Rotation-invariant Texture Classification by Spectral Transformation of Gabor Filter Features

摘要

Gabor filters have been proved to be powerful for texture analysis. Gabor filters are constructed as a filter bank in order to capture local textural properties characterized by different frequencies and spatial orientations. Conventionally, the set of power spectrum obtained by Gabor filters are used as texture features. This simple approach has drawbacks that, among others, fine tuning of parameters is required for optimal setting and good performance, and Gabor filter itself does not provide rotation-invariant texture classification scheme. In this paper, we propose a rotation-invariant texture classification method in which texture features are obtained with two processing stages: first, Gabor features are computed in a conventional way, and then Fourier transform is applied to the Gabor features to obtain rotation-invariant texture features. The resulting texture features allow for researchers to select outstanding components of textures and to visually inspect textural characteristics. The proposed method was compared with other methods and showed superior classification results.