Augmented multi-layer perceptron for rotation- and scale-invariant hand-written numeral recognition

An OCR system that can recognize hand-written numerals regardless of changes in rotation and scale is proposed. The system consists of two phases. In the first phase, a binary input image is transformed with complex-log mapping followed by the Fourier transform into a rotation- and scale-invariant image. Then the transformed image is fed into a multi-layer neural network, the weights of which are modified by the error-backpropagation algorithm to absorb slight shape distortions. The system was implemented and tested using hand-written numerals. High recognition rates of 90 to 95% were obtained. A method for improving performance is also suggested