Virtual license plate generation method and system based on generative adversarial network and storage medium

Abstract

The application discloses a virtual license plate generation method and system based on a generative adversarial network and a storage medium. The method comprises the following steps: generating an initial virtual license plate according to a license plate combination strategy, and automatically labeling the initial virtual license plate; performing multi-thread data enhancement on the labeled initial virtual license plate; classifying the initial virtual license plate after the multi-thread data enhancement, and performing style transfer processing on the generative adversarial network of a corresponding category according to a classification result to obtain a stylized license plate; if the stylized license plate meets a preset condition, outputting the stylized license plate as a final virtual license plate; otherwise, performing targeted secondary data enhancement processing on the stylized license plate in a traffic scene, and outputting the final virtual license plate after the secondary data enhancement processing. The application can quickly generate a virtual license plate with a realistic and real license plate effect.

Description

Technical Field

[0001] This invention belongs to the field of license plate generation and recognition technology, specifically relating to a virtual license plate generation method, system, and storage medium based on generative adversarial networks. Background Technology

[0002] With the rapid development of the economy and society, the number of motor vehicles is constantly increasing. To improve vehicle management efficiency and reduce labor costs, accurate license plate recognition plays a crucial role. Current research and practice show that the performance of deep learning algorithms depends on the quality and scale of the training data; therefore, deep learning-based license plate recognition algorithms require a large amount of high-quality data. However, collecting and labeling high-quality data requires significant time, money, and manpower. Furthermore, there are many types of license plates, including small vehicle plates, large vehicle plates, large double-layer plates, new energy vehicle plates, police plates, military plates, and double-layer military plates. Relatively rare license plates, such as those from remote areas, further increase the difficulty of sample collection. More importantly, data collection is even more challenging in complex scenarios such as excessive darkness, overexposure, large tilt angles, motion blur, low resolution, occlusion, and extreme weather conditions, yet these scenarios are crucial sources for license plate recognition algorithms. Additionally, the collected and labeled license plate data raises privacy and legal issues. To address these problems, effective virtual license plates can save significant human and material resources.

[0003] To reduce costs, existing technologies use computer graphics scripts to synthesize license plate images according to font, color, and composition rules. However, these synthesized images are not as accurate as real photographs. License plate recognition models trained using this approach do not perform well on real data, and the models are prone to learning incorrect knowledge. Early work also used computer image technology to manually synthesize license plate images according to rules regarding color, font, and character content. During the synthesis process, common data augmentation techniques were added, such as random cropping, scaling, rotation, random brightness, contrast, chroma, and saturation, and adding dirt or occlusion to make the synthesized image more closely resemble a real license plate image. However, these manually synthesized license plate images struggle to simulate the many details of real-world license plate images, resulting in overly simplistic image quality. This leads to the recognition model learning incorrect knowledge, failing to uncover the essential features of the license plate image data, thus reducing the model's generalization ability and resulting in poor performance on real data. Summary of the Invention

[0004] One of the objectives of this invention is to provide a method for generating virtual license plates based on generative adversarial networks, which can quickly generate virtual license plates that look like real license plates.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for generating virtual license plates based on generative adversarial networks (GANs), the method comprising:

[0007] An initial virtual license plate is generated based on a license plate combination strategy, and the initial virtual license plate is automatically labeled.

[0008] Perform multi-threaded data augmentation on the labeled initial virtual license plates;

[0009] The initial virtual license plates after multi-threaded data augmentation are classified, and the corresponding generative adversarial network is selected for style transfer processing based on the classification results to obtain stylized license plates.

[0010] If the stylized license plate meets the preset conditions, the stylized license plate is output as the final virtual license plate; otherwise, a traffic scene is selected to perform targeted secondary data augmentation on the stylized license plate, and the final virtual license plate is output after the secondary data augmentation.

[0011] Several alternative methods are provided below, but they are not intended as additional limitations on the overall solution above. They are merely further additions or optimizations. Provided there are no technical or logical contradictions, each alternative method can be combined individually with respect to the overall solution above, or multiple alternative methods can be combined with each other.

[0012] Preferably, the license plate combination strategy is a combination of province, city, and license plate type.

[0013] Preferably, the step of generating the initial virtual license plate according to the license plate combination strategy includes:

[0014] The license plate base and font are determined according to the license plate type;

[0015] Based on the distribution rules of license plate fonts, license plate numbers corresponding to the province and city are generated on the license plate base plate.

[0016] Preferably, the multi-threaded data enhancement includes one or more of the following: perspective transformation, affine transformation, license plate deformation, motion blur, weather transformation, edge cropping, overexposure, low contrast processing, yin-yang license plate processing, smudge processing, and partial occlusion processing.

[0017] Preferably, the multi-threaded data augmentation of the labeled initial virtual license plate includes performing different combinations of data augmentation processing on all or part of the labeled initial virtual license plate.

[0018] Preferably, the initial virtual license plates after multi-threaded data augmentation are classified into six types based on their background color: blue, black, yellow, white, green, and yellow-green.

[0019] Preferably, the step of selecting a generative adversarial network corresponding to the classification result for style transfer processing includes:

[0020] The DSMAP algorithm is used for style transfer processing of initial virtual license plates of type blue, green and black. The AdaIN algorithm is used for style transfer processing of initial virtual license plates of type white. The SIN algorithm is used for style transfer processing of initial virtual license plates of type yellow and yellow-green.

[0021] Preferably, in the stylization transfer process, the initial virtual license plate is used as the content image, and the real license plate photo is used as the style image.

[0022] This invention provides a virtual license plate generation method based on generative adversarial networks (GANs). It performs personalized, hierarchical style transfer based on various license plate types to achieve a more realistic license plate effect. In the first step of generating the virtual license plate, this method considers all types of license plates, essentially covering all license plate types. Then, it performs image enhancement and GAN style transfer based on the license plate type and rare actual scenarios, such as yin-yang license plates, damaged license plates, and obscured license plates. Image enhancement is performed first, followed by style transfer, and then further image enhancement can be performed based on the effect to meet different scenario requirements, making the generated virtual license plate closely resemble a real license plate.

[0023] The second objective of this invention is to provide a virtual license plate generation system based on generative adversarial networks, which can quickly generate virtual license plates that look like real license plates.

[0024] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0025] A virtual license plate generation system based on generative adversarial networks, the virtual license plate generation system based on generative adversarial networks comprising:

[0026] The license plate generation and labeling module is used to generate initial virtual license plates according to the license plate combination strategy and automatically label the initial virtual license plates;

[0027] The multi-threaded data augmentation module is used to perform multi-threaded data augmentation on the labeled initial virtual license plates.

[0028] The stylization transfer module is used to classify the initial virtual license plates after multi-threaded data augmentation, and select the corresponding generative adversarial network for style transfer processing based on the classification results to obtain stylized license plates.

[0029] The secondary data augmentation module is used to determine whether the stylized license plate meets the preset conditions. If so, the stylized license plate is output as the final virtual license plate. Otherwise, a traffic scene is selected to perform targeted secondary data augmentation on the stylized license plate, and the final virtual license plate is output after the secondary data augmentation.

[0030] This invention provides a virtual license plate generation system based on generative adversarial networks (GANs). It performs personalized, hierarchical style transfer based on various license plate types to achieve a more realistic license plate effect. In the first step of generating the virtual license plate, this solution considers all types of license plates, essentially covering all license plate types. Then, it performs image enhancement and GAN style transfer based on the license plate type and rare actual scenarios, such as yin-yang license plates, damaged license plates, and obscured license plates. Image enhancement is performed first, followed by style transfer, and then a second image enhancement can be performed based on the effect to meet the needs of different scenarios, making the generated virtual license plate closely resemble a real license plate.

[0031] The third objective of this invention is to provide a computer-readable storage medium that, upon execution, quickly generates a virtual license plate that looks realistic.

[0032] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0033] A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to execute a virtual license plate generation method based on a generative adversarial network. Attached Figure Description

[0034] Figure 1 This is a flowchart of the virtual license plate generation method based on generative adversarial networks of the present invention;

[0035] Figure 2 These are schematic diagrams of 35 different license plates generated by this invention;

[0036] Figure 3 This is a schematic diagram illustrating the stylization transfer processing of this invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0039] To address the issue that existing license plate classifications are incomplete, leading to inconsistent license plate generation, this embodiment provides a virtual license plate generation method based on generative adversarial networks to solve the aforementioned problem.

[0040] like Figure 1 As shown, the virtual license plate generation method based on generative adversarial networks in this embodiment includes:

[0041] Step 1: Generate initial virtual license plates according to the license plate combination strategy, and automatically label the initial virtual license plates.

[0042] Through research and analysis, existing Chinese license plates are classified into 35 categories. Different types of license plates have differences in font, background color, texture, and size. Therefore, more personalized and intelligent learning is particularly important during stylistic transfer. Previous technologies have two problems: first, the classification of license plates is not comprehensive enough; second, all license plates are learned in the same way, resulting in inconsistent results. This dirty data will greatly affect the accuracy of license plate recognition.

[0043] Therefore, in this embodiment, when generating the initial virtual license plate, different license plate combination strategies can generate license plates such as... Figure 2 The 35 different types of license plates shown basically cover all existing types of license plates, solving the problem of insufficient data on rare license plates. It should be noted that... Figure 2 There are multiple background colors for license plates and different font colors for license plate numbers. Due to color limitations in the attached image, these cannot be shown, but they correspond to actual license plates.

[0044] The license plate combination strategy involves combining province, city, and license plate type. It allows for customization to synthesize data strategies with multiple provinces, license plate types, and city-level license plates, ensuring the generation of license plate data for various tasks. When generating initial virtual license plates, one or more provinces, cities, and license plate types can be selected to maximize customization and meet dataset requirements.

[0045] Of course, for greater flexibility, in another embodiment, the license plate combination strategy can also be a combination of one or more of the following: province, city, license plate type, and specific content. For example, it can generate a license plate from a specified province, or generate a license plate from a specified city, or generate a license plate of a specified type, or generate a license plate with specific content. The specific content can be letters or numbers that must be included in the license plate number. The letters or numbers can be part or all of the license plate number. For example, if the specific content is 3, then a license plate with the number 3 in the license plate number will be generated.

[0046] In this embodiment, the initial virtual license plate is generated by selecting the license plate background and corresponding license plate font based on the license plate type, and then generating the license plate number corresponding to the province and city level on the license plate background according to the license plate font distribution rules. Specifically: the license plate background and license plate font are determined according to the license plate type; and license plate numbers corresponding to the province and city level are generated on the license plate background according to the license plate font distribution rules.

[0047] Since each type of license plate corresponds to a specific license plate background and font, in this embodiment, the license plate background and font are determined automatically by the program based on the license plate type, i.e., selected according to existing rules. Furthermore, to facilitate subsequent use of sample data, this embodiment also includes automatic license plate labeling when automatically generating virtual license plates, using the license plate number as a label to tag each initial virtual license plate.

[0048] Step 2: Perform multi-threaded data augmentation on the initial virtual license plates after labeling.

[0049] This embodiment employs various data augmentation methods, including perspective transformation, affine transformation, license plate deformation, motion blur, weather transformation, edge cropping, overexposure correction, low contrast correction, yin-yang license plate correction, defacement correction, and partial occlusion correction. One or more of these methods can be selected for multi-threaded processing. Multi-threaded processing ensures efficient license plate generation, achieving a speed of hundreds of thousands per hour. Data augmentation aims to reduce the data distribution differences between virtual and real license plates, thereby ensuring better style transfer performance of the GAN (Generative Adversarial Network).

[0050] Specifically, when performing multi-threaded data augmentation on the labeled initial virtual license plates, different combinations of data augmentation processing are performed on all or part of the labeled initial virtual license plates according to the user's needs. In this embodiment, in order to balance the number of virtual license plates after data augmentation, different combinations of data augmentation processing methods can be selected to be performed on all or part of the initial virtual license plates.

[0051] For example, if the expected data volume for yin-yang license plates and damaged license plates is 1:0.5, then when processing yin-yang license plates, all initial virtual license plates can be selected for processing, and when performing lossless processing, half of the initial virtual license plates can be selected for processing.

[0052] Step 3: Take the initial virtual license plates after multi-threaded data augmentation, classify them, and select the corresponding generative adversarial network for style transfer processing based on the classification results to obtain stylized license plates.

[0053] Different GAN networks have varying learning abilities for different textures. In this embodiment, 35 license plates are divided into six types based on their background color: blue, black, yellow, white, green, and yellow-green.

[0054] With the development of deep learning, many researchers have used GAN networks to further transfer the style of artificially synthesized license plates. However, a single GAN network often cannot guarantee that it will learn enough features for each type of license plate, resulting in serious distortion of the generated license plates.

[0055] like Figure 3 As shown, to improve the stylization effect, this embodiment uses the DSMAP algorithm for style transfer processing of initial virtual license plates of type blue, green, and black; the AdaIN algorithm for style transfer processing of initial virtual license plates of type white; and the SIN algorithm for style transfer processing of initial virtual license plates of type yellow and yellow-green. Furthermore, in the stylization transfer processing, the initial virtual license plate is used as the content image, and the real license plate photograph is used as the style image.

[0056] The common feature of the three style transfer algorithms is that they divide the loss function into two parts: a style loss function and a content loss function. However, the network design of each algorithm is very different.

[0057] The AdaIN algorithm uses the variance and mean of each channel of the feature map to represent style features. The content map and style map are integrated through the AdaIN layer. Since the content map, after being processed by virtual license plates and image enhancement, is already closer to a real license plate as input, the weight α of the style map is increased, as shown in the following formula:

[0058]

[0059] Where x is the content map, y is the style map, σ() represents the variance of the image, and μ() represents the mean of the image.

[0060] The DSMAP algorithm finds a mapping that maps content features in a shared latent space to a specific content space for each domain. This remapping of content features, combined with the style of the target domain, can achieve better results.

[0061] The SIN algorithm generates a pyramid from a single image, capturing statistical information within the image sequentially from coarse to fine and from large to small, generating high-quality samples with the same visual characteristics. Each GAN is responsible for the distribution information at different scales of the image, and the generated sample from the previous one will be used as the input of the next GAN network. Therefore, only one image is needed for training, which is very effective for scarce data.

[0062] Different style transfer algorithms are selected based on the license plate type, and various image enhancement methods are added before and after the GAN algorithm to realize various traffic scenarios, thereby generating more realistic virtual license plates.

[0063] Step 4: If the stylized license plate meets the preset conditions, output the stylized license plate as the final virtual license plate; otherwise, select a traffic scene to perform targeted secondary data augmentation on the stylized license plate, and output the final virtual license plate after secondary data augmentation.

[0064] After stylization, this embodiment can further construct traffic scenes according to actual task requirements, such as road sections with high frequency of sunlight changes or road sections with many damaged license plates, and perform image enhancement again to meet actual needs.

[0065] When determining whether a stylized license plate meets the preset conditions, it can be done by judging whether the degree of distortion of the stylized license plate is less than a threshold, or by judging whether the number of a certain category (e.g., yin-yang license plate) in the stylized license plate meets the threshold. That is, the preset conditions in this embodiment can be selected according to actual needs.

[0066] Furthermore, when constructing traffic scenarios, targeted traffic scenarios can be built based on the preset conditions set earlier. For example, if it is determined that there are not enough yin-yang license plates, road sections with high frequency of sunlight changes can be selected as traffic scenarios. During secondary data augmentation, stylized license plates can be processed to obtain a virtual license plate dataset that meets the requirements.

[0067] This solution can showcase different effects in four stages: First, the initial virtual license plate effect is displayed, achieving a preliminary simulation of the license plate font and color; second, the effect after multi-threaded data augmentation, including deformation, low light, edge cropping, different weather environments, noise, motion blur, overexposure, dirt, and uneven lighting causing uneven license plates, further simulating the environment of the license plate and reducing the distribution difference between the virtual and real license plates; third, the effect after style transfer, using deep learning methods to learn the style and content differences between the virtual and real license plates, thereby achieving a realistic license plate effect; fourth, the effect after secondary data augmentation according to requirements, resulting in a more complete and realistic virtual license plate in terms of content and effect.

[0068] This embodiment primarily employs personalized, layer-by-layer style transfer based on various license plate types to achieve a more realistic license plate effect. Previous solutions mainly simulated common license plate types and scenarios. However, the current limitation on license plate recognition accuracy lies mainly in special scenarios and rare license plates. Therefore, this solution considers 35 license plate types in the first step of generating the virtual license plate, essentially covering all license plate types. Then, image enhancement and GAN style transfer are performed based on the license plate type and actual rare scenarios, such as yin-yang license plates, damaged license plates, and obscured license plates. Image enhancement is performed first, followed by style transfer, and then further image enhancement can be performed based on the results to meet the needs of different scenarios.

[0069] Testing revealed that, using the same license plate recognition algorithm, the accuracy rate of the original real data (500,000 records) on a 16,000 test set was 98.9%; the accuracy rate of the original data (500,000 records) plus the 500,000 records generated by this solution on the 16,000 test set was 99.9%. Analysis showed that this embodiment primarily addresses the issues of rare provinces and special scenarios previously encountered.

[0070] In another embodiment, a virtual license plate generation system based on generative adversarial networks is also provided, comprising:

[0071] The license plate generation and labeling module is used to generate initial virtual license plates according to the license plate combination strategy and automatically label the initial virtual license plates;

[0072] The multi-threaded data augmentation module is used to perform multi-threaded data augmentation on the labeled initial virtual license plates.

[0073] The stylization transfer module is used to classify the initial virtual license plates after multi-threaded data augmentation, and select the corresponding generative adversarial network for style transfer processing based on the classification results to obtain stylized license plates.

[0074] The secondary data augmentation module is used to determine whether the stylized license plate meets the preset conditions. If so, the stylized license plate is output as the final virtual license plate. Otherwise, a traffic scene is selected to perform targeted secondary data augmentation on the stylized license plate, and the final virtual license plate is output after the secondary data augmentation.

[0075] Specific limitations regarding the virtual license plate generation system based on generative adversarial networks (GANs) can be found in the limitations of the virtual license plate generation method based on GANs mentioned above, and will not be repeated here. Each module in the aforementioned virtual license plate generation system based on GANs can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0076] In another embodiment, a computer-readable storage medium is provided, the computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the above-described virtual license plate generation method based on generative adversarial networks.

