An Interdiscilinary Approach to Visual Pattern Detection: Biologically Plausible Pattern Localisation and Parameter Estimation on Visual Data

A key challenge of developing a generic visual pattern detection system is the handling of variability in natural images. Estimation of multiple parameters that describe the pose of objects relative to a previously captured view or model in the images typically requires a search for an optimum in a high dimensional search space. Inspired from a controversial parallel search mechanism in the recent literature, the problem is tackled through a new neural circuitry model called Monte Carlo Map Seeking Circuit (MC-MSC). This replaces the regular sampling of transformation parameters in the original Map Seeking Circuit with a probabilistic sampling approach. Another novelty of this work is the ''queuing'' approach which serialises the search by a small amount and increases the performance considerably. This serialisation approach can also be considered as a rough estimation to the ''attentional mechanisms'' known to exist in primate vision strategy.