Iraqi Plate Number Recognition Using Single Value Decomposition (SVD)

Distinguishing car plate numbers is an important topic of researchers' concern, which assures the process required to be high speed and acceptable accuracy, with the need to access the database and verify it if there was a problem and give a warning if it is necessary. A method is proposed in this paper to distinguish the plate of Iraqi vehicles (new forms), which will prevail in the end, depends on the pre-processing of the image and apply some filters such as (median filter) as well as improving the image before starting the proposed method, which relies on a normalization in the horizontal and vertical direction process and then segment the image into regions. Information is extracted from each region, such as the area that defines the type of vehicle if it is a governmental, private, taxi, or others. The region that characterizes the city as well as Arabic and English, numbers region is segmented and then transform with single value decomposition (SVD) on the image and get features that will send to the database for identification. The proposed method has given a percentage of accuracy of about 90.4 % in the process of discrimination with significant time complexity on average using K-Nearest Neighbor K-NN classifier. It possible to implement the proposed method with application of emergency alarm, to give a warning alert.


Introduction
Several procedures are merged in most algorithms for number plate localization, which results in a long computational and accordingly considerable execution; applying less and simpler algorithms may reduce this process [1].The results highly depend on the quality of image, since the procedures reliability severely degrades in the case of noisy, complex pictures that contain a large number of details.The various procedures, unfortunately, barely offer a solution for this problem, the only solution is a precise camera adjustment [2].In other words, the car should be photographed in a way that the size of the number plate is as big as possible and the environment is excluded as possible.Since the optimum moment of exposure can hardly be guaranteed, the the Basis of Edge finding [3].These algorithms rely on the observation that number plates usually appear as high contrast areas in the image [4].

Literature Survey
There are several works deal with licenses plate recognition such as: • Dr. Yaduvir Singh and Mukesh Kumar [5] suggest a real-time method which recognizes license plates at gate based on extracting the license plate from a single image, isolating the characters of the plate and identifying the individual characters.
• Assist.Prof. Dr. Loay E. George Nada N. Kamal [6] suggest a recognition method from three steps; first stage is image binarization and segmentation, the second stage is license plate localization be determined, and the third stage is the license plate recognition distribution based on templet matching.[7] suggest a recognition method that compares the license plate with the database of plates that are law enforcement.Plates.It depends on isolation of character, Number that captured under different circumstances such as shadow.It based on back propagation neural network.

Single Value Decomposition
The singular-value decomposition (SVD) is a matrix factorization by decomposition of Eigen value and Eigen vector of the matrix.Symmetric matrix has positive eigenvalues with respect to any matrix via an extension of the polar decomposition.It is useful in many applications such as: signal processing and statistics [8], [9].SVD takings a rectangular matrix of data, The SVD theorem is: Where   is  , U are orthogonal,    is  pxp ,V are orthogonal SVD characterizes an expansion of the original data in a coordinate system where the covariance matrix is diagonal.SVD signifies an extension of original data using the covariance matrix is diagonal.

Affine Transform
Normalizing of the image is concerning the affine transform.The affine transform called generalized complex moments computed in polar coordinates and their behavior is analyzed in recognition of symmetrical objects [7,8], used factorization of the second order moment matrix to define the normalization constraints.

Image Plate Dataset
The dataset used in proposed method collect by capturing sense of vehicles.for a final decision.

Proposed Method
The Proposed method applied on Iraqi plate number to recognize the plate information.The information are: number, city, and the type of plate.
The proposed method is based on a set of phases; each phase has a special function for giving the car plate information as shown in figure 1. Vol

Image Normalization
The affine method used for normalization, it a simpler way of normalization to the affine transformation, which is based both on complex moments as well as traditional geometric.Also, the method is well defined for objects that have n-fold rotation symmetry, which is its main advantage as shown in figure 5.

Image RGB to Gray conversion
The image is changed to image with grayscale format, and then the median filter is applied to remove the noise.As shown above, the original image is converted to an image with grayscale format which has a high contrast.Now, there is a necessity for identifying the location of the number plate horizontally in which row it's present.The numbers and letters are placed in the same row (i.e., at identical vertical levels) which leads to frequent changes in the horizontal intensity for discovering the horizontal changes of the intensity since it is expected for the rows that contain the number plate are to exhibit many sharp variations.The Horizontal intensity graph is as follows, with the peaks indicating high contrast regions in the image.

Plate-Color Recognition
In this phase, the color of plate region needs to be recognized and it depends on the histogram of three color band RGB and specifies the maximum color in the range of each color in the histogram as shown in figure 7.

Plate City Recognition
The city segment is sent to check with dataset have all city names and return the city matching.
The matching applied to store database.Highlighting the name of the city takes place by sending the image to the database is the return of an identification number representing the city or a particular number of government cars; as shown in figure 8.

Segments SVD transform
This phase is one of the important phases, the image of segment number are transformed using SVD that produce three matrix U, S and V. the middle matrix S contains only diagonal values and the norm of these values are calculated to generate a number that represents a feature which will be sent to database that contains the characteristics of all features.After recognizing the total information, a query sends to another database to ensure the possibility of suspicion required number as shown in figure 10, while tables 1, 2 and 3 represent an example of U, S and V matrix respectively.

Experimental Test Results
The implementation of the proposed methods is done on the PC that has processor is Intel (R) COR(TM), i5-2630QM CPU @2.00 GHz 2.00 GHz, the operating system is windows 8 (64 bits x64) and RAM is 4GB.Using MATLAB 2014 is used as software tools.Experiments have performed to test the proposed method and to measure the time-complexity and accuracy of proposed method.The input images are resized in (160x320 pixel) size colored images.
The number of samples used in this paper is 84 vehicles, different weather and under various illumination conditions and distance.The success accuracy rate is 90.4 % and the average time required to detect the plate is 116.7808Sec milliseconds for training as shown in table 5. Some samples of dataset show in Figure 9 that applied in proposed method.
The collection of images used from a different position and different distances.It used about 84 images in the proposed method.Image Classification Image classification process used pixel based or feature based to distinguish between images or between regions in the same image.Two main classification methods are Supervised Classification and Unsupervised Classification.In supervised classification, there are predefined classes of images called training set.Some of predefined of images used for testing.There are a different kind of supervised classification algorithm such as decision tree method, Bayesian method, or K-NN etc.… K-NN classification algorithm calculates a distance of the input class to each class in the dataset and get the nearest odd number of classes (5, 7, or 9 etc..)

Figure 4 :
Figure 4: histogram equalization of plate image

Figure 5 :
Figure 5: affine transform on plate image

ConclusionAFigure 9 :
Figure 9: images plate used in proposed method

Table 4 :
Feature extraction samples

Table 5 :
Feature extraction samples