License plate recognition system and method based on image technology and medium

Abstract

The invention relates to the technical field of image recognition, in particular to a license plate recognition system and method based on an image technology and a medium. The method comprises the following steps: acquiring area sensing data and a camera image set, and performing deformation effect compensation to obtain an environment compensation image set; performing image diffusion reverse enhancement on the environment compensation image set to obtain a license plate area enhanced image set; performing character region high-dimensional topological mapping based on the license plate region enhanced image set to obtain a character segmentation matrix; extracting character morphological characteristics according to the character segmentation matrix, and performing character recognition on the character morphological characteristics to obtain a character recognition result; and carrying out cross-character semantic compensation on the character recognition result to obtain a semantic compensation license plate character vector, and carrying out multi-target cross verification on the semantic compensation license plate character vector to obtain a license plate recognition result. According to the invention, the accuracy and robustness of license plate recognition can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a license plate recognition system, method and medium based on image technology. Background Art

[0002] License Plate Recognition (LPR) technology, as an important component of the Intelligent Transportation System (ITS), has been widely used in many fields, including vehicle management, electronic toll collection, traffic enforcement, and parking management. The core of LPR technology lies in automatically extracting the license plate area, segmenting characters, and identifying license plate numbers through computer vision, pattern recognition, and deep learning algorithms, thereby achieving automated management of target vehicles. Early LPR methods mainly relied on traditional image processing techniques, using edge detection, grayscale projection, and connected region analysis to detect the license plate area, and using OCR (Optical Character Recognition) technology to recognize license plate characters. However, such methods have certain limitations when facing complex environments. For example, changes in lighting, motion blur, occlusion, and license plate damage can seriously affect recognition accuracy.

[0003] First, traditional image processing-based license plate detection methods typically use edge detection, color analysis, or morphological processing to locate the license plate area. These methods can achieve high detection accuracy under ideal conditions, but in practice, the license plate edge is often unclear due to lighting conditions, shadows, reflections, and complex backgrounds, resulting in high rates of false detection or missed detection. Furthermore, some license plate colors are close to the background color (such as a yellow license plate against a yellow background), making it difficult for traditional methods to accurately segment the license plate area.

[0004] Secondly, traditional methods for license plate character segmentation rely primarily on vertical projection analysis or connected component analysis. However, when faced with damaged, deformed, or occluded license plates, the spacing between characters may change, leading to segmentation errors. For example, stains on a license plate may connect two characters, mistaking them for a single unit, while ambiguous characters may be mis-segmented or lost, affecting subsequent character recognition. Summary of the Invention

[0005] Based on this, it is necessary for the present invention to provide a license plate recognition system, method and medium based on image technology to solve at least one of the above technical problems.

[0006] To achieve the above purpose, a license plate recognition method based on image technology includes the following steps:

[0007] Step S1: Acquire regional sensing data and a camera image set, and perform deformation effect compensation on the camera image set based on the regional sensing data to obtain an environment-compensated image set;

[0008] Step S2: performing adaptive license plate region division on the environment compensation image set to obtain a license plate region image set; performing image diffusion reverse enhancement on the license plate region image set to obtain a license plate region enhanced image set;

[0009] Step S3: performing character boundary morphology perception based on the license plate region enhanced image set to obtain a character boundary image set; performing high-dimensional topological mapping of the character region based on the character boundary image set to obtain a character segmentation matrix;

[0010] Step S4: extracting character morphological features according to the character segmentation matrix, and performing character recognition on the character morphological features to obtain a character recognition result;

[0011] Step S5: Perform cross-character semantic compensation on the character recognition result to obtain a semantically compensated license plate character vector, and perform multi-target cross-validation on the adjacent semantically compensated license plate characters to obtain a license plate recognition result.

[0012] By combining regional sensing data with a camera image set, the present invention adaptively segments the license plate area and compensates for deformation effects, significantly improving the detection accuracy of license plate areas under complex environmental conditions. The generation of an environmentally compensated image set effectively reduces license plate image deformation caused by factors such as lighting, shadows, and reflections, enhancing the robustness of license plate detection. Furthermore, the image diffusion inverse enhancement process optimizes the visual features of the license plate area, making it clearer in low-contrast or blurred conditions and facilitating subsequent character extraction. Character boundary morphology perception based on the license plate area enhanced image set enables more precise identification of license plate character boundaries, particularly in cases where characters are stained or deformed, thereby improving the accuracy of character segmentation. High-dimensional topological mapping of the character area further enhances the accuracy of the character segmentation matrix, ensuring complete character segmentation, especially when character spacing is irregular or partially occluded, providing high-quality data support for subsequent character recognition. Efficient character recognition relies on the extracted character morphological features, which facilitate accurate identification of individual characters in the license plate, particularly in complex situations where traditional methods struggle, such as stained or blurred characters. The character recognition results undergo cross-character semantic compensation, which can effectively correct character errors caused by misidentification or morphological similarity, further improving the accuracy of license plate recognition. Multi-objective cross-validation can verify the accuracy of character recognition and ensure the validity of the recognition results. In particular, when there is uncertainty in character recognition, it can improve the reliability of recognition through multiple verification methods, ultimately obtaining highly accurate license plate recognition results. Overall, the entire process effectively overcomes the limitations of traditional methods in complex environments by combining environmental compensation, image enhancement, topological mapping, feature extraction, and semantic compensation, significantly improving the accuracy and robustness of license plate recognition.

[0013] Optionally, step S1 specifically includes:

[0014] Step S11: acquiring regional sensor data and a camera image set, and performing data preprocessing on the regional sensor data and the camera image set respectively to obtain the regional sensor data to be analyzed and the camera image set to be analyzed;

[0015] Step S12: classifying the sensor data of the area to be analyzed into sensor types to obtain environmental sensor data and camera sensor data;

[0016] Step S13: performing illumination contrast compensation on the camera image set to be analyzed according to the environmental sensing data to generate an illumination compensated image set;

[0017] Step S14: performing camera distortion correction on the illumination compensation image set based on the camera sensing data to obtain a distortion corrected image set;

[0018] Step S15: Estimate motion blur and deformation errors according to the distortion-corrected image set, and perform inverse deformation compensation based on the estimation results to obtain an environment-compensated image set.

