Competitive neural network for affine invariant pattern recognition

This paper presents a layered neural network which consists of clusters of competitive cells for affine invariant pattern recognition. The first layer outputs affine invariant patterns via winner-take-all competition between input cells whose weight vectors are equivalent to each other in a cluster. Responding to affine-different patterns, the first layer outputs different patterns when the weight vectors of the input cells in all clusters involve a sufficient number of linearly independent affine-different vectors. Thus, responding to the output patterns of the first layer, the second-layer is capable of affine invariant pattern recognition via winner-take-all competition between the input cells whose weight vectors are generated by the output patterns of the first layer responding to input patterns to be recognized. Similarly, the network with appropriate weight vectors has the ability of pattern recognition invariant to transformations in a subclass of affine transformations. The shift invariance via both the Fourier transform and neocognitron relates to the present network. The performance is examined by computer simulation.