Balqabekova L.U., Utebaeva D.Z.
Almaty University of Power Engineering and Telecommunications
Research
of possibilities of character recognition using neural networks in MATLAB.
Among
the various configurations of artificial neural networks (NN), there are some
in the classification of which the principle of learning, strictly speaking,
does not fit into any supervised learning or learning without a teacher. In
such networks, the weights of synapses are computed only once before the start
of operation of the network based on the information about the data being
processed, and all the learning network is reduced by this calculation. On the
one hand, the presentation of a priori information can be viewed as helping
teachers, but on the other - in fact, just a network of stores samples before
coming to the input real data, and cannot change their behavior, so to speak of
a feedback loop with the "teacher "is not necessary. Out of networks
with a similar logic of the most well-known Hopfield network and the network
Hamming, which are commonly used for the organization of associative memory. Further, we will speak of them.
Hopfield
networks consist of a single layer of neurons, the number of which is at the
same time the number of inputs and outputs of the network. Each neuron is
connected with all other synapses neurons and synapse has one input through
which the input signals. Output signals are usually formed on axons.
Problem
to be solved by the network as an associative memory, usually stated as
follows. Known for a set of binary signals (images, sound, and other data
describing the characteristics of certain objects or processes) that are
considered exemplary. The network must be able to from any non-ideal signal
applied to its input, select ("recall" on partial information)
corresponding to the sample (if available) or "an opinion" that the
input does not match any of the samples. In general, any signal can be
described as a vector X = {xi: i = 0 ... n-1}, n - the number of neurons in the
network and the dimension of the input and output vectors (Figure 1). Each
element xi is either +1 or -1. Denote the vector describing the k-th sample
through Xk, and its components, respectively, - xik, k = 0 ... m-1, m - the
number of samples. When the network will recognize (or "remember") of
any sample on the basis of the data presented to it, its outputs will contain
his name, that is, Y = Xk, where Y - the vector of output values of the network
(Figure 2): Y = {yi: i = 0, ... n-1}. Otherwise output vector does not coincide
with any one exemplary. [1]
As an
example of the network we have considered a simple function (Y = X2) and used
the software package MATLAB (Figure 1). [2] In the MATLAB environment task is
performed using the nntool, which performs the following functions:
-
Formation of numeric arrays of master images used as training;
-
Preparation of the data needed to create neural networks;
- The
creation of neural networks, configure training of neural networks and training
of neural networks.
Figure 1 - Hopfield network
diagram
If, for
example, the signals are some images that will be displayed in graphical form
data from the output of the network, you will see a picture is identical to one
of the standard (if successful) or "free improvisation" of the network
(in case of failure).
Figure 2 – The output data of network
Master
image of each character is represented by a column vector [N, 1], the number of
elements N is equal to the number of features (in other words, N - dimension of
the feature space).
Character
recognition results presented in Figure 2 and show good recognition using
neural networks even in strong distortion (parameter p> 0,1). For an
objective assessment of the performance of neural networks requires a
calculation of the probability characteristics of recognition. With proper
selection of the parameters of the network training and the use of not less
than 100 training patterns can obtain the probability of correct recognition of
characters of the order of 0.6 ... 0.9 (depending on the type of recognizable
characters) in the distortion parameter p = 0,1 ... 0,2.
We
discussed in this article about the network Hamming and we have considered a
simple function. We took its character recognition results by program package
MATLAB.
THE REFERENCES:
1. ru.wikipedia.org
2.
Fedotov A.V., Ph.D., associate professor «Modelirovanie neironnih setei v
matlab»