[0019] The present invention pre-processes the regional sensing data and camera image set separately to ensure that the input data has good quality in subsequent analysis, reduce noise and redundant information, and thus improve the overall performance of the recognition system. By classifying the regional sensing data, the characteristics of the environment and camera sensor are further clarified, making data processing more accurate, thus laying the foundation for subsequent illumination contrast compensation and camera distortion correction. Illumination compensation is performed on the camera image set to be analyzed, effectively eliminating the impact caused by changes in ambient lighting conditions, especially in high-light or low-light environments, enhancing the contrast of the image, making the license plate area clearer and easier to identify. The image after illumination compensation is further corrected by camera distortion to correct the geometric distortion caused by the camera lens, ensuring that the geometric structure of the image is closer to the real world, which is crucial for subsequent license plate area detection and avoids false detection due to distortion. By estimating motion blur and deformation errors and performing inverse deformation compensation on the image, the blurring caused by vehicle movement or other environmental factors is successfully reduced, making the license plate characters clearer. Deformation inverse compensation technology improves image robustness under complex conditions, providing more stable support for the application of license plate recognition systems in various traffic environments. Overall, these steps effectively optimize image quality and enhance the recognition accuracy of license plate recognition systems in dynamic and complex environments. They also provide more reliable input data for subsequent steps such as license plate region extraction and character recognition.

[0020] Optionally, step S14 is specifically as follows:

[0021] Step S141: extracting camera posture features based on camera sensor data corresponding to the camera image set to obtain camera posture feature data;

[0022] Step S142: obtaining camera calibration data, and correlating the camera calibration data with camera posture feature data to integrate camera imaging parameters to obtain a camera imaging parameter matrix;

[0023] Step S143: performing a preliminary perspective transformation on the illumination-compensated image set, extracting global feature points and local feature points, and estimating the perspective distortion error based on the feature point matching results of the global feature points and the local feature points to obtain a perspective distortion parameter matrix;

[0024] Step S144: performing nonlinear optical compensation pixel resampling based on the perspective distortion parameter matrix and the camera imaging parameter matrix, and performing pixel mapping on the illumination compensation image set according to the compensation pixels to obtain a preliminary correction image set;

[0025] Step S145: performing local gradient direction sharpening on the preliminary corrected image set to obtain a distortion corrected image set.

[0026] By extracting camera posture features and integrating camera imaging parameters, the present invention can more accurately understand the camera's positioning and posture information, helping to determine the position and orientation of the license plate in the image. This process effectively improves the accuracy of the image's geometric information, making subsequent image transformation and correction more reliable. By acquiring and correlating camera calibration data with camera posture feature data, the camera imaging parameter matrix is further optimized, ensuring the image's consistency with the actual physical world. This is crucial for geometric correction of images, especially in the presence of distortion, ensuring the image's realism. By integrating these imaging parameters, subsequent image processing, such as perspective transformation and optical compensation, can be performed efficiently, effectively avoiding image distortion caused by camera perspective issues. During the initial perspective transformation of the illumination-compensated image set, global and local feature points are extracted and matched to successfully estimate the perspective distortion error. This process ensures that license plate information can be precisely aligned at different shooting angles, eliminating errors caused by perspective differences and improving the visual consistency of the image. Combining the perspective distortion parameter matrix with the camera imaging parameter matrix, nonlinear optical compensation pixel resampling effectively adjusts pixel distribution, restores the image's true structure, and avoids distortion-induced image distortion, thereby improving the recognizability of license plate characters. Local gradient sharpening makes the distortion-corrected image sharper, with clearer details and enhanced edge information of the license plate characters. This process helps improve edge recognizability in the license plate area, especially in low-contrast or blurred conditions. It effectively highlights the morphological features of the license plate characters, ensuring more precise subsequent character segmentation and recognition, and overall improving the robustness and accuracy of license plate recognition.

[0027] Optionally, step S15 is specifically as follows:

[0028] Step S151: performing motion blur detection on the distortion-corrected image set based on the distortion-corrected image set to obtain a blurred area image set;

[0029] Step S152: performing deconvolution on the blurred area image set to obtain a blurred repaired image set;

[0030] Step S153: performing deformation error estimation on the blurred inpainted image set, identifying geometric distortions between the blurred inpainted images, and thereby generating a deformation error matrix;

[0031] Step S154: performing inverse geometric correction of deformation on the blurred inpainted image set according to the deformation error matrix, reconstructing the geometrically distorted region in the blurred inpainted image set, and obtaining a deformation corrected image set;

[0032] Step S155: performing global color adaptive adjustment on the deformation-corrected image set to obtain an environment-compensated image set.

[0033] This invention uses motion blur detection based on a distortion-corrected image set to help identify and locate blurred areas in an image caused by vehicle motion or camera shake. Accurately identifying blurred areas avoids misjudgment or loss of critical information, ensuring that subsequent inpainting operations are performed only on necessary areas, significantly improving inpainting efficiency and accuracy. Deconvolution inpainting further addresses image blur. This step uses an inverse operation to restore image details, sharpening blurred areas. Compared to traditional deblurring techniques, deconvolution inpainting effectively preserves the image's structural features, particularly the edges of license plate characters, making the license plate information visually clearer and easier to read, facilitating subsequent character recognition accuracy. Deformation error estimation is performed on the inpainted image set to detect geometric distortion caused by the inpainting process and accurately generate a deformation error matrix. This process ensures geometric consistency, ensuring that the inpainted image maintains a high degree of geometric consistency with the actual license plate, avoiding misidentification issues caused by distortion. By performing inverse geometric correction on the deformation error, geometrically distorted areas in the image can be reconstructed, restoring the license plate's original geometry, thereby ensuring image realism and accuracy. Global color adaptation further enhances the visual quality of images, maintaining consistent color across varying lighting and environmental conditions. This ensures natural, realistic color in both bright and dim environments, helping the license plate recognition algorithm accurately determine the outline and content of license plate characters, thereby improving recognition reliability and accuracy.

