CN100414560C - License Plate Extraction Method Based on Wavelet Transform and Wright Transform - Google Patents

Abstract

A vehicle brand picking-up method bases on Wavelet transform and Radon transform commutation. Switch the really color picture to gray-level picture. Calculate the vertical grads of originality picture and bases the result of Wavelet transform of its plane projection curve to wide orientation about the vehicle brand. Take plane Radon transform to the vehicle brand after wide orientation and locate and delete up and down frames of vehicle brand. Take vertical Radon transformation to the vehicle brand deleted the up frame and down frame and locate and delete left frame and right frame of vehicle brand. Consolidate the colors to get one vehicle brand without frame and having identifiable number. It picks up vehicle brand effectively, quickly and precise in common condition and set veracious identify foundation to the word of vehicle brand.

Description

License Plate Extraction Method Based on Wavelet Transform and Wright Transform

technical field

The invention belongs to the technical field of intelligent transportation, in particular to a license plate extraction method in a complex background in the license plate recognition technology.

Background technique

Intelligent transportation system is currently a frontier research topic in the field of transportation in the world. Developed countries have proposed and implemented a series of research plans, the core of which is to use information technology, communication technology, computer technology, control In-depth transformation of traditional transportation systems, such as technology, to improve the efficiency of system resource use, system security, reduce resource consumption and environmental pollution, and my country is no exception in this respect. The license plate recognition technology in the intelligent transportation system mainly realizes automatic recognition of driving vehicle license plates, thereby completing functions such as automatic toll collection, unmanned parking management, traffic control at important intersections, and tracking of illegal vehicles, so as to save manpower and funds , while improving the efficiency of traffic management. Today, with the improvement of computer performance and the development of image processing technology, the license plate recognition system has become increasingly mature. See literature: Da-Shan Gao, Jie Zhou, Car license plates detection from complex scene, Signal Processing Proceedings, 5th International Conference, 21-25 Aug.2000 and literature: Shyang-Lih Chang, Li-Shien Chen, YunChung Chung, Sci -Wan Chen, Automatic license plate recognition, Intelligent Transportation Systems, IEEE Transactions on, March 2004.

In the license plate recognition system, license plate extraction is a key technology. Its purpose is to segment the license plate image from a digital image with a complex background, which requires high recognition rate and strong environmental adaptability. However, due to the complexity of the license plate background and the diversity of license plate features, there is still no completely general intelligent license plate location method so far. Most of the positioning methods are limited to the solution of a certain side problem, such as the tilt of the license plate, the interference of light, the influence of noise, etc., and there is still a relatively large distance from practical application. Therefore, how to combine the existing research results and take into account the working ability of the existing equipment, so that our license plate recognition system has good performance and recognition speed is the direction of our current research. For details, see Zhi-Bin Huang, Yan-Feng Guo, Classifier fusion-based vehicle license plate detection algorithm, Machine Learning and Cybernetics, 2003 International Conference on, 2-5 Nov.2003 and Yoshimori, S. Mitsukura, Y. Fukumi, M. Akamatsu, N. Khosal, R. License plate detection system in rainy days, Computational Intelligence in Robotics and Automation., 16-20 July 2003.

The license plate extraction method based on wavelet transform and Radon transform (abbreviation: Radon transform) may solve this problem, because the high-frequency information in the license plate area in the whole image is obvious and concentrated, and wavelet transform can be used to effectively extract and highlight These high-frequency information, but the license plate extracted at this time also includes the license plate frame, the existence of the license plate frame will increase the difficulty of the subsequent license plate character segmentation and recognition, in order to make the subsequent character segmentation and recognition process The difficulty can be maximized To minimize the extent, Wrighten transform is performed on the license plate image located after wavelet transform, and the four borders of the license plate can be located, so that the four borders are deleted, leaving only the recognizable character area, so that the subsequent character segmentation and recognition becomes relatively easy.

Among them, the so-called wavelet refers to some small waves based on wavelets, which have variable frequency and limited duration, and wavelet transform refers to performing some mathematical operations on these small waves and the signal to be processed. The Raiden transform is a geometric transformation. A series of mathematical judgments can be performed on the result of the Raiden transform of an image to detect whether the image contains a straight line and the position of the straight line. Finally, the license plate refers to the number installed on each license plate to identify the identity of the vehicle, which is standardized and generally hung on the front or rear of the vehicle.

The commonly used license plate extraction methods are:

① The license plate extraction method based on the color features of the license plate. It distinguishes the license plate area from the background area by extracting the special color features of the license plate that are different from other areas, thereby extracting the license plate. The advantage is that the color information of the image is fully utilized; the disadvantage is that the speed is slow, and it may fail for those images with very rich colors. For details, see the literature Zhu Wei-gang; Hou Guo-jiang; Jia Xing; A study of locating vehicle license plate based on color feature and mathematical morphology. Signal Processing, 26-30 Aug.2002,

② The method of license plate extraction using Hough transform. It searches for the position of the license plate by extracting the straight line of the license plate frame. The disadvantage is that the speed is slow, and the positioning is not accurate, and there are many interference areas. For details, see the literature K.M.Kim, B.J.Lee, K.Lyou.The automatic coefficient and Hough transform.Journal of Control, Automatic and System Engineering.3(5):511-519, 1997,

