Abstract
Content
- Introduction
- 1. Theme urgency
- 2. Goal and tasks of the research
- 3. Stages and methods of image processing and vehicles license plates recognition
- 3.1 Methods of image processing at localization of a license plate
- 3.2 Segmentation of a license plate
- 3.3 Methods of pattern recognition
- 3.4 Character recognition
- Conclusion
- References
Introduction
Today an important aspect of road safety and control is identification of cars by their license plate registration. License plate recognition systems have various applications, such as motor transport enterprises, control of entry to the territory of the enterprise and movement of vehicles at sites with limited access, gas stations, speed control, parkings [1].
1. Theme urgency
The urgency of the problem of vehicles license plates recognition is that with every day increases a need for automation of the control of entry to the territory of enterprises, speed control and determination of violation of traffic rules.
Currently, there are a lot of license plate detection systems, but not all of them are high-quality and reliable products. Systems with high speed and accuracy of recognition are commercial, classified and expensive, which does not allow to carry out their mass introduction [1].
It is important to determine the most effective methods of image processing and license plates recognition, as well as the creation of new combined or modified methods for high-quality recognition of system in real time.
2. Goal and tasks of the research
The goal of this masters work is to research and analysis of methods of image processing and vehicles license plates recognition, the development of new combined or modified methods.
Achieving this goal requires the solution of the following tasks:
- Analysis of methods and implementation of image processing at localization of a license plate.
- Analysis of segmentation methods.
- Research of pattern recognition methods.
- Development and research of algorithms for character recognition.
- The choice of methods of recognition and assessment of their effectiveness.
- Development of a system for image processing from surveillance cameras for the vehicle license plate recognition.
The problem of license plates recognition is divided into three stages: image preprocessing, segmentation, character recognition.
Image preprocessing is necessary for improvement of visual quality of the image and, finally, for localization of a license plate.
Segmentation is carried out to highlight the symbols for further recognition by the chosen method.
As part of the master's work is planned to investigate the methods of vehicles license plates recognition in order to find an efficient algorithm, and implement an information system that solves above problems.
4. Stages and methods of image processing and vehicles license plates recognition
3.1 Methods of image processing at localization of a license plate
The inputs are the images from the camera are often noisy, blurred, therefore is carried out preprocessing of them by smoothing and rank filters (Gaussian and median filters) to eliminate the additive and impulsive noise.
Smoothing and rank filters
A two-dimensional discrete Gaussian function with zero mean is used in image processing:
The filter, which is constructed on its basis, serves for smoothing. In this case, manipulations carried out with a parameter σ².
At a median filtration is used a two-dimensional window (mask filter), which has a central symmetry. The center of a window is located at the current point of filtering. The form of a window can be varied. Aperture sizes are optimized in the course of the analysis process, depend on image detail.
Values of the elements of the working selection are ordered on increase. The median is an element which occupies a central position in the sequence. If the central value is noise emission, the filter will provide its suppression.
Edge (contours) detection
Edge detection is carried out by means of gradient filters of the first and second order. Apply, for example, the second order filter – LoG-filter (Marr-Hildreth), which works by checking the image with Laplacian Gaussian function. It combines the edge detection with smoothing. The filter G mask is used:
The result of the LoG filter is presented in Fig. 1.
а) б)
Figure 1 – а) the original image, b) with allocated edges
3.2 Segmentation of a license plate
Segmentation algorithm searches for the coordinates of alleged symbols in a localized area. At the first stage of segmentation likelihood estimates of accessory of pixels to lines of characters are calculated. Result of the first stage is an array likelihood estimates EZhxw.
The second stage is to calculate the vector of mean likelihood estimates on columns:
The vector A reveals the gaps of background between the characters that appear in the form of extremes, if we consider A as a function.
At the third stage are detected vertical dividers are the edges between the characters:
Dev = {dev1,dev2,…,devk}, где k is quantity of the found dividers.
At the fourth stage is carried out refinement of the vertical and horizontal edges characters. Threshold value of brightness is calculated for an image fragment between two delimiters [13].
The morphological erosion is applied to simplification of a problem of license plate character recognition. The goal is to obtain the skeleton of character (the image width of 1 pixel).
Erosion of a binary image A by structuring element B is denoted by A θ B and is given by expression:
The structural element B passes through all the pixels of the image A. If the single pixel B coincides with a single pixel A, we perform a logical disjunction of the center pixel element with the corresponding pixel in the output image. (Fig. 2).
а) b)c)
Figure 2 – а) a binary image A, b) the structural element B, c) a refinement by the structural element of the image B²
As a result of erosion operation all objects smaller than the structural element are erased, and the objects are connected by thin lines are disconnected, and the sizes of all objects are decreased [14].
