A license plate search method and system for a parking management system

Abstract

The application relates to a license plate search method and system for a parking management system and belongs to the technical field of parking management. The method comprises the following steps: S1: collecting vehicle images in a parking management system coverage area; S2: performing license plate accurate positioning on a standardized vehicle image, obtaining license plate accurate positioning coordinates, and intercepting a predetermined license plate area image in the standardized vehicle image by using the license plate accurate positioning coordinates; and S3: inputting the predetermined license plate area image into a pre-trained convolutional neural network model. The application adopts a license plate accurate positioning mode to improve the accuracy and efficiency of license plate positioning in a complex scene, extracts license plate multi-region deep features by using a pre-trained convolutional neural network model, combines PCA dimension reduction processing, introduces cosine similarity calculation to realize accurate matching, combines similarity threshold screening and sorting, simultaneously sets an auxiliary search mode, solves the search failure problem caused by license plate recognition errors and incomplete information, and effectively improves search flexibility and a search success rate.

Description

Technical Field

[0001] This invention relates to a license plate search method and system for a parking management system, belonging to the field of parking management technology. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of motor vehicle ownership, the scale and number of parking lots are constantly expanding, gradually upgrading from traditional small open-air parking lots to large multi-level parking lots and smart parking lots. The difficulty of parking management is also increasing exponentially. As the core technology of intelligent parking management systems, license plate recognition technology has been widely used in all aspects of parking lot processes, including vehicle entry, exit, billing, and vehicle location.

[0003] In practical parking management scenarios, license plate recognition technology can generally meet normal usage needs, but it still has certain shortcomings: license plate search typically uses a single feature extraction method, such as extracting only the character features or texture features of a single area of ​​the license plate. This results in poor adaptability to complex scenarios. When the license plate is partially obscured, damaged, or subject to environmental interference, feature extraction is incomplete, leading to a significant decrease in search accuracy and efficiency. Furthermore, the license plate search function of existing parking management systems largely relies on the accurate input of the complete license plate number. If the user cannot remember the complete license plate number, or if the license plate recognition is incomplete or contains errors, accurate searching is impossible, resulting in extremely poor search flexibility and a relatively low search success rate. Therefore, this invention provides a license plate search method and system for parking management systems. Summary of the Invention

[0004] The technical problem this invention aims to solve is that using a single feature extraction method for license plate search results in poor adaptability to complex scenarios, significantly reducing search accuracy and efficiency. License plate search relies on the accurate input of complete license plate numbers, which leads to problems such as license plate recognition errors and search failures due to incomplete information. The search flexibility is extremely poor, and the search success rate is relatively low.

[0005] To solve the above-mentioned technical problems, the present invention provides a license plate search method for a parking management system. The proposed technical solution includes the following steps:

[0006] S1: Collect vehicle images within the coverage area of ​​the parking management system, preprocess the vehicle images to obtain standardized vehicle images;

[0007] S2: Perform precise license plate localization on the standardized vehicle image, obtain the precise license plate localization coordinates, and use the precise license plate localization coordinates to extract the predetermined license plate area image from the standardized vehicle image;

[0008] S3: Input the image of the predetermined license plate area into the pre-trained convolutional neural network model, extract the features of the fully connected layer, and combine them in a preset order to form an initial feature vector;

[0009] S4: Use a dimensionality reduction model to reduce the dimensionality of the initial feature vector to obtain a target feature vector of a predetermined dimension. Then, associate the target feature vector with the corresponding vehicle information and store it in the database of the parking management system.

[0010] S5: Receive the license plate search request input by the user. The license plate search request includes the target license plate image or target license plate feature information. Extract the target feature vector based on the license plate search request.

[0011] S6: Calculate the cosine similarity between the target feature vector and all target feature vectors stored in the database, filter out matching feature vectors based on the similarity threshold, retrieve vehicle information associated with the matching feature vectors, and complete the license plate search.

[0012] S7: When the license plate recognition result does not match or the search results are not found, the auxiliary search mode is activated. The auxiliary information is received from the user, including vehicle color, vehicle type, and entry time period. The search range is narrowed by combining the auxiliary information, and step S6 is executed again to complete the license plate search.

[0013] Furthermore, in step S1, the preprocessing includes image denoising, size normalization, grayscale correction, and illumination compensation. The size normalization is used to adjust the vehicle image to a preset fixed size, and the illumination compensation uses an adaptive histogram equalization algorithm to eliminate the influence of uneven illumination on the image.

[0014] Furthermore, in step S2, the precise license plate localization adopts a combination of color features and template matching, specifically including: S21: converting the standardized vehicle image from RGB color space to HSV color space, performing mask binarization based on the color threshold of the license plate background color to obtain a binarized image; S22: performing morphological opening and closing operations on the binarized image, extracting connected components and calculating the minimum bounding moment, and selecting connected components with an aspect ratio between 2 and 4 as candidate license plate regions; S23: if the number of candidate license plate regions is 0, using license plate templates from a preset template library to perform template matching on the lower half of the standardized vehicle image to obtain candidate license plate regions; S24: performing character segmentation on the candidate license plate regions, counting the number of characters after segmentation, and selecting regions with a character count that meets a preset standard as the final license plate region, and determining its precise localization coordinates; S25: using the precise localization coordinates to crop the images of the global license plate region and the four vertex regions as predetermined license plate region images, and adjusting all predetermined license plate region images to a preset size.

[0015] Furthermore, in step S3, the training process of the convolutional neural network model includes: S31: collecting license plate images under different weather conditions, lighting conditions, and angles as training samples, preprocessing and labeling the training samples, with the label being the corresponding license plate number; S32: constructing a convolutional neural network model, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, using Softmax as the loss function; S33: inputting the training samples into the convolutional neural network model for training, adjusting the parameters of the convolutional layer and the fully connected layer using stochastic gradient descent until the model loss value meets the preset conditions, thus obtaining a pre-trained convolutional neural network model; S34: extracting the output of the fully connected layer of the model after training as license plate features, and combining them in the order of global license plate features, top left corner features, top right corner features, bottom left corner features, and bottom right corner features to form an 800×1 dimensional initial feature vector.

[0016] Furthermore, in step S4, the dimensionality reduction model adopts the PCA dimensionality reduction model to reduce the initial 800×1 dimension feature vector to a target feature vector of 100×1 dimension; the vehicle information includes license plate number, entry time, entry location, vehicle appearance features and parking status.

[0017] Furthermore, in step S6, the formula for calculating the cosine similarity is: ,in, For the target feature vector, The feature vectors are stored in the database, and sim is the cosine similarity. The similarity threshold is preset to 0.85. Feature vectors with sim ≥ 0.85 are selected as matching results. If there are multiple matching results, they are sorted by sim value from largest to smallest and then output.

[0018] A license plate search system for a parking management system, comprising the following components:

[0019] The image acquisition module is used to acquire vehicle images within the coverage area of ​​the parking management system and transmit the acquired vehicle images to the preprocessing module.

[0020] The preprocessing module is used to perform noise reduction, size normalization, grayscale correction and illumination compensation on the vehicle image to obtain a standardized vehicle image.

[0021] The license plate positioning module is used to perform precise license plate positioning on standardized vehicle images, obtain the precise license plate positioning coordinates, capture the image of the predetermined license plate area, and adjust it to a preset size.

[0022] The feature extraction module is used to input the image of the predetermined license plate area into a pre-trained convolutional neural network model, extract the features of the fully connected layer, and combine them to form an initial feature vector;

[0023] The dimension reduction storage module is used to perform dimension reduction processing on the initial feature vector using the PCA dimension reduction model to obtain the target feature vector, and then associate the target feature vector with the corresponding vehicle information and store it in the database.

[0024] The search request receiving module is used to receive the license plate search request input by the user and extract the target feature vector from the search request;

[0025] The similarity calculation module is used to calculate the cosine similarity between the target feature vector and all target feature vectors in the database, and to filter the matching results based on the similarity threshold.

[0026] The results output module is used to retrieve vehicle information associated with the matching feature vectors, sort them by similarity, and output them to complete the license plate search.

[0027] Furthermore, the license plate localization module includes a color localization unit, a template matching unit, a candidate filtering unit, and a region cropping unit. The color localization unit converts a standardized vehicle image into an HSV color space and obtains candidate license plate regions through color thresholding and morphological processing. The template matching unit performs template matching using a preset template library to obtain candidate license plate regions when the color localization unit fails to obtain them. The candidate filtering unit performs character segmentation and character count on the candidate license plate regions to filter out the final license plate regions that meet preset standards and determines the precise localization coordinates. The region cropping unit crops a predetermined license plate region image based on the precise localization coordinates and adjusts its size.

[0028] Furthermore, it also includes a model training module and an auxiliary search module; the model training module is used to collect training samples, train and optimize the parameters of the convolutional neural network model to obtain a pre-trained convolutional neural network model; the auxiliary search module is used to receive auxiliary information input by the user when there are no search results or the identification does not match, to narrow the search range and assist in completing the license plate search.

[0029] The beneficial effects of this invention are:

[0030] This paper proposes a precise license plate localization method that combines color features with template matching to improve the accuracy and efficiency of license plate localization in complex scenarios. It uses a pre-trained convolutional neural network model to extract deep features of multiple regions of the license plate, combines PCA dimensionality reduction processing, introduces cosine similarity calculation to achieve accurate matching, and combines similarity threshold filtering and sorting. At the same time, it sets up an auxiliary search mode to solve the problem of search failure caused by license plate recognition errors and incomplete information, effectively improving search flexibility and search success rate. Attached Figure Description

[0031] Figure 1This is a flowchart of the overall license plate search process of the present invention.

[0032] Figure 2 This is a flowchart of the image preprocessing process of the present invention.

[0033] Figure 3 This is a flowchart of the license plate precision positioning process of the present invention.

[0034] Figure 4 This is a flowchart of the feature extraction and dimensionality reduction process of the present invention.

[0035] Figure 5 This is a flowchart of the similarity matching process of the present invention.

[0036] Figure 6 This is a system module connection diagram of the present 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. Example 1

[0038] like Figure 1-5 As shown, this embodiment of the invention provides a license plate search method for a parking management system, including the following steps:

[0039] S1: Collect vehicle images within the coverage area of ​​the parking management system, and preprocess the vehicle images to obtain standardized vehicle images. Specifically, high-definition license plate recognition cameras deployed at parking lot entrances and exits and lanes collect vehicle images with a resolution of 1920×1080 pixels and a frame rate of 25fps. The collected vehicle images are preprocessed as follows: Gaussian blur algorithm (convolution kernel size of 3×3) is used to remove image noise, the image size is normalized to 800×600 pixels, grayscale deviation is corrected by grayscale correction algorithm, and illumination compensation is performed by adaptive histogram equalization algorithm to eliminate the influence of backlight and low light on the image, thus obtaining standardized vehicle images.

[0040] S2: Perform precise license plate localization on standardized vehicle images, obtain precise license plate localization coordinates, and use these coordinates to crop a predetermined license plate area image. Precise license plate localization employs a combination of color feature and template matching, specifically including:

[0041] S21: Convert the standardized vehicle image from RGB color space to HSV color space, set the color threshold for the blue license plate (H: 100-140, S: 40-255, V: 40-255), perform mask binarization processing, and obtain a binarized image;

[0042] S22: Perform morphological opening operations (3×3 erosion kernel, 3×3 dilation kernel) on the binarized image to remove small noise, and then perform closing operations (3×3 dilation kernel, 3×3 erosion kernel) to fill holes, extract connected components and find the minimum circumscribed moment, and select connected components with aspect ratios between 2 and 4 as candidate license plate regions.

[0043] S23: If no candidate license plate area is obtained (e.g., the license plate is white and similar to the body color), call the white license plate template (labeled with a sunny day) in the preset template library, perform template matching on the lower half of the standardized vehicle image, and use the normalized square matching method to obtain the candidate license plate area;

[0044] S24: Perform character segmentation on the candidate license plate area, use the projection method to segment the characters, count the number of characters after segmentation, and select the area with 7 characters as the final license plate area, and determine its precise positioning coordinates, such as the upper left corner coordinates (200, 300) and the lower right corner coordinates (400, 360).

[0045] S25: Extract the images of the global license plate area, the upper left corner area, the upper right corner area, the lower left corner area, and the lower right corner area, and adjust all the area images to 224×224 pixels to serve as the predetermined license plate area image.

[0046] S3: Input the image of the predetermined license plate area into a pre-trained convolutional neural network model, extract features from the fully connected layers, and combine them to form an initial feature vector; the training process of the convolutional neural network model includes:

[0047] S31: Collect license plate images under different weather conditions, lighting conditions, and angles as training samples. Preprocess and label the training samples, with the label being the corresponding license plate number.

[0048] S32: Construct a convolutional neural network (CNN) model, which includes an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer, using Softmax as the loss function;

[0049] S33: Input the training samples into the convolutional neural network model for training, and adjust the parameters of the convolutional layer and the fully connected layer by the stochastic gradient descent method until the model loss value meets the preset conditions to obtain the pre-trained convolutional neural network model.

[0050] S34: Extract the output of the fully connected layer of the model after training as the license plate feature, and combine them in the order of the global license plate feature, the upper left corner feature, the upper right corner feature, the lower left corner feature, and the lower right corner feature to form an initial feature vector of 800×1 dimension.

[0051] Specifically, the convolutional neural network model includes an input layer, 5 convolutional layers, 2 pooling layers, 2 fully connected layers, and 1 output layer; the input layer receives a predetermined license plate area image of 224×224 pixels, the convolutional layer uses the ReLU activation function, the pooling layer uses max pooling (pooling kernel 2×2), and the output dimension of the fully connected layer is 160 dimensions; input 5 predetermined license plate area images into the model respectively, extract the fully connected layer features (160 dimensions) of each area, and combine them in the order of the global feature, the upper left corner feature, the upper right corner feature, the lower left corner feature, and the lower right corner feature to form an initial feature vector of 800×1 dimension.

[0052] S4: Use a dimensionality reduction model to perform dimensionality reduction processing on the initial feature vector to obtain a target feature vector of a predetermined dimension, and associate and store the target feature vector with the corresponding vehicle information in the database of the parking management system. In step S4, the dimensionality reduction model uses the PCA dimensionality reduction model to reduce the 800×1-dimensional initial feature vector to a 100×1-dimensional target feature vector; the vehicle information includes the license plate number, the entry time, the entry location, the vehicle appearance feature, and the parking status; specifically, use the PCA dimensionality reduction model to perform dimensionality reduction processing on the initial feature vector to obtain a 100×1-dimensional target feature vector, and associate and store the target feature vector with the corresponding vehicle information (license plate number: Yue A12345, entry time: 2026-04-01 10:00, entry location: Parking space No. 10 in Area A, vehicle color: black, vehicle type: sedan, parking status: not yet exited) in the database.

[0053] S5: Receive a license plate search request input by the user. The license plate search request includes a target license plate image or target license plate feature information, and extract the target feature vector based on the license plate search request; specifically, receive a license plate search request input by the user, the user uploads a blurred license plate photo (target license plate), and perform the processing of steps S1-S4 on this photo to extract the target feature vector.

[0054] S6: Calculate the cosine similarity between the target feature vector and all the target feature vectors stored in the database, screen out the matching feature vectors according to the similarity threshold, and retrieve the vehicle information associated with the matching feature vectors to complete the license plate search. The calculation formula of the cosine similarity is: , where is the target feature vector, The database stores feature vectors, and sim represents the cosine similarity. The similarity threshold is preset to 0.85. Feature vectors with sim ≥ 0.85 are selected as matching results. If multiple matching results exist, they are sorted by sim value from largest to smallest and then output. Specifically, the cosine similarity between the target feature vector and all target feature vectors in the database is calculated. The preset similarity threshold is 0.85. Three matching results with sim ≥ 0.85 are selected and sorted by sim value from largest to smallest (0.92, 0.88, and 0.86, respectively). The corresponding vehicle information is retrieved, and the search results are output.

[0055] S7: When the license plate recognition result does not match or the search yields no results, the auxiliary search mode is activated. The system receives at least one piece of auxiliary information from the user's input: vehicle color, vehicle type, and entry time. This auxiliary information is used to narrow the search range, and step S6 is executed again to complete the license plate search. Specifically, if the license plate photo entered by the user is too blurry and the search yields no results, the auxiliary search mode is activated. The user inputs auxiliary information: vehicle color is black, vehicle type is sedan, and entry time is 2026-04-01 09:00-11:00. The system filters out 10 vehicle information entries that meet the criteria based on the auxiliary information, recalculates the cosine similarity of the target feature vectors of these 10 entries, filters out matching results, and completes the search. Example 2

[0056] like Figure 6 As shown, this embodiment of the invention provides a license plate search system for a parking management system, including an image acquisition module, a preprocessing module, a license plate localization module, a feature extraction module, a dimensionality reduction and storage module, a search request receiving module, a similarity calculation module, a result output module, a model training module, and an auxiliary search module;

[0057] The image acquisition module uses a 2-megapixel high-definition license plate recognition camera, which is deployed at the entrance and exit of the parking lot and in the lane. It supports a wide dynamic range of 120dB, IP54 waterproof and dustproof, and acquires vehicle images in real time and transmits them to the preprocessing module.

[0058] The preprocessing module uses an embedded processing chip to denoise, normalize, correct grayscale, and compensate for illumination in vehicle images, and outputs standardized vehicle images.

[0059] The license plate localization module includes a color localization unit, a template matching unit, a candidate filtering unit, and a region cropping unit. The color localization unit realizes license plate localization based on the HSV color space, the template matching unit realizes template matching localization, the candidate filtering unit filters out the final license plate area that meets the conditions, and the region cropping unit crops the image of the predetermined license plate area and adjusts its size.

[0060] The feature extraction module is deployed on the edge controller. It inputs the image of the predetermined license plate area into the pre-trained convolutional neural network model, extracts the features of the fully connected layer, and combines them to form an initial feature vector.

[0061] The dimensionality reduction storage module uses the PCA dimensionality reduction model to reduce the dimensionality of the initial feature vector to obtain the target feature vector, which is then associated with vehicle information and stored in a distributed database, supporting local caching and cloud backup.

[0062] The search request receiving module supports image uploads, character inputs, and other methods to receive user search requests and extract target feature vectors.

[0063] The similarity calculation module uses parallel computing to calculate the cosine similarity between the target feature vector and the feature vectors in the database, and then filters the matching results.

[0064] The results output module sorts the matching results by similarity and supports reverse vehicle search navigation;

[0065] The model training module collects training samples, trains and optimizes the convolutional neural network model, and supports online parameter updates;

[0066] The auxiliary search module receives auxiliary information input by the user, narrows down the search scope, and assists in completing the search.

[0067] The parking lot equipment layout and network topology are as follows: one set of high-definition license plate recognition camera and barrier gate are deployed at each entrance and exit. The parking space area is equipped with ultrasonic parking space detectors in groups of 25 spaces. An edge controller and switch are deployed in the duty room. Two self-service terminals are set up in the parking lot. The management personnel manage the system through the management backend. The edge controller is connected to the cloud server through the switch to realize data synchronization and remote management. The equipment adopts Ethernet networking to form a star topology, which is convenient for future expansion.

[0068] In summary, compared with the prior art, the present invention has the following advantages:

[0069] A license plate localization method combining color features and template matching is adopted to improve the accuracy and efficiency of license plate localization in complex scenarios, and is applicable to parking management scenarios under different weather and lighting conditions.

[0070] By using a pre-trained convolutional neural network model to extract deep features of multiple regions of license plates, and combining it with PCA dimensionality reduction processing, the dimensionality of the feature vector is reduced while ensuring the feature representation capability, thereby reducing storage costs and the complexity of similarity calculation, resulting in high search flexibility and improved efficiency of license plate search.

[0071] Cosine similarity calculation is introduced to achieve accurate matching. Combined with similarity threshold filtering and sorting, the accuracy of search results is ensured. At the same time, an auxiliary search mode is set up to narrow down the search range by auxiliary information such as vehicle color, vehicle type, and entry time, so as to solve the problem of search failure caused by license plate recognition errors and incomplete information, and improve the search success rate.

[0072] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A license plate search method for a parking management system, characterized in that: Includes the following steps: S1: Collect vehicle images within the coverage area of ​​the parking management system, preprocess the vehicle images to obtain standardized vehicle images; S2: Perform precise license plate localization on the standardized vehicle image, obtain the precise license plate localization coordinates, and use the precise license plate localization coordinates to extract the predetermined license plate area image from the standardized vehicle image; S3: Input the image of the predetermined license plate area into the pre-trained convolutional neural network model, extract the features of the fully connected layer, and combine them in a preset order to form an initial feature vector; S4: Use a dimensionality reduction model to reduce the dimensionality of the initial feature vector to obtain a target feature vector of a predetermined dimension. Then, associate the target feature vector with the corresponding vehicle information and store it in the database of the parking management system. S5: Receive the license plate search request input by the user. The license plate search request includes the target license plate image or target license plate feature information. Extract the target feature vector based on the license plate search request. S6: Calculate the cosine similarity between the target feature vector and all target feature vectors stored in the database, filter out matching feature vectors based on the similarity threshold, retrieve vehicle information associated with the matching feature vectors, and complete the license plate search. S7: When the license plate recognition result does not match or the search results are not found, the auxiliary search mode is activated. The auxiliary information is received from the user, including vehicle color, vehicle type, and entry time period. The search range is narrowed by combining the auxiliary information, and step S6 is executed again to complete the license plate search.

2. The license plate search method for a parking management system according to claim 1, characterized in that: In step S1, the preprocessing includes image denoising, size normalization, grayscale correction, and illumination compensation. Size normalization is used to adjust the vehicle image to a preset fixed size, and illumination compensation uses an adaptive histogram equalization algorithm to eliminate the influence of uneven illumination on the image.

3. The license plate search method for a parking management system according to claim 1, characterized in that: In step S2, the precise license plate localization employs a combination of color features and template matching, specifically including: S21: Converting the standardized vehicle image from RGB color space to HSV color space, performing mask binarization based on the color threshold of the license plate background color to obtain a binarized image; S22: Performing morphological opening and closing operations on the binarized image, extracting connected components and calculating the minimum bounding moment, and selecting connected components with an aspect ratio between 2 and 4 as candidate license plate regions; S23: If the number of candidate license plate regions is 0, using license plate templates from a preset template library to perform template matching on the lower half of the standardized vehicle image to obtain candidate license plate regions; S24: Performing character segmentation on the candidate license plate regions, counting the number of characters after segmentation, and selecting regions with a character count that meets a preset standard as the final license plate region, determining its precise localization coordinates; S25: Using the precise localization coordinates, cropping the global region and the four vertex regions of the license plate as predetermined license plate region images, and adjusting all predetermined license plate region images to a preset size.

4. The license plate search method for a parking management system according to claim 1, characterized in that: In step S3, the training process of the convolutional neural network model includes: S31: collecting license plate images under different weather conditions, lighting conditions, and angles as training samples, preprocessing and labeling the training samples, with the label being the corresponding license plate number; S32: constructing a convolutional neural network model, which includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, using Softmax as the loss function; S33: inputting the training samples into the convolutional neural network model for training, adjusting the parameters of the convolutional layer and the fully connected layer using stochastic gradient descent until the model loss value meets the preset conditions, thus obtaining a pre-trained convolutional neural network model; S34: extracting the output of the fully connected layer of the model after training as license plate features, and combining them in the order of global license plate features, top left corner features, top right corner features, bottom left corner features, and bottom right corner features to form an 800×1 dimensional initial feature vector.

5. A license plate search method for a parking management system according to claim 4, characterized in that: In step S4, the dimensionality reduction model adopts the PCA dimensionality reduction model to reduce the initial feature vector of 800×1 dimension to the target feature vector of 100×1 dimension; the vehicle information includes license plate number, entry time, entry location, vehicle appearance features and parking status.

6. The license plate search method for a parking management system according to claim 1, characterized in that: In step S6, the formula for calculating the cosine similarity is: ,in, For the target feature vector, The feature vectors are stored in the database, and sim is the cosine similarity. The similarity threshold is preset to 0.

85. Feature vectors with sim ≥ 0.85 are selected as matching results. If there are multiple matching results, they are sorted by sim value from largest to smallest and then output.

7. A license plate search system for a parking management system, characterized in that, include: The image acquisition module is used to acquire vehicle images within the coverage area of ​​the parking management system and transmit the acquired vehicle images to the preprocessing module. The preprocessing module is used to perform noise reduction, size normalization, grayscale correction and illumination compensation on the vehicle image to obtain a standardized vehicle image. The license plate positioning module is used to perform precise license plate positioning on standardized vehicle images, obtain the precise license plate positioning coordinates, capture the image of the predetermined license plate area, and adjust it to a preset size. The feature extraction module is used to input the image of the predetermined license plate area into a pre-trained convolutional neural network model, extract the features of the fully connected layer, and combine them to form an initial feature vector; The dimension reduction storage module is used to perform dimension reduction processing on the initial feature vector using the PCA dimension reduction model to obtain the target feature vector, and then associate the target feature vector with the corresponding vehicle information and store it in the database. The search request receiving module is used to receive the license plate search request input by the user and extract the target feature vector from the search request; The similarity calculation module is used to calculate the cosine similarity between the target feature vector and all target feature vectors in the database, and to filter the matching results based on the similarity threshold. The results output module is used to retrieve vehicle information associated with the matching feature vectors, sort them by similarity, and output them to complete the license plate search.

8. A license plate search system for a parking management system according to claim 7, characterized in that: The license plate localization module includes a color localization unit, a template matching unit, a candidate filtering unit, and a region extraction unit. The color localization unit is used to convert a standardized vehicle image into an HSV color space and obtain candidate license plate regions through color thresholding and morphological processing. The template matching unit is used to perform template matching using a preset template library to obtain candidate license plate regions when the color localization unit fails to obtain candidate license plate regions. The candidate filtering unit is used to perform character segmentation and character count on the candidate license plate area, filter out the final license plate area that meets the preset standard, and determine the precise positioning coordinates; the area cropping unit is used to crop the image of the predetermined license plate area according to the precise positioning coordinates and adjust the size.

9. A license plate search system for a parking management system according to claim 7, characterized in that: It also includes a model training module and an auxiliary search module; the model training module is used to collect training samples, train and optimize the parameters of the convolutional neural network model to obtain a pre-trained convolutional neural network model; the auxiliary search module is used to receive auxiliary information input by the user when there are no search results or the recognition does not match, to narrow the search range and assist in completing the license plate search.

Images