③ License plate extraction method based on morphological operators. It extracts the license plate by enhancing the edge features of the license plate by using operations such as expansion and erosion. Its disadvantage is that quite a lot of interference areas will be found for those areas with relatively rich edge features, which will affect the effect of license plate extraction. For details, see M. Shridhar, J.W.Miller. Recognition of license plate image, In Processings of International Conference on Document Analysis and Recognition, 17-20, 1999,

The above three methods for license plate extraction all have a common feature: these methods are all aimed at a specific environment, and the extracted high-frequency information is limited. Once the environment changes, the extraction accuracy will fluctuate greatly, which will degrade the performance of the entire license plate recognition system. Therefore, it lacks adaptability to various environments, and the robustness of the method is not good. The environment we are talking about here refers to the environment in which the original image of the license plate is collected. Different environments, such as weather, time, the background of the vehicle, and the movement of the vehicle itself, etc., the collected images are different in terms of resolution and texture features. Often there are considerable differences. High-frequency information refers to the local detail features of the image, and the local detail features of the license plate area are more obvious than other areas.

Contents of the invention

The task of the present invention is to provide a license plate extraction method based on wavelet transform and Wright transform. By adopting the method of the present invention, the license plate can be extracted effectively, quickly and accurately in common environments, laying the foundation for the accurate identification of license plate characters. . The general environment includes: sunny day, rainy day, license plate level, license plate tilt, vehicle stationary, vehicle moving, daytime, night and so on.

In order to facilitate the description of the content of the present invention, a definition of terms is first made here:

1. License plate: a rectangular number plate installed in each vehicle to identify the identity of the vehicle. The number plate has four borders, up, down, left, and right. The frame contains 7 characters. There are uniform production standards. It is generally hung on the front of the vehicle or At the rear, the license plate standards for vehicles with different uses are different.

2. Grayscale image: An image that only contains brightness information without any color information.

3. Gray scale judgment: an image processing method for judging whether the currently read image belongs to a true-color image. Here, it refers to judging the captured vehicle image. If the vehicle image matrix contains three primary color components, the currently read vehicle image is a true color image, otherwise it is a grayscale image.

4. Grayscale conversion: a means of converting a color image into a grayscale image, using the formula f(i,j)=0.114*I(i,j,1)+0.587*I(i,j,2)+ 0.299*I(i, j, 3) can convert a grayscale image into a grayscale image, where i represents the row coordinate value of the image, j represents the column coordinate value of the image, and f(i, j) represents the converted The gray value of the pixel in row i and column j in the grayscale image, I(i, j, 1), I(i, j, 2) and I(i, j, 3) represent the color image respectively R, G, and B component values of the pixel at row i and column j.

5. Vertical gradient: In order to highlight the vertical detail information of the original vehicle image, extract the edge, and subtract the gray value of the previous pixel from the gray value of the next pixel in each row to obtain the vertical gradient value. The calculation formula is g _V (i, j)=|f(i, j+1)-f(i, j)|, where i represents the row coordinate value of the image, j represents the column coordinate value of the image, f(i, j) represents the gray value of the pixel in row i, column j, f(i, j+1) represents the gray value of the pixel in row i, column j+1, and g _V (i, j) represents The corresponding vertical gradient value.

6. Gradient horizontal projection: a method of transforming the image gray gradient value in two-dimensional space into one-dimensional space through a function transformation, the function is transformed into $T_h\left(i\right)=\underset{j=1}{\overset{\mathrm{no}}{\mathrm{\Σ}}}{g}_V(i,j),$ Where g _V (i, j) represents the vertical gradient value of row i and column j, n represents the total number of columns of the vertical gradient image matrix, and T _H (i) is the projection value of row i.

7. Gaussian filtering: a well-known smoothing filtering method, which can make the curve smoother by filtering the curve through a filter function transformation. Here it refers to filtering the gradient horizontal projection curve. The filter function is $T_h^{\&prime;}\left(i\right)=\frac{1}{k}\{T_h\left(i\right)+\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}[T_h(i-j)h(j,\mathrm{\&sigma;})+T_h(i+j)h(j,\mathrm{\&sigma;})\left]\right\}.$ Among them, T′ _H (i) is the gradient horizontal projection value after filtering, the range of i is from 1 to n, n represents the height of the image, that is, the total number of rows of the image matrix; w represents the width of the smooth area, here take 8; $h(j,\mathrm{\&sigma;})=\exp \left(\frac{{-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051002165100181.png"\; state="\{\{state\}\}">j</figure-callout>}}^2}{2{\mathrm{\&sigma;}}^2}\right)$ is a Gaussian function, $k=2\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}h(j,\mathrm{\&sigma;})+1,$ σ represents the mean square error of the grayscale image, here 0.05; T _H (i) represents the i-th gradient horizontal projection value, _TH (ij) represents the (ij) gradient horizontal projection value T _H (i+j ) represents the (i+j)th gradient horizontal projection value.

8. Wavelet transform: Here, it specifically refers to performing one-dimensional wavelet transform on the gradient horizontal projection curve, so as to extract curve features that are beneficial to license plate location from the gradient horizontal projection curve. The specific transformation is, first, construct a Gaussian function $g_{\mathrm{the\; s}}\left(x\right)=\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp \left(\frac{{-x}^2}{2{\mathrm{\&sigma;}}^2}\right),$ Among them, σ represents the mean square error of the Gaussian function, which is 0.5 here; then, calculate the first-order derivative of the Gaussian function, and use the first-order derivative as the wavelet function; finally, perform mathematical convolution operation on the wavelet function and the gradient horizontal projection curve , so as to obtain a wavelet transform curve.

