This project aims to show a brief description about the C4.5 algorithm, used to create Univariate Decision Trees. We also talk about Multivariate Decision Trees, their process to classify instances using more than one attribute per node in the tree. We try to discuss how they work, and how to implement the algorithms that build such trees, including examples of Univariate and Multivariate results.
Remote sensing images with large sizes are usual. They also include several spectral channels, increasing the volume of information. To get valuable information from data automatically, computers need higher amounts of memory and efficient processing techniques. Segmentation is a key technique to deal with remote sensing. It identifies regions in images. Therefore, it deals with large amounts of information. Even with current computational power, some image sizes exceed the memory limits, which need different solutions. An alternative to overcome such limits is to employ divide and conquer strategy, splitting the image into tiles, and segmenting each one individually. However, arises the problem of merging neighboring tiles and keeping the homogeneity in such regions. In this work, we propose an alternative to create the tiles, by defining noncrisp borders between tiles, but adaptive borders for the tiles. By applying our method, we avoid the postprocessing of neighboring regions, and therefore speed up the final segmentation.
This project implements in MATLAB, a technique for face recognition, using the Principal Component Analysis (PCA), through Eigenfaces. With a training set of images, the system becomes able to recognize new faces, in test images.
The Expectation-Maximization algorithm is an iterative procedure which can be shown to converge to a (local) maximun of the marginal a posteriori probability function p(theta|x) = p(x|theta) p(theta), without the need to explicitly manipulate the marginal likelihood p(x|theta). The classification methods are used to clusterize areas in the Earth surface that show similar meanings. I developed a set of Classes for TerraLib Library, that realizes the EM (Expectation-Maximization) method for image non-supervisioned classification. This code, and an article explaining the method are available for download.
Using JAVA 3D and JAMA library for matrix calculations, this project presents a computational system that is able to simulate human movements in a virtual face, modelled through Computer Graphics techniques. This face will be represented through a polygon mesh, defined by a set of points (vertices) in the 3D space, inter connected by edges. There\'s lots of approaches to animate a virtual face, classified in two main groups: the Physics-Based models and the Parametric ones. The parametric approach using Radial Basis Functions were choosed and implemented in this project. The main idea is based on the insertion of control points in the virtal face, and trough the movement of this points, generate the animation of the neiborhood, in a influence region defined based on heuristics extracted from the human anatomy. More information at http://facialdas.sf.net/.
The aim of this project is to simulate and Fuzzy logic application in the control of an automobile, comparing the results with the PID control, using Flash language. The results are shown in the link [ swf ], in the right.
GeoDMA is a plugin for TerraView software, used for geographical data mining. With a remote sensing image, the user can perform segmentation, attributes extraction, normalization and supervised/unsupervised classification. More information at http://www.dpi.inpe.br/geodma/.
Classification is the process of image information extraction to recognize patterns and homogeneous objects. The classification methods are used to clusterize areas in the Earth surface that show similar meanings. I developed a set of Classes for TerraLib Library, that realizes the Isoseg method for image classification. This code, and an article explaining the method are available for download.
This project implements a Text-To-Speech system through the use of concatenative synthesis, as an application layer for the MBROLA system. This system is Free Software, using PHP, and is hosted in http://fttss.sourceforge.net/.
This project implements a series of Octave scripts for Artificial Neural Networks. The input/output must be binary so that can become generic for every type of data. The algorithms support several parameters for neural network tunning and configuration. The learning model is the supervised through Back-Propagation approach, with weight correction using the delta rule. The stop criteria are stabilished by a maximum of iterations, or a threshold over the found errors. The purpose of this algorithm is to provide a supervised classifier to be applied into digital images.
Image segmentation is one of the most important tasks in Digital Image Processing. It is used in several scientific areas, such as character recognition, image detection and classification. Segmentation is being applied also to remote sensing data, since the sensors present better resolution nowadays. This work proposes a methodology for re-segmentation of high resolution urban imagery, shape and graph based. The input is formed by a set of images and polygons resultant from an over-segmentation. Adjacent regions, or polygons, are connected in a graph structure, and a graph search is performed, looking for rectangular shapes, present on roofs and buildings. For other urban objects, such as streets or trees, other heuristic are employed. To perform such task, a previous operation is necessary. The classification is done over the set of polygons, using the Self Organizing Maps (SOM) algorithm. The polygons are then classified and connected in a Region Adjacency Graph (RAG), according their topology. Using the RAG, the algorithm tries to fit rectangles for building and roof classes, and other heuristic-based approaches for merging polygons from remaining classes. The result of the algorithm is a new set of polygons that best fits the urban environment. Some results are shown and discussed, as a way to prove the accuracy of the proposed work.
The Kohonen\'s maps, knonw as Self Organizing Maps (SOM) can be employed to unsupervised image classification. An algorithm for TerraLib library was developed for this purpouse, taken as input an attribute-vector, that flows through a set of neurons. Such neurons, iteratively, simulate the congnitive process of the human brain in order to find clusters in the input data. The links in the side presents one video with an example application, using a Graphical User Interface (GUI) developed in MATLAB. The source code for the TerraLib class is available at http://www.terralib.org/. Besides the video, you can also donwload the source code of the GUI, that runs for MATLAB v7.0 or superior.
