To solve the tasks of characters recognition we have developed the optical characters recognition system on basis of neural networks methods. The developed system consists of subsystem of raster image preprocessing and actually of recognition subsystem. Subsystem of raster image preprocessing allows improving the quality of recognition. It includes the following tools:

1. Gaussian filter
2. Sobel operator
3. binarization of original raster image on basis of adaptive threshold

The implemented characters recognition subsystem refers to pattern recognition systems. Algorithm of recognition uses multilayer neural network to recognize a symbol.

The advantages of the developed system are:

1. rather high stability against image defects
2. high processing speed of input data
3. ability to learn

The disadvantages of the system on the other hand are:

1. bad quality of symbol recognition, which font differs from the standard one a lot
2. absence of the tool to remove gradient image background.

Taking in account what was mentioned above we are currently developing an improved algorithm of recognition system. The algorithm is based on representation of each class of symbols as a set of some unique elements and as a relation between these elements. Thus each symbol is represented as a set of unique features, which are invariant with respect to the print. The features selected in the original symbol are being compared with the features of standard symbols and the belonging of a symbol to a certain class is determined.
As a result of our work we will develop a library, which will be able to recognize texts in different languages on the images of various type including newspaper ads and image titles in video file. This library will be integrated with the Real Time Video Stream Recognition System.