9. Peak: A characteristic of a value on a curve; a curve value at which it is greater than both the immediately preceding and immediately following curve values.

10. Trough: A characteristic of values on a curve; where a curve value is smaller than both the immediately preceding and immediately following curve values.

11. License plate candidate area: some areas with certain characteristics in an image, these features are, first of all, it is a rectangle or parallelogram with a certain size, and the rectangle or parallelogram cannot be too large to occupy The entire vehicle image should not be too small to be visible to the naked eye. Usually, it is a rectangular or parallelogram area with a length of about 440mm and a height of about 140mm; secondly, the area inside the rectangle must have characters; finally, the rectangle is located in a vehicle image middle and lower part. Normally, the license plate in a vehicle image is a rectangle, but the license plate area may be deformed into a parallelogram due to various reasons when the vehicle image is captured.

12. Horizontal Raiden transform: Raiden transform is performed in the horizontal direction. Raiden transform is a geometric transformation that can be used to detect straight lines in an image, perform Raiden transform within a certain horizontal angle range, and then use the transformation result The straight line in the horizontal direction in the image can be detected, and here it is specifically used to detect the upper and lower borders in the rough-cut license plate image. If the horizontal direction is defined as 0 degrees, the rough-cut license plate is subjected to Wright transformation every 1 degree within the range of -5 degrees to 5 degrees, and 11 Wright transformation curves will be obtained after the transformation.

13. Vertical Raiden transform: Carry out Raiden transform in the vertical direction. Here it is used to detect the left and right borders in the license plate image. The vertical direction is defined as 90 degrees, and every 1 in the range of 85 degrees to 95 degrees Carry out Wright transformation on the license plate whose upper and lower borders have been deleted, and obtain 11 Wright transformation curves after transformation.

14. Linear crest: the shape is a line or a very narrow crest, here specifically refers to the linear crest in the horizontal and vertical Radon transformation of the license plate, which corresponds to a straight line in the license plate image.

15. Grayscale flip: For the transformation of the grayscale image, subtract each grayscale value from 255, convert black to white, and white to black.

A kind of new license plate extracting method based on wavelet analysis and Wright transform provided by the invention, it comprises the following steps:

Step 1. Install the camera device at an appropriate location at a highway intersection, toll booth or parking lot, and collect images after the vehicle enters the camera range to obtain the original image containing the license plate image;

Step 2: Carry out grayscale judgment on the original image of the vehicle obtained in step 1, if the original image of the vehicle captured is a grayscale image, no processing is performed; if the original image of the vehicle captured is a true color image, then Carry out grayscale conversion on the original image of the vehicle to obtain a grayscale image containing the license plate; the specific method is to use the formula f(i,j)=0.114*I(i,j,1)+0.587*I(i,j, 2) +0.299*I(i, j, 3) for conversion, where i represents the row position of the image, j represents the column position of the image, and f(i, j) represents the i-th row and jth in the converted grayscale image The gray value of the pixel in the column, * is the symbol of the multiplication operation, I(i, j, 1), I(i, j, 2) and I(i, j, 3) respectively represent the i-th row in the color image R, G, and B component values of pixels in column j;

Step 3. Carry out vertical gradient calculation to the grayscale image obtained in step 2, obtain a vertical grayscale gradient map that contains the vehicle grayscale image of license plate; Concrete method is to adopt formula

g _V (i, j)=|f(i, j+1)-f(i, j)| for vertical gradient calculation, where i represents the row coordinate value of the image, j represents the column coordinate value of the image, f(i , j) represents the gray value of the pixel in row i, column j, f(i, j+1) represents the gray value of the pixel in row i, column j+1, g _V (i, j) Indicates the vertical gradient value of row i and column j;

Step 4. Gradient horizontal projection is carried out to the vertical grayscale gradient map obtained in step 3 to obtain a rough gradient horizontal projection curve; the calculation formula of gradient horizontal projection is $T_h\left(i\right)=\underset{j=1}{\overset{\mathrm{no}}{\mathrm{\&Sigma;}}}{g}_V(i,j),$ Among them, g _V (i, j) represents the vertical gradient value of the i-th row j column, n represents the total number of columns of the vertical gradient image matrix, and T _H (i) is the projection value of the i-th row;

Step 5 performs Gaussian filtering on the rough gradient horizontal projection curve obtained in step 4 to obtain a smooth gradient horizontal projection curve; the filtering method is to use the formula $T_h^{\&prime;}\left(i\right)=\frac{1}{k}\{T_h\left(i\right)+\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}[T_h(i-j)h(j,\mathrm{\&sigma;})+T_h(i+j)h(j,\mathrm{\&sigma;})\left]\right\}$ Filtering; wherein T′ _H (i) is the gradient horizontal projection value after filtering, and the range of i is from 1 to n, and n represents the height of the vehicle image; w represents the width of the smooth area, which is 8 here; $h(j,\mathrm{\&sigma;})=\exp \left(\frac{{-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="C20051002165100181.png"\; state="\{\{state\}\}">j</figure-callout>}}^2}{2{\mathrm{\&sigma;}}^2}\right)$ is a Gaussian function, $k=2\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}h(j,\mathrm{\&sigma;})+1,$ σ represents the mean square error of the grayscale image; T _H (i) represents the i-th gradient horizontal projection value, _TH (ij) represents the (ij)-th gradient horizontal projection value, T _H (i+j) represents the ( i+j) gradient horizontal projection values;

