A method and device for detecting multi-direction license plates

Abstract

The present invention relates to a method and apparatus for detecting multi-direction license plates. The method includes the following steps: S1. Extract features from the collected and processed image to obtain a first feature; detect the first feature to obtain the predicted position of the license plate; S2. Expand the predicted position of the license plate to obtain an expanded area; S3. Using the expanded area as a reference, process the first feature to obtain a second feature; S4. Correct and re-detect the second feature to obtain the position of the multi-direction license plate. The present invention provides a license plate detection method from coarse to fine to improve the detection performance of license plates of different sizes. The present invention uses an anchor-free mechanism and a coarse detection method to extract and process the features of the license plate in the image, and uses a position correction module to improve the accuracy of license plate detection. In the position correction module, a calculation method based on key point regression of four vertices is used to locate the multi-direction license plate. A large number of experiments show the effectiveness and scalability of the method of the present invention.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and particularly relates to a method and device for detecting multi-direction license plates. Background Art

[0002] The intelligent transportation system is an important development direction of the future transportation system. The intelligent license plate recognition system is an important part of the intelligent transportation system. As a pre-module of the license plate recognition system, automatic license plate detection is also particularly important. Traditional license plate detection methods achieve license plate detection by designing features to detect the edges of license plates, but this method has very large limitations, can only adapt to very small scenarios, and has a relatively slow speed. With the unprecedented increase in computing power and data in recent years, deep learning has been used in various fields.

[0003] Currently, deep learning-based license plate detection methods all ignore the foreground-background class imbalance problem caused by small-sized license plates, which is caused by the fact that small-sized license plates often only occupy a small part of the image, which also leads to poor detection performance of their small-sized license plates. In some existing methods, the entire image is directly analyzed to detect the license plate position, ignoring the foreground-background class imbalance problem caused by small-sized license plates, often resulting in many missed detections; there are also some methods that first detect the vehicle position and then locate the license plate in the vehicle area, relying on the effect of the vehicle detector. If the vehicle is partially blocked, the vehicle may not be detected and thus the license plate may not be detected; moreover, if large vehicles such as large trucks and buses are encountered, the license plate will still only occupy a small part of the vehicle area, which will require complex post-processing and still does not solve the problem of missed detections of license plates of large vehicles. Summary of the Invention

[0004] The purpose of the present invention is to solve the foreground-background class imbalance problem in the detection task caused by small-sized license plates in the prior art for road traffic scenarios, which leads to the problem of missed detections of small-sized license plates. The present invention aims to provide a method and device for detecting multi-direction license plates to solve the above problems existing in the prior art.

[0005] The above technical object of the present invention will be achieved by the following technical solutions.

[0006] A method for detecting multi-direction license plates, the method comprising the following steps:

[0007] S1. Extract features from the collected and processed image to obtain a first feature; detect the first feature to obtain the predicted position of the license plate;

[0008] S2. Expand the predicted position of the license plate to obtain an expanded area;

[0009] S3. Using the expanded area as a reference, process the first feature to obtain a second feature;

[0010] S4. Correct and re-detect the second feature to obtain the position of the multi-direction license plate.

[0011] In the above aspects and any possible implementation manners, a further implementation manner is provided. The collected and processed picture is obtained by proportionally scaling and normalizing the collected picture according to a pre-specified size.

[0012] In the above aspects and any possible implementation manners, a further implementation manner is provided. Specifically, S3 includes:

[0013] S31. Expand the predicted position of the license plate outward by 1 to 4 times to obtain an expanded area, then crop the corresponding cropped area from the first-layer feature map of the first feature, and perform scaling processing on the cropped area;

[0014] S32. Repeat step S31 to obtain all scaled cropped areas corresponding to the predicted position of the license plate from the first-layer feature map of the first feature, and splice all the cropped areas to obtain the second feature.

[0015] In the above aspects and any possible implementation manners, a further implementation manner is provided. Specifically, S4 includes: extracting the second feature to obtain a third feature, processing the third feature to obtain a 2X2X9-dimensional vector, and calculating the position of the multi-direction license plate from the vector.

[0016] The present invention also provides a multi-direction license plate detection device, which includes a rough detection module, a region expansion and feature selection module, and a position correction module.

