Early - Visual Features as Determinants of PerceivedTexture

This thesis is about the perception of texture similarity and the procedures underlying its computation. A model of texture similarity was engineered using a collection of simulated early-visual operations. These operations include oriented, multi-scale linear ltering; rectifying non-linearity; divisive normalization; low-order stochastic modeling; recursive cascading; and whitening. The rst aim of this thesis was to illuminate the gap between these low-level operations and high-level judgements of perceived texture similarity in humans. The second aim was to assess the merits and limitations of early-visual operations as a practical basis for image retrieval. A taxonomy of early-visual models were used to retrieve natural textures from the Brodatz database. The fraction of correct textures found versus the total textures retrieved was used as an objective performance criterion. The most accurate early-visual model achieved a recall accuracy on par with the best known SAR 29] model, dramatically outperforming benchmark models by Haralick 15], Tamura 43], and Kabir 23]. In addition, early-visual texture models exhibited very favorable cost/performance characteristics: one model attained a recall accuracy of 80% using only 8 oating-point numbers to represent each texture in the database. Previous work has shown that humans categorize natural textures along three orthogonal dimensions of periodicity, directionality, and complexity 39]. A pilot psy-chophysical experiment was performed to test how well early-visual features could