Step 6 performs wavelet transformation on the smooth gradient horizontal projection curve obtained in step 5 to obtain a wavelet transformation curve; the wavelet transformation method is: first, construct a Gaussian function $g_{\mathrm{the\; s}}\left(x\right)=\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp \left(\frac{{-x}^2}{{2\mathrm{\&sigma;}}^2}\right),$ Where x is the independent variable of the Gaussian function, and σ represents the mean square error of the Gaussian function, which is taken as 0.5 here; then, the first-order derivative of the Gaussian function is calculated to obtain $w\left(x\right)=\frac{-x}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}^3}\exp \left(\frac{{-x}^2}{{2\mathrm{\&sigma;}}^2}\right),$ The first-order derivative w(x) is used as a wavelet function; finally, the wavelet function and the gradient horizontal projection curve are subjected to mathematical convolution operation to obtain a wavelet transformation curve;

Step 7. Normalize the wavelet transform curve obtained in step 6, so as to obtain a curve with a curve value between -1 and 0; the specific method is, at first, obtain the maximum value max and the minimum value of the wavelet transform curve min; then, using the formula ${\mathrm{the\; s}}^{\&prime;}\left(x\right)=\frac{(\mathrm{the\; s}\left(x\right)-\min )}{(\max -\min )}+1$ Carry out normalization, wherein s (x) is the wavelet transformation curve value before normalization, and x is the coordinate value of curve, and s' (x) is the curve value of the wavelet transformation after normalization processing;

Step 8. Carry out gradient horizontal projection curve scanning to the wavelet transform curve obtained in step 7, obtain the positional coordinates of two smaller valley values in the wavelet transform curve; Concrete method is: find valley from the starting point of curve, record simultaneously The position coordinates of the lower trough value, so search until the end of the curve;

Step 9. Utilize the valley position coordinates obtained in step 8 to carry out the rough positioning operation of the license plate, and obtain the position coordinates of one or two license plate candidate areas in the vehicle image containing the license plate; the specific method is: according to the curve provided in step 8 The position coordinates of the smaller trough value are used to search for the position coordinates of the left and right peaks adjacent to the above-mentioned trough, and the position coordinates of the left peak correspond to the abscissa of the upper boundary of the license plate in the vehicle image matrix containing the license plate. Here, top_boundary is used to indicate the abscissa of the upper boundary position of the license plate, and the position coordinate of the peak on the right corresponds to the abscissa of the lower boundary of the license plate in the vehicle image containing the license plate. Here, bot_boundary is used to indicate the lower boundary position of the license plate the abscissa;

Step 10 uses the position information obtained in step 9 to carry out rough cutting of the license plate to obtain one or two license plate candidate areas; the specific method is, at first, define a matrix mask whose elements are all zero, and the number of rows and columns of the mask are respectively equal to The number of rows and columns of the vehicle grayscale image matrix containing the license plate is the same; secondly, all the elements of the columns between the top_boundary row and the bot_boundary row in the matrix mask are set to 1, where top_boundary indicates that the license plate contains The abscissa value of the upper boundary in the vehicle grayscale image of the license plate, bot_boundary represents the abscissa value of the lower boundary of the license plate in the vehicle image; then, compare the matrix mask with the elements with the same coordinate value in the vehicle grayscale image matrix After multiplication, the elements in the non-license plate candidate area in the vehicle grayscale image matrix are all zero, and the element values in the license plate candidate area remain unchanged; finally, define the matrix I_cut whose elements are all zero, and the number of rows of the matrix I_cut It is bot_boundary-top_boundary, the number of columns is equal to the number of columns of the vehicle grayscale image matrix, and the non-zero elements in the result matrix of multiplying the matrix mask and the vehicle grayscale image matrix are assigned to I_cut, and the matrix I_cut represents the license plate candidate area;

Step 11. Carry out horizontal Radon transformation to the license plate candidate area obtained in step 10, obtain 11 Radon transformation curves; Concrete processing is: define the horizontal direction as 0 degree, within the scope of -5 degree to 5 degrees, every 1 degree performs Wrighten transformation on the rough-cut license plate, and 11 Wrighten transformation curves will be obtained after transformation;

Step 12 utilizes the width values of 11 Raiden transformation curves in step 11 to determine the coordinate values of the upper and lower borders of the license plate in the license plate image; the specific method is: compare the widths and widths of 11 Raiden transformation curves obtained in step 12 The transformation angle corresponding to the smallest Raiden transformation curve is the inclination angle of the license plate, which is recorded as R_hdegree. At this time, the coordinate value of the first wave peak on the Raiden transformation curve is greater than a certain value, and the coordinate value of the linear peak corresponds to the upper frame of the license plate in the license plate The abscissa in the image is marked as R_top, and the coordinate value of the last peak value of the linear peak greater than a certain value on the Raiden transformation curve corresponds to the vertical coordinate of the lower frame of the license plate in the license plate image and is recorded as R_bot;

