Abstract
A shape recognition system is presented that learns novel shapes (like letters) in one shot and recognizes them even if they are distorted, changed in size, shifted in position, fragmented or disrupted by ‘noisy’ contours. A shape is represented as a field of local orientations, which is analogous to an optical flow field. This orientation field is obtained by looking at the contour propagation flow of a shape through a set of orientation columns and it is thus called the ‘propagation field’. Recognition occurs firstly by determining the local orientations of the input shape, followed by matching them against the propagation field of each shape. A neural architecture is presented that performs this template matching process. It consists of three separate components: 1) a ‘propagation map’ that propagates contours, which is employed for learning and partly for recognition; 2) a set of orientation columns that determine local orientations in the propagation map; 3) a stack of ‘shape maps’, in which each map encodes a separate propagation field: a shape map consists of a 2D field of neurons with each neuron having the same number of synapses as there are distinguishable orientations. Each synapse (in each shape map) receives input from its spatially corresponding orientation cell. To learn a novel shape, a new shape map is employed and only those synaptic weights are turned on, whose corresponding orientations are stimulated during contour propagation in the propagation map. During recognition, the shape map with the most similar propagation field is activated the highest and its population activity signals the presence of a shape. Due to the wide-spread representation of a shape, recognition is enormously robust and occurs without error. The recognition system has already been presented in book form (Rasche 2005, The Making of a Neuromorphic Visual System, Springer, New York) but is here publicly presented the first time.
This work was supported by grants from the National Institute of Mental Health (1 R03 MH59845) and The PennsylvaniaState University, both to Michael J Wenger