Hand written letter recognition with neural networks

Summary form only given, as follows. Neural networks that can recognize 36 handwritten alphanumeric characters are studied. Thin line letters, in 32*32 binary arrays, are used as the input pattern. The system is built from two major units, a three-layered preprocessing unit and a recognition unit. Shift, scale, and deformation tolerance in recognition are provided through reprocessing. Three learning paradigms including an error backpropagation learning, a simple perceptron learning, and a competitive learning are examined and compared.>