Step 13 uses the license plate inclination angle and the coordinate information of the upper and lower borders obtained in step 12 to correct the license plate and delete the upper and lower borders of the license plate; the specific processing is: first, correct the license plate, if the inclination angle R_hdegree of the license plate is not zero, then the license plate is considered to be Slanted, rotate the license plate image -R_hdegree to correct the license plate; then define a matrix R_hmask with zero elements, the total number of columns and the total number of rows of the matrix are respectively equal to the total number of columns and the total number of rows of the matrix I_cut, and the matrix R_hmask is located at R_top The elements between R_bot and R_bot are set to 1, and the elements with the same coordinate values in the matrix R_hmask and the matrix I_cut are multiplied to obtain a multiplication result matrix R_htbmask; finally, a matrix R_hcut with zero elements is defined, and the total number of rows of the matrix For R_bot-R_top, the total number of columns is the same as the total number of columns of the matrix R_htbmask, assign the non-zero elements in the matrix R_htbmask to the matrix R_hcut, and the matrix R_hcut is the image matrix corresponding to the license plate image whose upper and lower borders of the license plate have been deleted;

Step 14. Carry out vertical Raiden transformation to the image matrix obtained in step 13, which has deleted the upper and lower borders of the license plate, to obtain 11 Raiden transformation curves; the specific processing is: define the vertical direction as 90 degrees, at 85 degrees to 95 degrees Within the range of , Wrighten transform is performed on the license plate whose upper and lower borders have been deleted every 1 degree, and 11 Wrighten transformation curves are obtained after the transformation;

Step 15 utilizes the peak values of the 11 Raiden transformation curves obtained in step 14 to determine the left and right borders of the license plate. The specific method is: first, find a Raiden transformation curve with the smallest width; Right search for a linear peak with a peak value greater than a certain value. The coordinate value of this peak is the vertical coordinate of the left frame of the license plate whose upper and lower borders have been deleted in the license plate image matrix. R_left, and then search for a A linear peak with a peak value greater than a certain value, the coordinate value of this peak is the vertical coordinate of the right border of the license plate image matrix in which the upper and lower borders of the license plate have been deleted is marked as R_right;

Step 16 utilizes the coordinate values of the left and right borders of the license plate that has deleted the upper and lower borders obtained in step 15 to delete the left and right borders of the license plate; the specific method is, at first, define a matrix R_vmask whose elements are zero, its total row number and total The number of columns is equal to the total number of rows and columns of the matrix R_hcut, and the elements of the matrix R_vmask located in the R_left column and the R_right column are set to 1; The result matrix R_vlrmask; finally define a matrix R_vcut whose elements are all zero, the total number of columns of the matrix is R_right-R_left, the total number of rows is equal to the total number of rows of the matrix R_vlrmask, and the non-zero elements in the matrix R_vlrmask are assigned to R_vcut, the matrix R_vcut It is to delete the license plate image matrix of the four borders of the upper, lower, left, and right sides;

Step 17. Carry out the license plate color normalization process to the accurate license plate image obtained in step 16, obtain a license plate image with a black background and a white number that contains a recognizable number; The number of white pixels and the number of black pixels; secondly, compare the number of black pixels and the number of white pixels, if the number of white pixels is the majority, then perform grayscale flipping;

Through the above steps, we have extracted the license plate image from the original image containing the license plate.

It should be noted:

1. The vertical gradient calculation in step 3 is because the pixel values in the license plate area change faster and more concentrated than the pixel values in the non-license plate area, and are more obviously concentrated in the vertical gradient.

2. The gradient horizontal projection in step 4 is because the license plate area must be located in a small area with a rather high gradient density value, so we first perform gradient horizontal projection to locate the horizontal candidate area of the license plate.

3. After performing wavelet transformation in step 6, the values of the gradient horizontal projection curve will all become negative values, so the original peak on the curve becomes a valley after wavelet transformation, and the original valley on the curve after wavelet transformation becomes It becomes a peak, so when searching for peaks and valleys, searching for the trough of the curve after wavelet transformation is equivalent to searching for the peak of the curve before wavelet transformation. Similarly, searching for the peak of the curve after wavelet transformation is equivalent to searching for the peak of the curve before wavelet transformation trough of the curve. In addition, the gradient horizontal projection curve before wavelet transformation has been processed by a Gaussian filter, and the curve at this time is relatively smooth, but the gradient horizontal projection curve after wavelet transformation is smoother and removes some of the curve before transformation. False peaks and valleys are more conducive to the subsequent license plate location processing, which is another advantage of using wavelet transform.

4. When searching for troughs in step 8, not all troughs that meet the trough conditions are adopted. Only those troughs whose value is less than a certain value and whose width satisfies a certain value range are considered to be troughs. The certain value ranges that are satisfied are empirical values obtained from experiments. In the search, only the trough with the smallest trough value and the trough with the second smallest trough value are taken. Usually, the trough with the smallest trough value corresponds to the license plate area, but sometimes the smallest trough value is caused by strong interference in the background of the vehicle image The corresponding area is not the license plate area, but the area corresponding to the trough that becomes the second smallest value of the trough value is the license plate area, but if too many trough positions are extracted, more license plate candidate areas will be introduced in the subsequent license plate cutting process. To increase the difficulty of processing, it is reasonable to choose the two troughs with the smallest trough value through the summary of experiments.

5. In the rough location calculation of the license plate in step 9, when searching for the left and right peaks adjacent to the valley that satisfies the conditions, only those peaks that meet a certain value range are considered as true peaks, and the small fluctuations on the curve lead to false peaks. The existence of the wave peak and the certain value range that the true wave peak satisfies are empirical values obtained through experiments. In addition, what is obtained in the rough positioning operation is the rough position of the license plate in the vehicle image containing the license plate. Since this step is a rough positioning operation of the license plate, there may be one or two position coordinates of the license plate candidate area.

