:: The basic idea of work ::
The object of research of the dissertation work is the interactive program system of character recognition, it’s internal structure. Artificial neural networks lie at the heart of system.
The task of recognition of patterns or graphic samples belongs to the class of NP-tasks. It requires search of new solution techniques. At processing of the document image it is possible to select some stages – selection of the fields containing prospective graphic images, recognition of graphic images, recognition results checking. At each stage especial processing methods are used [1].
As primary sources of the information for licence plate are used, as a rule, video- and a photo-images. Thus on them can be objects of an any kind [5]. Therefore before the direct analysis of the represented subjects it is necessary to execute a number of the preliminary operations, allowing to receive the image of objects itself without extraneous images [6].
For an effective employment of that images different approaches to decomposition of data model are used, allowing to present the general model as set of hierarchically interconnected more simple models of a different level of hierarchy [4].
One of the most widespread methods of the decision of these problems is contour segmentation.
Methods of contrast segmentation are used in many areas where objects on analyzed images possess the big complexity that causes high requirements to reliability, accuracy and reliability of research results. Use of computer equipment and mathematical methods in this area allows not only to accelerate process of material processing, but also to raise accuracy of research results [4].
Fast development of digital technique recently opens new opportunities in realization of these methods. The speedup of computer equipment allows to use complex algorithms, and owing to appearance of high-resolution color television gauges it is possible to receive and process color images. New technical opportunities allow to expand considerably a circle of researches, open new ways of the problems decision, that regards the images analysis [2].
The most typical method of contrast segmentation: a method of the certain window and a method of chain codes.
Boundary is the contrast area of the image containing sharp distinction of brightness between two adjacent pixels (as a rule, it is object boundaries). There is a set of various methods of boundaries selection [8].
At a preparatory stage of the certain window method researchers find the areas containing necessary contrast (high or, on the contrary, low). Further the window is created on the assumption of the provisional sizes and the form of required object, and it is considered quantity of sides in "suspicious" areas. If it is in the set range - the object is allocated. The range of quantity of sides gets out experimentally.
The selected image of licence plate represents a two-dimensional monochrome signal. It is divided into set of areas (patterns) of real characters images by contour boundaries. Examples of patterns are shown on fig. 1.
Figure 1 – Licence plate characters
Properties of artificial neural networks allow to use them at a stage of selected licence plates recognition (fig. 2, 3).
Figure 2 – Visualization of selection of licence plates' images. Animation (Ulead GIF Animator 5), 720x288 px, 111 Kb, 4 frames with delay 150 msec between frames; the quantity of playback cycles is limited 10th.
Figure 3 – Binarized bit image of the entry character and desirable result of recognition – a character serial number
However it is required to select type of a neural network and its architecture for qualitative solution of the task of recognition of digits and Russian characters. They should provide correct reading of the information from the image and it’s analysis for the purpose of automobile license plate identification even in the presence of interferences.
Within the limits of the carried out research matching of the most widespread sorts of neural networks from the point of view of recognition efficiency is fulfilled.
The analysis has shown that neural network with back propagation training algorithm and a Kohonen network give good results.
The research object is the architecture of a neural network with back propagation training algorithm (fig. 4).
Figure 4 - Multilayered perceptron
The analysis of the received results has shown that the three-layer neural network with back propagation training algorithm (15x15x41) with sigmoid activation function possesses the best recognizing ability for this class of printing characters.
Kohonen neural networks [3] are for visualization and the initial ("prospecting") analysis of the data, first of all [9].
During the analysis other modification in which we take more than one neuron-winner has been used, and it has allowed to reduce time of training and to raise quality of recognition.
|