[0017] Among them, the rough detection module is used to extract features from the collected and processed picture to obtain a first feature; detect the predicted position of the license plate from the first feature, and transmit the predicted position of the license plate to the region expansion and feature selection module;

[0018] The region expansion and feature selection module expands the received predicted position of the license plate to obtain an expanded area, uses the expanded area as a reference, processes the first feature to obtain a second feature, and transmits the second feature to the position correction module;

[0019] The position correction module corrects and re-detects the second feature to obtain the position of the multi-direction license plate.

[0020] For the aspects and any possible implementation manners described above, a further implementation manner is provided, where the collected and processed picture is obtained by proportionally scaling and normalizing the collected picture according to a pre-specified size.

[0021] For the aspects and any possible implementation manners described above, a further implementation manner is provided, where the rough detection module includes a feature extraction module and an anchor-free detection module connected to each other, and the feature extraction module extracts the first feature from the collected and processed picture.

[0022] The anchor-free detection module respectively detects license plates of different sizes in the first feature, obtains a vector with a certain spatial dimension, and calculates the vector to obtain the predicted positions of license plates of different sizes.

[0023] For the aspects and any possible implementation manners described above, a further implementation manner is provided, where an optimization objective function is set in the rough detection module for detecting the difference between the position of the real license plate in the image and the position of the predicted license plate; an optimization objective function is also set in the position correction module for detecting the difference between the position of the real license plate in the image and the position of the multi-direction license plate.

[0024] For the aspects and any possible implementation manners described above, a further implementation manner is provided, where the position correction module includes a multi-direction license plate detection module. The position correction module extracts the third feature from the second feature, inputs the third feature into the multi-direction license plate detection module for processing to obtain a 2X2X9-dimensional vector, and calculates the vector to obtain the position of the multi-direction license plate.

[0025] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the detection method for multi-direction license plates according to the present invention is implemented.

[0026] The beneficial technical effects of the present invention

[0027] The multi-direction license plate detection method and device provided by the embodiments of the present invention, the method includes the following steps: S1. Extract features from the collected and processed image to obtain the first feature; detect the first feature to obtain the predicted position of the license plate; S2. Expand the predicted position of the license plate to obtain an expanded area; S3. Using the expanded area as a reference, process the first feature to obtain the second feature; S4. Correct and re-detect the second feature to obtain the position of the multi-direction license plate. The present invention provides a license plate detection method from coarse to fine to improve the detection performance of license plates of different sizes. The present invention uses an anchor-free mechanism and a coarse detection method to extract and process the features of the license plate in the image, and uses a position correction module to improve the accuracy of license plate detection. The position correction module adopts a calculation method based on key point regression of four vertices to locate the multi-direction license plate. A large number of experiments show the effectiveness and scalability of the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings, where:

[0029] Figure 1 It is a schematic flow chart of the method in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the drawings and specific embodiments, but the embodiments of the present invention are not limited thereto.

[0031] As Figure 1 shown, a multi-direction license plate detection method of the present invention, the method includes the following steps: S1. Extract features from the collected and processed image to obtain the first feature; detect the first feature to obtain the predicted position of the license plate;

[0032] S2. Expand the predicted position of the license plate to obtain an expanded area;

[0033] S3. Using the expanded area as a reference, process the first feature to obtain the second feature;

[0034] S4. Correct and re-detect the second feature to obtain the position of the multi-direction license plate.

[0035] Preferably, in the embodiments of the present invention, the collected and processed image is obtained by scaling and normalizing the collected image according to a pre-specified size.

[0036] Preferably, in the embodiments of the present invention, S3 specifically includes:

[0037] S31. Expand the predicted position of the license plate outward by 1 to 4 times to obtain an expanded area, then crop the cropping area corresponding to the expanded area from the first-layer feature map, and scale the cropping area;

[0038] S32. Repeat step S31 to obtain all scaled cropping areas corresponding to the predicted position of the license plate from the first-layer feature map, and splice all the cropping areas to obtain the second feature.

