CIRCULAR NEIGHBORHOOD AND 1-D DFT FEATURES FOR TEXTURE CLASSIFICATIONAND SEGMENTATION

The authors introduce a texture descriptor that utilises circular neig hbourhoods and l-D discrete Fourier transforms to obtain rotation-inva riant features. Since rotating an image does not change the intensitie s of its pixels but shifts them circularly, rotation-invariant feature s can be realised if the relationship between circular motion and spat ial shift Is established. For each individual circular neighbourhood c entred at every pixel, a number of input sequences are formed by the i ntensities of pixels on concentric rings of various radii measured fro m the centre of each neighbourhood. Fourier transforming the sequences would generate coefficients whose magnitudes are invariant to rotatio n. Features extracted from these magnitudes were used in various class ification and segmentation experiments. These features outperformed th ose of the circular simultaneous autoregressive model in classifying r otated images and those of the wavelet transform and the Gaussian Mark ov random field in classifying unrotated images of 30 classes. They al so showed superior performance to those of the CSAR in several rotatio n-invariant segmentation experiments.