[0077] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0078] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0079] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for generating a virtual license plate based on a generative adversarial network, characterized in that, The virtual license plate generation method based on generative adversarial networks includes: An initial virtual license plate is generated based on a license plate combination strategy, and the initial virtual license plate is automatically labeled. Perform multi-threaded data augmentation on the labeled initial virtual license plates; The initial virtual license plates after multi-threaded data augmentation are classified, and the corresponding generative adversarial network is selected for style transfer processing based on the classification results to obtain stylized license plates. If the stylized license plate meets the preset conditions, the stylized license plate is output as the final virtual license plate. The determination of whether the stylized license plate meets the preset conditions is to determine whether the number of preset categories in the stylized license plate meets the threshold. Otherwise, the stylized license plate is subjected to targeted secondary data augmentation processing based on the traffic scene constructed according to the preset conditions. After the secondary data augmentation processing, the final virtual license plate is output. The classification of the initial virtual license plates after multi-threaded data augmentation includes: classifying the initial virtual license plates into six types based on their background color: blue, black, yellow, white, green, and yellow-green. The step of selecting the corresponding category of generative adversarial network for style transfer processing based on the classification result includes: using the DSMAP algorithm for style transfer processing of initial virtual license plates of type blue, green and black, using the AdaIN algorithm for style transfer processing of initial virtual license plates of type white, and using the SIN algorithm for style transfer processing of initial virtual license plates of type yellow and yellow-green. In the stylization transfer process, the initial virtual license plate is used as the content map, and the real license plate photo is used as the style map. The AdaIN algorithm uses the following formula: wherein, is the output of the AdaIN algorithm, is the content image, is the style image, denotes the variance of an image, denotes the mean of an image, is the weight of the style image. 2.The virtual license plate generation method based on a generative adversarial network according to claim 1, wherein, The license plate combination strategy is a combination of province, city, and license plate type. 3.The virtual license plate generation method based on a generative adversarial network according to claim 2, wherein, The process of generating an initial virtual license plate based on a license plate combination strategy includes: The license plate base and font are determined according to the license plate type; Based on the distribution rules of license plate fonts, license plate numbers corresponding to the province and city are generated on the license plate base plate.

4. The virtual license plate generation method based on generative adversarial networks as described in claim 1, characterized in that, The multi-threaded data enhancement includes one or more of the following: perspective transformation, affine transformation, license plate deformation, motion blur, weather transformation, edge cropping, overexposure, low contrast processing, yin-yang license plate processing, smudge processing, and partial occlusion processing.

5. The virtual license plate generation method based on generative adversarial networks as described in claim 4, characterized in that, The multi-threaded data augmentation of the labeled initial virtual license plate includes performing different combinations of data augmentation processing on all or part of the labeled initial virtual license plate.

6. A virtual license plate generation system based on generative adversarial networks, characterized in that, The virtual license plate generation system based on generative adversarial networks includes: The license plate generation and labeling module is used to generate initial virtual license plates according to the license plate combination strategy and automatically label the initial virtual license plates; The multi-threaded data augmentation module is used to perform multi-threaded data augmentation on the labeled initial virtual license plates. The stylization transfer module is used to classify the initial virtual license plates after multi-threaded data augmentation, and select the corresponding generative adversarial network for style transfer processing based on the classification results to obtain stylized license plates. The classification of the initial virtual license plates after multi-threaded data augmentation includes: classifying the initial virtual license plates into six types based on their background color: blue, black, yellow, white, green, and yellow-green. The selection of the corresponding generative adversarial network for style transfer processing based on the classification results includes: using the DSMAP algorithm for initial virtual license plates of type blue, green, and black; using the AdaIN algorithm for initial virtual license plates of type white; and using the SIN algorithm for initial virtual license plates of type yellow and yellow-green. In the stylization transfer process, the initial virtual license plate is used as the content map, and the real license plate photo is used as the style map. The AdaIN algorithm uses the following formula: in, This is the output of the AdaIN algorithm. For content images, For style illustrations, Represents the variance of the image. Represents the mean of the image. The weights of the style map; The secondary data augmentation module is used to determine whether the stylized license plate meets the preset conditions. If so, the stylized license plate is output as the final virtual license plate. The determination of whether the stylized license plate meets the preset conditions is to determine whether the number of preset categories in the stylized license plate meets the threshold. Otherwise, the stylized license plate is subjected to targeted secondary data augmentation processing based on the traffic scene constructed according to the preset conditions. After the secondary data augmentation processing, the final virtual license plate is output.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method of any one of claims 1-5.

Images