Equip a Gaussian-vector-field feature extracting mechanism to an MLP for optical character recognition
Tai-Wen Yue, Yu-Jen Chang, Chien-Wu Tsai · 2002
To successfully apply a multilayered perceptron neural network (NN) to pattern recognition, the feature vectors fed into the NN must contain rich representative information so that the NN is able to distinguish the patterns belonging to different classes. For optical character recognition (OCR), the feature vectors, hence, must be endowed with a distortion insensitive property. In the paper, the authors propose a 5-layer perceptron (3 hidden layers) for OCR. One hidden layer is dedicated to extract the so-called Gaussian-vector-field (GVF) feature, which is insensitive to patterns deformed in shapes, of input characters. The other two hidden layers perform hyperregion encoding and decoding functions. A traditional error-propagation learning algorithm is used to train the NN for classifying hand-written numeric characters. Simulation result shows that the MLP can tolerate a large degree of pattern distortion. Furthermore, the size of the MLP is quite small when compared with the other approaches.