[0039] Specifically, the method of the present invention is to detect multi-direction license plates in a traffic scene. The multi-direction license plate is, in an actual scene, due to reasons such as the shooting angle, the license plate in the acquired image is not a regular rectangle but an arbitrary quadrilateral similar to a parallelogram. For the convenience of subsequent processing, it is called a multi-direction license plate. The multi-direction license plate detection device of the present invention includes three modules: a rough detection module, a region expansion and feature selection module, and a position correction module. Among them, the rough detection module detects the rough position of the license plate and uses an anchor-free detection module to improve the recall ability for small license plate detection; the region expansion and feature selection module (ERM) expands the license plate region according to the result of the rough detection module and obtains low-level visual features and then inputs them into the position correction module; the position correction module (location refinement module) further extracts and analyzes the features output by the region expansion and feature selection module to further correct the position of the license plate and achieve the four-point detection of the multi-direction license plate.

[0040] The rough detection module of the present invention is used to estimate the general area where the license plate is located in the entire image. The rough detection module includes a feature extraction module and an anchor-free detection head module.

[0041] The collected images are scaled to a predefined size proportionally, where the longest side of the image is 320 pixels. The above processing helps to reduce the overall computational amount, and the proportional scaling can maintain the original shape of the license plate in the scaled image. Then, the scaled images are normalized. Each pixel in the normalized image is divided by 256 and then sent to the feature extraction module, where feature extraction is performed to obtain the first feature. The first feature is a multi-layer feature map composed of the first-layer feature map, the second-layer feature map, the third-layer feature map... the nth-layer feature map, where n is greater than or equal to 2. The first-layer feature map includes the region closest to the shape of the license plate. The size of the license plate includes large license plates, medium license plates, and small license plates according to different vehicle categories. Correspondingly, the first feature includes high-level features containing large license plates, middle-level features containing medium license plates, and low-level features containing small license plates, and the high-level features, middle-level features, and low-level features exist in each layer of the feature map. The first feature obtained by the feature extraction module is sent to an anchor-free detection module for detection.

[0042] The anchor-free detection module includes three components arranged in parallel: a low-level component, a middle-level component, and a high-level component. The input of the low-level component is the low-level feature in the first feature; the input of the middle-level component is the middle-level feature in the first feature; the input of the high-level component is the high-level feature in the first feature. The outputs of the three components are all the predicted license plate positions. The three components analyze and predict the positions for the features in different layers of the first feature. Therefore, when designing the optimization objective, it is necessary to allocate features containing license plates of different sizes to different components of the anchor-free detection module, that is, allocate the high-level features containing large license plates to the high-level component, the middle-level features containing medium license plates to the middle-level component, and the low-level features containing small license plates to the low-level component. Such an operation is more conducive to detection.

[0043] Preferably, the present invention can also implement the allocation of features containing license plates of different sizes to different components of the anchor-free detection module in the following way. First, the position G=(x, y, w, h)∈ of the real license plate (corresponding to the license plate predicted by the present invention) in the defined and collected image is obtained. Among them, the variables x and y respectively represent the coordinates of the center point of the circumscribed rectangle of the real license plate; the variables w and h respectively represent the width and height of the circumscribed rectangle. The longest side of the circumscribed rectangle of the license plate in the collected image is used to determine which component in the anchor-free detection module the image is assigned to: First, calculate the longest side of the circumscribed rectangle of the real license plate, that is, max(w, h), and then divide it by the length L of the first feature. If both max(w, h) / L > m j and max(w, h) / L < m j+1, then the image will be assigned to the j-th component of the anchor-free detection module. In the method of the present invention, j = 1, 2, 3. When j = 1, it represents the low-level component; when j = 2, it is the middle-level component; when j = 3, it is the high-level component. Correspondingly, m1 = 0, m2 = 0.25, m3 = 0.5, m4 = +∞. In order to assign different license plates to different components for learning, the longest side of the bounding rectangle of the license plate in the collected image is used to assign to different components in the anchor-free detection module. Correspondingly, if the license plate in an image is assigned to the j-th component, then the three points closest to the center point of the license plate are regarded as positive sample points.

[0044] Preferably, the present invention uses an anchor-free detection module to predict and locate the approximate positions of license plates of different sizes. After being processed by the anchor-free detection module, each component outputs a vector F of dimension H×W×5, where H and W are the length and width of the output vector of this layer component respectively, that is, F ∈ R H×W×5 : The position P = (P x , P y , P w , P h ) and confidence S of the predicted license plate can be calculated through the vector F. The following five results are calculated:

[0045] P x = 2×(σ(F(0)) - 0.5)

[0046] P y = 2×(σ(F(1)) - 0.5)

[0047] P w = σ(F(2))×W / 8

[0048] P h = σ(F(3))×H / 8

[0049] S = σ(F (4) )

[0050] Among them, σ represents the Sigmoid activation function; P x , P y are the positions of the center point of the bounding rectangle of the predicted license plate respectively, and P w , P h are the width and height of the bounding rectangle of the predicted license plate respectively; F (1) , F (2) , F (3) , F (4) , F (5) ∈ R H×W×1 are the first to fifth dimensional vectors of the output H×W×5 respectively.

[0051] Therefore, the predicted position of a small-sized license plate is obtained by calculating with the vector F output by the low-level component, the predicted position of a medium-sized license plate is obtained by calculating with the vector F output by the middle-level component, and the predicted position of a large-sized license plate is obtained by calculating with the vector F output by the high-level component.

[0052] The position of the real license plate in the acquired image and the position of the predicted license plate obtained through the foregoing calculation have both been acquired; next, a loss function is constructed to detect the difference between these two positions.

[0053] In the anchor-free detection module, this optimization objective function (or loss function) is set, so that the anchor-free detection module has better detection performance for the objects in the picture. The points within the license plate area in the picture are called positive sample points, and the optimization object of the optimization objective function is the positive sample points. By optimizing, if more positive sample points can be matched, it means that the object of the license plate existing in the picture can be detected better. Since the small license plate area on the picture is small in size, therefore, the number of positive sample points included in the small license plate area is relatively small. In the present invention, the three points closest to the center point of the small license plate are selected as positive samples. The optimization objective function (or loss function) enables the anchor-free detection module to predict the license plate position, especially the small license plate position, more accurately.

[0054] To optimize the detection effect, the loss function is constructed as follows:

[0055]

[0056] where k represents three different components in the anchor-free detection module, k takes 1, 2, 3, k = 1 represents the low-level component, k = 2 represents the middle-level component, and k = 3 represents the high-level component; S is the confidence level calculated previously, T is the sample category, if it is a positive sample, T = 1, otherwise T = 0; L cls represents the category loss, which is used to supervise the predicted confidence level S; G is the position of the real license plate in the acquired image, and P is the predicted position of the license plate; N pos represents the number of license plates in the acquired picture; L loc represents the position loss, which is used to supervise the predicted license plate position; λ is the weight coefficient, which is used to balance the magnitudes of the category loss and the position loss. Preferably, in the method of the present invention, λ takes the value of 0.3.

[0057] Preferably, in the embodiment of the present invention, the region expansion and feature selection module is connected to the coarse detection module. The region expansion and feature selection module processes the predicted position P of the license plate obtained by the coarse detection module, and then obtains the features of the corresponding region and sends them to the position correction module.

[0058] The region expansion and feature selection module expands the predicted position P of the license plate detected by the coarse detection module by 1 - 4 times along the longest side to obtain an expanded region. In this invention, 2 times is preferably used. Then, it crops out the cropping region corresponding to this expanded region from the first - layer feature map in the first feature, and removes other regions in the first - layer feature map, thus cutting off other regions that are non - license plates as the background, and scales this cropping region to 32 * 32 pixels. According to the above - mentioned processing method, all scaled cropping regions corresponding to the predicted position P of the license plate are obtained from the first - layer feature map, and then all the cropping regions are stitched together to obtain the second feature. In this process, the non - license plate regions as the background in the first - layer feature map are discarded. The region expansion and feature selection module takes the expanded region obtained by expanding the predicted position P of the license plate by 2 times as a reference to obtain the cropping region from the first - layer feature map of the first feature. Since the size of the first - layer feature map is basically the same as the collected image and contains low - dimensional information features, it is convenient for the subsequent acquisition of multi - direction license plates.

[0059] The position correction module. The input of the position correction module is the second feature output by the region expansion and feature selection module. After being processed by the foregoing modules, the license plate will occupy most of the region of the second feature, and the foreground (license plate) - background (non - license plate) imbalance problem is greatly alleviated. Since the second feature is selected and stitched from the first - layer feature map and contains only low - dimensional information features, it is necessary to extract the second feature again to obtain the third feature with higher dimensions. Then, the third feature is input into the multi - direction license plate detection module, and this multi - direction license plate detection module is a sub - module of the position correction module. After processing the third feature, this multi - direction license plate detection module obtains a 2X2X9 - dimensional vector M, that is, M ∈ R 2×2×9 .

[0060] Using this vector to calculate the position of the multi - direction license plate: P′=(p′ x1 , p′ y1 , p′ x2 , p′ y2 , p′ x3 , p′ y3 , p′ x4 , p′ y4 ) and confidence S′,

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] S’ = σ(M(8))

[0070] Wherein, respectively represent the coordinates of the four vertices of the multi-direction license plate; (M (0) , M (1) , M (2) , M (3) , M (4) , M (5) , M (6) , M (7) , M (8) ) ∈ R 2×2×9 .

[0071] An optimization objective function (loss function) is also set in the position correction module, and the formula is as follows:

[0072] loss LRM = L loc (G, P′) × T′ + λ′L cls (S′, T′)

[0073] Wherein, G is the defined true position of the license plate, and P′ is the position of the multi-direction license plate. T’ is the sample category. If it is a positive sample, T’ = 1, otherwise T’ = 0. In this module, the point closest to the center point of the license plate is regarded as the positive sample point; L loc is the position loss, which is used to supervise the learning of the four-point position; L cls is the classification loss, which is used to constrain the foreground confidence S′, and λ′ is the weight coefficient, which is used to balance the magnitudes of the classification loss and the position loss. In the method of the present invention, the value of λ′ is 1.

[0074] Preferably, the present invention also provides an optimization objective function for the entire invention: loss = lossloss CDM + λ1 * loss LRM , and λ1 is usually set to 0.02. The entire optimization objective function is used to optimize the entire invention simultaneously to find the best position of the multi-direction license plate.

[0075] The second feature in the present invention: It is cropped and selected from the first-layer feature map of the first feature, and only contains the features of the license plate and a small part of the background information features around it, but they are all low-dimensional visual information features.

[0076] The third feature: The high-dimensional features of the license plate extracted from the second feature.

[0077] The foregoing description has shown and described several preferred embodiments of the present invention. However, as previously mentioned, it should be understood that the present invention is not limited to the forms disclosed herein, and should not be regarded as excluding other embodiments. Instead, it can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept of the application described herein through the above teachings or the techniques or knowledge in the relevant field. Any alterations and changes made by those skilled in the art without departing from the spirit and scope of the present invention shall fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for detecting multi-directional license plates, characterized in that: The method comprises the following steps: S1. Extract features from the collected and processed image to obtain a first feature; detect the first feature to obtain the predicted position of the license plate. The first feature is a multi-layer feature map consisting of a first-layer feature map, a second-layer feature map, a third-layer feature map...an n-th-layer feature map, where n is greater than or equal to 2. The first-layer feature map includes the area closest to the shape of the license plate. The first feature includes high-level features of large license plates, middle-level features of medium license plates, and low-level features of small license plates, and the high-level features, middle-level features, and low-level features exist in each layer of feature map. The first feature mentioned by the feature extraction module of the multi-directional license plate detection device is sent to the anchor-free frame detection module of the multi-directional license plate detection device for detection. Row detection, the anchor-free frame detection module includes three components arranged in parallel: a low-level component, a middle-level component, and a high-level component. Among them, the input of the low-level component is the low-level feature in the first feature; the input of the middle-level component is the middle-level feature in the first feature; the input of the high-level component is the high-level feature in the first feature. The outputs of the three components are all predicted license plate positions. The three components analyze and predict the positions of the features of different layers in the first feature. The predicted position of the small-sized license plate is calculated by the vector F output by the low-level component, the predicted position of the medium-sized license plate is calculated by the vector F output by the middle-level component, and the predicted position of the large-sized license plate is calculated by the vector F output by the high-level component. S2. Expand the predicted position of the license plate to obtain an expanded area; S3. Using the expanded area as a reference, processing the first feature to obtain a second feature, specifically comprising: S31. expanding the predicted position of the license plate outward by 1 to 4 times to obtain an expanded area, then cropping a cropped area corresponding to the expanded area from the first-layer feature map, and scaling the cropped area; S32. Repeat step S31 to obtain all scaled cropped areas corresponding to the predicted license plate positions from the first-layer feature map of the first feature, and concatenate all cropped areas to obtain the second feature; S4. Correct and re-detect the second feature to obtain the positions of the multi-directional license plates.