[0034] Optionally, the adaptive division of the license plate area in step S2 is specifically as follows:

[0035] Perform edge detection on the environmental compensation image set, extract image contour features, and obtain image area contour feature data;

[0036] The foreground area of the license plate image is screened according to the image area contour feature data to obtain a preliminary license plate candidate area set;

[0037] Perform morphological filtering on the preliminary license plate candidate region set and calculate the aspect ratio and connectivity features of each candidate region;

[0038] Based on the preset license plate size rules, set the size constraints of a single license plate and multiple license plates according to the license plate size rules;

[0039] According to the size constraint of a single license plate, the aspect ratio and connectivity characteristics of each candidate region are used to screen the license plate region, and a single license plate candidate region set is obtained;

[0040] Calculate the regional contrast distribution of the candidate region set of a single license plate to obtain regional gradient change data, and use the regional gradient change data to perform regional superpixel segmentation to obtain a single license plate region image set;

[0041] Connected component analysis is performed based on the aspect ratio and connectivity characteristics of each candidate region, and low-density regions are filtered to obtain a preliminary set of multiple license plate candidate regions;

[0042] Performing license plate area screening on the preliminary multiple license plate candidate area sets according to the multiple license plate size constraints to obtain multiple license plate area sets;

[0043] Perform superpixel segmentation on multiple license plate region sets and perform region segmentation density equalization to obtain multiple license plate region image sets;

[0044] The single license plate region image set and multiple license plate candidate region sets are merged to obtain a license plate region image set.

[0045] The present invention effectively improves the ability to accurately detect and identify license plate regions through precise processing and screening of image regions. During image processing, edge detection is first used to extract image contour features. This allows for accurate identification of license plate edge information, helping to separate the license plate region from complex backgrounds. Extracting image region contour features allows for better identification of the license plate's position within the image and provides a solid foundation for subsequent license plate screening. Foreground region screening of license plate images based on contour feature data quickly removes irrelevant background information from the image, ensuring that only regions that may be license plates are retained. Morphological filtering further refines candidate regions, removing noise and enhancing the morphological features of the license plate region. Calculating the aspect ratio and connectivity features of candidate regions facilitates precise screening of license plate regions, avoiding misidentifications caused by deformation or occlusion. Pre-set license plate size rules provide dimensional constraints for the screening process, ensuring that candidate regions meet standard license plate dimensions, thereby eliminating inappropriate regions and improving the accuracy of license plate region recognition. By screening license plate regions based on their aspect ratio and connectivity features, misidentified regions are effectively eliminated, ensuring that the final selected candidate regions meet the actual morphological features of the license plate. By calculating the contrast distribution of the region and utilizing regional gradient change data for superpixel segmentation, the extraction of the license plate region can be further optimized in detail, resulting in a clearer segmented image and facilitating subsequent character recognition. Low-density region filtering based on connected component analysis effectively removes low-quality regions caused by license plate occlusion, damage, or other reasons, improving the accuracy of multiple license plate detection. Region screening based on multiple license plate size constraints ensures that the final selected license plate regions meet the expected license plate size, avoiding false detections. During the processing of multiple license plate region image sets, superpixel segmentation and region segmentation density balancing operations further optimize the image quality of the license plate region, making the license plate region more distinct and facilitating subsequent recognition and analysis. By merging the data of a single license plate region image set with multiple license plate candidate region sets, various information of the license plate image is integrated to obtain a precise license plate region image set, providing clear and accurate image data for subsequent character recognition and license plate parsing.

[0046] Optionally, the image diffusion inverse enhancement in step S2 is specifically:

[0047] The local gradient change of the license plate area image set is calculated to obtain the local gradient change rate of the pixel point, and the diffusion factor range is set to [0,1] to extract the image diffusion factor from the local gradient change rate of the pixel point;

[0048] The reverse diffusion calculation is performed on the license plate area image set according to the image diffusion factor, and the iterative calculation is performed for [50,100] rounds. The diffusion range is set to [0.1,0.5] to obtain the reverse diffusion image set.

[0049] Perform edge detection on the reverse diffusion image set, set the regional contrast stretching factor to [1.2.2.0] to calculate the local edge strength, and perform regional contrast nonlinear stretching based on the local edge strength results to obtain the diffusion reverse enhancement image set;

[0050] The license plate area image set and the diffusion inverse enhanced image set are adaptively fused. The scale range is set to [3, 5, 7] to calculate the multi-scale clarity score. The image set fusion weight is adjusted according to the multi-scale clarity score result to obtain the license plate area enhanced image set.

[0051] The present invention improves the detail extraction and enhancement of license plate area images through local gradient change calculation and image diffusion processing. By calculating the local gradient change rate, subtle detail changes in the license plate area image can be captured, facilitating the extraction of license plate outline information in complex backgrounds or low-contrast conditions. Extraction of the image diffusion factor further optimizes local image detail processing, ensuring that the key features of the license plate are retained during the image diffusion process. Inverse diffusion calculation based on the image diffusion factor, through multiple rounds of iterative calculations, can effectively restore image details lost due to illumination, blur, or distortion. By setting a reasonable diffusion range during the diffusion process, the local contrast and detail level of the image can be further enhanced, making the license plate area clearer and facilitating subsequent recognition. In addition, edge detection and setting regional contrast stretching factors effectively enhance the contrast of important areas in the image, increasing the salience of the license plate edges, and thus enhancing the recognizability of the license plate characters. Local edge strength calculation and regional contrast nonlinear stretching can further enhance the clarity of the license plate area, making character edges clearer, improving the visual effect of the image, and providing more accurate input data for character segmentation and recognition. The generation of a diffusion-inverse enhanced image set effectively enhances image detail, providing higher-quality image data for subsequent license plate character extraction. The license plate region image set is adaptively fused with the diffusion-inverse enhanced image set. A multi-scale clarity score is used to assess image clarity at different scales, thereby weighting the image fusion. This process effectively integrates information from different image processing stages, optimizing the overall quality of the license plate image set and ensuring high clarity of the fused image at all scales, thereby improving the accuracy and stability of license plate recognition.

