CN111695565A

CN111695565A - Automobile mark accurate positioning method based on road barrier fuzzy image

Info

Publication number: CN111695565A
Application number: CN202010539510.1A
Authority: CN
Inventors: 刘秀萍; 扆亮海
Original assignee: Jingmen Huiyijia Information Technology Co ltd
Current assignee: Jingmen Huiyijia Information Technology Co ltd
Priority date: 2020-06-14
Filing date: 2020-06-14
Publication date: 2020-09-22

Abstract

According to the automobile mark accurate positioning method based on the road checkpoint fuzzy image, the area where the automobile mark is located is accurately positioned through the image processing method of four stages according to the obvious image characteristics of the automobile mark area, the image information of the automobile mark is obtained, the problem that the road checkpoint photographing equipment is required to provide a clearer image in the prior art is solved, even if the provided image is fuzzy, the automobile mark identification and positioning accuracy is still high, a test picture in an experiment is made of materials under different illumination and meteorological conditions, the comprehensive identification rate reaches 91%, the judgment of the automobile mark positioning processing effect is accurate, the accuracy and the timeliness of an algorithm are good, the method has robustness and high efficiency, the method can adapt to the conditions of various automobile brands, and has universality and transportability and huge market popularization and application potential.

Description

Automobile mark accurate positioning method based on road barrier fuzzy image

Technical Field

The invention relates to a method for processing a fuzzy text image in a sharpening way, in particular to a method for accurately positioning an automobile mark based on a road barrier fuzzy image, and belongs to the technical field of fuzzy text image processing.

Background

Automobile sign positioning is an important link for automobile information capture and collection in the field of intelligent traffic development, the requirements of people on data information quality are more strict at present, and the information acquisition must meet higher integrity and authenticity. Regional traffic control systems, license plate recognition and search systems, automobile speed detection systems, pedestrian passage red light running recognition and detection systems and the like are applied to many cities and are in initial scale. The current intelligent road traffic field is mainly the application of image processing related technologies, and essentially adopts the process of acquiring various image information in a complex road environment by using road image acquisition equipment, collecting the data together in the background, analyzing and integrating the data. With the high-speed development and evolution of the internet of things and the internet, the road traffic field tends to be intelligent and convenient. It is expected that image analysis processing technology in the field of intelligent roads is more and more popular and more important, and the development momentum is more and more powerful.

The prior art carries out system research and wide application on an intelligent road system, a video monitoring method mainly based on a video image processing technology is modularized and integrated and industrialized, a mature product is used for monitoring targets, a tracking algorithm is mature relatively, but the change of an application scene may cause results with different precision, in addition, the algorithms have some disadvantages, the accuracy of the algorithms is influenced by complex road conditions, and the accuracy and the robustness of the algorithms under different road conditions and environments need to be improved. Along with the continuous development of economy, the number of motor vehicles is increasing day by day, a plurality of illegal behaviors are generated, the phenomena of road regulation violation such as vehicle modification on the road, license plate applying, fake vehicle license plate applying and the like are forbidden frequently, so the vehicle information identification technology in an intelligent road system is also paid attention gradually, and the method becomes a main research and application direction. The automobile information identification technology has high practical value for guaranteeing the safety of automobiles, assisting traffic management, reducing road accidents and the like.

The proposal of accurate positioning of the automobile mark is generated according to the current requirements, and the comprehensive management burden of roads is heavier and heavier along with the increase of road automobiles, and the method is mainly embodied in three aspects of road management and control, road inspection, traffic police law enforcement and the like. The automobile sign information embodied as the significant features of the automobile type is gradually emphasized, the identification of the automobile sign can obviously promote and optimize the road management, particularly the law enforcement and inspection of the road, for example, some fake license plates, unlicensed license plates and fake license plates are checked, whether the automobile is illegal or not must be identified by simultaneously searching the feature consistency of other methods on the basis of accurately positioning the number plate of the automobile, and the positioning and identification of the automobile sign can play a significant role

The prior art provides some automobile mark identification and positioning methods, including automobile mark identification based on principal component analysis combined with edge invariant moment features, automobile mark identification method for solving edge features combined with image feature invariant moment, adaptive mark identification of multi-feature combination, and the like. Most of the recognition algorithms in the prior art can only be used for standard automobile sign training and recognition, if a training model is used for testing an incomplete automobile sign or a slightly interfered automobile sign, the recognition rate can be rapidly reduced, however, a series of problems can be encountered in real work, road inspection and police officers just want to rapidly locate and search the type of the automobile sign through an automobile sign classification algorithm, if each automobile image needs to be intercepted or cut by an artificial automobile sign, the problem is obviously half the effort, and the automobile sign accurate location algorithms in the prior art mainly comprise the following types: one is that the automobile mark is extracted based on the positioning of the number plate of the vehicle, and the algorithm has the advantages of high efficiency and high positioning speed and has the defect that the positioning precision is seriously influenced due to high external interference; the other type is an automobile mark model based on the existing automobile mark recognition, and has the advantages that the recognition and the positioning can be completed at one time, and the defects are mainly reflected in poor algorithm accuracy and timeliness. In addition to the two main flow methods, the prior art also provides a positioning strategy depending on the position relationship, wherein a small number of small cars with similar car mark sizes, such as the public, Toyota and the like, are mainly researched, and the fixed position belgium relationship among the car mark position, the car number plate position and the car lamp position is found, so that the positioning is realized. However, the algorithm is only effective for certain brands of automobiles, and the actual application cannot be only under the conditions of the automobiles, so that the method cannot be popularized and applied due to inherent defects.

