Robustness of image pyramids under structural perturbations

Abstract Image pyramids have been used by many investigators as computational structures for multiresolution image processing and analysis. We have subjected such pyramids to various structural perturbations and investigated their effects on the functions of the pyramid. The perturbations ranged from adding Gaussian noise to the weights of the generating kernel, to generating a hierarchy of completely irregular tessellations of the image field. We have shown that homogeneous parts of the low resolution representations of the input image may be recovered by renormalizing the corrupted weights. Multiresolution algorithms transposed to irregular (stochastic) structures exhibited only a small decrease in performance. We conclude that pyramidal algorithms are robust and are only weakly dependent on the underlying structure. We propose that some of these pyramidal algorithms may also serve as computational models for perceptual phenomena.