[0052] Optionally, step S3 specifically includes:

[0053] Step S31: performing Sobel operator edge detection on the license plate area enhanced image set, extracting character outline features, and performing morphological closing operation on the character outline features to divide the connected regions and obtain a connected region image set;

[0054] Step S32: calculating the aspect ratio, area and rectangle fitting of the connected regions according to the connected region image set, and filtering the character candidate region set according to the calculation results of the aspect ratio, area and rectangle fitting of the connected regions;

[0055] Step S33: performing boundary enhancement on the character candidate region set, and performing superpixel segmentation on the boundary-enhanced character candidate region set to extract character region structural features to obtain a character boundary image set;

[0056] Step S34: integrating the geometric features of the character region based on the character boundary image set to obtain the character region outline system features, and performing spatial relationship topological structure modeling based on the character region outline system features to obtain the character region topological structure model;

[0057] Step S35: performing high-dimensional feature projection on the character region topology structure model to obtain character region topology mapping data, and performing character segmentation based on the character region topology mapping data to obtain a character segmentation matrix.

[0058] The present invention uses the Sobel operator for edge detection, effectively extracting the character outline features within the license plate area and enhancing the edge information in the image, which provides an important basis for subsequent character segmentation and morphological analysis. The application of morphological closing operations further optimizes the integrity of the character outline, ensuring that the character area is accurately retained in the division of connected regions, thereby effectively avoiding false detection or missed detection due to noise or other interference factors. Calculating the aspect ratio, area, and rectangular fit of the connected regions helps to screen candidate regions that meet the characteristics of the license plate characters. This step ensures that the geometric characteristics of the character area are accurately analyzed, thereby effectively eliminating areas that do not meet the standards, reducing interference from irrelevant areas, and improving the accuracy and efficiency of character recognition. The set of character candidate regions is further subjected to boundary enhancement and superpixel segmentation. This process not only enhances the visibility of the character boundaries, but also extracts the structural features of the character area, making the character details clearer and laying the foundation for subsequent geometric morphological analysis. In the process of integrating the geometric morphological features of the character area, multiple character features are effectively fused, providing sufficient information for subsequent spatial relationship topological structure modeling. By establishing a topological structure model of the character area, the spatial relationship between characters can be better captured, making the character segmentation process more accurate and robust. Ultimately, the character region topology mapping data generated based on high-dimensional feature projection not only helps further improve the segmentation accuracy of the character region, but also enables the character segmentation matrix to more effectively separate characters, ensuring accurate license plate number recognition. This entire process, by gradually optimizing image features and region selection, significantly improves license plate character recognition, especially robustness against complex backgrounds or blurred environments.

[0059] Optionally, the cross-character semantic compensation in step S5 is specifically as follows:

[0060] According to the statistical distribution characteristics of the character sequence in the character recognition results, the arrangement pattern threshold is set to 0.8 to identify the arrangement pattern between characters and obtain the character arrangement pattern data;

[0061] The co-occurrence window size is set to 3 to 7 characters, the co-occurrence probability between characters is calculated for the character arrangement pattern data, and based on the co-occurrence probability calculation result, the confidence score range is set to [0, 1] to estimate the character probability combination confidence score;

[0062] Perform character semantic matching on the character recognition results based on the character probability combined confidence score, filter and mark low-confidence characters, and obtain a character semantic association matrix;

[0063] Combined with the character morphological features, the morphological correction threshold is set to 0.6 to correct the misrecognition of visually similar characters for the low-confidence characters in the character semantic association matrix, and the semantically compensated license plate character vector is obtained.

[0064] The present invention effectively filters out the arrangement patterns between characters by setting a reasonable arrangement pattern threshold, so that the character recognition process can specifically process characters with different arrangements, thereby improving the adaptability to complex license plate layouts. By calculating the co-occurrence probability between characters, the system can comprehensively analyze the correlation between characters, effectively avoiding recognition errors caused by incorrect combinations or arrangements between characters, and further improving the accuracy of the overall recognition results. This co-occurrence probability calculation process also provides accurate data support for subsequent confidence scoring, allowing the system to more scientifically evaluate the reliability of character recognition. After setting a reasonable confidence scoring range, the system can effectively filter low-confidence characters based on the scoring results, avoiding the influence of erroneous characters, and ensuring the accuracy of the recognition results. By performing semantic matching on low-confidence characters, the system can identify and correct potential misidentified characters, thereby generating high-quality character sequences. This process further optimizes the semantic relationship between characters through the establishment of a semantic association matrix, so that character recognition does not only stay at the morphological level, but also goes deeper into the semantic level of the characters. Combining the morphological features of the characters and using a morphological correction threshold to correct visually similar misidentifications of low-confidence characters can effectively address recognition errors caused by similar character appearances. This step corrects misidentified characters through visual similarity analysis, thereby improving character recognition accuracy, especially in cases where the license plate characters are irregular or damaged. Ultimately, the resulting semantically compensated license plate character vector integrates multiple aspects of feature correction, making the license plate character recognition results more accurate and robust, and able to effectively cope with various interference factors such as overlapping characters, blurred or complex backgrounds, etc.

[0065] Optionally, this specification further provides a license plate recognition system based on image technology, which is used to execute the license plate recognition method based on image technology as described above. The license plate recognition system based on image technology includes:

[0066] An environmental compensation module is used to obtain regional sensing data and a camera image set, and to compensate for deformation effects of the camera image set based on the regional sensing data to obtain an environmental compensation image set;

[0067] The license plate area enhancement module is used to perform adaptive license plate area division on the environment compensation image set to obtain a license plate area image set; perform image diffusion reverse enhancement on the license plate area image set to obtain a license plate area enhanced image set;

[0068] The character segmentation module is used to perform character boundary morphology perception based on the license plate area enhanced image set to obtain a character boundary image set; perform high-dimensional topological mapping of the character area based on the character boundary image set to obtain a character segmentation matrix;

[0069] The character recognition module is used to extract character morphological features according to the character segmentation matrix, and perform character recognition on the character morphological features to obtain character recognition results;

[0070] The license plate recognition result verification module is used to perform cross-character semantic compensation on the character recognition results to obtain the semantically compensated license plate character vector, and perform multi-target cross-validation on the adjacent semantically compensated license plate characters to obtain the license plate recognition result.

[0071] The license plate recognition system based on image technology of the present invention can implement any one of the license plate recognition methods based on image technology of the present invention, and is used to combine the operations and signal transmission media between various modules to complete the license plate recognition method based on image technology. The internal modules of the system cooperate with each other, thereby effectively overcoming the limitations of traditional methods in complex environments and significantly improving the accuracy and robustness of license plate recognition.