2. The multi-directional license plate detection method according to claim 1, characterized in that: The collected and processed picture is obtained by scaling and normalizing the collected picture according to a predetermined size.

3. The multi-directional license plate detection method according to claim 1, characterized in that: The S4 specifically includes: extracting the second feature to obtain a third feature, processing the third feature to obtain a 2X2X9-dimensional vector, and calculating the vector to obtain the position of the multi-directional license plate.

4. A multi-directional license plate detection device, characterized in that: The device includes a coarse detection module, a region expansion and feature selection module and a position correction module, wherein the coarse detection module includes a feature extraction module and an anchor-free frame detection module that are interconnected, and the feature extraction module is used to extract features from the collected and processed image to obtain a first feature, wherein the first feature is a multi-layer feature map consisting of a first-layer feature map, a second-layer feature map, a third-layer feature map...an n-th-layer feature map, wherein n is greater than or equal to 2, the first-layer feature map includes an area closest to the shape of the license plate, the first feature includes high-layer features of large license plates, middle-layer features of medium license plates and low-layer features of small license plates, and the high-layer features, middle-layer features and low-layer features are all present in each layer of feature map, the first feature obtained by the feature extraction module is sent to the anchor-free frame detection module for detection, and the anchor-free frame detection ... The detection module includes three components arranged in parallel: a low-level component, a middle-level component and a high-level component, wherein the input of the low-level component is the low-level feature in the first feature; the input of the middle-level component is the middle-level feature in the first feature; the input of the high-level component is the high-level feature in the first feature, and the outputs of the three components are all predicted license plate positions. The three components analyze and predict positions based on features of different layers in the first feature. The predicted position of a small-sized license plate is obtained by calculating the vector F output by the low-level component, the predicted position of a medium-sized license plate is obtained by calculating the vector F output by the middle-level component, and the predicted position of a large-sized license plate is obtained by calculating the vector F output by the high-level component; the first feature is detected to obtain the predicted position of the license plate, and the predicted position of the license plate is transmitted to the region expansion and feature selection module; The region expansion and feature selection module expands the received predicted position of the license plate to obtain an expanded region, uses the expanded region as a reference, processes the first feature to obtain a second feature, and transmits the second feature to the position correction module, specifically comprising: expanding the predicted position of the license plate outward by 1 to 4 times to obtain an expanded region, then cropping a cropping region corresponding to the expanded region from the first-layer feature map, and scaling the cropping region; repeatedly processing the first-layer feature map of the first feature to obtain all scaled cropping regions corresponding to the predicted position of the license plate, and splicing all cropping regions to obtain the second feature; The position correction module corrects and re-detects the second feature to obtain the position of the multi-directional license plate.

5. The multi-directional license plate detection device according to claim 4, characterized in that: The collected and processed picture is obtained by scaling and normalizing the collected picture according to a predetermined size.

6. The multi-directional license plate detection device according to claim 4, characterized in that: The coarse detection module is provided with an optimization objective function for detecting the difference between the position of the real license plate in the image and the predicted position of the license plate; the position correction module is also provided with an optimization objective function for detecting the difference between the position of the real license plate in the image and the position of the multi-directional license plate.

7. The multi-directional license plate detection device according to claim 4, characterized in that: The position correction module includes a multi-directional license plate detection module, wherein the position correction module extracts the second feature to obtain a third feature, inputs the third feature into the multi-directional license plate detection module for processing to obtain a 2X2X9-dimensional vector, and calculates the vector to obtain the position of the multi-directional license plate.

8. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the method for detecting multi-directional license plates according to any one of claims 1 to 3 is implemented.

Images