Circular neighbourhood features for texture classification

Four of the most powerful rotation invariant texture descriptors are the CSAR, the wavelet transform with hidden Markov model, Gabor filters and the Laplacian pyramid filters. Even though all of these techniques reported high recognition rates, most of them require a large number of training samples from various orientations to produce good class representations. The CSAR is the only method that can be sufficiently trained with unrotated images to classify rotated images successfully. This paper introduces a new texture descriptor that performs well for rotated or unrotated texture images without being trained with rotated images. A review of circular neighbourhood is given and an examination of the 1-D discrete Fourier transform property is provided. This is followed by a discussion of our rotation invariant features including the experimental results given.