Circular neighbourhood and 1-D DFT features for texture classification and segmentation

The authors introduce a texture descriptor that utilises circular neighbourhoods and l-D discrete Fourier transforms to obtain rotation-invariant features. Since rotating an image does not change the intensities of its pixels but shifts them circularly, rotation-invariant features can be realised if the relationship between circular motion and spatial shift Is established. For each individual circular neighbourhood centred at every pixel, a number of input sequences are formed by the intensities of pixels on concentric rings of various radii measured from the centre of each neighbourhood. Fourier transforming the sequences would generate coefficients whose magnitudes are invariant to rotation. Features extracted from these magnitudes were used in various classification and segmentation experiments. These features outperformed those of the circular simultaneous autoregressive model in classifying rotated images and those of the wavelet transform and the Gaussian Markov random field in classifying unrotated images of 30 classes. They also showed superior performance to those of the CSAR in several rotation-invariant segmentation experiments.