A Rotation Invariant Image Descriptor based on Radon Transform

In this paper, we proposed a rotation invariant image descriptor based on Radon transform (RT) and energy operator. Radon transform captures the directional features of the pattern image by projecting the pattern onto different orientation slices, and its most attractive ability is to transform rotational components to circular shift components. Meanwhile, the energy operator can remove the circular shift components. Therefore, the proposed descriptor is invariant with orientations. Moreover, the time complexity of traditional RT is O(N). In order to hasten the procedures, an acceleration strategy was introduced based on projection slice theorem, so that a RT only needs O(NlogN) multiplications. The experiments demonstrate the anti-noising and rotation invariant abilities of the proposed descriptor, and a classification test on 20 distinct natural textures selected from Brodatz’s album shows that the proposed descriptor is superior to wavelet packet analysis, combined invariant feature, and radon transform plus multiscale analysis in respect to classification accuracy and computation time.