Extração de atributos de forma e seleção de atributos usando algoritmos genéticos para a classificação de regiões / Shape feature extraction and feature selection using genetic algorithm for region classification

AUTOR(ES)
DATA DE PUBLICAÇÃO

2007

RESUMO

Discrimination power of earth surface targets has been continuously bettered with the increasing remote sensors technological advances. Urban systems applications are one of the areas that recently took benefit of these advances, particularly when using high resolution remote sensing data. In this scale, however, it is often necessary to take into account the shape of the objects to achieve a proper identification when separating objects with the same spectral signature. In this work five new shape features have been considered, in addition to those already existing in a system being developed for region classification. In such system, high feature space dimensionality can be readily achieved, which can jeopardize the image classification, due to information redundancy and high processing costs. This work also uses genetic algorithms (GA) as search strategy for feature selection and dimensionality reduction. These algorithms are based on living creatures evolution mechanisms. In order to reach the best solution, the Jeffrerys-Matusita (JM) distance is used as quality criterion. It is considered the best solution the one which presents the greatest JM average distance between pairs of classes. Two distinct genetic algorithms were tested: binary GA and permutation GA. In order to check the efficiency of the proposed methodology, tests were performed with a synthetic image and a Quickbird image with 0,6m of spatial resolution. In such tests 46 features were obtained from training samples. Feature subsets of size 1, 2, 3 and 4, were selected and in all of them the permutation AG showed a better performance than binary GA. With these selected features, kappa coefficient from 0,8647 to 0,9927 were reached. This research showed the importance of shape features for intra-urban classification; also showed an excellent performance of GA algorithms as search strategy, once, using it, a similar quality as exhaustive search was reached, at a fraction of the cost. This work, in a broad perspective, presented the importance of shape feature extraction and the need of an adequate feature selection procedure in image processing systems devoted to region classification.

ASSUNTO(S)

extração de atributos feature extraction classificação de imagens atributos de forma processamento de imagens algoritmos genéticos region classification pattern recognitio genetic algorithms seleção de atributos shape features computer science sensoriamento remoto remote sensing image processing feature selection reconhecimento de padrões

Documentos Relacionados