Hypotheses for Image Features, Icons and Textons

Abstract

We review ideas about the relationship between qualitative description of local image structure and quantitative description based on responses to a family of linear filters. We propose a sequence of three linking hypotheses. The first, the Feature Hypothesis, is that qualitative descriptions arise from a category system on filterresponse space. The second, the Icon Hypothesis, is that the partitioning into categories of filter response space is determined by a system of iconic images, one associated with each point of the space. The third, the Texton Hypothesis, is that the correct images to play the role of icons are those that are the most likely explanations of a vector of filter responses. We present results in support of these three hypotheses, including new results on 2-D 1st order structure.

Texture Description Through Histograms of Equivalent Patterns

“…It was only recently, in fact, that Griffin et al [16,29] did very clearly set into evidence this concept establishing a direct link between detection of local features and space partitioning. The main difference with the HEP-which deals directly with image patches without previous filtering-is that the method presented in Ref.…”

“…Unser [111] affirmed that, theoretically, any transform of the original image patches satisfying some optimality criteria should always improve the efficiency, when compared with the original representation. Griffin and Lillholm [29] established a relationship between the quantitative description provided by a family of linear filters and the qualitative description of local image structure. Furthermore, certain classes of filters (e.g.…”

Statistics and category systems for the shape index descriptor of local 2nd order natural image structure

“…We adopt the general notion that an alphabet can be found through a partitioning [29,26,31,35,54,60,15,17] of response space into categories of qualitatively similar structures. We will, however, refine this and rather discuss the partitioning of a factored response space.…”

“…One is based on the idea of categories being defined as Voronoi cells around prototypes [12]. The second approach is inspired by Koenderink [29,26,31] and refined through Geometric Texton Theory [15,17].…”

“…We follow recent trends [53,23,46,19] and use the van Hateren Natural Stimuli Collection as a example of a substantial collection of natural images. Specifically, we only use 1220 of the images from the database -see [17] for details. The overall procedure is otherwise identical to the validation of the Gaussian cases described above.…”