[0072] Optionally, this specification also provides a computer-readable storage medium having a computer program stored thereon, which implements the license plate recognition method based on image technology as described above when the computer program is executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0074] Figure 1 A schematic flow chart of the steps of the license plate recognition method based on image technology of the present invention;

[0075] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0076] Figure 3 Detailed step flow diagram of step S3 in the present invention;

[0077] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0078] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0079] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0080] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0081] To achieve this, please refer to Figures 1 to 3 The present invention provides a license plate recognition method based on image technology, which includes the following steps:

[0082] Step S1: Acquire regional sensing data and a camera image set, and perform deformation effect compensation on the camera image set based on the regional sensing data to obtain an environment-compensated image set;

[0083] In this embodiment, regional sensing data is collected by regional sensors (such as lidar, temperature sensor, humidity sensor, etc.), and a series of image sets captured by cameras are obtained. The deformation effect of the camera image is compensated by the physical features in the regional sensing data (such as regional deformation, environmental impact, etc.). Specifically, based on the regional sensor data and the image calibration model, the deformation error of the camera image is calculated, and the corresponding mathematical model (such as affine transformation or perspective transformation) is applied to perform geometric correction on the image. The compensation process uses quadratic curve fitting technology to perform pixel-level compensation to obtain an environmental compensation image set. The compensated image uses an interpolation algorithm to fill in the missing pixel information to maintain the consistency of the image content. The interpolation step size in the specific compensation method is 2 pixels to ensure the restoration and integrity of the image details.

[0084] Step S2: performing adaptive license plate region division on the environment compensation image set to obtain a license plate region image set; performing image diffusion reverse enhancement on the license plate region image set to obtain a license plate region enhanced image set;

[0085] In this embodiment, the compensated image set undergoes adaptive segmentation of the license plate region. Using an edge detection-based segmentation algorithm (such as Canny edge detection), edge information is first extracted from the image. Then, a dynamic threshold selection method is used to extract the license plate region boundary. Specifically, the adaptive threshold selection method suppresses image noise, making the license plate region boundary more defined. Subsequently, a deep learning algorithm (such as a convolutional neural network (CNN)) is used to further improve the accuracy of the segmentation, leveraging the image's illumination, contrast, and texture characteristics. This yields a set of license plate region images. For image diffusion inverse enhancement, the local gradient of the image is first calculated, with the local gradient change rate threshold set to 0.2. An image diffusion factor is then calculated based on this gradient information. The inverse diffusion algorithm is then applied to the license plate region image, with 20 iterations, to enhance the image's edge details. During the diffusion process, the enhancement effect is particularly pronounced at the edges, improving image clarity. Finally, the image set is weighted fused using a multi-scale clarity score to yield the final enhanced license plate region image set.

[0086] Step S3: performing character boundary morphology perception based on the license plate region enhanced image set to obtain a character boundary image set; performing high-dimensional topological mapping of the character region based on the character boundary image set to obtain a character segmentation matrix;

[0087] In this embodiment, based on the license plate area enhanced image set, character boundary morphology perception is first performed. By adopting an edge detection algorithm based on gradient information (such as the Sobel operator), the boundary information of the characters is extracted in the enhanced image, and the local intensity of the character edge is calculated. The generation of the character boundary image set is based on the local contrast information of the image, and the contrast intensity threshold is set to 0.3. The part with higher edge clarity will be retained first. Then, a high-dimensional topological mapping method based on topological data analysis (TDA) is used to segment the character area. The TDA method can map high-dimensional data to a low-dimensional space, thereby helping to understand the structural relationship of the character area. In this process, a topological persistence algorithm is used to analyze the shape of the character area, and a character segmentation matrix is obtained by calculating the topological persistence map. The dimension of the matrix is 128×128, which represents the topological relationship between every two points in the character area. The value of the character segmentation matrix is generated by calculating the distance, similarity and angle information between characters. The range of the matrix value is [0,1], where the larger the value, the closer the relationship between the characters.

[0088] Step S4: extracting character morphological features according to the character segmentation matrix, and performing character recognition on the character morphological features to obtain a character recognition result;

[0089] In this embodiment, the morphological features of the characters are extracted from the generated character segmentation matrix. The character morphological feature extraction is performed using the HOG (Histogram of Directed Gradients) algorithm, and the specific steps include calculating the gradient direction and amplitude of the character area, and extracting the geometric information of the character through these features. The morphological features include the aspect ratio, stroke density, and edge thickness of the character, which are used for further character recognition. Then, a deep neural network (such as CNN) is applied to train and recognize the extracted morphological features. The input of the neural network is a character image of 256×256 pixels, and the output is a predefined character category. The network structure is a 5-layer convolution layer and a 2-layer fully connected layer. The size of the convolution kernel is 3×3, the pooling layer uses maximum pooling, and the pooling window is 2×2. The output of CNN is classified by the softmax activation function, and the number of character categories is 36 (including numbers and letters). After the character recognition result is generated, its accuracy is evaluated by cross-validation, and the recognition accuracy threshold is set to 95% to ensure the reliability of the final recognition result.

[0090] Step S5: Perform cross-character semantic compensation on the character recognition result to obtain a semantically compensated license plate character vector, and perform multi-target cross-validation on the adjacent semantically compensated license plate characters to obtain a license plate recognition result.

[0091] In this embodiment, cross-character semantic compensation is performed on the character recognition results. The statistical distribution characteristics of the character sequence are used to calculate the arrangement pattern between characters, and they are matched in combination with the confidence in the character recognition results. Specifically, common character combination patterns are identified by calculating the co-occurrence probability between characters (such as based on the n-gram model). The co-occurrence probability is calculated by the sliding window method, and the window size is set to 5 to calculate the joint probability between characters in the same license plate image. The character probability combination confidence score calculated based on the co-occurrence probability is used to evaluate the semantic consistency of the characters. For low-confidence characters, the confidence threshold is set to 0.7. Characters below this value will be marked as low-confidence characters and enter the semantic correction stage. Next, the character semantic association matrix is used to correct the low-confidence characters. The construction of the semantic association matrix is based on the similarity measure of characters. The cosine similarity is used to calculate the similarity between characters. The dimension of the matrix is 36×36. For each low-confidence character, the most similar character is selected for replacement by comparing its similarity to other characters. The corrected character is then compared with the character's morphological features to ensure visual similarity, resulting in the final semantically compensated license plate character vector. The accuracy of the license plate recognition results is ensured through multi-objective cross-validation. K-fold cross-validation is used in this cross-validation, with a K value of 5 to ensure robustness and stability. The final license plate recognition result is output as a license plate character sequence.

