Perceptual texture similarity estimation

This thesis evaluates the ability of computational features to estimate perceptual texture similarity. In the first part of this thesis, we conducted two evaluation experiments on the ability of 51 computational feature sets to estimate perceptual texture similarity using two different evaluation methods, namely, pair-of-pairs based and retrieval based evaluations. These experiments compared the computational features to two sets of human derived ground-truth data, both of which are higher resolution than those commonly used. The first was obtained by free-grouping and the second by pair-of-pairs experiments. Using these higher resolution data, we found that the feature sets do not perform well when compared to human judgements. Our analysis shows that these computational feature sets either (1) only exploit power spectrum information or (2) only compute higher order statistics (HoS) on, at most, small local neighbourhoods. In other words, they cannot capture aperiodic, long-range spatial relationships. As we hypothesise that these long-range interactions are important for the human perception of texture similarity we carried out two more pair-of-pairs experiments, the results of which indicate that long-range interactions do provide humans with important cues for the perception of texture similarity. In the second part of this thesis we develop new texture features that can encode such data. We first examine the importance of three different types of visual information for human perception of texture. Our results show that contours are the most critical type of information for human discrimination of textures. Finally, we report the development of a new set of contour-based features which performed well on the free-grouping data and outperformed the 51 feature sets and another contour type feature set with the pair-ofpairs data.