The synthesis is based on the assumption that ``matching the pointwise statistics of the pyramid representation does matching some of the spatial statistics of the reconstructed image'' [46]. However, as shown in the experimental results, the method produces satisfactory results on stochastic textures with mainly short-range statistical signal dependence but fails on quasi-periodic or regular textures involving large scale features. This indicates that the chosen statistics (the histogram of a pyramid level) are not sufficient for representing especially long-range spatial features. Another main drawback is that the convergence of the iterative synthesis has not been theoretically justified.
Nevertheless, the pyramid-based method is a very first effort on applying non-parametric technique for texture synthesis. The novelty of this method is in the idea of synthesising texture by matching image signal statistics.
