Catálogo de publicaciones - libros

The fastest and most straightforward way of communication for mankind is the voice. Therefore, the best way to interact with computers should be the voice too. That is why at the moment men are searching new ways to interact with computers. This interaction is improved if the words spoken by the speaker are organized in Natural Language.

In this article, it is proposed a model to recover information from databases through queries in Spanish Natural Language using the voice as the way of communication. This model incorporates a Hybrid Intelligent System based on and a Kohonen Self-Organizing Map (SOM) to recognize the present phonemes in a word through time. This approach allows us to remake up a word with speaker independence. Furthermore, it is proposed the use of a compiler with type 2 grammar according to the to support the syntactic and semantic structure in Spanish language. Our experiments suggest that the Spoken Natural Language improves notably the Human-Computer interaction when compared with traditional input methods such as: mouse or keybord.

In this paper, a new 3D reconstruction approach in neuro-vision system is presented. Firstly, RBF network (RBFN) is used to provide effective methodologies for solving camera calibration and stereo rectification problems. RBFN works mainly in two aspects: (1) a RBFN is adopted to learn and memorize the nonlinear relationship in stereovision system; (2) another RBFN is trained to search the correspondent lines in two images such that stereo matching could be performed in one dimension. Secondly, a new matching method based on Hopfield neural network (HNN) is presented. The energy function is built on the basis of uniqueness, compatibility and similarity constraints. It is then mapped onto a 2-D neural network for minimization, whose final stable state indicates the possible correspondence of the matching units. The depth map can be acquired through performing the above operation on the all epipolar lines. Experiments have been performed on common stereo pairs and the results are accurate and convincing.

In order to meet the demand of real time nonuniformity correction of Infrared Focal Plane Array (IRFPA), a new algorithm based on neural network (NN) is proposed. Comparing to the traditional NN algorithm, the new algorithm uses weighted average of four adjacent pixels value to calculate the expected output value, and uses weighted average of four adjacent corrected output value to replace the finally corrected value. With the help of high performance image-processing system based on TMS320C6201 DSP, the nonuniformity of IRFPA is real time corrected. In this paper the mathematical model of the new NN algorithm is established, and the specific hardware correction process is expatiated. At last experimental results are given out. The results show that new NN algorithm is more satisfying to increase the quality of the corrected image.

This paper proposes a novel direct 3-D reconstruction methodology namely Shape-From-Shading with specular reflectance using wavelet networks. The thought of this approach is to optimize a proper reflectance model by learning the parameters of wavelet networks. Hybrid reflection models which contain diffuse reflectance and specular reflectance are used to formulate reflectance map equation because they are prone to reality. The approach uses wavelet networks as a parametric representation of the unknown surface to be reconstructed. After the orientation expressed by the parametric form of the surface is substituted into hybrid reflection model, the shape from shading problem is formulated as minimization of the total intensity error function over the network weights. Gradient-descent method is used to update the parameters of wavelet networks. The heights of the surface can then be obtained from the wavelet networks after supervised learning. Experiments on both synthetic and real images demonstrate the performance of the proposed SFS method.

In this paper, a space partition method called “Label Constrained Graph Partition” (LCGP) is presented to solve the Sample-Interweaving-Phenomenon in the original space. We first divide the entire training set into subclasses by means of LCGP, so that the scopes of subclasses will not overlap in the original space. Then “Most Discriminant Subclass Distribution” (MDSD) criterion is proposed to decide the best partition result. At last, typical LDA algorithm is applied to obtain the feature space and the RBF neural network classifier is utilized to make the final decision. The computer simulations and comparisons are given to demonstrate the performance of our method.

Mean shift tracking fails when the velocity of target is so large that the target’s window kernel in the previous frame can not cover the target in the current frame. Combination of mean shift and single Kalman filter also fails when the target’s velocity changed suddenly. To deal with the problem of tracking image target that has large and changing velocity, an efficient image tracking method integrated mean shift and double model filters is proposed. Two motion models can switch each other by using a probabilistic likelihood. Experiment results show the method integrated mean shift and double model filters can successfully keep tracking target, no matter the target’s velocity is large or small, changing or constant, with modest requirement of computation resource.

In this paper, a human action recognition method using a hybrid neural network is presented. The method consists of three stages: preprocessing, feature extraction, and pattern classification. For feature extraction, we propose a modified convolutional neural network (CNN) which has a three-dimensional receptive field. The CNN generates a set of feature maps from the action descriptors which are derived from a spatiotemporal volume. A weighted fuzzy min-max (WFMM) neural network is used for the pattern classification stage. We introduce a feature selection technique using the WFMM model to reduce the dimensionality of the feature space. Two kinds of relevance factors between features and pattern classes are defined to analyze the salient features.

This paper proposes a novel parallel RBFNN (Radial Basis Function Neural Network) classifier based on S-transform for recognition and classification of PQ (Power Quality) disturbances. S-transform is used to extract feature vectors, while the constructed parallel RBFNN classifier is used to recognize and classify PQ disturbances according to the extracted feature vectors. The parallel RBFNN classifier consists of eight sub-networks, each of which is only able to recognize one type of disturbance. In order to improve the convergence performance of RBFNN and optimize the number of hidden layer nodes, a dynamic clustering algorithm which clusters all training samples to determine the number of hidden layer nodes is proposed. Simulation and test results demonstrate that the method proposed to recognize and classify PQ disturbances is correct and feasible, and that the RBFNN classifier based on the dynamic clustering algorithm has a faster convergence speed and a higher correct identification rate.

The study of small target recognition in low SNR (Signal Noise Ratio) is the key problem about processing of forward-looking infrared (FLIR) images information. Eight features of objects based on IR radiation characteristics and wavelet-based are presented. These features are used to a radial basis function (RBF) network as input for learning and classification. The propose recognition algorithm is invariant to the translation, rotation, and scale channel of a shape. Experiments by real infrared images and noisy images are performed, and recognition results show that the method is very effective.

Bayesian subspace analysis (BSA) has been successfully applied in data mining and pattern recognition. However, due to the use of probabilistic measure of similarity, it often needs much more projective vectors for better performance, which makes the compression ratio very low. In this paper, we propose a novel 2D Bayesian subspace analysis (2D-BSA) method for face recognition at high compression ratios. The main difference between the proposed 2D-BSA and BSA is that the former adopts a new Image-as-Matrix representation for face images, opposed to the Image-as-Vector representation in original BSA. Based on the new representation, 2D-BSA seeks two coupled set of projective vectors corresponding to the rows and columns of the difference face images, and then use them for dimensionality reduction. Experimental results on ORL and Yale face databases show that 2D-BSA is much more appropriate than BSA in recognizing faces at high compression ratios.