Computational requirements for classification of textured scenes and object contour shape are complimentary to each other. Texture classification requires measurment of average local oriented energy over multiple scales using a reasonably large patches of the scene. Contour classification and identification, on the other hand, requires computation of sharp contrast or texture boundaries and their interrelationships. A neural model is being devloped, based on the current knowledge of the laminar organization of primary visual cortices, which decomposes image information at different scales of interaction -- local for texture discrimination, and long range for object contour computation. A unified percept is then generated by fusing these individually computed scene attributes.

Textured input is classified locally by a distributed ARTMAP network using multi-scale oriented filter-bank respones. The classifier prototypes for textures are then used to reduce variability in the filter-bank activities to facilitate figure boundary computation. After long-range horizontal completion, this constitutes the input to a shape recognition system in the next cortical region. Boundary and surface based attention reduce feature mixing (not shown above!) at the texture boundaries, improving texture classification performance and resulting in better figure boundary localization. The classification output shown above is without the feedback in place. Newer results to come