[0092] Optionally, step S1 specifically includes:

[0093] Step S11: acquiring regional sensor data and a camera image set, and performing data preprocessing on the regional sensor data and the camera image set respectively to obtain the regional sensor data to be analyzed and the camera image set to be analyzed;

[0094] In this embodiment, data is collected through a variety of sensors (such as temperature and humidity sensors, light sensors, air pressure sensors, camera sensors including inertial measurement units, etc.), and the collection frequency is set to 10 times per second. The data range covers the temperature (10°C to 40°C), humidity (10%RH to 90%RH), and light intensity (0 to 1000lux) of the environment. The camera image set comes from a high-definition camera with a resolution of 1920×1080. The image format of the collected images is JPEG, and the number of images is 1000. In order to remove the noise in the original data, the regional sensor data is first denoised, and the Kalman filter is used for data smoothing. The process noise covariance matrix of the filter is set to 0.01, and the measurement noise covariance matrix is set to 1. For the camera image set, Gaussian blur is used for preprocessing, the kernel size is set to 5×5, and the standard deviation is 1.5 to remove high-frequency noise in the image. After data preprocessing, the regional sensor data to be analyzed and the camera image set to be analyzed are obtained.

[0095] Step S12: classifying the sensor data of the area to be analyzed into sensor types to obtain environmental sensor data and camera sensor data;

[0096] In this embodiment, the sensor data of the area to be analyzed is first classified by sensor type. By analyzing the timestamp and sensor ID of the data, the collected sensor data is divided into two categories: environmental sensor data (temperature, humidity, air pressure) and camera sensor data (light intensity and exposure time captured by the camera, etc.). In order to achieve automatic classification, support vector machine (SVM) is used for classification. The kernel function of SVM selects Gaussian kernel, the C parameter is set to 1, and the γ value is set to 0.5. The data is classified by the trained model, and the sensor data is finally divided into two categories: environmental sensor data and camera sensor data, each category of data contains 1000 samples. Environmental sensor data and camera sensor data will be used for subsequent image processing and sensor data analysis respectively to ensure the efficiency and accuracy of data processing.

[0097] Step S13: performing illumination contrast compensation on the camera image set to be analyzed according to the environmental sensing data to generate an illumination compensated image set;

[0098] In this embodiment, illumination contrast compensation is performed based on environmental sensing data (especially illumination intensity data). The exposure of the image is estimated by the illumination intensity information in the environmental sensing data. Assuming that the illumination intensity range is 0 to 1000 lux, if the illumination intensity is lower than 200 lux, it is considered that the image has a problem of insufficient illumination, and if it is higher than 800 lux, it is considered that the image has a problem of excessive illumination. Based on this information, illumination compensation is performed on the camera image set to be analyzed. The compensation method is based on histogram equalization technology, which optimizes the contrast of the image by adjusting the pixel value distribution of the image. During the image processing process, the minimum contrast value is set to 1.2 and the maximum contrast value is set to 3.0 to adjust the brightness and contrast of the image. The compensated image set ensures that the key detail areas of the image are enhanced in the case of uneven illumination by applying a local contrast adjustment algorithm, and finally obtains an illumination compensated image set, and the image brightness uniformity is improved by about 25%.

[0099] Step S14: performing camera distortion correction on the illumination compensation image set based on the camera sensing data to obtain a distortion corrected image set;

[0100] In this embodiment, camera distortion correction is performed on the image set after illumination compensation. First, the distortion model of the camera is estimated by the distortion parameters (including focal length, principal point position, and distortion coefficient) in the camera sensing data. Assuming that the distortion coefficients of the camera are k1=-0.2, k2=0.1, p1=0.05, and p2=-0.05, these coefficients can be obtained by camera calibration. Correction is performed using the pinhole model and the radial distortion model, and the least squares method is used to optimize the distortion parameters. The reprojection error during the correction process is set to less than 0.5 pixels. During the specific correction, the image is subjected to an inverse distortion transformation, and the correct geometric shape of the image is restored through an affine transformation. The corrected image set is distortion-free, and the straight edges and perspective effects of the image are restored.

[0101] Step S15: Estimate motion blur and deformation errors according to the distortion-corrected image set, and perform inverse deformation compensation based on the estimation results to obtain an environment-compensated image set.

[0102] In this embodiment, the motion blur and deformation error in the image are estimated based on the distortion-corrected image set. Motion blur is caused by the relative movement of the camera during the shooting process, and the deformation error is the distortion of the image shape caused by environmental factors (such as air flow, object movement, etc.). In order to estimate motion blur, the image is analyzed using a frequency domain-based Wiener filtering method. Let the original clear image be I(x,y) and the blurred image be B(x,y). The blurring process can be modeled as convolution: B(x,y)=I(x,y)*H(x,y)+N(x,y); where H(x,y) is the motion blur kernel and N(x,y) is noise. Perform a two-dimensional Fourier transform (FFT) on the blurred image to obtain the frequency spectrum: F B (u,v)=F I (u,v)·F H (u,v)+F N (u, v). By analyzing the directionality of the Fourier spectrum, the direction angle and length of the blur are estimated: the direction angle θ is searched in the range of 0°-180°, with a step size of 1°. The direction of maximum energy concentration is detected by Radon transform, and the motion direction θ = 47° is obtained. The motion blur length L is estimated using the Fourier spectrum autocorrelation method, and L is set to 8 pixels. In the frequency domain, deconvolution is performed using Wiener filtering to restore the clear image: Where K = 0.001 is a regularization parameter to prevent noise amplification. The deformation error is estimated using a matching algorithm based on image transformation. The image is divided into several small blocks, and the deformation degree of each block is calculated by block matching. The deformation error threshold is set to 0.05, and the area exceeding this value will be compensated. You can also select two adjacent frames I t and I t+1, use the Farneback optical flow algorithm to calculate the pixel motion vector: V(x,y) = (u(x,y),v(x,y)). Where u(x,y) and v(x,y) represent the pixel offset in the horizontal and vertical directions, respectively. Set the window size to 15×15, the sampling step to 2 pixels, and the maximum number of iterations to 20. Calculate the local deformation intensity using the optical flow field: Set the deformation threshold T d =5 pixels, if D(x,y)>T d It is believed that there is deformation error in this area. The deformation inverse compensation adopts the image-based inverse diffusion compensation method. Through iterative calculation, the anisotropic diffusion equation can be used to restore the image: The diffusion coefficient c(x,y) is determined by the motion blur direction and the deformation error intensity: Where σ = 3 is the smoothing coefficient. The maximum number of iterations is set to 15, with a step size of 0.01. The image gradient change after each iteration is calculated until the gradient stabilizes (gradient change rate < 0.001). Ultimately, through iterative compensation, motion blur and deformation errors are eliminated, resulting in a high-quality set of environmentally compensated images.

