Spatial Duality : An Evolutionary Approach to Texture Classification

Maxim Shoshani
Tuesday, 4.4.2006, 11:30
Taub 337

Texture Classification attracts attention in most diverse fields of research. A significant number of them concern relationships between surface processes and imagery patterns/ textures (urban morphology, cells' biology , ecology, geomorphology, etc.). Such patterns evolve over time and each of their imagery realization is a snapshot reflecting a stage in a process. However, a full imagery description of spatial processes is frequently lacking due to technical limitations on temporal and spatial resolutions and extents. In a recent study we have employed an implicit reconstruction technique based on progressive gray level partitions, capable of extracting information regarding pattern evolution from a single image. The result is a transformation of an image' textures into sequence of binary (black and white) images representing evolution of patch patterns. Modeling, characterizing and monitoring patch pattern dynamics through their spread/ dissolve, expansion/contraction and agglomeration/ break down is highly active area of research. However, most of this research is restricted to foreground (e.g., white) patterns rather than to relationships between patterns of foreground and background (e.g., black) phenomenon. Spatial Duality approach concerns modeling mutual changes in pattern characteristics of these two complementary units along sequences of images representing the surface processes. These characteristics are determined by the use of metrics (such as Shannon Weiner fragmentation index), and consequently there are formed two sequences of metrics : one for the foreground patterns and the other for the background patterns. Mutual changes in sequences of properties of foreground and background patches' arrangements are closely linked to texture types in the image dismantled into progressive partitions. Analysis of these properties and of relationships between selected foreground/ background patterns' metrics indicated that they are highly related to deriving forces explaining the evolution of the observed textures. Implementation of this type of analysis to 20 natural and artificial textures (not representing formation processes) revealed that most of them are uncorrelated and may serve for discriminating between texture types. Application of clustering techniques provided meaningful information regarding similarities and dissimilarities between texture types. The combination of implicit reconstruction of evolutionary sequences of binary maps from a single image together with spatial duality offers a new conceptual framework for dealing with texture classification. The implementation of the methodology is relatively simple and shows high potential in extracting meaningful information from images of both natural and artificial scenes. .

Back to the index of events