Pattern recognition by graph foresting transform with non-smooth connectivity functions

Abstract

This project addresses the pattern recognition problem by extending recent techniques in graphs, which have been used successfully in the context of image processing, by the usage of non-smooth connectivity functions in the framework of the Image Foresting Transform (IFT). The intended techniques will allow a broader view than that currently supported by the recent works involving the classification of patterns by Optimum-Path Forest (OPF). Classifiers generated by OPF have shown promising results in different applications, where its effectiveness is greater than or equivalent to that obtained by neural networks and support vector machines, being tens to thousands of times faster than both. Theoretical aspects of the developed methods will be evidenced through software for creating multidimensional synthetic bases, which will allow the execution of various methods in various scenarios for the purposes of comparative analysis. (AU)