3.3 Methods of pattern recognition
Basic methods of recognition: correlation, based on decision-making by criterion of proximity with the templates; indicative and syntactic are the least labor intensive.
Multiple-correlation method at completely set template is the most noiseproof and labor-consuming comes down to exhaustive search in the space of signals.
The algorithms based on a method of partial correlations are highly resistant to random and local noises. In this case, partial correlation coefficients that obtained for the individual template fragments in signal space, can be seen as signs.
Recognition by method of two-dimensional correlation is to find the most similar pair of images: template B0i – input image B, i = 1,..,n0. The minimum value of a measure corresponds to the maximum similarity ρ(B,B0i). Recognition in the conditions of geometric distortions is based on the following rule:
where G is a group of alleged distortions; dВ is the threshold value, which is introduced because of the noise in the input image.
Similarity measures between the images can be determined by one of the following formulas:
where B(x,y) and B0(x,y) are mean values of brightness of images B and B0 accordingly; N is quantity of points of area D.
The template images must correspond to a dissimilarity condition:
where n0 is quantity of template images; d0 is the threshold value.
Value signal/noise is defined as follows:
where B(x,y) is mean brightness of images; d is the standard deviation of Gaussian noise.
Indicative and syntactic methods are based on statistical and determined approaches. Sign selection is a major difficulty in the indicative methods. Rules: a) a class of signs of images may vary only slightly, and b) signs of images of different classes should vary significantly, and c) a set of signs to be as low as possible, because their reliability depends on the amount and complexity of processing.
Syntactic methods are based on receiving of the structural and linguistic signs when the image is broken into derivative elements (signs). Rules of compounds of these elements are introduced which are identical to the template and input image. The analysis of grammar provides decision-making [15].
Figure 3 – A visual representation of image processing and vehicle license plate recognition
(Animation: 10 frames, 5 cycles of repetition, 132 kilobytes)
3.4 Character recognition
The problem of license plate recognition is solved by specific methods, which include the methods of classifiers and neural networks.
There are three basic types of classifiers:
- template;
- indicative;
- structural.
Template classifiers converts the original image of character into a set of points, and then placing it on the templates available in the database system. A template that has the least differences is required. They give a sufficiently high accuracy of defective character recognition. The disadvantage is the inability to recognize the font, at least a little different from the mortgaged to the system.
Figure 4 – The generalized algorithm of the template classifier
Indicative classifiers for each symbol is calculated set of numbers (signs) and compare these sets. The disadvantage is an irreversible loss of part of information about the character at the stage of feature extraction.
Structural classifiers store information about the character topology. Structural methods represent an object as a graph, the elements of the input object are nodes, and the spatial relationships between them are arcs. The disadvantage is the difficulties at defective characters recognition [16].
At the solution of a problem of character recognition are widely used the following types of neural networks: multi-layer neural network, Hopfield neural network, Kohonen neural networks, etc.
The architecture of multi-layer neural networks (MNN) consists of serially connected layers where the inputs of each layer neuron connected to all the neurons of the preceding layer, and outputs connected to all the neurons of the next layer.
Hopfield neural network is single-layered and fully connected, its outputs are connected to the inputs. Unlike MNN is relaxation, i.e. being set to an initial state, operates until it reaches a stable state, which will be its output value.
Kohonen neural networks provide a topological ordering of the input space images. They allow continuously to display the input n-dimensional space in the m-dimensional output [17].
A key aspect of the neural network is its training that comes down to the definition of connections (synapses) between neurons and the establishment of the strength of these connections (weight coefficients). Neural network training algorithms are simplistically reduced to determination the dependence between the weight coefficient of the coupling of the two neurons on the number of examples to prove this dependence. The most widespread neural network training algorithm is back-propagation algorithm. The objective function should to provide minimizing of a square of a mistake in the training by all the examples:
where Ti is the preset value of an output sign on i-th example; yi is the calculated value of an output sign on i-th example [18].
Figure 5 – Multi-layer neural networks with back-propagation algorithm [19]
The essence of back-propagation algorithm is as follows:
- Specify randomly small initial values of the weights of connections of neurons.
- For all training pairs «values of the input signs – the value of the output sign» to calculate the output of the network (Y).
- Execute recursive algorithm starting from the output nodes towards the first hidden layer, until it reaches the minimum level of mistake [18].
Conclusion
Research and definition of the most suitable to a specific task methods of image processing and characters segmentation will allow reduce the requirements for the recognition, which is important in the problem of vehicles license plates recognition.
