Texture features based on Fourier transform and Gabor filters: an empirical comparison
Résumé
This paper presents an empirical comparison of two texture descriptors reported in recent publications. The first one is based on Discrete Fourier Transform (DFT) and the other is based on the Gabor filters. The two are compared for texture recognition and retrieval. To have deeper insight in to the role of neighbourhood properties and the filter banks in the texture description, extensive experimentation is performed over the image sets containing noisy and rotated variants of the textures from Brodatz album. A method for estimating rotation variance (RV) of a texture descriptor is also presented, which gives an idea of how rotation-sensitive is a certain texture descriptor. The results establish that the DFT-based features outperform the features based on Gabor filters in noiseless conditions whereas the later outperform otherwise.