Adding image descriptors to the PEx-Image tool

• Grant number: 08/11096-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: March 01, 2009
• End date: February 28, 2010
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: João Do Espirito Santo Batista Neto
• Grantee: Renan Ribeiro da Silva
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

An exciting task in visualization is that of multi dimensional data projection by means of techniques capable of mapping a m-dimensional space into another of smaller dimension, typically 2 ou 3. Such projection must encompass, in the projected space, the most relevant data relations. This is the main goal of a broader project (PEx - Projection Explorer), a resulting effort of CNPq Projeto Universal (15/2007, # 484256/2007-6), entitled MultiVis. PEx-Image is an extension of such project and aims to produce image projections from features computed over such data. Although some projections have revealed consistent groupings for many image data sets, the current evaluation is subjective, solely based on the observer's point of view. When images are projected close to each other, then one can infer that the set of features responsible for the projection is consistent. Otherwise, another feature set should be desirable. An alternative means to provide a more accurate and reliable outcome over such projections is the introduction of quantitative elements capable to measure the quality of the features employed. This can be achieve if a pattern recognition behavior is added to the process. These elements are: a) more representative image descriptors; b) ability to classify data as well as the already available projection scheme and c) means to reduce the numbe of features, which is capable of selecting among the entire set of features those that best represent the image data. The work which aims at integrating both elements has already been started in our research group and has currently been investigated by MSc and PhD students.This research project is part of the integration process just described, in which pattern recognition elements will be brought into multi-dimension data projection toll PEx-image. Its main goals are: a) implementation of new image descriptors or features; b) generation of mechanisms which will ease the task of integrating the new descriptors to the PEx-Image's graphical interface, originally written in Java. As a result, we hope to enhance the quality of the projections so far available, as the addition of new image descriptors (shape, color and texture) will be more natural and quicker and, also, more easily integrated into the PEx-Image tool.