[0103] Optionally, step S14 is specifically as follows:

[0104] Step S141: extracting camera posture features based on camera sensor data corresponding to the camera image set to obtain camera posture feature data;

[0105] In this embodiment, the camera sensor data includes three-axis gyroscope data, accelerometer data, and magnetometer data to obtain the camera's posture information in three-dimensional space. Specifically, the image frames collected by the camera are first time-synchronized, and the accelerometer and gyroscope data are fused using a Kalman filter to reduce sensor noise interference. Then, the rotation matrix of the camera relative to the reference coordinate system is calculated. and Thus, the camera posture feature data P = (R, T) is generated. R is the rotation matrix of the camera, which is used to describe the rotation state of the camera in three-dimensional space. It converts the point in the world coordinate system to the camera coordinate system. T is the translation vector of the camera, which is used to describe the position of the camera in three-dimensional space. It represents the coordinates of the origin of the world coordinate system in the camera coordinate system. The rotation matrix R is obtained by the quaternion q = (q w ,q x ,q y ,q z ) is converted to Euler angle representation, satisfying:

[0106] Step S142: obtaining camera calibration data, and correlating the camera calibration data with camera posture feature data to integrate camera imaging parameters to obtain a camera imaging parameter matrix;

[0107] In this embodiment, the camera internal parameter matrix K is obtained through the camera management platform and the Zhang Zhengyou calibration method, which is in the form of where f x , f y is the focal length, c x , c y are the optical center coordinates. Next, the camera's relative transformation matrix H is calculated under multiple viewing angles using a feature matching method (such as ORB or SIFT). Combined with the pose feature data P, a nonlinear least squares solution is performed using the Levenberg-Marquardt optimization method to reduce cumulative error. Finally, the calibration data and pose feature data are integrated to obtain the camera imaging parameter matrix C, which is expressed as: C = K[R|T]; where [R|T] is the extrinsic parameter matrix.

[0108] Step S143: performing a preliminary perspective transformation on the illumination-compensated image set, extracting global feature points and local feature points, and estimating the perspective distortion error based on the feature point matching results of the global feature points and the local feature points to obtain a perspective distortion parameter matrix;

[0109] In this embodiment, a preliminary perspective transformation is performed on the illumination compensation image set, and bilinear interpolation is used to resample the image to obtain a normalized perspective image. Subsequently, the global key point set G = {g1, g2, ..., g m} and local key point set L={l1,l2,…,l n The FLANN (Fast Library for Approximate Nearest Neighbors) matching algorithm is used to calculate the global feature point matching pairs, and the RANSAC (Random Sample Consensus) algorithm is used to remove abnormal matching points to obtain the homography matrix: The feature point drift is calculated by the optical flow method, and the perspective distortion error is calculated by combining the homography matrix H to construct the perspective distortion parameter matrix: Among them, d x , d y is the perspective distortion scaling factor, t x , t y is the pixel displacement correction amount.

[0110] Step S144: performing nonlinear optical compensation pixel resampling based on the perspective distortion parameter matrix and the camera imaging parameter matrix, and performing pixel mapping on the illumination compensation image set according to the compensation pixels to obtain a preliminary correction image set;

[0111] In this embodiment, the nonlinear optical compensation pixel resampling is performed using the view distortion parameter matrix D and the camera imaging parameter matrix C. First, the pixel coordinate transformation after distortion correction is calculated: p′=CDp; where p is the original pixel coordinate, p ′ ∈ R⁻¹ are the compensated pixel coordinates. B-Spline interpolation is used for pixel mapping to transform the illumination-compensated image set into the new coordinate system, generating a preliminary corrected image set. To reduce boundary artifacts, Gaussian kernel convolution is used for edge smoothing, with a kernel size of 5×5.

[0112] Step S145: performing local gradient direction sharpening on the preliminary corrected image set to obtain a distortion corrected image set.

[0113] In this embodiment, local gradient direction sharpening is performed on the preliminary corrected image set. First, the corrected image is subjected to edge detection using the Sobel operator: The gradient magnitude is then calculated: Combined with Laplace filtering for sharpening: Among them, λ is set to 0.5 to avoid over-sharpening. Finally, a set of distortion-corrected images is generated.

[0114] Optionally, step S15 is specifically as follows:

[0115] Step S151: performing motion blur detection on the distortion-corrected image set based on the distortion-corrected image set to obtain a blurred area image set;

[0116] In this embodiment, based on the distortion correction image set I c First, we use gradient direction analysis and Fourier transform to detect motion blur. Gradient direction analysis uses the Sobel operator to calculate the horizontal gradient G of the image. x and vertical gradient G y , calculate the gradient amplitude And construct the gradient direction histogram (GOH). If the gradient ratio in a certain direction is significantly higher than that in other directions, it indicates that there may be motion blur in that area. At the same time, perform 2D Fourier transform on the image. Calculate the power spectral density |F(u,v)| 2 If the high-frequency component is significantly attenuated, it is further confirmed that the area is blurred. Combining the gradient and spectrum analysis results, the Otsu threshold method is used to generate the blurred area mask M b And optimize the boundary through morphological closing operation to extract the fuzzy area image set I b .

