Fig. 3.

PS statistics used in the texture synthesis. A similar description can be found in McDermott and Simoncelli (51). The statistics are grouped into seven categories represented by different colors. The terms used for each group are based on those used by Freeman et al. (7). The uppermost image represents the input texture. Marginal statistics were directly computed from that texture and include statistics from the luminance histogram, including mean, variance, skewness, and kurtosis; in the minPS, only the skewness was incorporated. The image was convolved using Gabor-like filters with different scales and orientations to generate “Responses of linear filters.” From these responses, the correlations between spatially neighboring filters (Linear cross position) and the correlations between filters with neighboring scales (Linear cross scale) were computed. The Responses of linear filters were then converted to “Responses of energy filters” by taking the amplitudes of the responses. From those, average amplitudes of filter outputs (Spectral), correlations between filters with neighboring orientations (Energy cross orientation), correlations between spatially neighboring filters (Energy cross position), and correlations between filters with neighboring scales (Energy cross scale) were extracted.