|
In my master's work the problem of pattern recognition with use of artificial neural networks is researched. A practical part of the given work is creation of the program - realizations of a neural network (in a simple case perceptron), and training of neural network on set of images (the graphic raster files containing the images of letters of Russian alphabet).
|
|
|
Last 10 years there is an active development of analytical systems of new type. In their basis - the technologies of an artificial intelligence, which simulating natural processes, such as activity of neurons of a brain or process of natural selection. The most popular and checked up from these technologies are neural networks and genetic algorithms. The first commercial realizations on their basis have appeared in 80th years and have received a wide circulation in the developed countries.
|
|
|
Neural networks in general sense are imitations of a brain, therefore with their help successfully solve various "indistinct" problems - pattern recognition, speech recognition, the hand-written text recognition, revealing of laws, classification, forecasting. In such problems, where traditional technologies are powerless, neural networks often jut out as a unique effective technique of the decision. Genetic algorithms are a special technology for search of optimum decisions which is successfully applied in various areas of a science and business. In these algorithms the idea of natural selection among living organisms in the nature is used, therefore they are called genetic. Genetic algorithms often are applied together with neural networks, allowing to create extremely flexible, fast and effective tools of the analysis of data.
|
|
|
By means of neural networks very big circle of problems - recognition of speech, of a sound, diagnostics in medicine, the analysis and forecasting of economic parameters is solved. Research of a problem of recognition of letters is carried out the master's work, planned to write the corresponding program, allowing to make recognition. Novelty of the master's work are the methods used in preparation of images to recognition, and also the algorithms used in the program. Also the separate purpose is reduction of time of training of a network and increase in speed of work of the program.
|
|
|
Pattern recognition is generally reference of object to any set of images. In our case all variants of a writing of the separate letter concern to one set.
|
|
|
In general the problem of pattern recognition consists of two parts: training and recognition. Training is carried out by display of separate objects with the indication of their accessory to this or that image. As a result of training the recognizing system should get ability to react identical reactions to all objects of one image and various - on all objects of various images. It is very important, that process of training should come to the end only by displays of final number of objects without any other helps. As objects of training there can be either pictures, or other visual images (letter), or the various appearance of an external world, for example sounds, states of an organism at the medical diagnosis, the status of technical object in control systems, etc. It is important, that during training objects and their accessory to an image are specified only. Training is followed with process of recognition of new objects which characterizes actions of already trained system. Automation of these procedures also makes a problem of training of pattern recognition. In that case when the person himself solves or thinks out, and then imposes to the machine a rule of classification, the problem of recognition is solved partially as a main problems (training) the person take upon oneself.
|
|
|
Problems of classification (such as recognition of letters) is very bad for algorithmization. And if in case of recognition of letters the correct answer is obvious to us in advance, in more complex practical problems the learned neural network acts as the expert, who owns wide experience and is capable to answer of a difficult question.
|
|
|
|
|
|
The list of the used literature
|
- NeuroProject - Обучение. Учебник
- Курсовая работа Смирнова Евгения
- И. В. Заенцев. Нейронные сети: основные модели (Учебное пособие к курсу "Нейронные сети")
|
|
