Discriminating Structure and Random Texture Images with Binary Fourier Spectrum for Invariant Texture Retrieval

Because of the diversity of the textures, pre-classification before the retrieval seems very necessary and important. We divided texture images into two categories: structure and random, based on binary Fourier spectrum and Shannon entropy as a pre-classification stage. Gabor wavelets is especially useful for texture analysis because of the tunable property of its scale and orientation, but it is sensitive to the directional change of a query image. Scale-invariance Gabor representation is used here after the adjustment step which can reduce the rotation variance of the query image to achieve rotation-invariance, scale-invariance and translation-invariance task in the first class. For the images in the latter class we use a combined rotation and scale invariant Gabor representation. The features we used are much less than conventional Gabor wavelets and therefore it is less time-consuming not only in feature extraction but also in similarity measurement. Simulation results using Brodatz database clearly show that our proposed rotation-invariant, scale-invariant and translation-invariant methods give rise to high retrieval performances.