Classification of region using shape feature and feature selection / Classificação de regiões usando atributos de forma e seleção de atributos

AUTOR(ES)
DATA DE PUBLICAÇÃO

2005

RESUMO

With the steady increase in the number of features available from remote sensing sources, there is a growing necessity to reduce the complexity of the classification task. When data dimensionality is very high, a search strategy should be used to select the subset of features that gives the minimum classification error, considering the limited size of training data. Particularly when one deals with the very complex environment that is the urban scene, shape feature extraction is necessary to distinguish different classes of objects which have similar spectral signature. Large number of features and limited number of training samples is the common situation in urban studies which implies the application of feature selection methods adapted to region classification. This work proposes shape feature extraction and application of feature selection methods in systems of image processing. Firstly, we extracted 38 shape and texture features for each training sample. The selected feature subset by each criterion was classified based on greater average Jeffrerys-Matusita (JM) distance. In order to determine the efficiency of the methodology, tests were developed in synthetic and Landsat-TM images. The Sequential Forward Feature Selection (SFS), Sequential Backward Feature Selection (SBS) and NEXKSB (Next subset of an n-Set) were tested to reduce the data dimensionality. These three different feature selection methods were used and an accuracy of 100% was achieved by using 2 features in both synthetic and Landsat-TM images. As there is no a deterministic relation between feature selection methods and classification error, it is possible to conclude that all search strategies should be used to narrow the number of choices assessments based on classification error. This work showed the importance of the shape feature extraction and the necessity of the application feature selection methods in systems of image processing.

ASSUNTO(S)

region classification seleção de atributos reconhecimento de padrões feature extraction pattern recognition image processing extração de atributos processamento de imagens classificação por região feature selection

ACESSO AO ARTIGO

Documentos Relacionados