Электронная библиотека АГНИ

Character recognition involves the translation of handwritten or typewritten text captured by a scanner into machine-editable text. This technology has long been used by libraries to make old documents available electronically. The task is made difficult by variations in how the characters are drawn. Neural networks are particularly adept at overcoming this problem, because they can learn to extract the features that are most relevant to the recognition of the characters. Our goal hence is to create a neural network, as shown in Figure 3, that takes a picture of a character as an input and tells us which letter of the alphabet it represents.

For recognition of the alphabet (26 characters) is chosen package Neural Network Toolbox (NNT), running MATLAB [1], in which you can create linear and radial basis networks, multilayer perceptrons, self-organizing and recurrent networks, and design of control systems dynamic processes. By the means of recognition, digitizes each character uses a network of two layers, not including input from n (10) neurons in the hidden layer and the p (26) in the output neurons (one letter).Schematically, the considered network can be represented diagram shown in Figure 2. This is a two-layer network. Activation function will put a logarithmic sigmoid function, which is useful because the output vectors contain elements with values between 0 and 1, which then conveniently converted to a Boolean algebra [2].

Fig. 2.Two-layer neural network.

The sequence of training:

1. To select the input image from the training set with subsequent submission input vector to the input network;

2. Sizing off the network;

3. Determination of the difference between the network and the vector of goals;

4. Minimizing the error by adjusting the weights and repeat the preceding paragraphs.

228