Thesis

  1. CLARIMAR JOSE COELHO, 25/02/99
  2. NILTON CORREIA DA SILVA, 09/07/1999
  3. VÂNIA CRISTINA DE ABREU, 13/09/1999
  4. ANDRÉ MACHADO CARICATTI, 14/04/2000
  5. ISABELA N. F. DE QUEIROZ, 22/03/2001
  6. CARLOS ROBERTO PORFÍRIO JUNIOR, 21/06/2001
  7. CANDIDO GUERRERO SALGADO, 10/01/2002
  8. WESLEY MARTINS TELES,18/07/2003
  9. OSMAR QUIRINO DA SILVA. 30/09/2003
  10. SORAIA SILVA PRIETCH, 13/02/04
  11. MARCOS VINICIUS PINHEIRO DIB, 03/09/2004
  12. MANQI WU, 15/04/2005
  13. DANIELA PEREIRA ALVES, 09/11/2006
  14. BUENO BOGES SOUZA, 13/06/2008
  15. SANDRO CARLOS VIEIRA, 21/07/2008
  16. ANTONIO CARLOS DE ARRUDA JUNIOR, 08/12/2009
  17. CICERO ROBERTO FERREIRA DE ALMEIDA, 26/08/2010
  18. ANTONIO CRESPO, 26/10/2010
  19. CHARLES HENRIQUE GONÇALVES SANTOS, 15/08/2011 
  20. ZHANG JIANYA, 09/03/2012
  21. LEONARDO LUIZ BARBOSA VIEIRA CRUCIOL, 08/03/2012

1. CLARIMAR JOSE COELHO, 25/02/99

Identification with Image Analysis in "'' Multiresolution Wavelet Systems Based Reasoning

The representation of a high-quality image is done by a lot of bits. In a photographic quality image, for example, may be required up to 100 million bits. To store, process, analyze and transmit large volumes of images are required high rates of compression. In general, the image processing is not done with the images in its pure state, but is necessary to reduce the large amount of bits that are used for their representation. One of the mathematical processes more successful in this task is to transform "wavelet". Systems Based Reasoning (CBR's) offer great advantages and facilities in various application domains such as planning, design and diagnosis. We will present a Case Based Reasoning system hybrid that combines the advantages of racioínio systems RBC's with the power of mathematical transforms "wavelets" for image processing in an application to rainfall estimation of rainfall. The processed "wavelets" will be used to: eliminate noise, reduce the number of bits and retain the most important physical information of the images.

 

2. NILTON CORREIA DA SILVA, 09/07/1999

Implementation of Self-Parallel networks Srganizing Map to map weather radar images

This innovative work has relevance for the use of neural networks and parallel processing in the classification process of meteorological images (important process in the task of predicting more precipitation). The algorithm presented in this paper uses the benefits that a parallel processing environment has to help in the process of training, testing and use of unsupervised neural networks ("Self-Organizing Maps"). This feature of the learning algorithm results in a single network trained by him. In this work called "Once Learning".

With this type of training can effectively parallelize the process of training, testing and use of neural networks, which means a great saving processing time. This effect is of great importance in the field of Neural Networks as the processing of large numbers of data is a routine task in the subprocess training, testing and utilizaçãodas networks.

This work also contributes directly to the area of image processing weather, because with this novoalgoritmo, neural networks classifier (used in detecting rainfall radar and satellite images) can analyze a much larger number of images (in a unit of time) in a parallel environment.

Through this research, we developed a prototype (in Matlab 5.0) that simulates, in a serial environment, the algorithm Parallel Self-Organizing Map for mapping meteorological images.

 

3. VÂNIA CRISTINA DE ABREU

Developing a Methodology Distribution Travel with Application of Fuzzy Logic

4. ANDRÉ MACHADO CARICATTI

Talkers Recognition in Portuguese with Neural Network Models and Gaussian

In this thesis we study the problem of recognition of speakers, or specifically checking announcers. Seen as a subset of the signal processing, and therefore divided into steps signal acquisition, parameter extraction and classification, working with parameters formed by mel-frequency cepstral coefficients and their rates of change, called coefficients delta-mel- frequency cepstral. Considering the objectives of the study, formed a database from 14 people, providing locutions of the Portuguese language, indoors, though imperfect, is considered noise and echo. Altogether, three systems were implemented, being a dependent text, and two other independent text. In the first, we used multilayer neural networks with learning by back-propagation, calculating correlations between mel-frequency cepstral coefficients, then composing the input values. When performing recognition of text independent broadcasters, neural networks have been applied self-organizing maps (SOM - self-organizing maps), together with calculation of the relative entropy models between talkers and reference talkers unknown. On this occasion, tests were conducted in order to scale the optimal number of neurons for the output layer, reaching 225, and the length of phrases used in training models, getting around 48 seconds to reach the correct recognition in 50% of attempts. Finally, using Gaussian mixture models (GMM - Gaussian Mixture Models) was appreciated, served as compared to the implementation of SOM's, while also obtained estimates of the length of phrases to form models of speakers being sufficient in this case samples approximately 20 seconds to 32 models for speaker mixtures, coming to successful recognition in 64% of experiments