A Neural System for Scale and Orientation Invariant Correspondence Finding

Proposing A Neural Mechanism. Transformation-tolerant recognition seems to still pose a hard challenge for artificial vision systems, as opposed to their biological counterparts. We present a neurally-plausible mechanism to establish local feature correspondences between object images at different scales and orientations while using and explicitly representing the information about the given transformations. The fundamental functionality is based on a network of cortical macrocolumns supporting the Dynamic Link Architecture (DLA, [1]). This architecture can dynamically route information flowing from one domain to another, adapting the initial all-to-all connectivity of processing units to match potential transformations. As result, the mapping, constructed between the local features of the input domain and the model domain, provides solution for the correspondence problem, while the transformation detection mechanism in control domain signals for encountered scale and rotation (see the figure). Both tasks are accomplished in a distributed, parallel fashion, assisting rapid processing.