3-LAYER NEURAL NETWORK MODELS FOR ROTATED PATTERNS RECOGNITION

This paper proposes two types of neural net models for learning and recognizing rotated patterns. They are :(1) a rotation invariant learning model; (2) a copy-learning model. These are all 3-layer neural nets. Their learning is based on the error back-propagation algorithm. The first one has the same recognition ability as Widrow's model while featuring simple learning and recognition mechanisms, making it easy to build a net. The second one employs the efficient use of copying, making it possible to recognize rotation angles, which cannot be recognized with conventional models. Each of the two models features a simple net structure of compact size, and a substantially reduced time, required for learning and recognition. The effectiveness of the proposed models has been verified through experiments.