6. The certain value greater than the linear peak in steps 12 and 15 is an empirical value obtained through experiments.

The present invention first preprocesses the original image, and converts the true color image into a grayscale image; secondly, calculates the vertical gradient of the original image and performs rough positioning of the license plate according to the wavelet transformation result of the horizontal projection curve of the vertical gradient, thereby roughly positioning a or Two license plate candidate regions. Carry out horizontal Raiden transformation on the roughly located license plate to locate and delete the upper and lower borders of the license plate. At the same time, if the license plate is tilted, it will be corrected; And delete the left and right borders of the license plate; finally, perform color normalization on the license plate area where the four borders have been deleted, and finally output a license plate image containing only recognizable numbers. The present invention adopts the combination of rough positioning and precise positioning, adopts the method of wavelet transform in the rough positioning, adopts the method of Wright transform in the precise positioning, and performs the normalization processing of brightness and color after the precise positioning, and finally obtains A license plate with a recognizable number and no plate frame with a uniform background and number color. By adopting the license plate extraction method of the present invention, the license plate can be effectively, quickly and accurately extracted in ordinary environments, and the foundation is laid for accurate recognition of license plate characters.

The innovation of the present invention is:

1. Use wavelet transform to roughly locate the license plate. Wavelet transform has the characteristics of extracting high-frequency characteristics in the signal. Using wavelet transform to process the gradient vertical projection curve after a Gaussian filter can extract the peak on the curve to the greatest extent. The trough has a certain smoothing function for the curve, so that the license plate area can be roughly positioned quickly and accurately.

2. Use the Wrighten transform to accurately locate the license plate, because the wavelet transform can only be used to roughly locate the license plate. Due to noise and background interference, the roughly located license plate also includes the license plate frame and some surrounding non-license plate areas. At the same time, due to some The reason may be that the license plate is tilted, which will make the subsequent character segmentation and recognition more difficult. Therefore, the Wright transform is used to detect and delete the frame of the license plate and correct the license plate so that it is in a horizontal position. The final result is a character only An image of a license plate containing a recognizable number.

3. A method combining coarse and precise license plate positioning is proposed in order to increase the speed of the method and meet the real-time requirements of the license plate recognition system. Firstly, at most two license plate candidate regions are roughly extracted by using the vertical projection of the vertical gradient map of the original image, and then the license plate is precisely located by applying the Wrighten transform method to these candidate regions.

Description of drawings

Figure 1 is a schematic diagram of the original image containing the license plate;

Among them, mark 1 represents the windshield of the vehicle, mark 2 represents the hood of the vehicle, mark 3 represents the area where lights, license plates and bumpers are installed, mark 4 represents the license plate, mark 5 represents the wheels, and mark 4 represents the area In the license plate, capital letters A, B, C, D, E, F and G represent the first, second, third, fourth, fifth, sixth and seventh of the license plate respectively numbers, where the second and third numbers are slightly spaced apart.

Fig. 2 is the gradient vertical projection curve after the Gaussian filter of the present invention;

Wherein, the value on the ordinate represents the value of the gradient vertical projection curve after Gaussian filtering and normalization, and the value on the abscissa represents the coordinate value of the gradient vertical projection curve after Gaussian filtering and normalization.

Fig. 3 is the gradient vertical projection curve of the present invention through wavelet transform

Wherein, the value on the ordinate represents the value of the gradient vertical projection curve after wavelet transformation and normalization, and the value on the abscissa represents the coordinate value of the gradient vertical projection curve after wavelet transformation and normalization.

Fig. 4 is the license plate image schematic diagram that the present invention finally obtains;

Among them, A, B, C, D, E, F and G represent the first, second, third, fourth, fifth, sixth and seventh numbers of the license plate respectively.

Fig. 5 is a schematic flow chart of the present invention

Detailed ways:

Adopt method of the present invention, at first at the entrance of expressway, toll booth and other any suitable positions adopt the original image of photographing device to automatically photograph vehicle; Secondly, the original image of vehicle photographed is input to the program written in Matlab language as original data Automatic processing in the software without any manual intervention; finally a license plate image containing a recognizable number is obtained. A total of 300 color vehicle images taken on the spot were used as the original data, 295 of which were roughly positioned, with a positioning accuracy rate of 98.33%, and 285 were accurately positioned, with a positioning accuracy rate of 95%. The license plate image only takes 40ms.

To sum up, the method of the present invention can quickly and accurately locate the license plate image containing the recognizable number from the provided original vehicle image.

Claims (1)

1. vehicle license plate extraction method based on wavelet analysis and Lei Deng conversion is characterized in that it comprises the following step:

Step 1. is installed on the highway crossing with camera head, and image acquisition is carried out after in vehicle enters image pickup scope in the appropriate location in charge station or parking lot, obtains containing the original image of license plate image;

Resulting vehicle original image carries out the gray scale judgement in the step 2 pair step 1, if taken vehicle original image is the gray-scale map image, does not then handle; If taken vehicle original image is a true color image, then the vehicle original image is carried out gradation conversion, obtain the gray level image that a width of cloth comprises car plate; Concrete grammar for adopt formula f (i, j)=0.114*I (i, j, 1)+0.587*I (i, j, 2)+0.299*I (i, j, 3) changes, the line position of i presentation video wherein, the column position of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of i in the gray level image after the expression conversion, * be the multiplying symbol, I (i, j, 1), I (i, j, 2) and I (i, j, 3) represent the R of the pixel of the capable j row of i in the coloured image respectively, G, the value of B component;