TerraAIDA is a package of image processing operators used with InterIMAGE system, for automatic remote sensing image interpretation. The operators include segmentation, atrithmetic functions, and classification, and are developed using the free GIS library Terralib. More information at http://www.dpi.inpe.br/terraaida/.
This project aims to show a brief description about the C4.5 algorithm, used to create Univariate Decision Trees. We also talk about Multivariate Decision Trees, their process to classify instances using more than one attribute per node in the tree. We try to discuss how they work, and how to implement the algorithms that build such trees, including examples of Univariate and Multivariate results.
Remote sensing images with large sizes are usual. They also include several spectral channels, increasing the volume of information. To get valuable information from data automatically, computers need higher amounts of memory and efficient processing techniques. Segmentation is a key technique to deal with remote sensing. It identifies regions in images. Therefore, it deals with large amounts of information. Even with current computational power, some image sizes exceed the memory limits, which need different solutions. An alternative to overcome such limits is to employ divide and conquer strategy, splitting the image into tiles, and segmenting each one individually. However, arises the problem of merging neighboring tiles and keeping the homogeneity in such regions. In this work, we propose an alternative to create the tiles, by defining noncrisp borders between tiles, but adaptive borders for the tiles. By applying our method, we avoid the postprocessing of neighboring regions, and therefore speed up the final segmentation.
This project implements in MATLAB, a technique for face recognition, using the Principal Component Analysis (PCA), through Eigenfaces. With a training set of images, the system becomes able to recognize new faces, in test images.
The Expectation-Maximization algorithm is an iterative procedure which can be shown to converge to a (local) maximun of the marginal a posteriori probability function p(theta|x) = p(x|theta) p(theta), without the need to explicitly manipulate the marginal likelihood p(x|theta). The classification methods are used to clusterize areas in the Earth surface that show similar meanings. I developed a set of Classes for TerraLib Library, that realizes the EM (Expectation-Maximization) method for image non-supervisioned classification. This code, and an article explaining the method are available for download.
Using JAVA 3D and JAMA library for matrix calculations, this project presents a computational system that is able to simulate human movements in a virtual face, modelled through Computer Graphics techniques. This face will be represented through a polygon mesh, defined by a set of points (vertices) in the 3D space, inter connected by edges. There\'s lots of approaches to animate a virtual face, classified in two main groups: the Physics-Based models and the Parametric ones. The parametric approach using Radial Basis Functions were choosed and implemented in this project. The main idea is based on the insertion of control points in the virtal face, and trough the movement of this points, generate the animation of the neiborhood, in a influence region defined based on heuristics extracted from the human anatomy. More information at 
The aim of this project is to simulate and Fuzzy logic application in the control of an automobile, comparing the results with the PID control, using Flash language. The results are shown in the link [ swf ], in the right.
GeoDMA is a plugin for 
Classification is the process of image information extraction to recognize patterns and homogeneous objects. The classification methods are used to clusterize areas in the Earth surface that show similar meanings. I developed a set of Classes for TerraLib Library, that realizes the Isoseg method for image classification. This code, and an article explaining the method are available for download.
This project implements a Text-To-Speech system through the use of concatenative synthesis, as an application layer for the MBROLA system. This system is Free Software, using PHP, and is hosted in 
This project implements a series of Octave scripts for Artificial Neural Networks. The input/output must be binary so that can become generic for every type of data. The algorithms support several parameters for neural network tunning and configuration. The learning model is the supervised through Back-Propagation approach, with weight correction using the delta rule. The stop criteria are stabilished by a maximum of iterations, or a threshold over the found errors. The purpose of this algorithm is to provide a supervised classifier to be applied into digital images.
Image segmentation is one of the most important tasks in Digital Image Processing. It is used in several scientific areas, such as character recognition, image detection and classification. Segmentation is being applied also to remote sensing data, since the sensors present better resolution nowadays. This work proposes a methodology for re-segmentation of high resolution urban imagery, shape and graph based. The input is formed by a set of images and polygons resultant from an over-segmentation. Adjacent regions, or polygons, are connected in a graph structure, and a graph search is performed, looking for rectangular shapes, present on roofs and buildings. For other urban objects, such as streets or trees, other heuristic are employed. To perform such task, a previous operation is necessary. The classification is done over the set of polygons, using the Self Organizing Maps (SOM) algorithm. The polygons are then classified and connected in a Region Adjacency Graph (RAG), according their topology. Using the RAG, the algorithm tries to fit rectangles for building and roof classes, and other heuristic-based approaches for merging polygons from remaining classes. The result of the algorithm is a new set of polygons that best fits the urban environment. Some results are shown and discussed, as a way to prove the accuracy of the proposed work.
The Kohonen\'s maps, knonw as Self Organizing Maps (SOM) can be employed to unsupervised image classification. An algorithm for TerraLib library was developed for this purpouse, taken as input an attribute-vector, that flows through a set of neurons. Such neurons, iteratively, simulate the congnitive process of the human brain in order to find clusters in the input data. The links in the side presents one video with an example application, using a Graphical User Interface (GUI) developed in MATLAB. The source code for the TerraLib class is available at 
TerraAIDA is a package of image processing operators used with