Representação de imagens através de grafos utilizando o algoritmo Split-And-Merge combinado com descritores de cor e textura

Abstract

This master thesis presents the use of graphs to image representation. There are several approaches based on graphs in computer imaging, such as image retrieval based on content, tumor detection in medical imaging and image recognition, which is one of the most important area of the image proccessing research. When an image is recognized by its comparison to a model and both are represented by graphs, the matching graph problem is then caracterized, where the criteria for correspondence is defined as a set of values to measure the similarity between the model and the image to be recognized. An interface for the manipulation of images and their corresponding graphs, called ImGraph, was implemented as a undergraduation computer science course final project. This tool is able to create attributed graphs that represent the images (graphs) and their features (attributes). This dissertation consists of an extension of an known automatic segmentation algorithm investigation with color and texture descriptors. The combination of color and texture characteristics is analysed for the segmentation algorithm chosed with the objetctive of improving the image representation by attributed graphs. The split-and-merge segmentation approach is used, recursively spliting the image into smaller homogeneous regions and then joining them according to some similarity measure, considered in this work as probabilistic distances. It was also analyzed the combination of texture descriptor LBP (Local Binary Pattern) with the color descriptor, on the premise that some regions of the image can be better described not only by a specific descriptor, but also by a combination of them. Also, to reduce the size of the RGB color space of three dimensions to one dimension the PCA (Principal Components Analysis) method was used. As another contribution of this thesis, two mathematical functions defined from the attributes of the graphs for the similarity computing were proposed and investigated. All the cited contributions were added to the ImGraph, for its improvement. Experiments were conducted using a high resolution aerial space image of the Jucu River, in the state of Espírito Santo.