Resulting gray level image carries out VG (vertical gradient) calculating in the step 3. pair step 2, obtains a vertical shade of gray figure who includes the vehicle gray level image of car plate; Concrete grammar is for adopting formula g _V(i, j)=| f (i, j+1)-f (i, j) | carry out VG (vertical gradient) and calculate, the row-coordinate value of i presentation video wherein, the row coordinate figure of j presentation video, f (i, j) gray-scale value of the pixel of the capable j row of expression i, f (i, the j+1) gray-scale value of the pixel of the capable j+1 row of expression i, g _V(i, j) the capable j of expression i row the VG (vertical gradient) value;

Resulting vertical shade of gray figure carries out the gradient horizontal projection in the step 4. pair step 3, obtains a coarse gradient horizontal projection curve; The computing formula of gradient horizontal projection is $T_H\left(i\right)=\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{g}_V(i,j),$ G wherein _V(n represents total columns of VG (vertical gradient) image matrix, T for i, j) the VG (vertical gradient) value of the capable j row of expression i _H(i) be the capable projection value of i;

Resulting coarse gradient horizontal projection curve carries out gaussian filtering in the step 5 pair step 4, obtains a level and smooth gradient horizontal projection curve; Filtering method is for adopting formula $T_H^{\&prime;}\left(i\right)=\frac{1}{k}\{T_H\left(i\right)+\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}[T_H(i-j)h(j,\mathrm{\&sigma;})+T_H(i+j)h(j,\mathrm{\&sigma;})\left]\right\}$ Carry out filtering; T ' wherein _H(i) be filtered gradient horizontal projection value, the variation range of i is from 1 to n, and n represents the height of vehicle image; W has represented the width size of smooth region, gets 8 herein; $h(j,\mathrm{\&sigma;})=\exp \left(\frac{-j^2}{2{\mathrm{\&sigma;}}^2}\right)$ Be Gaussian function,

$k=2\underset{j=1}{\overset{w}{\mathrm{\&Sigma;}}}h(j,\mathrm{\&sigma;})+1,$ σ represents the mean square deviation of gray level image; T _H(i) i gradient horizontal projection value of expression,

T _H(i-j) expression (i-j) individual gradient horizontal projection value, T _H(i+j) expression (i+j) individual gradient horizontal projection value;

Resulting level and smooth gradient horizontal projection curve carries out wavelet transformation in the step 6 pair step 5, obtains a wavelet transformation curve; Small wave converting method is: at first, construct a Gaussian function $g_s\left(x\right)=\frac{1}{\sqrt{2\mathrm{\&pi;}}\mathrm{\&sigma;}}\exp \left(\frac{-x^2}{2{\mathrm{\&sigma;}}^2}\right),$ Wherein x is the Gaussian function independent variable, and σ represents the mean square deviation of Gaussian function, gets 0.5 herein; Then, the first order derivative of calculating this Gaussian function obtains $w\left(x\right)=\frac{-x}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}^3}\exp \left(\frac{-x^2}{2{\mathrm{\&sigma;}}^2}\right),$ This first order derivative w (x) as wavelet function; At last this wavelet function and gradient horizontal projection curve are carried out the mathematics convolution algorithm, thereby obtain a wavelet transformation curve;

Resulting wavelet transformation curve carries out normalized in the step 7. pair step 6, thereby obtains the curve of a curve values between-1 to 0; Concrete grammar is at first, to obtain wavelet transformation curve maximal value max and minimum value min; Then, utilize formula $s^{\&prime;}\left(x\right)=\frac{(s\left(x\right)-\min )}{(\max -\min )}+1$ Carry out normalization, wherein s (x) is the wavelet transformation curve values before the normalization, and x is the coordinate figure of curve, and s ' is the curve values of the wavelet transformation after the normalized (x);

Resulting wavelet transformation curve carries out gradient horizontal projection curved scanning in the step 8. pair step 7, obtains minimum trough value and time position coordinates of small echo valley in the wavelet transformation curve; Concrete grammar is: seek the trough that satisfies certain condition from the starting point of curve, by relatively obtaining the position coordinates of minimum trough value and time small echo valley, satisfy certain condition and be meant that the trough value satisfies certain span less than certain value and trough width;

Step 9. utilizes that resulting wave trough position coordinate carries out the computing of car plate coarse positioning in the step 8, obtains the position coordinates of one or two license plate candidate area in including the vehicle image of car plate; Concrete grammar is: the minimum trough value of the curve that provides according to step 8 and time small echo valley position coordinates search for this trough of stating next-door neighbour about the position coordinates of two crests, the position coordinates correspondence of left side crest be the horizontal ordinate of the coboundary of car plate in including the vehicle image matrix of car plate, the horizontal ordinate of representing the position, coboundary of car plate herein with top_boundary, the right crest the position coordinates correspondence be the horizontal ordinate of the lower boundary of car plate in including the vehicle image of car plate, represent the horizontal ordinate of the lower boundary position of car plate herein with bot_boundary;