The automobile mark identification and positioning method in the prior art mainly comprises the following categories, namely a morphological processing method, which effectively ensures the integrity of an automobile mark area, can effectively avoid the error influence caused by position relation errors under the condition that the positioning of a vehicle number plate has deviation, and mainly overcomes the defects that if phenomena such as light reflection exist in an image subjected to morphological processing, a plurality of large-block-shaped areas possibly exist, the automobile mark area needing to be positioned exists in the areas, so that the method is completely ineffective; secondly, the feature recognition search is directly matched and positioned, the positioning side of the automobile mark is mainly recognized, the optimal automobile mark matching position is determined through high-precision recognition rate, the method can achieve good recognition effect on the basis of a large number of samples, if a training positive sample is insufficient, the similar automobile marks are difficult to distinguish, if a training negative sample is too little, a non-automobile mark area or an incomplete area is judged to be an automobile mark part, the workload is huge in the early stage, and the recognition sample has harsh requirements; and thirdly, the automobile mark is accurately positioned based on deep learning, but the portability of the actual model is not good, and the requirements on the quantity and the quality of training samples are high.

The research and application of the prior art on automobile mark positioning are improved day by day, and a plurality of advanced image processing and mode recognition methods play an important role in automobile mark positioning, but the methods in the prior art generally need road level photographing equipment to provide a clearer image, and if the image is fuzzy, the accuracy of automobile mark recognition positioning is greatly reduced, the reliability is seriously reduced, and the utilization value is lost. In order to realize an accurate positioning system of the automobile sign with high efficiency, good portability and strong practicability, further research and development are necessary.

In the prior art, automobile marks with part of special conditions are difficult to position and mainly comprise the following three conditions: firstly, the edge information of the radiator around the automobile mark is very obvious, and the edge has the condition of diversity in all directions; secondly, the automobile picture at night has obvious light reflection phenomenon in an automobile mark area, which causes great interference to the binary image, the saliency image and related edge information; and thirdly, some automobile marks are special in shape or position, some automobile marks are positioned on the surface of an automobile, and some automobile marks are large or long in shape, so that even a binary image subjected to image morphology processing cannot be completely closed, and the problem of incomplete automobile mark positioning area is caused. The accurate positioning technology for the road barrier automobile mark developed by the invention mainly faces the three problems which need to be well processed.

In summary, the present invention is intended to solve the following problems in view of some of the drawbacks of the prior art:

firstly, some automobile mark recognition and positioning methods proposed in the prior art include automobile mark recognition based on principal component analysis combined with edge invariant moment features, an automobile mark recognition method for solving edge features combined with image feature invariant moment, and adaptive mark recognition of multi-feature combination, most of which can only be used for standard automobile mark training and recognition, if the automobile mark is incomplete or slightly interfered, the recognition rate can be rapidly reduced, however, a series of problems can be encountered in real work, road inspection and police officers just want to rapidly position and search automobile mark types through an automobile mark classification algorithm, and if each automobile image needs manual automobile mark interception or clipping, the method is obviously both twice as successful and almost has no use and popularization value.

Secondly, the positioning precision is seriously influenced due to large external interference by extracting automobile marks on the basis of vehicle number plate positioning in the prior art; the accuracy and timeliness of the algorithm are poor based on the existing automobile mark model identified by the automobile mark. The prior art also provides a positioning strategy depending on a position relationship, wherein a fixed position belgium relationship among automobile mark positions, automobile number plate positions and automobile lamp positions of a few automobile types is mainly researched, so that the positioning is realized. However, the algorithm is only effective for certain brands of automobiles, and the actual application cannot be only under the conditions of the automobiles, so that the method cannot be popularized and applied due to inherent defects, and is very limited.

Thirdly, in the prior art, if phenomena such as light reflection exist in the images subjected to morphological processing, a plurality of large block-shaped areas may exist, and the automobile mark areas needing to be positioned exist in the areas, so that the method is completely ineffective; the feature recognition search in the prior art is directly matched and positioned, a good recognition effect can be achieved on the basis of a large number of samples, if a training positive sample is insufficient, similar automobile marks are difficult to distinguish, if a negative sample is too few, a non-automobile mark area or an incomplete area can be judged to be an automobile mark part, the early-stage workload is huge, harsh requirements are placed on recognition samples, the algorithm is too high in complexity and too low in speed, the automobile mark is accurately positioned based on deep learning, the transportability of an actual model is too poor, and the number and the quality of the training samples are also highly required.

Fourthly, the method in the prior art generally needs the road checkpoint photographing equipment to provide a clearer image, if the image is fuzzy, the accuracy of the identification and positioning of the automobile mark can be greatly reduced, the reliability is seriously reduced, the utilization value is lost, the high-definition image can not be obtained frequently due to the limitation of various conditions in the environment, sometimes even a fuzzy image can be realized, the prior art has almost no method for accurately positioning the automobile mark of the fuzzy image, and the high-efficiency and strong-practicability accurate positioning of the automobile mark can not be realized.

Fifthly, the automobile mark under the special condition is difficult to locate in the prior art, and the following three conditions are mainly adopted: firstly, the edge information of the radiator around the automobile mark is very obvious, and the edge has the condition of diversity in all directions; secondly, the automobile picture at night has obvious light reflection phenomenon in the automobile mark area, which causes great interference; and thirdly, some automobile marks are special in shape or position, some automobile marks are positioned on the surface of an automobile, and some automobile marks are large or long in shape, so that even a binary image subjected to image morphology processing cannot be completely closed, and the problem of incomplete automobile mark positioning area is caused. The prior art cannot deal with the three problems. The judgment of the automobile mark positioning processing effect is not accurate, and the automobile mark positioning processing method has no robustness and high efficiency.

