Análise de Desempenho de Classificadores Baseados em Redes Neurais, Máquinas de Vetores de Suporte e Florestas de Caminhos Ótimos para o Reconhecimento de Dígitos Manuscritos

Aparecido Nilceu MaranaJoão Paulo PapaGiovani Chiachia

The optical character recognition (OCR) is animportant tool in computer vision. Several researcheshave been motivated by the development of systems thatcan automatically recognize digital text. In this context,an interesting problem is the automatic reading ofhandwritten text, in which several applications can beaddressed. In such a way, this work aims to compare therobustness of four important classifiers with respect tohandwritten numeric digits recognition: artificial neuralnetwork using multilayer perceptron (ANN-MLP),support vector machines with radial basis function (SVMRBF)and linear (SVM-LINEAR) kernels and optimumpathforest (OPF), which is a novel graph-basedclassifier recently introduced in the literature. We are thefirst to introduce the OPF in the OCR research field aswell to perform a detailed study about the SVM-basedand ANN-MLP classifiers' performance in this context.We used a dataset composed by 2000 handwrittensamples, in which each digit was represented by itsfeature vector composed by 72 features extracted by anon homogeneous sampling from the Hough space. Theexperiments showed that all classifiers achieved goodperformance, which can demonstrate the robustness ofthe features extracted from the Hough space. Highlightsto the SVM-RBF classifier accuracy and to the OPFclassifier training speed, which was much faster.

Caso o link acima esteja inválido, faça uma busca pelo texto completo na Web: Buscar na Web

Biblioteca Digital Brasileira de Computação - Contato:
     Mantida por: