrophic Interference and preserves the topology of Clustered data in changing environments. The model consists in receptive fields of self organizing maps. Each Receptive Field projects high-dimensional data of the input space onto a neuron position in a low-dimensional output space grid by dynamically adapting its structure to a specific region of the input space.

Furthermore the model automatically finds the number of maps and prototypes needed to successfully adapt to the data. The model has the capability of both growing its structure when novel clusters appears and gradually forgets when the data volume is reduced in its receptive fields.

Finally we show the capabilities of our model with experimental results using synthetic sequential data sets and real world data.