Disclosure of Invention

Aiming at the defects of the prior art, the automobile mark accurate positioning method based on the road checkpoint fuzzy image, provided by the invention, accurately positions the area where the automobile mark is located through the image processing method of four stages according to the obvious image characteristics of the automobile mark area, obtains the image information of the automobile mark, solves the problem that the road checkpoint photographing equipment is required to provide a clearer image in the prior art, has high automobile mark identification and positioning accuracy and guaranteed reliability even if the provided image is fuzzy, and can well position the automobile mark area by using the obvious characteristic and the edge characteristic combined criterion of the automobile mark area through experimental comparison on the basis of accurate license plate positioning, wherein the test picture in the experiment is from materials under different illumination and meteorological conditions, the comprehensive identification rate reaches 91 percent, and the judgment of the automobile mark positioning processing effect is accurate, the algorithm has good accuracy and timeliness, the method has robustness and high efficiency, can adapt to the conditions of various automobile brands, and has universality and transportability and great market popularization and application potential.

In order to achieve the technical effects, the technical scheme adopted by the invention is as follows:

the automobile mark accurate positioning method based on the road barrier fuzzy image accurately positions the area where the automobile mark is located through an image processing method of four stages according to the obvious image characteristics of the automobile mark area to obtain the image information of the automobile mark;

the extracted significant image characteristics of the automobile sign area comprise edge characteristics, color characteristics, morphological characteristics and position relation characteristics, and the image processing method in four stages comprises coarse positioning of the vehicle number plate area, coarse positioning of the automobile sign area, fine positioning of the automobile sign, extraction of the outline of the automobile sign and a feedback correction mechanism for positioning the automobile sign;

the method comprises the following steps that a vehicle license plate positioning algorithm based on color features and edge features is adopted for vehicle license plate area rough positioning and is combined with digital morphology processing, the algorithm for positioning the vehicle logo is comprehensively judged by utilizing the obvious features of the area above the vehicle license plate in combination with the edge features in the vehicle logo area rough positioning, the interference around the automobile logo is removed through the binarization and morphology processing of an obvious image in the vehicle logo area fine positioning, the accurate position of the vehicle logo is determined through searching an external contour, and the automobile logo area after positioning is subjected to judgment and verification through an automobile logo positioning feedback correction mechanism;

in the coarse positioning of the vehicle license plate region, firstly, an automobile moving target is positioned, the automobile in at least one image of three frames is ensured to be positioned in the middle of the image, the coarse positioning of the vehicle license plate region positions and segments the region containing the vehicle license plate part in the image by an image processing method, and the region is laid for the detection, classification and identification of subsequent characters; according to the difference between the vehicle number plate area and the vehicle body area, the vehicle number plate is in a blue bottom white character or a yellow bottom black character, the shapes of the vehicle number plates are unified into a rectangle with the same specification and size, the coarse positioning of the vehicle number plate area adopts a detection method of color characteristics and edge characteristics, and digital morphological processing is assisted;

in some level images, shooting time, shooting place and camera model data of the level images exist above and below the level images, and the rough positioning of the vehicle number plate area needs to be subjected to image cutting processing, and the uppermost end area and the lowermost end area are subtracted.

The method for accurately positioning the automobile mark based on the road barrier fuzzy image further comprises the following specific steps of adopting a vehicle license plate positioning algorithm based on color features and edge features for coarse positioning of a vehicle license plate area and combining with digital morphology processing:

firstly, roughly positioning a vehicle license plate region based on color characteristics, converting an image into a specific color space by adopting segmentation based on the color space, only reserving a region containing yellow and blue, performing AND operation on the yellow and blue images, and dividing the processed images into two conditions, namely that the color of a vehicle body is blue or yellow and the region of the vehicle body is other colors;

secondly, adopting edge feature assistance, utilizing wavelet analysis to preprocess the vehicle license plate image, extracting longitudinal texture features, namely edge features, of a character region of the vehicle license plate image, and obtaining contour information of a certain region.

The invention relates to an automobile mark accurate positioning method based on a road barrier fuzzy image, which comprises the following steps of obtaining contour information of a block-shaped area, searching an external quadrangle of the area, obtaining one or more rectangular areas of the area, according to a self-adaptive binarization result, referring to a plurality of characters which are fixedly spaced on a vehicle number plate, verifying that one transverse area has a plurality of light and shade jumps, and adding a position relationship for auxiliary judgment;

in the digital morphology processing of the vehicle license plate contour extraction, firstly, the corrosion operation is carried out on a vehicle license plate candidate area by adopting the mask size of 2 multiplied by 2, the independent noise interference is removed, and the section with the cavity is closed by adopting the expansion processing of the structural element of 6 multiplied by 6, so that the vehicle license plate area becomes a closed communication area.

The invention discloses an automobile sign accurate positioning method based on a road checkpoint blurred image, which further adopts edge features in the processes of rough positioning of a vehicle number plate area and rough positioning of an automobile sign, adopts the process of searching a section of continuous image point set with the largest difference with surrounding pixels in an image by edge detection, adopts a gradient first derivative operator for edge feature extraction, and defines the gradient of an image g (x, y) at the position (x, y) as follows:

θ＝arctan(A_y/A_x)

A_x、A_yis the absolute value of the component of the gradient in the x and y directions, respectively, the maximum value approximates the magnitude of the gradient, theta is the angle between the gradient vector and the x axis,

A_x≈g(x，y)-g(x-1，y)

A_y≈g(x，y)-g(x，y-1)

the convolution template for the gradient operator is:

carrying out edge detection by using a Sobel operator, carrying out convolution operation on the image by using a MASK sliding frame for coordinates [ i, j ], more obviously representing pixels with obvious image edge gradient after the operation is finished, weakening the rest areas,

A_i＝[g(i-1，j+1)+2g(i，j+1)+g(i+1，j+1)]-[g(i-1，j-1)+2g(i，j-1)+g(i+1，j-1)]

A_j＝[g(i-1，j-1)+2g(i-1，j)+g(i-1，j+1)]-[g(i+1，j-1)]+2g(i+1，j)+g(i+1，j+1)]

the gradient tension value is calculated by the formula:

A[g(i，j)]＝|A_i|+|A_j|

the Sobel operator can select the horizontal direction to obtain the horizontal edge of the image and the vertical direction to obtain the vertical edge of the image by using two directions,

and applying the coordinate information of the vehicle number plate in the positioning of the automobile mark, and roughly positioning the automobile mark by using the position of the vehicle number plate.

