Using Face Images to Investigate the Differences between LDA and SVM Separating Hyper-planes

Carlos E. ThomazPaulo S. RodriguesGilson A. Giraldi

Statistical separating hyper-planes have played an important role of characterizing differences between a reference group of patterns and the population under investigation. In this paper, we compare LDA-based and SVM separating hyper-planes for extracting group-differences between images. Our experimental results on face images indicate that SVM is a more robust technique for classification than the LDA-based method. However, the hyper-planes found by both approaches are distinct from the point of view of detecting discriminative information from images. The experiments of this study support the hypothesis that extracting and classifying group differences are two different problems. Variations within each group captured by the LDA-based method, although not important for classification, are very useful for extracting discriminative information. In contrast, the SVM discriminative direction that focuses on the data at the boundary of the classes extracts group differences that are less perceivable on the original image space. This is an important distinction between the two approaches, particularly in extracting discriminative information from groups of patterns where ground truth might be unknown, such as in medical image analysis of a particular brain disorder.

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: