Training method and system of license plate recognition model

By adjusting the image and pixel values ​​of the license plate recognition model and optimizing the gradient descent, the problem of the reduction in recognition accuracy in actual applications of the traditional license plate recognition model is solved, and higher recognition accuracy and adaptability are achieved.

CN119027958AInactive Publication Date: 2024-11-26SHENZHEN GUARDSHEN TECH CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411063787.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-26
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

In actual application of traditional license plate recognition models, due to factors such as lighting conditions, angle changes and occlusion, the recognition accuracy decreases. Due to the diversity and complexity of the data set, the fine-tuning process is easily affected by data deviations and model overfitting. .

Method used

By retrieving license plate recognition model and related training information from the database, retraining and adjustments are combined with new image data, including image adjustment processing, pixel value adjustment and gradient descent optimization until the recognition accuracy reaches the preset threshold.

Benefits of technology

The recognition accuracy and adaptability of the license plate recognition model are improved, the accuracy and stability in practical applications are ensured, and performance degradation caused by time or environmental changes are avoided.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027958A_ABST
    Figure CN119027958A_ABST
Patent Text Reader

Abstract

The invention provides a license plate recognition model training method and system, and the method comprises the steps: calling a license plate recognition model from a database, and obtaining the training information of the license plate recognition model; calling image data corresponding to the current license plate type from a database, and performing image adjustment processing on the image data to obtain adjusted image data; and taking the adjusted image data as sample data to train the license plate recognition model until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold value, calling the license plate recognition model and related training information from a database, and performing retraining and adjustment in combination with new image data. The accuracy and adaptability of license plate recognition can be further improved, and the performance and effect in practical application are ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention proposes a training method and system for a license plate recognition model. The invention relates to the technical field of model training, and in particular to the training of a license plate recognition model. Background Art

[0002] With the development of intelligent transportation and big data technology, license plate recognition technology has been widely used in traffic management, intelligent transportation systems, vehicle safety monitoring and other fields. The license plate recognition model is one of the key technologies to achieve automatic license plate recognition, and its performance and accuracy directly affect the operation effect of the entire system. Traditional license plate recognition models are usually trained and optimized on fixed data sets, but in actual applications, due to factors such as lighting conditions, angle changes, and occlusion, the quality and clarity of license plate images will be greatly affected, resulting in a decrease in recognition accuracy. In actual applications, due to the diversity and complexity of data sets, the fine-tuning process may be affected by problems such as data bias and model overfitting. Summary of the invention

[0003] The present invention provides a training method and system for a license plate recognition model to solve the problem that traditional license plate recognition models are usually trained and optimized on a fixed data set, but in actual applications, due to factors such as lighting conditions, angle changes, and occlusion, the quality and clarity of the license plate image will be greatly affected, resulting in a decrease in recognition accuracy. In actual applications, due to the diversity and complexity of the data set, the fine-tuning process may be affected by data bias, model overfitting, and other issues:

[0004] The present invention provides a method and system for training a license plate recognition model, wherein the method for training the license plate recognition model comprises:

[0005] Retrieve the license plate recognition model from the database to obtain the training information of the license plate recognition model;

[0006] Retrieving image data corresponding to the current license plate type from a database, and performing image adjustment processing on the image data to obtain adjusted image data;

[0007] The license plate recognition model is trained using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold.

[0008] Furthermore, the license plate recognition model is retrieved from the database to obtain the training information of the license plate recognition model, including:

[0009] Retrieving a license plate recognition model and historical training records corresponding to the license plate recognition model from a database;

[0010] Obtaining the last license plate recognition model training, i.e., the optimized recognition accuracy, from the historical training record corresponding to the license plate recognition model;

[0011] The recognition accuracy rate after the last license plate recognition model training, i.e. optimization, is used as the recognition accuracy rate threshold.

[0012] Further, the image data corresponding to the current license plate type is retrieved from the database, and image adjustment processing is performed on the image data to obtain adjusted image data, including:

[0013] Retrieving image data corresponding to the current license plate type from the database as original image data;

[0014] Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data;

[0015] Using a pixel value adjustment model to adjust the pixel values ​​of the original image data to obtain adjusted pixel values;

[0016] Obtaining a clarity value corresponding to the adjusted pixel value, and determining whether the clarity value is lower than the clarity threshold;

[0017] If the clarity value corresponding to the adjusted pixel value is not lower than the clarity threshold, continue to use the pixel value adjustment model to adjust the pixel value of the original image data until the clarity value is lower than the clarity threshold, and use the image data corresponding to the pixel value whose clarity value is lower than the clarity threshold as the adjusted image data.

[0018] Furthermore, extracting pixel values ​​of the original image data and using the pixel values ​​to obtain a definition threshold corresponding to the original image data includes:

[0019] Extracting pixel values ​​of a plurality of the original image data;

[0020] The pixel deviation value corresponding to each original image data is set using the pixel value of the original image data.

[0021] The pixel deviation value corresponding to each original image data is combined with the pixel value of the original image data to set a plurality of clarity thresholds corresponding to the original image data as a whole.

[0022] Further, the license plate recognition model is trained using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold, including:

[0023] Performing preliminary training on the license plate recognition model using the original image data, and obtaining a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold;

[0024] When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model;

[0025] Determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, determine that the license plate recognition model has completed training;

[0026] When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent.

[0027] Furthermore, the training system of the license plate recognition model includes:

[0028] An information acquisition module is used to retrieve the license plate recognition model from the database and obtain the training information of the license plate recognition model;

[0029] An image processing module is used to retrieve image data corresponding to the current license plate type from a database, and perform image adjustment processing on the image data to obtain adjusted image data;

[0030] The model adjustment module is used to train the license plate recognition model using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold.

[0031] Furthermore, the information acquisition module includes:

[0032] A training module, used to retrieve a license plate recognition model and historical training records corresponding to the license plate recognition model from a database;

[0033] An accuracy acquisition module, used to obtain the recognition accuracy of the last license plate recognition model training, i.e., after optimization, from the historical training records corresponding to the license plate recognition model;

[0034] The recognition accuracy rate after the last license plate recognition model training, i.e. optimization, is used as the recognition accuracy rate threshold.

[0035] Furthermore, the information acquisition module includes:

[0036] A clarity threshold acquisition module is used to retrieve image data corresponding to the current license plate type from the database as original image data;

[0037] Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data;

[0038] A pixel adjustment module, used to adjust the pixel value of the original image data using a pixel value adjustment model to obtain the adjusted pixel value;

[0039] Obtaining a clarity value corresponding to the adjusted pixel value, and determining whether the clarity value is lower than the clarity threshold;

[0040] If the clarity value corresponding to the adjusted pixel value is not lower than the clarity threshold, continue to use the pixel value adjustment model to adjust the pixel value of the original image data until the clarity value is lower than the clarity threshold, and use the image data corresponding to the pixel value whose clarity value is lower than the clarity threshold as the adjusted image data.

[0041] Furthermore, the clarity threshold acquisition module includes:

[0042] A pixel deviation value acquisition module, used for extracting pixel values ​​of a plurality of the original image data;

[0043] The pixel deviation value corresponding to each original image data is set using the pixel value of the original image data.

[0044] The clarity threshold acquisition module is used to set a plurality of clarity thresholds corresponding to the original image data as a whole by using the pixel deviation value corresponding to each original image data in combination with the pixel value of the original image data.

[0045] Furthermore, the information acquisition module includes:

[0046] A comparison module, used to perform preliminary training on the license plate recognition model using the original image data, and obtain a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold;

[0047] When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model;

[0048] An adjustment module, used to determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than a preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, it is determined that the license plate recognition model has completed training;

[0049] When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent.

[0050] Beneficial effects of the present invention:

[0051] The present invention proposes a training method and system for a license plate recognition model. Through retraining and adjustment, the recognition accuracy of the license plate recognition model can be further improved, ensuring accuracy and stability in practical applications. By adjusting and processing different image data, the model can better adapt to different practical application scenarios. Through model retraining, it can be ensured that the model always maintains optimal performance and avoids performance degradation due to the passage of time or environmental changes. By retrieving the license plate recognition model and related training information from the database, and combining retraining and adjustment with new image data, the accuracy and adaptability of license plate recognition can be further improved, ensuring performance and effects in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the composition of a training method for a license plate recognition model. DETAILED DESCRIPTION

[0053] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0054] In one embodiment of the present invention, a method and system for training a license plate recognition model are provided. The method for training a license plate recognition model comprises:

[0055] Retrieve the license plate recognition model from the database to obtain the training information of the license plate recognition model;

[0056] Retrieving image data corresponding to the current license plate type from a database, and performing image adjustment processing on the image data to obtain adjusted image data;

[0057] The license plate recognition model is trained using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold.

[0058] The working principle of the above technical solution is as follows: First, retrieve the trained license plate recognition model from the database. This model usually contains the relevant algorithms and parameters for license plate recognition. At the same time, obtain the training information of the license plate recognition model from the database. This information may include training parameters, training data, and training results. According to the type of the current license plate, retrieve the corresponding image data from the database. Then, adjust the image data, use the adjusted image data as sample data, and retrain the license plate recognition model. Through continuous iteration and optimization, the recognition accuracy of the model is gradually improved. After each training, the recognition accuracy of the model will be evaluated. When the recognition accuracy exceeds the preset recognition accuracy threshold, stop training; otherwise, continue training until the conditions are met.

[0059] The effect of the above technical solution is: through retraining and adjustment, the recognition accuracy of the license plate recognition model can be further improved to ensure accuracy and stability in practical applications. By adjusting and processing different image data, the model can better adapt to different practical application scenarios. Through model retraining, it can be ensured that the model always maintains the best performance and avoids performance degradation caused by the passage of time or environmental changes. By retrieving the license plate recognition model and related training information from the database, and combining it with new image data for retraining and adjustment, the accuracy and adaptability of license plate recognition can be further improved to ensure performance and effect in practical applications.

[0060] In one embodiment of the present invention, the license plate recognition model is retrieved from a database to obtain training information of the license plate recognition model, including:

[0061] Retrieving a license plate recognition model and historical training records corresponding to the license plate recognition model from a database;

[0062] Obtaining the last license plate recognition model training, i.e., the optimized recognition accuracy, from the historical training record corresponding to the license plate recognition model;

[0063] The recognition accuracy rate after the last license plate recognition model training, i.e. optimization, is used as the recognition accuracy rate threshold.

[0064] The working principle of the above technical solution is as follows: first, retrieve the trained license plate recognition model and its corresponding historical training records from the database. The historical training records may contain information such as parameters, time, and recognition accuracy of previous trainings.

[0065] From the historical training records, find the optimized recognition accuracy rate after the last license plate recognition model training. This accuracy rate can be used as the recognition accuracy threshold for subsequent model training.

[0066] The effect of the above technical solution is that by using historical training records, the system can set a reasonable recognition accuracy threshold based on the last optimized training in the past. Such threshold setting is more reliable and avoids inaccurate or inappropriate thresholds caused by arbitrary settings. As time goes by and new data is added, the accuracy in the historical training records may change. By regularly checking and updating the recognition accuracy threshold, the system can adaptively adjust the threshold to ensure that it always matches the current performance of the model. If the new data is similar to the historical data, decisions based on historical data can reduce unnecessary model training, thereby saving computing resources and time. By referring to the best recognition accuracy in the historical training records, the system can set a relatively high threshold, thereby pushing the model towards higher recognition accuracy. In short, by using the historical training records of the license plate recognition model to determine the recognition accuracy threshold, the system can make decisions based on past experience to ensure that the training and adjustment of the model are always based on best practices. This helps to improve the accuracy of the model, save resources and ensure continuous optimization of the system.

[0067] In one embodiment of the present invention, image data corresponding to the current license plate type is retrieved from a database, and image adjustment processing is performed on the image data to obtain adjusted image data, including:

[0068] Retrieving image data corresponding to the current license plate type from the database as original image data;

[0069] Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data;

[0070] The pixel value adjustment model is used to adjust the pixel value of the original image data to obtain the adjusted pixel value; wherein the specific structure of the pixel value adjustment model is as follows:

[0071]

[0072] Among them, T zm is the pixel adjustment value, X ys is the original pixel value of the original image data, ΔC t To reduce the original image size by default, C y is the original image size, ΔF t To reduce the original image resolution, F y is the original image resolution, ΔL t To adjust the brightness of the original image, L y is the original image brightness, ΔR t To reduce the original image sharpness, R y is the original image sharpness, ΔB tTo lower the saturation of the original image, B y is the original image saturation, ΔS t To preset the original image color temperature, S y is the color temperature of the original image, α is the adjustment coefficient, and its value range is 0-2;

[0073] Obtaining a clarity value corresponding to the adjusted pixel value, and determining whether the clarity value is lower than the clarity threshold;

[0074] If the clarity value corresponding to the adjusted pixel value is not lower than the clarity threshold, continue to use the pixel value adjustment model to adjust the pixel value of the original image data until the clarity value is lower than the clarity threshold, and use the image data corresponding to the pixel value whose clarity value is lower than the clarity threshold as the adjusted image data.

[0075] The working principle of the above technical solution is as follows: the image data corresponding to the current license plate type is retrieved from the database as the original image data. Pixel values ​​are extracted from the original image data, and the clarity threshold of the original image data is calculated based on these pixel values. The pixel value adjustment model is used to adjust the pixel values ​​of the original image data. This involves scaling or other forms of adjustment of the pixel values. The adjusted pixel values ​​obtain the corresponding clarity values, which are then compared with the clarity threshold of the original image data. If the adjusted clarity value is not lower than the threshold, the pixel value adjustment model is continued to be used to adjust the original image data until the clarity value is lower than the threshold. The image data corresponding to the pixel values ​​whose clarity values ​​are lower than the threshold are used as the final adjusted image data.

[0076] The effect of the above technical solution is that by adjusting the pixel value, the clarity of the original image can be enhanced, making it easier to identify and analyze. The ratio of the change in size, resolution, brightness, sharpness, saturation and color temperature relative to the preset value can be calculated. The overall calculation formula can calculate the appropriate pixel value adjustment value combined with the adjustment coefficient considering multiple factors, which enhances the flexibility, comprehensiveness and accuracy of pixel adjustment. By calculating the clarity threshold based on the characteristics of the original image data, the system determines a suitable clarity standard through calculation to ensure that the clarity adjustment of the image is neither excessive nor too conservative. Since iterative adjustment is continued only when the adjusted clarity value is not lower than the threshold, unnecessary calculations can be reduced and resource utilization can be optimized. Through continuous pixel value adjustment, it can be ensured that the final output image data reaches a relatively appropriate level in clarity. In short, this method extracts the pixel value of the original image data, calculates the clarity threshold, uses the pixel value adjustment model to iteratively adjust, and finally outputs the image data that meets the clarity requirements, which helps to improve the accuracy and efficiency of license plate recognition. The pixel value adjustment model set by this method can enhance the recognition accuracy of the license plate recognition model and the recognition sensitivity to various image influencing factors.

[0077] In one embodiment of the present invention, extracting pixel values ​​of the original image data and using the pixel values ​​to obtain a definition threshold corresponding to the original image data includes:

[0078] The pixel deviation value corresponding to each original image data is set using the pixel value of the original image data.

[0079] The pixel deviation value corresponding to each original image data is combined with the pixel value of the original image data to set a plurality of clarity thresholds corresponding to the original image data as a whole.

[0080] The working principle of the above technical solution is:

[0081] Extracting pixel values ​​of a plurality of the original image data;

[0082] The pixel deviation value corresponding to each original image data is set using the pixel value of the original image data. The pixel deviation value corresponding to the original image data is obtained by the following formula:

[0083]

[0084] Among them, P czi is the pixel deviation value corresponding to the i-th original image data, j is the number of original image data, Z S is the recognition accuracy threshold, X ZVi is the recognition accuracy corresponding to the pixel value of the i-th original image data, Y XS is the preset pixel value, X Si is the pixel value of the i-th original image data, U XS is the average pixel value of all original image data, Z XSis the pixel value corresponding to the recognition accuracy threshold;

[0085] The pixel deviation value corresponding to each original image data is combined with the pixel value of the original image data to set a plurality of clarity thresholds corresponding to the original image data as a whole. The clarity threshold is obtained by the following formula:

[0086]

[0087] Among them, Q yz is the clarity threshold, q y is the initial preset definition, j is the number of original image data, X Si is the pixel value of the i-th original image data, P czi is the pixel deviation value corresponding to the i-th original image data.

[0088] The effect of the above technical solution is: extract multiple pixel values ​​from the original image data. For each original image data, calculate the deviation of its pixel value. This can be based on the difference between pixels, the deviation of pixel values ​​from the average value, or other statistical characteristics related to pixels. Based on the calculated pixel deviation, set a corresponding pixel deviation value for each original image data, and calculate the pixel deviation value by the formula in the pixel deviation value calculation formula. The ratio of the difference between the pixel value corresponding to the preset pixel value, the average pixel value and the recognition accuracy threshold and the actual pixel value to the corresponding total pixel value can be calculated. It can be used to set a more suitable clarity. The pixel deviation value corresponding to each original image data is combined with the pixel value of the original image data, and the clarity threshold corresponding to each original image data as a whole is calculated by calculating the combined data through the clarity threshold. The total percentage of the difference between the single pixel value of all original image data and the corresponding pixel deviation value compared to the total number can be calculated, and the overall calculation formula can provide an accurate basis for the subsequent image clarity evaluation. Through the above technical effects, this method can provide more accurate pixel deviation values ​​and clarity thresholds for the original image data, thereby providing a basis for subsequent image adjustment and optimization. The clarity threshold is calculated and set based on the data calculated by the pixel deviation value, so that the clarity threshold setting takes into account the factor of pixel deviation, making the clarity threshold more adaptable to the license plate recognition type, greatly enhancing the rationality of the clarity threshold design.

[0089] In one embodiment of the present invention, the license plate recognition model is trained using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold, including:

[0090] Performing preliminary training on the license plate recognition model using the original image data, and obtaining a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold;

[0091] When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model;

[0092] Determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, determine that the license plate recognition model has completed training;

[0093] When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent.

[0094] The working principle of the above technical solution is: use the original image data to perform preliminary training on the license plate recognition model, and obtain a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold;

[0095] When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model;

[0096] Determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, determine that the license plate recognition model has completed training;

[0097] When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent. The adjusted gradient is obtained by the following formula:

[0098]

[0099] Among them, T dtx is the adjusted gradient, T cs is the initial gradient value of the license plate recognition model, M is the true value of the license plate recognition model, is the recognition value of the license plate recognition model, is the license plate recognition model parameter, V ss is the actual convergence speed of the license plate recognition model, Vys is the preset convergence speed of the license plate recognition model, X ss is the actual learning speed of the license plate recognition model, X ys is the preset learning speed of the license plate recognition model, W sx is the actual training error data of the license plate recognition model, W yx The preset training error data for the license plate recognition model.

[0100] The effect of the above technical solution is: use the original image data to perform preliminary training on the license plate recognition model. Through training, obtain the preliminary training results of the license plate recognition model and its corresponding recognition accuracy. According to the preliminary training results, calculate and record the first recognition accuracy. If the first recognition accuracy exceeds the preset recognition accuracy threshold, enter the next stage; otherwise, it may be necessary to further adjust or retrain the model. When the first recognition accuracy exceeds the threshold, use the adjusted image data to perform secondary training on the license plate recognition model. The goal of this training is to optimize the performance of the model and obtain a higher recognition accuracy. After the secondary training, calculate and record the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model. If the recognition accuracy threshold corresponding to the secondary training result is not lower than the preset recognition accuracy threshold, it is determined that the license plate recognition model has completed training. If it is lower than the preset threshold, further adjustment or optimization is required. When the secondary training result does not meet the requirements, the license plate recognition model is adjusted using the gradient descent method. Through the calculation formula of the adjusted gradient The loss function of the license plate recognition model can be calculated by using the formula The learning speed, convergence speed and ratio of the training error data of the license plate recognition model to the preset data can be calculated, and the learning speed, convergence speed and influence coefficient of the training error data on the license plate recognition model can be obtained. Gradient descent is an optimization algorithm that reduces the model error by gradually adjusting the parameters of the model, thereby improving the performance of the model. If the performance of the model still does not meet expectations after gradient descent adjustment, the above process can be repeated until the preset recognition accuracy threshold is met or the preset upper limit of the number of training times is reached. Through the above technical effects, the system can realize the preliminary training, secondary training and gradient descent-based adjustment of the license plate recognition model. This process ensures that the model is continuously optimized and adjusted before reaching the preset recognition accuracy threshold, thereby improving the accuracy and performance of the model.

[0101] In one embodiment of the present invention, the training system of the license plate recognition model includes:

[0102] An information acquisition module is used to retrieve the license plate recognition model from the database and obtain the training information of the license plate recognition model;

[0103] An image processing module is used to retrieve image data corresponding to the current license plate type from a database, and perform image adjustment processing on the image data to obtain adjusted image data;

[0104] The model adjustment module is used to train the license plate recognition model using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold.

[0105] The working principle of the above technical solution is as follows: the information acquisition module retrieves the trained license plate recognition model from the database. This model usually contains the relevant algorithms and parameters for license plate recognition. The image processing module obtains the training information of the license plate recognition model from the database. This information may include training parameters, training data, and training results. According to the type of the current license plate, the corresponding image data is retrieved from the database. The model adjustment module adjusts the image data and uses the adjusted image data as sample data to retrain the license plate recognition model. Through continuous iteration and optimization, the recognition accuracy of the model is gradually improved. After each training, the recognition accuracy of the model is evaluated. When the recognition accuracy exceeds the preset recognition accuracy threshold, the training is stopped; otherwise, the training continues until the conditions are met.

[0106] The effect of the above technical solution is: through retraining and adjustment, the recognition accuracy of the license plate recognition model can be further improved to ensure accuracy and stability in practical applications. By adjusting and processing different image data, the model can better adapt to different practical application scenarios. Through model retraining, it can be ensured that the model always maintains the best performance and avoids performance degradation caused by the passage of time or environmental changes. By retrieving the license plate recognition model and related training information from the database, and combining it with new image data for retraining and adjustment, the accuracy and adaptability of license plate recognition can be further improved to ensure performance and effect in practical applications.

[0107] In one embodiment of the present invention, the information acquisition module includes:

[0108] A training module, used to retrieve a license plate recognition model and historical training records corresponding to the license plate recognition model from a database;

[0109] An accuracy acquisition module, used to obtain the recognition accuracy of the last license plate recognition model training, i.e., after optimization, from the historical training records corresponding to the license plate recognition model;

[0110] The recognition accuracy rate after the last license plate recognition model training, i.e. optimization, is used as the recognition accuracy rate threshold.

[0111] The working principle of the above technical solution is as follows: first, retrieve the trained license plate recognition model and its corresponding historical training records from the database. The historical training records may contain information such as parameters, time, and recognition accuracy of previous trainings.

[0112] From the historical training records, find the optimized recognition accuracy rate after the last license plate recognition model training. This accuracy rate can be used as the recognition accuracy threshold for subsequent model training.

[0113] The effect of the above technical solution is that by using historical training records, the system can set a reasonable recognition accuracy threshold based on the last optimized training in the past. Such threshold setting is more reliable and avoids inaccurate or inappropriate thresholds caused by arbitrary settings. As time goes by and new data is added, the accuracy in the historical training records may change. By regularly checking and updating the recognition accuracy threshold, the system can adaptively adjust the threshold to ensure that it always matches the current performance of the model. If the new data is similar to the historical data, decisions based on historical data can reduce unnecessary model training, thereby saving computing resources and time. By referring to the best recognition accuracy in the historical training records, the system can set a relatively high threshold, thereby pushing the model towards higher recognition accuracy. In short, by using the historical training records of the license plate recognition model to determine the recognition accuracy threshold, the system can make decisions based on past experience to ensure that the training and adjustment of the model are always based on best practices. This helps to improve the accuracy of the model, save resources and ensure continuous optimization of the system.

[0114] In one embodiment of the present invention, the information acquisition module includes:

[0115] A clarity threshold acquisition module is used to retrieve image data corresponding to the current license plate type from the database as original image data;

[0116] Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data;

[0117] The pixel adjustment module is used to adjust the pixel values ​​of the original image data using a pixel value adjustment model to obtain the adjusted pixel values; wherein the specific structure of the pixel value adjustment model is as follows:

[0118]

[0119] Among them, T zm is the pixel adjustment value, X ys is the original pixel value of the original image data, ΔC t To reduce the original image size by default, C y is the original image size, ΔFt To reduce the original image resolution, F y is the original image resolution, ΔL t To adjust the brightness of the original image, L y is the original image brightness, ΔR t To reduce the original image sharpness, R y is the original image sharpness, ΔB t To lower the saturation of the original image, B y is the original image saturation, ΔS t To preset the original image color temperature, S y is the color temperature of the original image, α is the adjustment coefficient, and its value range is 0-2;

[0120] Obtaining a clarity value corresponding to the adjusted pixel value, and determining whether the clarity value is lower than the clarity threshold;

[0121] If the clarity value corresponding to the adjusted pixel value is not lower than the clarity threshold, continue to use the pixel value adjustment model to adjust the pixel value of the original image data until the clarity value is lower than the clarity threshold, and use the image data corresponding to the pixel value whose clarity value is lower than the clarity threshold as the adjusted image data.

[0122] The working principle of the above technical solution is as follows: the image data corresponding to the current license plate type is retrieved from the database as the original image data. Pixel values ​​are extracted from the original image data, and the clarity threshold of the original image data is calculated based on these pixel values. The pixel value adjustment model is used to adjust the pixel values ​​of the original image data. This involves scaling or other forms of adjustment of the pixel values. The adjusted pixel values ​​obtain the corresponding clarity values, which are then compared with the clarity threshold of the original image data. If the adjusted clarity value is not lower than the threshold, the pixel value adjustment model is continued to be used to adjust the original image data until the clarity value is lower than the threshold. The image data corresponding to the pixel values ​​whose clarity values ​​are lower than the threshold are used as the final adjusted image data.

[0123] The effect of the above technical solution is: by adjusting the pixel value, the clarity of the original image can be enhanced, making it easier to identify and analyze. By calculating the clarity threshold based on the characteristics of the original image data, the system determines a suitable clarity standard through calculation to ensure that the clarity adjustment of the image is neither excessive nor too conservative. Since iterative adjustment is only continued when the adjusted clarity value is not lower than the threshold, this can reduce unnecessary calculations and optimize resource utilization. Through continuous pixel value adjustment, it can be ensured that the final output image data reaches a relatively appropriate level in clarity. In short, this method extracts the pixel values ​​of the original image data, calculates the clarity threshold, uses the pixel value adjustment model to iteratively adjust, and finally outputs image data that meets the clarity requirements, which helps to improve the accuracy and efficiency of license plate recognition.

[0124] In one embodiment of the present invention, the clarity threshold acquisition module includes:

[0125] A pixel deviation value acquisition module, used for extracting pixel values ​​of a plurality of the original image data;

[0126] The pixel deviation value corresponding to each original image data is set using the pixel value of the original image data.

[0127] The clarity threshold acquisition module is used to set a plurality of clarity thresholds corresponding to the original image data as a whole by using the pixel deviation value corresponding to each original image data in combination with the pixel value of the original image data.

[0128] The working principle of the above technical solution is as follows: a pixel deviation value acquisition module is used to extract pixel values ​​of a plurality of the original image data;

[0129] The pixel deviation value corresponding to each original image data is set using the pixel value of the original image data. The pixel deviation value corresponding to the original image data is obtained by the following formula:

[0130]

[0131] Among them, P czi is the pixel deviation value corresponding to the i-th original image data, j is the number of original image data, Z S is the recognition accuracy threshold, X ZVi is the recognition accuracy corresponding to the pixel value of the i-th original image data, Y XS is the preset pixel value, X Si is the pixel value of the i-th original image data, U XS is the average pixel value of all original image data, Z XS is the pixel value corresponding to the recognition accuracy threshold;

[0132] The definition threshold acquisition module is used to set a plurality of definition thresholds corresponding to the original image data as a whole by using the pixel deviation value corresponding to each original image data and the pixel value of the original image data. The definition threshold is obtained by the following formula:

[0133]

[0134] Among them, Q yz is the clarity threshold, q y is the initial preset definition, j is the number of original image data, X Si is the pixel value of the i-th original image data, P czi is the pixel deviation value corresponding to the i-th original image data.

[0135] The effect of the above technical solution is: the pixel deviation value acquisition module extracts multiple pixel values ​​from the original image data. For each original image data, the deviation of its pixel value is calculated. This can be based on the difference between pixels, the deviation of the pixel value from the average value or other pixel-related statistical characteristics. Based on the calculated pixel deviation, a corresponding pixel deviation value is set for each original image data. This value can be used for subsequent image adjustment and clarity evaluation. The clarity threshold acquisition module combines the pixel deviation value corresponding to each original image data with the pixel value of the original image data, and calculates the clarity threshold corresponding to each original image data as a whole through the clarity threshold calculation of the combined data. For multiple original image data, their corresponding clarity thresholds are calculated and set respectively. This provides a basis for subsequent image adjustment and clarity evaluation. Through the above technical effects, the system can provide more accurate pixel deviation values ​​and clarity thresholds for the original image data, thereby providing a basis for subsequent image adjustment and optimization.

[0136] In one embodiment of the present invention, the information acquisition module includes:

[0137] A comparison module, used to perform preliminary training on the license plate recognition model using the original image data, and obtain a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold;

[0138] When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model;

[0139] An adjustment module, used to determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than a preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, it is determined that the license plate recognition model has completed training;

[0140] When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent.

[0141] The working principle of the above technical solution is as follows: the comparison module is used to perform preliminary training on the license plate recognition model using the original image data, and obtain the first recognition accuracy corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy exceeds the recognition accuracy threshold; after the first recognition accuracy exceeds the recognition accuracy threshold, the license plate recognition model is retrained using the adjusted image data to obtain the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model; the adjustment module is used to determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, it is determined that the license plate recognition model has completed training; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent. Among them, the adjusted gradient is obtained by the following formula:

[0142]

[0143] Among them, T dtx is the adjusted gradient, T cs is the initial gradient value of the license plate recognition model, M is the true value of the license plate recognition model, is the recognition value of the license plate recognition model, is the license plate recognition model parameter, V ss is the actual convergence speed of the license plate recognition model, V ys is the preset convergence speed of the license plate recognition model, X ss is the actual learning speed of the license plate recognition model, X ys is the preset learning speed of the license plate recognition model, W sx is the actual training error data of the license plate recognition model, W yx The preset training error data for the license plate recognition model.

[0144] The effect of the above technical solution is: the license plate recognition model is preliminarily trained using the original image data. Through training, the preliminary training results of the license plate recognition model and its corresponding recognition accuracy are obtained. According to the preliminary training results, the first recognition accuracy is calculated and recorded. If the first recognition accuracy exceeds the preset recognition accuracy threshold, the next stage is entered; otherwise, the model may need to be further adjusted or retrained. When the first recognition accuracy exceeds the threshold, the license plate recognition model is trained again using the adjusted image data. The goal of this training is to optimize the performance of the model and obtain a higher recognition accuracy. After the secondary training, the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is calculated and recorded. If the recognition accuracy threshold corresponding to the secondary training result is not lower than the preset recognition accuracy threshold, it is determined that the license plate recognition model has completed training. If it is lower than the preset threshold, further adjustment or optimization is required. When the secondary training result does not meet the requirements, the license plate recognition model is adjusted using the gradient descent method. Gradient descent is an optimization algorithm that reduces the error of the model by gradually adjusting the parameters of the model, thereby improving the performance of the model. If the performance of the model still does not meet expectations after gradient descent adjustment, the above process can be repeated until the preset recognition accuracy threshold is met or the preset upper limit of training times is reached. Through the above technical effects, the system can realize the initial training, secondary training and gradient descent-based adjustment of the license plate recognition model. This process ensures that the model is continuously optimized and adjusted before reaching the preset recognition accuracy threshold, thereby improving the accuracy and performance of the model.

[0145] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for training a license plate recognition model, characterized in that: The training method of the license plate recognition model includes: Retrieve the license plate recognition model from the database to obtain the training information of the license plate recognition model; Retrieving image data corresponding to the current license plate type from a database, and performing image adjustment processing on the image data to obtain adjusted image data; The license plate recognition model is trained using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold.

2. The method for training a license plate recognition model according to claim 1, characterized in that: Retrieve the license plate recognition model from the database and obtain the training information of the license plate recognition model, including: Retrieving a license plate recognition model and historical training records corresponding to the license plate recognition model from a database; Obtaining the last license plate recognition model training, i.e., the optimized recognition accuracy, from the historical training record corresponding to the license plate recognition model; The recognition accuracy rate after the last license plate recognition model training, i.e. optimization, is used as the recognition accuracy rate threshold.

3. The method for training a license plate recognition model according to claim 1, characterized in that: Retrieving image data corresponding to the current license plate type from a database, and performing image adjustment processing on the image data to obtain adjusted image data, including: Retrieving image data corresponding to the current license plate type from the database as original image data; Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data; Using a pixel value adjustment model to adjust the pixel values ​​of the original image data to obtain adjusted pixel values; Obtaining a clarity value corresponding to the adjusted pixel value, and determining whether the clarity value is lower than the clarity threshold; If the clarity value corresponding to the adjusted pixel value is not lower than the clarity threshold, continue to use the pixel value adjustment model to adjust the pixel value of the original image data until the clarity value is lower than the clarity threshold, and use the image data corresponding to the pixel value whose clarity value is lower than the clarity threshold as the adjusted image data.

4. The method for training a license plate recognition model according to claim 3, characterized in that: Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data, includes: Extracting pixel values ​​of a plurality of the original image data; Using the pixel values ​​of the original image data, setting the pixel deviation value corresponding to each original image data; The pixel deviation value corresponding to each original image data is combined with the pixel value of the original image data to set a plurality of clarity thresholds corresponding to the original image data as a whole.

5. The method for training a license plate recognition model according to claim 1, characterized in that: The license plate recognition model is trained using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold, including: Performing preliminary training on the license plate recognition model using the original image data, and obtaining a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold; When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model; Determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, determine that the license plate recognition model has completed training; When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent.

6. A training system for a license plate recognition model, characterized in that: The training system of the license plate recognition model includes: An information acquisition module is used to retrieve the license plate recognition model from the database and obtain the training information of the license plate recognition model; An image processing module is used to retrieve image data corresponding to the current license plate type from a database, and perform image adjustment processing on the image data to obtain adjusted image data; The model adjustment module is used to train the license plate recognition model using the adjusted image data as sample data until the recognition accuracy of the trained license plate recognition model exceeds a recognition accuracy threshold.

7. The training system for the license plate recognition model according to claim 6, characterized in that: The information acquisition module includes: A training module, used to retrieve a license plate recognition model and historical training records corresponding to the license plate recognition model from a database; An accuracy acquisition module, used to obtain the recognition accuracy of the last license plate recognition model training, i.e., after optimization, from the historical training records corresponding to the license plate recognition model; The recognition accuracy rate after the last license plate recognition model training, i.e. optimization, is used as the recognition accuracy rate threshold.

8. The training system for the license plate recognition model according to claim 6, characterized in that: The information acquisition module includes: A clarity threshold acquisition module is used to retrieve image data corresponding to the current license plate type from the database as original image data; Extracting pixel values ​​of the original image data, and using the pixel values ​​to obtain a definition threshold corresponding to the original image data; A pixel adjustment module, used to adjust the pixel value of the original image data using a pixel value adjustment model to obtain the adjusted pixel value; Obtaining a clarity value corresponding to the adjusted pixel value, and determining whether the clarity value is lower than the clarity threshold; If the clarity value corresponding to the adjusted pixel value is not lower than the clarity threshold, continue to use the pixel value adjustment model to adjust the pixel value of the original image data until the clarity value is lower than the clarity threshold, and use the image data corresponding to the pixel value whose clarity value is lower than the clarity threshold as the adjusted image data.

9. The training system for the license plate recognition model according to claim 8, characterized in that: The clarity threshold acquisition module comprises: A pixel deviation value acquisition module, used for extracting pixel values ​​of a plurality of the original image data; Using the pixel values ​​of the original image data, setting the pixel deviation value corresponding to each original image data; The clarity threshold acquisition module is used to set a plurality of clarity thresholds corresponding to the original image data as a whole by using the pixel deviation value corresponding to each original image data in combination with the pixel value of the original image data.

10. The training system for the license plate recognition model according to claim 6, characterized in that: The information acquisition module includes: A comparison module, used to perform preliminary training on the license plate recognition model using the original image data, and obtain a first recognition accuracy rate corresponding to the preliminary training result of the license plate recognition model until the first recognition accuracy rate exceeds the recognition accuracy rate threshold; When the first recognition accuracy rate exceeds the recognition accuracy rate threshold, performing secondary training on the license plate recognition model using the adjusted image data to obtain a recognition accuracy rate threshold corresponding to the secondary training result of the license plate recognition model; An adjustment module, used to determine whether the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than a preset recognition accuracy threshold; when the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is not lower than the preset recognition accuracy threshold, it is determined that the license plate recognition model has completed training; When the recognition accuracy threshold corresponding to the secondary training result of the license plate recognition model is lower than the preset recognition accuracy threshold, the license plate recognition model is adjusted by gradient descent.

Citation Information

Patent Citations

  • License plate recognition method and system combining sequence image super-resolution reconstruction

    CN104809461A

  • Method and device for license plate clearness based on generative adversarial network

    CN107590774A

  • Car washer safety monitoring system and method based on car scene AI intelligent identification

    CN116935308A

  • End-To-End Lightweight Method And Apparatus For License Plate Recognition

    US20190095730A1