The automobile sign accurate positioning method based on the road barrier fuzzy image further adopts an algorithm of comprehensively judging by combining the salient features of the area above the number plate of the automobile and the edge features for the coarse positioning of the automobile sign area, and the specific flow is as follows:

firstly, correctly positioning a vehicle number plate area to ensure that an automobile mark area above the vehicle number plate exists;

and secondly, taking an area with 3 times of height of the vehicle number plate and 1.2 times of width of the vehicle number plate as an effective area of the vehicle logo above the area of the vehicle number plate, carrying out horizontal and longitudinal Sobel filtering on the effective area of the vehicle logo, acquiring edge information, wherein the edge information of the vehicle logo is rich, and the edge information is still reserved in many directions through the Sobel filtering and is not limited to edges in two directions, and other areas and noise areas can be filtered.

The invention provides a method for accurately positioning an automobile mark based on a road checkpoint blurred image, and further provides another strategy for processing an automobile mark image with an excessively obvious transverse edge, a transverse radiator of an automobile has extremely strong remarkable characteristics, and after the parts are removed, the operating efficiency of a significance algorithm is higher; firstly, filtering an image at a transverse edge, only reserving the image at the transverse part, and performing difference on an original image and the original image, wherein the image after transverse filtering is a single channel, the original image is three channels, and a channel conversion process is needed;

after the processing, salient region detection is carried out on the image to effectively eliminate some transverse edge interference, the image is preprocessed in the process of roughly positioning the automobile mark region, and the extraction of image features and the calculation processing of the salient region are carried out on the basis;

converting the image into a single-channel gray image in the area above the number plate of the vehicle, respectively performing transverse Sobel filtering and color space conversion on the single-channel gray image, and performing subtraction operation on the original image and the filtered image to reduce the interference of a transverse heat dissipation device;

the saliency algorithm is based on the color brightness characteristics of the image, and a salient region is obtained by adopting a color image; the color of the area around the automobile mark is single, and the color significance can be approximately regarded as the representation of the brightness of the automobile mark area.

The automobile mark accurate positioning method based on the road checkpoint fuzzy image further comprises the steps of extracting the accurate position of an automobile mark from an automobile mark area coarse positioning image, and adopting image self-adaptive binarization, morphological processing and contour scanning methods; firstly, carrying out binarization processing on a coarse positioning area of the automobile mark, iteratively merging pixels in a certain interval of a pixel area into a class by using a clustering algorithm, and enhancing the pixels to a high-brightness pixel value;

the contour extraction steps are as follows: sequentially searching bright spot pixels from the left side to the right side and from the upper side to the lower side of the image, reserving the position coordinates of the bright spot pixels when one bright spot pixel is searched, starting to search whether the upper, lower, left and right adjacent areas exist or not by taking the position coordinates as a starting point, and sequentially recording the areas if the area exists, wherein the position of each point in the traversing process and the number of all traversed points are sequentially recorded according to the principle of from top to bottom and from left to right; in the process, when a certain point on the rightmost side cannot find a point on the right side, searching downwards, repeating the steps, searching upwards leftwards, and storing all traversed points in a dynamic array, wherein the points are all points of a required contour area;

after each point is determined by the method, the points are connected into an external contour by a curve for one time, and an external quadrangle is searched, only points with changed horizontal and longitudinal directions are reserved, namely the upper left and lower right points, and the external quadrangle of the area is obtained through the two position coordinates;

according to the method, the circumscribed rectangle of each closed interval is calculated, so that a plurality of circumscribed rectangle areas can be obtained, and a plurality of block areas in other positions or block areas containing incomplete automobile marks exist; the invention divides the areas into two types, including the automobile mark area and the area without the automobile mark area, and the work to be completed is to merge the areas containing the automobile mark as much as possible and abandon the areas without the automobile mark.

The automobile mark accurate positioning method based on the road barrier fuzzy image further comprises the steps that positioning areas of different algorithms are stored, positioning results of the different algorithms are compared, if the difference is not large, a result area of a significant model algorithm is reserved, and if the difference is large, matching search is carried out; the car logo positioning feedback correction mechanism mainly comprises the following steps:

training a car logo model, directly calling by referring to library functions of libsvm, selecting HOG features of a car logo area in the training, uniformly adjusting the size of the car logo to 32 × 32 pixels in the feature extraction process, forming each 8 × 8 pixel into a cell, forming each 2 × 2 cell into a block, wherein each cell has 8 features, each block has 4 × 8 × 32 features, 8 pixels are used as step length, 3 scanning windows are arranged in the horizontal direction, 3 scanning windows are arranged in the vertical direction, namely 32 × 32 images have 32 × 9 features in total, and the most voting scheme is used for the final result by means of a hierarchical multi-class support vector machine;

after training is finished, reading in the model and starting sliding search in a positioning interval, wherein the method trains 16 types of common automobile signs; the search matching is divided into two strategies, one is block search, the other is incremental variable search according to a certain pixel distance, and the return value of the search result of each area is recorded; in the block search model, the original positioning area is divided into a plurality of block areas, namely an upper block area, a lower block area, a left block area and a right block area, wherein each block area covers each other, the total coverage area is 20% larger than that of the original block area, and the recognition rate is further improved by recognizing the blocks.

Compared with the prior art, the invention has the advantages and innovation points that:

the method for accurately positioning the automobile mark based on the road checkpoint blurred image solves the problem that the prior art method generally needs the road checkpoint photographing equipment to provide a clearer image, even if the provided image is fuzzy, the accuracy of automobile mark identification and positioning is still high, the reliability is fully guaranteed, a high-definition image can not be obtained frequently due to the limitation of various conditions in an implementation environment, and sometimes even in the case of a fuzzy image, and a set of method for accurately positioning the automobile mark is provided for the road checkpoint blurred image, so that the efficient and practical automobile mark accurate positioning is realized, and the effectiveness, the practicability and the advancement of the method are verified through experiments.

The automobile mark accurate positioning method based on the road barrier blurred image provided by the invention is a special solution to the problem that the automobile mark under some special conditions is difficult to position in the prior art, and well solves the problem of difficulty in positioning the automobile mark: firstly, the edge information of the radiator around the automobile mark is very obvious, and the edge has diversity in all directions, so that the problem of difficulty in identification is solved; secondly, the automobile picture at night has obvious light reflection phenomenon in the automobile mark area, thereby causing great interference; and thirdly, some automobile marks are special in shape or position, some automobile marks are positioned on the surface of an automobile, and some automobile marks are large or long in shape, so that even a binary image subjected to image morphology processing cannot be completely closed, and the problem of incomplete automobile mark positioning area is caused. The judgment of the automobile mark positioning processing effect is accurate, and the automobile mark positioning processing method has robustness and high efficiency, and is very beneficial to the subsequent utilization of the accurate positioning of the automobile mark.

Thirdly, according to the salient image characteristics of the automobile mark region, the automobile mark region is accurately positioned through the image processing method of four stages to obtain the image information of the automobile mark, the automobile mark positioning is based on the accurate license plate positioning, the salient characteristics and the edge characteristics of the automobile mark region are combined to judge the criterion through experimental comparison, the automobile mark region can be well positioned, the test pictures in the experiment are from materials under different illumination and meteorological conditions, the comprehensive recognition rate reaches 91%, and the method has great use value.

The automobile mark accurate positioning method based on the road checkpoint fuzzy image effectively avoids the influence on positioning accuracy due to larger external interference, has good accuracy and timeliness of the algorithm, avoids the positioning realized by utilizing the fixed position ratio relationship between the automobile mark position of a few automobile types, the automobile number plate position and the automobile lamp position, is effective to all brands of automobiles, can adapt to the conditions of various automobile brands in practical application, has small limitation, universality and transportability, and has huge market popularization and application potential.

Fifthly, the method for accurately positioning the automobile mark based on the road checkpoint blurred image can be used for standard automobile mark training and recognition, and for incomplete automobile marks or slightly interfered automobile marks, the recognition rate can be kept at a higher level, so that the method is not used and has high popularization value. The method has the advantages of low algorithm complexity, easy realization, obvious effect, stable precision positioning quality of the automobile mark of the checkpoint fuzzy image, good quality control capability and service quality guarantee mechanism, and more robustness and high efficiency.

Drawings

FIG. 1 is a flow chart of an implementation of the method for accurately positioning the automobile mark based on the road barrier blurred image.

FIG. 2 is a schematic view of the rough location of the automotive marking area for determining the general area of the automotive marking according to the present invention.

FIG. 3 is a diagram illustrating comparison of results after processing by the four-stage image processing method according to the present invention.

Detailed Description

The following describes a technical solution of the method for accurately positioning an automobile mark based on a blurred image of a road level with reference to the accompanying drawings, so that those skilled in the art can better understand the method and can implement the method.

According to the automobile mark accurate positioning method based on the road checkpoint blurred image, the area where the automobile mark is located is accurately positioned through the image processing method of four stages according to the remarkable image characteristics of the automobile mark area, and the image information of the automobile mark is obtained; the extracted significant image characteristics of the automobile sign area comprise edge characteristics, color characteristics, morphological characteristics and position relation characteristics, and the image processing method in four stages comprises coarse positioning of the automobile number plate area, coarse positioning of the automobile sign area, fine positioning of the automobile sign, extraction of the outline of the automobile sign and a feedback correction mechanism for positioning the automobile sign.

FIG. 1 is a flow chart of an implementation of the method for accurately positioning the automobile sign based on the road barrier blurred image, the coarse positioning of the vehicle number plate area adopts a vehicle number plate positioning algorithm based on color features and edge features, and is combined with digital morphology processing, the coarse positioning of the automobile sign area utilizes an algorithm for comprehensively judging and positioning the automobile sign by combining the salient features of the area above the vehicle number plate and the edge features, the fine positioning of the automobile sign removes surrounding interference through binarization and morphology processing of the salient image, and the accurate position of the automobile sign is determined by searching for an external contour. In order to further reduce the positioning error, the invention judges and verifies the positioned vehicle logo area through a vehicle logo positioning feedback correction mechanism, thereby reducing the error caused by the problems of positioning offset of the vehicle number plate and the like.

Coarse positioning of vehicle license plate area

In order to obtain automobile information, firstly, an automobile moving target needs to be positioned, in order to strictly test the performance of the invention, the barrier image part selected by the invention is from a fuzzy image shot by a common barrier camera, the number of shot frames is not high actually, each frame is about 2.3 seconds apart, the condition that the automobile is in the middle position of the image in at least one frame of the three frames is ensured, and the interference of the surrounding automobile can be ignored, the invention is a static image, no tracking condition is considered, the automobile body area is not needed to be repeatedly positioned, the area containing the automobile license plate part in the image is positioned and divided by an image processing method for the rough positioning of the automobile license plate area, a cushion is made for the detection, classification and identification of subsequent characters, according to the difference between the automobile license plate area and the automobile body area, the automobile license plate is a white character with a blue bottom or a black character with a yellow bottom, the shapes of the automobile license plates are uniform and are rectangles with the same specification and size, the rough positioning of the, and the digital morphology processing is assisted, in some barrier images, shooting time, shooting place and shooting model data of the barrier images exist above and below the images, and the positioning effect of automobile signs can be greatly interfered, so that the rough positioning of the vehicle number plate area needs to perform image cutting processing, and the uppermost end area and the lowermost end area are subtracted.

The coarse positioning of the vehicle license plate area adopts a vehicle license plate positioning algorithm based on color characteristics and edge characteristics, and is combined with digital morphology processing. The method comprises the following specific steps:

firstly, roughly positioning a vehicle license plate region based on color characteristics, wherein the color change of the vehicle license plate region is small, and the vehicle license plate region has coherence and consistency, adopting color space-based segmentation to the first-step segmentation of the vehicle license plate region, converting an image into a specific color space, only reserving a region containing yellow and blue, performing AND operation on the yellow and blue images, filtering an interference region of non-common colors outside a vehicle body, and dividing the processed images into two conditions, namely that the color of the vehicle body is blue or yellow and the color of the vehicle body region is other colors;

secondly, if the single image feature of color is adopted to position the vehicle license plate area, the vehicle body area and the vehicle license plate area with the same main tone as the vehicle license plate area cannot be distinguished, so that edge feature assistance is needed, wavelet analysis is utilized to preprocess the vehicle license plate image, longitudinal texture features, namely edge features, of the character area of the vehicle license plate image are extracted, and outline information of a certain area is obtained. And (3) carrying out digital morphology processing on the image after binarization processing to enable the vehicle license plate area to approach a closed interval, forming the closed interval, and strengthening or weakening the periphery to enable the image to achieve the expected effect by utilizing the relation between the pixel of a certain point and the surrounding pixels. Because the ratio of the vehicle number plate area in the vehicle body area is small, an overlarge closure operator cannot be adopted to carry out corrosion operation on the binary automobile sign image.

The method and the device search the circumscribed quadrangle of the block area after acquiring the outline information of the block area, and acquire one or more rectangular areas of the block area. In order to ensure the accuracy, according to the result of self-adaptive binarization, one transverse area is selected to verify that a plurality of light and shade jumps exist on the vehicle license plate according to the prior condition, and the position relation is added for auxiliary judgment.

In the digital morphology processing of the vehicle license plate contour extraction, firstly, the mask size of 2 multiplied by 2 is adopted to carry out corrosion operation on a vehicle license plate candidate area, independent noise interference is removed, and the structural element expansion processing of 6 multiplied by 6 is adopted to close a section with a cavity, so that the area where the vehicle license plate is located forms a communicated area. Through the mathematical morphology processing, the vehicle number plate area becomes a closed communication area. Experiments show that the mask size of the parameters selected by the invention can completely reserve all areas of the vehicle license plate, and a relatively large mask area is adopted in expansion.

Finally, the uppermost and lowermost boundaries of the vehicle number plate area are found and positioned. The method selects the barrier fuzzy image, the comprehensive positioning rate can reach 94.68% by utilizing the method, most errors come from polluted vehicle license plates or unlicensed vehicles, and the method keeps certain tolerance for the ineligible forces.

The invention adopts edge characteristics in the processes of coarse positioning of a vehicle number plate area and coarse positioning of an automobile mark, the edge detection searches a section of continuous image point set with the maximum difference with surrounding pixels in an image, the edge characteristic extraction adopts a gradient first derivative operator, and the gradient of an image g (x, y) at the position (x, y) is defined as follows:

θ＝arctan(A_y/A_x)

A_x≈g(x，y)-g(x-1，y)

A_y≈g(x，y)-g(x，y-1)

the convolution template for the gradient operator is:

A_j＝[g(i-1，j-1)+2g(i-1，j)+g(i-1，j+1)]-[g(i+1，j-1)]]+2g(i+1，j)+g(i+1，j+1)]

the gradient tension value is calculated by the formula:

A[g(i，j)]＝|A_i|+|A_j|

the Sobel operator can select the horizontal direction to obtain the horizontal edge of the image and select the vertical direction to obtain the vertical edge of the image by using two directions.

The invention jointly applies a plurality of judgment strategies to improve the positioning accuracy of the vehicle license plate in the process of positioning the automobile mark, which is the basic guarantee of subsequent information acquisition. The coordinate information of the vehicle license plate also needs to be applied in the accurate positioning of the vehicle mark, and because most of the positions of the vehicle mark are right above the vehicle license plate, the positions of the vehicle license plate can be utilized to perform coarse positioning on the vehicle mark.

Coarse positioning of automobile mark area

The invention provides an algorithm for comprehensively judging and positioning automobile signs by combining the salient features of the areas above the number plates of the vehicles with the edge features, and after the areas of the number plates of the vehicles are obtained, the judgment standard for the correct positioning of the number plates of the vehicles is that the areas of the number plates of the vehicles covered by the external positioning rectangular frames of the number plates of the vehicles are not less than 75% and not more than 125% of the number plates of the vehicles, namely the areas of the automobile signs above the areas are well reserved. The obtained checkpoint image automobile information is summarized and analyzed, so that most automobile marks are located at a certain height right above the number plate of the automobile and have obvious edge characteristics and color characteristics, and the specific process of roughly positioning the automobile mark area comprises the following steps:

and secondly, taking a region with 3 times of height and 1.2 times of width of the vehicle number plate as an effective region of the vehicle logo above the vehicle number plate region, and carrying out horizontal and longitudinal Sobel filtering on the effective region of the vehicle logo according to the fact that the vehicle logo in a normal image has a positive relation with the dimension of the vehicle number plate to obtain edge information.

The invention also has another strategy for processing the car logo image with too obvious transverse edge, the car transverse radiator has extremely strong obvious characteristics, the area can be retained by the significance algorithm, which generates great interference to the extraction of the car logo outline, if the parts can be removed, the operation efficiency of the significance algorithm is higher.

Fig. 2 shows that the empirical parameter used in the present invention can completely cover the automobile mark as long as the correct positioning of the vehicle license plate can be ensured, and the complete automobile mark area can be positioned only by processing the area covering the complete automobile mark.

After the processing, the salient region detection is carried out on the image, and some transverse edge interference can be effectively eliminated. In the process of roughly positioning the automobile mark area, the image is preprocessed, and on the basis, the extraction of the image characteristics and the calculation processing of the salient area are carried out. Converting the image into a single-channel gray image in the area above the number plate of the vehicle, respectively performing transverse Sobel filtering and color space conversion on the single-channel gray image, and performing subtraction operation on the original image and the filtered image to further reduce the interference of a transverse heat dissipation device; the saliency algorithm uses a color image to find a salient region based on the color brightness characteristics of the image. In order to avoid the interference of the transverse bars, the image used for participating in the significance calculation is an image obtained by subtracting the original image from the transverse interference area, and the color significance can be approximately regarded as the representation of the light and shade of the automobile mark area due to the single color of the area around the automobile mark.

Thirdly, fine positioning of the automobile mark and extraction of the outline of the automobile mark

The rough positioning result is utilized to basically determine the approximate position of the automobile mark, the accurate position of the automobile mark is extracted from the rough positioning map of the automobile mark area, and the methods of image self-adaptive binarization, morphological processing and contour scanning are adopted. Firstly, carrying out binarization processing on a coarse positioning area of the automobile mark, iteratively merging pixels in a certain interval of a pixel area into a class by using a clustering algorithm, and enhancing the pixels to a high-brightness pixel value; the reason for binarizing the region by adopting the clustering algorithm is as follows: the clustering algorithm is used for binaryzation of the significant image, when the significant image is solved, the size of the significant region in the image processing is double-precision between 0 and 1, the automobile mark region cannot be well expressed by adopting a self-adaptive method, the image after the binaryzation processing is probably not a closed region, image morphological corrosion and expansion processing are further required to be further carried out, noise interference is removed, and the automobile mark region is in a block shape through a certain search strategy.

The contour extraction steps are as follows: sequentially searching bright spot pixels from the left side to the right side and from the upper side to the lower side of the image, reserving the position coordinates of the bright spot pixels when one bright spot pixel is searched, starting to search whether the upper, lower, left and right adjacent areas exist or not by taking the position coordinates as a starting point, and sequentially recording the areas if the area exists, wherein the position of each point in the traversing process and the number of all traversed points are sequentially recorded according to the principle of from top to bottom and from left to right; in the process, when a certain point on the rightmost side cannot find a point on the right side, a downward search is started, and the like, then a leftward search is carried out, and all the traversed points are stored in a dynamic array, wherein the points are all the points of a required contour area.

Because more or less discrete points exist around the automobile mark region, the automobile mark region can be considered as a small outline, and one-step advanced screening is needed.

After each point is determined by the method, the points are connected into an external contour by a curve at one time, and because the contours are various in shape, the method for searching the external quadrangle is adopted, only the points with changed horizontal and longitudinal directions, namely the upper left and lower right points, are reserved, and the external quadrangle of the area is obtained through the two position coordinates.

The invention calculates the circumscribed rectangle of each closed interval, and can obtain a plurality of circumscribed rectangle areas, and a plurality of block areas in other positions or block areas containing incomplete automobile marks exist. The invention only divides the areas into two types, including the automobile mark area and the area without the automobile mark area, the work to be completed is to merge the areas containing the automobile mark as much as possible and abandon the areas without the automobile mark, and the specific method is as follows:

firstly, combining the coordinates of the upper left position and the lower right position and the size of an external rectangle, when the center of the rectangular area is positioned in the rectangular area which is close to the symmetric axis of the vehicle license plate and does not deviate from the rectangular area with more than 40 pixels, the rectangular area is considered to be a correct candidate area, and rectangular frames on the rough positioning edge are excluded;

secondly, judging the size of the rectangular frame, searching the rectangular frames with the height of the vehicle number plate of +/-30 percent and the width of the vehicle number plate of +/-35 percent according to the judgment standard, if the rectangular frames meet the quantity standard, storing the rectangular frames in an array of a reserved matrix, if not, still judging the distance between the rectangular frames at two sides of the shaft by taking the center as the shaft, calculating the distance according to the center distance, merging the two rectangular areas when the center distance between the two rectangular frames is 9 to 16 percent of the width of the vehicle number plate, externally connecting the rectangular areas, namely, respectively reserving four positions of the top, the left, the bottom, the right and the top, then combining the four positions into a new diagonal vertex of the rectangle, determining a new merged rectangle, deleting the small rectangular area before merging, still searching whether the merged frame exists according to the method, if so, continuing and judging the difference between the height and the width of the size of the frame and the vehicle number plate of a certain ratio, or the height of the vehicle number plate is +/-30%, and the width of the vehicle number plate is +/-35%, and finally the combined area is stored in the candidate area.

When the edge of the automobile is extremely reflective, or there is no automobile mark at all, or the length of the automobile mark is very long, which exceeds the empirical value, the merged interval is too large or still not enough to cover the whole automobile mark area. And a car logo positioning feedback correction mechanism is added to improve the identification precision of the algorithm.

3.5. Car logo positioning feedback correction mechanism

The method only combines the window and the image entropy value search has an obvious problem, a large rectangular frame is selected for searching, the size and the position of each obtained barrier image are different no matter the same or different types of automobile signs, even if the same automobile is in different frames, the size of the automobile signs cannot be determined by a uniform scale due to different distances from a camera, and in order to ensure that an automobile sign area close to the standard is obtained, the invention designs an automobile sign positioning feedback correction mechanism, and a detailed flow chart is as follows:

in order to ensure that the positioning algorithm and the recognition algorithm have optimal fusion and accuracy and improve the accuracy of accurate positioning, the vehicle logo positioning feedback correction mechanism stores positioning areas of different algorithms, compares positioning results of different algorithms, reserves a significant model algorithm result area if the difference is not large, and performs matching search if the difference is large. The method mainly comprises the following steps:

the method comprises the steps of training an automobile mark model, directly calling by referring to library functions of libsvm, selecting HOG features of an automobile mark area in the training, uniformly adjusting the size of the automobile mark to 32 x 32 pixels in the feature extraction process, forming a cell by 8 pixels, forming a block by 2 cells, forming 8 features by 8 cells, forming 4 x 8-32 features in each block, taking 8 pixels as step length, forming 3 scanning windows in the horizontal direction and 3 scanning windows in the vertical direction. I.e., 32 x 32 images, for a total of 32 x 9-288 features. With the help of a hierarchical multiclass support vector machine, the most voting schemes are used for the final result.

After training is finished, the model is read in and sliding search is started in a positioning interval, and the method trains 16 common automobile signs. The search matching is divided into two strategies, one is block search, and the other is incremental variable search according to a certain pixel distance, and the return value of the search result of each area is recorded. The invention sets a high recognition probability critical value, and when the critical value of a certain type of automobile mark is judged to be smaller than the current set value, the result of the area is not recorded, so that the situation that a plurality of same labels are repeated and the critical values are close to each other is avoided.

Through the series of methods, the coarse positioning precision of the automobile mark can be further improved. In the block model, the original positioning area is divided into a plurality of block areas, namely an upper block area, a lower block area, a left block area and a right block area, wherein each block area covers each other, the total coverage area is 20% larger than that of the original block area, the recognition rate is further improved by recognizing the blocks, and incomplete results in the original positioning are avoided.

FIG. 3 is a diagram illustrating comparison of results after processing by the four-stage image processing method according to the present invention. The image group (A) is a result schematic diagram of the original image after the coarse positioning of the vehicle number plate area, (B) is a vehicle mark coarse positioning area selected according to an empirical value, (C) is divided into an upper part and a lower part, the upper part is an edge characteristic, the lower part is a remarkable characteristic, and (D) is a final vehicle mark coarse positioning result obtained by the method.

The method analyzes and develops the automobile mark accurate positioning method in road information acquisition under the large background of an intelligent road. And designing and realizing the automobile mark accurate positioning system by applying the significance characteristics. The automobile mark positioning method has the advantages that the positioning precision is further improved, the automobile mark can be positioned in a poor environment, and the anti-interference capability of the accurate positioning of the automobile mark is improved.

The automobile mark accurate positioning method based on the road checkpoint fuzzy image, provided by the invention, designs the overall architecture of an automobile mark accurate positioning module, divides the automobile mark accurate positioning module into modules, introduces the action and value of each module, adopts a series connection realization method, analyzes the realization strategy of each module, expounds the design realization idea, designs an optimal vehicle license plate positioning strategy aiming at the checkpoint fuzzy image, analyzes the robustness and the timeliness of the whole system, and is proved by experiments to have excellent performance, good stability and good accuracy and robustness of a detection and identification algorithm.

Claims

1. The automobile mark accurate positioning method based on the road barrier fuzzy image is characterized in that according to the remarkable image characteristics of an automobile mark area, the area where the automobile mark is located is accurately positioned through an image processing method in four stages, and the image information of the automobile mark is obtained;

2. The method for accurately positioning the automobile mark based on the road barrier blurred image as claimed in claim 1, wherein the vehicle number plate area rough positioning adopts a vehicle number plate positioning algorithm based on color features and edge features, and is combined with digital morphology processing, and the specific steps are as follows:

3. The method for accurately positioning the automobile mark based on the road barrier blurred image is characterized in that after contour information of a block-shaped area is obtained, a circumscribed quadrangle of the area is searched, one or more rectangular areas of the area are obtained, according to a self-adaptive binarization result, a plurality of characters with fixed intervals exist on a vehicle number plate, one transverse area is selected for verification, and a plurality of light and shade jumps are added, so that auxiliary judgment of a position relationship is carried out;

4. The method for accurately positioning the automobile mark based on the road barrier blurred image is characterized in that the edge features are adopted in the processes of rough positioning of the vehicle number plate area and rough positioning of the automobile mark, the edge detection is used for searching a continuous image point set with the largest difference with surrounding pixels in an image, the edge feature extraction adopts a gradient first derivative operator, and the gradient of an image g (x, y) at (x, y) is defined as follows:

θ＝arctan(A_y/A_x)

A_x≈g(x，y)-g(x-1，y)

A_y≈g(x，y)-g(x，y-1)

the convolution template for the gradient operator is:

the gradient tension value is calculated by the formula:

A[g(i，j)]＝|A_i|+|A_j|

5. The method for accurately positioning the automobile mark based on the road barrier blurred image as claimed in claim 1, wherein the rough positioning of the automobile mark area adopts an algorithm of comprehensive judgment by combining salient features of the area above the number plate of the automobile with edge features, and the specific flow is as follows:

6. The method for accurately positioning the automobile mark based on the road barrier blurred image is characterized in that another strategy is provided for processing the automobile mark image with the excessively obvious transverse edge, a transverse radiator of the automobile has extremely strong remarkable characteristics, and after the parts are removed, the operating efficiency of a significance algorithm is higher; firstly, filtering an image at a transverse edge, only reserving the image at the transverse part, and performing difference on an original image and the original image, wherein the image after transverse filtering is a single channel, the original image is three channels, and a channel conversion process is needed;

7. The automobile mark accurate positioning method based on the road checkpoint blurred image as claimed in claim 1, wherein the automobile mark accurate positioning and the automobile mark outline extraction are methods of extracting the accurate position of an automobile mark from an automobile mark area rough positioning map, and adopting image self-adaptive binarization, morphological processing and outline scanning; firstly, carrying out binarization processing on a coarse positioning area of the automobile mark, iteratively merging pixels in a certain interval of a pixel area into a class by using a clustering algorithm, and enhancing the pixels to a high-brightness pixel value;

8. The automobile mark accurate positioning method based on the road barrier blurred image is characterized in that the automobile mark positioning feedback correction mechanism stores positioning areas of different algorithms, compares the positioning results of the different algorithms, reserves a significant model algorithm result area if the difference is not large, and performs matching search if the difference is large; the car logo positioning feedback correction mechanism mainly comprises the following steps: