Use of Gabor filters for texture classification of digital images

Abstract

In this article various methodologies, based on the use of Gabor filters, are described and analysed for the extraction of texture features and the subsequent classification of aerial and satellite digital images. Images of urban, forest and agricultural areas were used, where the complexity of the terrain and the differences in vegetation density require the consideration of the existing texture features as a base for elaborating land use cartography. The use of Gabor filters is driven by the potential they have to isolate texture according to particular frequencies and orientations. The parameters that define a Gabor filter are its frequency, standard deviation and orientation. By varying these parameters, a filter bank is obtained that covers the frequency domain almost completely. Several alternatives have been studied for the application of Gabor filters: (a) the use of complete filter banks; (b) the sum of the filters of equal frequency; and (c) the selection of those filters that minimise, a priori, the classification error. From the application of filters in each of the three methods, a group of images is obtained that allow for the numeric quantification of textures in the image. The evaluation of the classification results shows that combining these textural variables with the multispectral information permits us to characterize the existing regions in the territory with more precision, using supervised digital image classification techniques.