Although performances in the high nineties are typically obtained for tasks such as texture segmentation and classification the same cannot be said of judging texture similarity where a classifier has to estimate the degree to which pairs of textures appear similar to human observers. In an investigation of 51 computational feature sets Dong et al. [1] showed that none of these managed to estimate similarity data derived from a population of human observers better than an average agreement rate of 57.76%. Coincidently, none of these computed higher order statistics (HOS) over large regions (≥ 19×19 pixels).We have discovered few methods that encode long-range, aperiodic characteristics of texture; however, it is well-known that such data are critical to human perception of imagery [2,3]. For instance, scrambling phase spectra (while leaving the power spectra intact) will often render imagery unintelligible to the human observer [3]. It is also well-known that humans are extremely adept at exploiting the long-range visual interactions evident in contour information [2,4]. Therefore, we designed an experiment with human observers in order to determine which of three different types of information (2nd-order statistics, local higher order statistics and contour information, see Figure 1) are more important for the perception of texture.Ten human observers were used in a 2AFC (two-alternative forced choice) scheme with 334 texture images drawn from the Pertex database [5]. In each trial the observer was required to compare an original texture image quarter and one variant image quarter ("variant" being one of either contour, power spectrum or randomized block) and decide whether the variant represented the original texture or not (50% of the time they did not). Different quarters of the same texture sample were used in order to prevent observers from performing pixel-wise comparisons. It was found that contour data is more important than local image patches, or 2nd-order global data, to human observers.Figure 1: Each of the three columns shows two images derived from the same texture sample (although not the same physical texture area). The upper row shows unprocessed images. The lower row shows, from left to right, the corresponding contour map, power spectrum image and randomized, blocked image.We therefore developed a contour-based feature set that exploits the long-range HOS encoded in the spatial distribution and orientation of contour segments. A contour is first fragmented into a set of equidistant segments and is then encoded using the spatial distribution and orientation of these segments. Note that images are first processed with the Canny edge detector [6] followed by a morphological erosion operator [7] in order to produce skeleton maps (see Figure 2 (b)). Connected component labelling [7] is performed on skeleton maps. Subsequently, the Moore-neighbour tracing algorithm with Jacob's stopping criteria [7] is applied to each contour and a sequence of points is obtained from each contour. Each contour is t...