Classificação supervisionada de padrões utilizando floresta de caminhos otimos / Supervised pattern classification using optimum path forest

AUTOR(ES)
DATA DE PUBLICAÇÃO

2008

RESUMO

Patterns are usually represented by feature vectors obtained from samples of a dataset, which can be fully, partially or non labeled. Depending on the amount of available information of these datasets, three kinds of pattern identification techniques can be applied: supervised, semi-supervised or non supervised. In this work, we addressed the supervised ones, which are characterized by the fully knowledge of the labels from the dataset samples, and we also proposed a novel idea for supervised pattern recognition based on Optimum-Path Forest (OPF), which models the pattern recognition problem as a graph, where the nodes are the samples and the arcs are defined by some adjacency relation. The most relevant samples (prototypes) are identified and a competition process between them is started, which try to offer optimum-path costs to the remaining dataset samples. We presented here two approaches, which differ from each other in the adjacency relation, path-cost function and the prototypes identification procedure. The first ones uses as the adjacency relation the complete graph and identify the prototypes in the boundaries of the classes, which offer optimum-path costs that are computed as been the maximum path arc-weight between these prototypes and the other dataset samples, in which the arc-weight is given by the distance between their feature vectors. In this case, the OPF algorithm tries to minimize these costs for each sample of the dataset. The other approach uses as the adjacency relation a k-nn graph and identifies the prototypes as the maxima of a probability density function, which is computed using the arc-weigths. The path-cost value is given by the lowest density value among it. The OPF algorithm now tries to maximize these costs. We also presented a generic learning algorithm, which tries to teach a classifier through its erros in a validation set, replacing the misclassified samples by other selected using some constraints. This process is repeated until an error criterion is satisfied. Comparisons with SVM, ANN-MLP, k-NN and BC classifiers were also performed, being the OPF similar to SVM, but much faster, and superior to the remaining classifiers

ASSUNTO(S)

image processing reconhecimento de padrões processamento de imagens pattern recognition inteligencia artificial artificial intelligence

ACESSO AO ARTIGO

Documentos Relacionados