Shape recognition using an invariant pulse code and a hierarchical, competitive neural network

The paper deals with the invariant recognition of patterns, and aims at developing (i) their pulse-coded representation; and (ii) an algorithm for their recognition. The proposed pattern encoder utilizes the properties of complex logarithmic mapping (CLM) (computed with reference to the center of gravity, CoG, of the shape), which maps rotation and scaling in its domain to shifts in its range. The encoder, then, invokes a pulse-encoding scheme similar to the one proposed by Dodwell [1] in order to handle these shifts, thereby generating pulse-codes invariant to scaling, rotation, and shift in the input shape. These pulses are then fed to a novel multi-layered neural recognizer which (i) invokes template matching with a distinctly implemented architecture; and (ii) achieves robustness (to noise and pattern deformation) by virtue of its overlapping strategy for code classification. The proposed encoder–recognizer (E–R), which is hardware implementable by a high-speed electronic switching circuit, can add new patterns on-line to the existing ones. The E–R is illustrated with experimental results. While human visual system has been the main motivation to the proposed model, no claim, however, has been made on its direct biological plausibility.