Multistaged self-organizing neural network with biologically inspired preprocessing features for rotation and scale invariant pattern recognition

The authors present a pattern recognition system that self-organizes to recognize objects by shape. The network uses a combination of a log polar image mapping and a neural network implementation of a dynamic thresholding mechanism, a translation invariant transformation, and a modified neocognitron to produce a network that can recognize learned patterns without regard to their rotational orientation or size. The network´s three layers perform the functionally disjoint tasks of preprocessing, invariance, and recognition. The network performed successfully on rotated and scaled images (except when the objects were very small) and had some tolerance to distortions and noise