Image analysis and scene segmentation usually have to cope with the problem of binding different features in a global unifying percept. Biological approaches to computer vision and image processing have evidenced the advantages of a delocalized representation of information through feature maps which mix topological contiguity with proximity in the feature space. By example, the peculiar dispositions of orientation-selective cells can be understood in terms of dimension-reducing mappings which translate neighborhood relations in the orientation-subspace in spatial neighborhood relations on the cortical surface . Thus, cortical maps are not only a repository of information about the features present in the image, but are the substrate on which to organize a locally based set of excitatory and inhibitory interconnections among features. The overall functionality of these systems will depend both on their structural schemata and on the "physics" of the interaction processes. In this framework, the spreading of features on the cortical surface poses, at some point, the problem of a global coding that encourages grouping processes among visual fetaures that belongs to a common object, and dissociates those features that belongs to distinct objects, though they be equally close together in the cortical map. After recent observations that synchronous oscillations in different regions of visual cortex reflect global properties of the stimulus, several cooperative neural networks have been proposed in which global properties of the image can be read in the correlated activity of neural populations . Usually such cooperative oscillating systems are based on: (i) non-linear self-catalyzing action with a delayed negative feedback, or (ii) cooperative-competive action between two populations acting through cross-coupled non-linear delay differential equations. In this paper we have analyzed the computational capabilities of continuous-distributed coupled diffusion maps governed by laterally-reacting non linear equations. Coupled diffusion maps can be interpreted as a phenomenological model for the processes that take place in the cortical surface. Such systems are based on (i) linear spatial diffusion components, and (ii) non-linear reaction components. The former allow for a weighted averaging on the input signal over a region of a certain extent (the receptive field), and, through anisotropic diffusion coefficients, lead to continuous distributions of oriented receptive fields similar to real orientation maps . The latter ignite cooperative processes. Specifically, activation in a point of the network sets in motion a local excitatory diffusion process and a spatially displaced inhibitory reaction through lateral clustered coupling. The non-linear competition between the two processes leads to a macroscopic organized dynamic behaviour that, under appropriate input conditions, results in oscillations. Superimposed dynamic coupling between localized portions of the orientation map, characterized by similar response properties, impose a long-range coherence and the whole system evolves throught different configurations of coherent states each representing features related to common textured regions. Hence, diffusive-reacting processes allow the systems to show phasic transitions from disordered states to ordered states or from one ordered state to a different one, and long-range connections synchronize such transitions on how coherent features are in the image. Averaging the pattern of activity over a window comparable in size with those of a complete set of orientations (hypercolumn) we gain istantaneous global perception supporting segmentation. Experiments conducted with artificial and natural textures validated the approach.

Back to index
last updated: 03, November 99 by