Evolvable Image Filter based on Rotation Invariant Pixel Correlations

Abstract

As one of the most typical nonlinear filters, weighted median filter has considerable filter behaviour by assigning weights. But each pixel has only one weight in the traditional weighted median filter. Correlations between pixel and other pixels under the filtering template are not considered. A novel rotation-invariant image filter model is proposed here which has a stable filtering performance when filtering on rotated and reversed images, symmetry of the correlations is utilized to reduce the complexity. Cartesian Genetic Programming is adopted to evolve image filter operators. Experiments about rotated images show that this rotation invariant model exhibits robust characteristics. Image denoising of different images show that it has better properties of edge or details preservation, efficient noise attenuation.