Heterogeneous pattern recognition and its applications in biology and nanotechnology
| Grant number: | 09/16206-1 |
| Support type: | Research Grants - Young Investigators Grants |
| Duration: | March 01, 2010 - February 28, 2014 |
| Field of knowledge: | Physical Sciences and Mathematics - Computer Science |
| Principal Investigator: | João Paulo Papa |
| Grantee: | João Paulo Papa |
| Home Institution: | Faculdade de Ciências (FC). Universidade Estadual Paulista (UNESP). Campus de Bauru. Bauru , SP, Brazil |
| Assoc. researchers: | Alexandre Xavier Falcão ; Aparecido Nilceu Marana |
| Associated grant(s): | 14/02424-5 - 22nd International Conference on Pattern Recognition,
AR.EXT 12/13988-1 - The 37th IEEE conference on local computer networks, AR.EXT |
| Associated scholarship(s): | 13/20387-7 - Hyperparameter optimization in deep learning arquitectures,
BE.PQ 12/06472-9 - Exploring contextual classification approaches for Optimum-Path Forest, BP.DR 11/14058-5 - Exploring sequential learning approaches for Optimum-Path Forest, BP.MS + associated scholarships - associated scholarships |
Abstract
Traditional pattern classifiers, such as Support Vector Machines (SVM) and neural networks, pay the price for an expensive training phase to achieve acceptable recognition rates in a test set. Thus, these techniques are inviable in situations that require a permanent data retraining, and mainly in which we have large datasets (interactive segmentation of magnetic resonance images of the brain and segmentation of ferrous allow samples obtained from high resolution metallographic images, for instance). Recently, a new pattern recognition algorithm called Optimum-Path Forest was proposed in the literature, which has been demonstrated to be superior than artificial neural networks and bayesian classifiers, and similar to SVM, but extremely faster (500x-1000x, depending on the dataset size). The OPF also received 3 prizes in 2009. The OPF classifier models the data classification task as a partition problem in a graph induced by the feature space into optimum-path trees (OPTs), in which each sample is stronger connected to the root of its tree than to any other root in this forest. Samples that belong to the same OPT receive the same label in the data classification process. Although OPF classifier has been used in several research topics in the last 2 years, such as remote sensing, computer vision (digital fingerprint and face recognition), automatic identification of human parasites and biomedical signal processing, there exists many others research areas that need to validate the OPF applicability. This research project has as the main goal a wide and complete study about OPF classifier, as well the development of its new variants, its implementation in GPU (Graphics Processing Unit), and to validate the OPF applicability in other research topics and in situations that require large datasets, which cannot be solved with both efficiency and effectiveness by the traditional pattern recognition methods, such as neural networks and SVM. This project also aims to apply OPF for object tracking and signal processing. Cooperations with several national and international research groups working with the same objective, i.e., to divulge and to validate the OPF classifier, will be addressed. Recall that all proposed works inside this project are innovative, due to the fact of each one of them to address one research topic that was not already explored by the OPF classifier. This research project address activities in several research levels, such as undergraduate and graduate studies. (AU)