Biomedical Engineering Reference
In-Depth Information
Besides the input layer of neurons, the network also contains the output layer of
neurons, which contains the masks of the names of the handwritten characters.
Between the input and output layers are the connections whose weights are changed
during training. Training is produced according to the results of the handwritten
character recognition. If the neural network recognizes characters correctly, the
weights do not change. But if the neural network makes errors, the weights of
the connections leading from the excited neurons of the input layer to the neurons of
the erroneously selected mask name decrease, and the weights of the connections
leading to the neurons of the correct mask increase.
The experimental results
For these experiments created and conducted with D. A. Rachkovskij, ten hand-
written words in English were selected, along with characters from the Latin alphabet
and handwritten digits. During the recognition of handwritten words, the probability
of correct recognition was 99-100% (for the familiar writer). During the recognition
of Latin characters and handwritten digits, this probability was somewhat lower
(98%). For the unfamiliar writer, the probability of handwritten word recognition
was approximately 80%, and after the optimization of the set of features, it was about
84%. Let us consider in detail the optimization of the feature set.
Optimization of the feature set
The purpose of these experimental investigations is to improve recognition rate
through optimization of the feature set. Ten features utilized for handwritten
character recognition during the first stage of studies were selected intuitively.
However, there is an approach that makes it possible to optimize the feature set.
For this approach, an extended feature set containing many more features than the
initial set is also formed intuitively. The problem is to select from the extended
feature set the subset that optimizes a certain function of the quality of recognition.
This problem was solved by the simulation of the evolution of biological species.
The method of evolutionary simulation, or, as it is now called, the evolutionary
algorithm, was proposed in the 1960s and at present is widely used by many authors
[
17
-
20
].
Each “individual” is placed in correspondence with the feature subset that the
“individual” is trained to recognize. Thereafter, the function of the quality of
recognition is calculated. The process of creating the “offspring of the individual”
is simulated using “mutations” and the process of “natural selection.” We did not
use crossing over in this work.
The feature set was optimized for the recognition of ten handwritten words
written by different people. The extended feature set contained 41 features that
are segments of lines and arcs of different length and radii having different
orientations. Each of the features was encountered in the words, but a priori it is