Within the conducted researches is executed:
- Based on an analysis of references are allocated basic stages and algorithms that can be used in a given system design. Necessary for solving this problem are the noise-canceling filtering, edge detection, skeletonization and segmentation, and the most appropriate methods of recognition are correlation methods, template methods and neural network.
- The software requirements were evaluated, search functionally similar software products was executed and the analysis of advantages and disadvantages was carried out.
- Images pre-processing was implemented to improve their quality and convenience of localization of license plate.
Further researches focused on the following aspects:
- A profound analysis of the basic methods of image processing and vehicles license plates recognition.
- The development of the new combined or modified methods for license plates recognition.
- Based on conducted research will be developed a functional system of vehicles license plates recognition, which can be used to check the efficiency of the chosen recognition methods.
References
- И.С. Личканенко. Методы обработки изображений и распознавания образов для задачи обнаружения номерных знаков транспортных средств // Информатика и компьютерные технологии – 2013. – Донецк: Донецкий национальный технический университет, 2013
- Авто-Инспектор – система распознавания автомобильных номеров [Электронный ресурс]. – Режим доступа: http://www.iss.ru/products/intelligent/auto/
- АПК «АВТОУРАГАН» [Электронный ресурс]. – Режим доступа: http://www.recognize.ru/ node/23
- Fahmy M.M.M., 1994, Automatic Number-plate Recognition : Neural Network Approach, Proceedings of VNIS’94 Vehicle Navigation and Information System Conference, 3 1 Aug-2 Sept, pp.291-296, 1994
- Nijhuis J.A.G. , Brugge Ter M.H., Helmholt K.A., Pluim J.P.W., Spaanenburg L., Venema L., Westenberg M.A., 1995, Car License PlateRecognition with Neural Networks and Fuzzy Logic, IEEE International Conference on Neural Networks, pp.2232-2236, 1995
- Lotufo R.A., Morgan A.D., and Johnson AS., 1990, Automatic Number-Plate Recognition, Proceedings of the IEE Colloquium on Image analysis for Transport Applications, V01.035, pp.6/1-6/6, February 16, 1990.
- Kim S.K., Kim D.W., and Kim H.J., 1996, A Recognition of Vehicle License Plate Using a Genetic Algorithm Based Segmentation, Proceedings of 3rd IEEE International Conference on Image Processing, V01.2., pp.661-664, 1996
- Hontani H., and Koga T., (2001), Character extraction method without prior knowledge on size and information, Proceedings of the IEEE International Vehicle Electronics Conference (IVEC’01), pp. 67-72.
- Система распознавания автомобильных номеров «НомерОк» [Электронный ресурс]. – Режим доступа: http://avtonomerok.com/
- Животченко О.В. Реферат по теме выпускной работы: Разработка компьютерной системы обработки изображений с камер видеонаблюдения для определения номерного знака транспортного средства. – Портал магистров ДонНТУ [Электронный ресурс]. – Режим доступа: http://masters.donntu.ru/2012/fknt/zhivotchenko/diss/index.htm
- Федоров А.В. Реферат по теме выпускной работы: Исследование методов контурной сегментации для построения системы оптического распознавания символов. – Портал магистров ДонНТУ [Электронный ресурс]. – Режим доступа: http://masters.donntu.ru/ 2010/fknt/fedorov/diss/index.htm
- Грузман И.С., Киричук В.С. Цифровая обработка изображений в информационных системах: Учебное пособие. – Новосибисрк: Изд-во НГТУ, 2002. – 352 c.
- Воскресенский Е.М., Царев В.А. Моделирование и адаптация систем распознавания текстовых меток на видеоизображениях. – Череповец: ИНЖЭКОН-Череповец. – 2009. – 154 с.
- Математическая морфология [Электронный ресурс]. – Режим доступа: http://habrahabr.ru/post/113626/
- Путятин Е.П. Нормализация и распознавание изображений [Электронный ресурс]. – Режим доступа: http://sumschool. sumdu.edu.ua/is-02/rus/lectures/pytyatin/pytyatin.htm
- Абраменко А. Компьютер читает [Электронный ресурс]. – Режим доступа: http://www.ocrai.narod.ru/fr.html
- Брилюк Д.В., Старовойтов В.В. Нейросетевые методы распознавания изображений [Электронный ресурс]. – Режим доступа: http://rusnauka.narod.ru/lib/author/briluk_d_b/1/
- Тельнов Ю.Ф. Интеллектуальные информационные системы / Московский государственный университет экономики, статистики и информатики. – М.: МЭСИ, 2004.
- Шитиков В.К., Розенберг Г.С., Зинченко Т.Д. Количественная гидроэкология: методы системной идентификации. – Тольятти: ИЭВБ РАН, 2003. – 463 с.