Step 10 is utilized in the step 9 resulting positional information to carry out the car plate rough lumber and is cut processing, obtains one or two license plate candidate area; Concrete grammar is, at first, the definition element is zero matrix mask, and the line number of mask is consistent with the line number and the columns of the vehicle gray level image matrix that includes car plate respectively with columns; Secondly, all be changed to 1 being positioned at top_boundary elements capable and bot_boundary all row in the ranks among the matrix mask, wherein top_boundary represents the abscissa value of the coboundary of car plate in including the vehicle gray level image of car plate, and bot_boundary represents the abscissa value of the lower boundary of car plate in vehicle image; Then, the element with same coordinate value in matrix mask and the vehicle gray level image matrix is multiplied each other, after multiplying each other, the element in the vehicle gray level image matrix in the non-license plate candidate area is zero, and the element value in the license plate candidate area remains unchanged; At last, the definition element is zero matrix I_cut, the line number of matrix I_cut is bot_boundary-top_boundary, columns and vehicle gray level image matrix column number equate, nonzero element in the matrix of consequence of matrix mask and vehicle gray level image matrix multiple is composed to I_cut, and matrix I_cut just represents license plate candidate area;

Resulting license plate candidate area carries out horizontal thunder and steps on conversion in the step 11. pair step 10, obtains 11 thunders and steps on transformation curve; Specifically be treated to: the definition horizontal direction is 0 degree, spends the car plate that every 1 degree rough lumber is cut out in the scopes of 5 degree-5 and carries out thunder and step on conversion, will obtain 11 thunders after the conversion and step on transformation curve;

The width value that step 12 utilizes 11 thunders in the step 11 to step on transformation curve is determined the coordinate figure of upper and lower side frame in license plate image of car plate; Concrete grammar is: resultant 11 thunders are stepped on the width of transformation curve in the comparison step 12, it is exactly that the angle of inclination of car plate is designated as R_hdegree that the thunder of width minimum is stepped on the pairing translation-angle of transformation curve, first crest value that this moment, thunder was stepped on the transformation curve is designated as R_top greater than the coordinate figure of the wire crest of certain value corresponding to the horizontal ordinate of car plate upper side frame in license plate image, thunder step on the transformation curve last crest value greater than the coordinate figure of the wire crest of certain value corresponding to the car plate lower frame in license plate image ordinate and be designated as R_bot;

Step 13 is utilized the license plate sloped angle that obtains in the step 12 and the coordinate information of upper and lower side frame, proofreaies and correct car plate and deletes the upper and lower side frame of car plate; Specifically be treated to: at first, proofread and correct car plate, if the angle of inclination R_hdegree of car plate is non-vanishing, thinks that then car plate tilts, thereby license plate image is rotated-R_hdegree proofreaies and correct car plate; Define an element then and be zero matrix R_hmask, total columns of this matrix and total line number equal total columns and the total line number of matrix I_cut respectively, the element of matrix R_hmask between R_top and R_bot is changed to one, and the element that has the same coordinate value among matrix R_hmask and the matrix I_cut is multiplied each other obtains a multiplied result matrix R_htbmask; At last, define an element and be zero matrix R_hcut, total line number of this matrix is R_bot-Rtop, total columns of total columns and matrix R_htbmask is identical, nonzero element among the matrix R_htbmask is composed to matrix R_hcut, and matrix R_hcut is exactly an image array of having deleted the license plate image correspondence of car plate upper and lower side frame;

The resulting image array of having deleted the car plate upper and lower side frame carries out vertical thunder and steps on conversion in the step 14. pair step 13, obtains 11 thunders and steps on transformation curve; Specifically be treated to: the definition vertical direction is 90 degree, spends in the scopes of 95 degree 85 and the car plate of deleting upper and lower side frame is carried out thunder steps on conversion every 1 degree, obtains 11 thunders after the conversion and steps on transformation curve;

Step 15 utilizes resultant 11 thunders in the step 14 to step on the crest value of transformation curve, determines the left and right side frame of car plate, and concrete grammar is: at first, find out a thunder of width minimum and step on transformation curve; On this curve, search for the wire crest of a crest value then from left to right greater than certain value, the coordinate figure of this crest is exactly that the ordinate of left frame in this license plate image matrix of having deleted the car plate of car plate upper and lower side frame is designated as R_left, turn left from the right side and search for the wire crest of a crest value greater than certain value, the coordinate figure of this crest is exactly that the ordinate of left frame in this license plate image matrix of having deleted the car plate upper and lower side frame is designated as R_right;

Step 16 is utilized the coordinate figure of the left and right side frame of the resulting car plate of having deleted upper and lower side frame in the step 15, the left and right side frame of deletion car plate; Concrete grammar is, at first, defines an element and is zero matrix R_vmask, and its total line number and total columns equate with total line number and the total columns of matrix R_hcut respectively, matrix R_vmask is positioned at the element that R_left row and R_right be listed as is changed to 1; Then the element that has the same coordinate value among matrix R_vmask and the matrix R_hcut is multiplied each other and obtain a matrix of consequence R_vlrmask; Define an element at last and be zero matrix R_vcut, total columns of this matrix is R_right-R_left, total line number of total line number and matrix R_vlrmask equates, nonzero element among the matrix R_vlrmask is composed to R_vcut, and matrix R_vcut has deleted the license plate image matrix of four frames up and down;

Resulting license plate image accurately carries out car plate color normalized in the step 17. pair step 16, but obtains a license plate image with black background and white number that includes identification number; Concrete grammar is at first, to add up white pixel number and black picture element number in the license plate area; Secondly compare black picture element number and white pixel number,, then carry out the gray scale upset if the white pixel number is most.

Images