[0117] Step S152: performing deconvolution on the blurred area image set to obtain a blurred repaired image set;

[0118] In this embodiment, for the obtained fuzzy area image set I b , the blind deconvolution method is used for deblurring. First, assume that the blurred image is generated by the motion blur kernel k(x,y), that is, I b =I s *k+n, where I s is the clear image, and n is the noise. The maximum a posteriori estimation (MAP) is combined with the Laplace sparse prior to solve K(x,y). In the deblurring process, the Richardson-Lucy deconvolution algorithm is used, and the iterative calculation is performed until the residual error converges. Finally, the non-local mean (NLM) denoising is used to reduce the high-frequency noise introduced in the deblurring process, and the blurred repaired image set I is obtained. r .

[0119] Step S153: performing deformation error estimation on the blurred inpainted image set, identifying geometric distortions between the blurred inpainted images, and thereby generating a deformation error matrix;

[0120] In this embodiment, since motion blur removal may cause slight deformation of the geometric structure of certain local areas, it is necessary to perform deformation error analysis on the repaired image. First, the scale-invariant feature transform (SIFT) is used to extract the key feature points of the image, and the feature point matching relationship is calculated. If the feature point matching deviation in certain areas of the blurred repaired image is large, it means that there is geometric distortion in the area. Secondly, the optical flow method is used to analyze the motion trajectory of the pixel points in the image sequence. If the optical flow field jumps violently or is discontinuous in certain areas, it indicates that the area may be affected by the deformation. In order to further quantify the deformation error, a deformation error matrix E is constructed, where E i,j Represents the degree of deformation of the image at the pixel (i, j). This matrix is obtained by calculating the comprehensive weight of the feature point offset, optical flow gradient change and local perspective error: where p i,j , p′ i,j are the positions of the feature points before and after deformation, is the optical flow gradient change, H i,j is the local perspective transformation matrix, I is the identity matrix, and α, β, and γ are weight coefficients. The resulting deformation error matrix is used to guide subsequent geometric correction.

[0121] Step S154: performing inverse geometric correction of deformation on the blurred inpainted image set according to the deformation error matrix, reconstructing the geometrically distorted region in the blurred inpainted image set, and obtaining a deformation corrected image set;

[0122] In this embodiment, geometric correction is performed on the affected image area based on the deformation error matrix. First, the inverse transformation mapping of the deformed area is calculated using bilinear interpolation and least squares optimization methods to restore the distorted area to its original geometric form. Then, the global deformed area is adjusted as a whole using affine transformation, and texture fusion is performed using the Poisson equation (Poisson Editing) to keep the repaired area consistent with the surrounding environment. For areas with larger deformation errors, a deep learning model (such as an image restoration network based on U-Net) is further used to generate more accurate geometric correction results. Finally, after inverse geometric correction processing, a set of repaired deformation-corrected images is obtained to ensure that the geometric structure of the entire image is consistent with the original scene.

[0123] Step S155: performing global color adaptive adjustment on the deformation-corrected image set to obtain an environment-compensated image set.

[0124] In this embodiment, since the image may experience color shift during the blur repair and deformation correction process, color adaptive adjustment is required to ensure that the color of the final image is consistent with the original scene. First, the image after deformation correction is converted to the LAB color space, the color histogram of the image is calculated, and matched with the color distribution of the input image to ensure color consistency. Then, the Retinex image enhancement method is used to perform brightness compensation on the local area to enhance the contrast and visibility of the image. In addition, the overall brightness distribution is adjusted based on the Gamma correction method to ensure that both dark and bright details are optimized. During the color adjustment process, the influence of ambient light must also be considered. If the image has obvious color cast (such as bluish or reddish), the adaptive white balance algorithm is used for correction to ensure that the final output image color is natural and close to the real environment. After color adaptive adjustment, an environmentally compensated image set is finally obtained to ensure that the clarity, geometry and color of the image are in the best state.

[0125] Optionally, the adaptive division of the license plate area in step S2 is specifically as follows:

[0126] Perform edge detection on the environmental compensation image set, extract image contour features, and obtain image area contour feature data;

[0127] In this embodiment, the Canny edge detection algorithm is used to extract image edge features for the image set after environmental compensation to obtain the contour information of the objects in the scene. First, the input image is Gaussian filtered with the Gaussian kernel size set to 1.2. Then, the gradient amplitude G and direction θ of the image are calculated: Among them G x and G yThe horizontal and vertical gradients of the image are calculated using the Sobel operator. Then, non-maximum suppression is performed to remove non-edge pixels and a double threshold T is set. low =50 and T high = 150 for edge connection to ensure effective extraction of the contour features of the license plate area. Finally, the image area contour feature data is obtained for subsequent license plate foreground screening.

[0128] The foreground area of the license plate image is screened according to the image area contour feature data to obtain a preliminary license plate candidate area set;

[0129] In this embodiment, based on the extracted image contour feature data, the connected domain analysis method is used to screen the foreground area. First, a morphological closing operation is performed on the binary image after edge detection (the structure element size is set to 5×5) to eliminate small gaps and enhance connectivity. Then, the Flood Fill algorithm is used to mark the connected areas, screen candidate areas with an area greater than 500 pixels, and calculate their minimum enclosing rectangle (MinAreaRect), and record the aspect ratio and direction angle of the rectangle. According to the rectangular characteristics usually presented by license plates, candidate areas with an aspect ratio of 2.0≤R≤6.0 and an angle of -15°≤θ≤15° are screened, and finally a preliminary set of license plate candidate areas is formed.

[0130] Perform morphological filtering on the preliminary license plate candidate region set and calculate the aspect ratio and connectivity features of each candidate region;

[0131] In this example, the selected license plate candidate regions are further subjected to morphological analysis, using morphological closing to further smooth edges and performing skeletonization to analyze regional connectivity. The aspect ratio R of each candidate region is calculated, and the number of connected components C within the region is counted. The following connectivity constraints are set: 3 ≤ C ≤ 15, 2.5 ≤ R ≤ 5.5. If a candidate region meets these requirements, it is retained; otherwise, it is discarded. Ultimately, a set of candidate regions with high connectivity and morphological characteristics consistent with license plate characteristics is obtained.

[0132] Based on the preset license plate size rules, set the size constraints of a single license plate and multiple license plates according to the license plate size rules;

[0133] In this embodiment, license plate size rules are set, taking into account the national standard license plate sizes (blue, yellow, and new energy vehicle license plates). Constraints are established: Single license plate size constraints: license plate width W is within the range of 100 ≤ W ≤ 400 pixels, and license plate height H is within the range of 30 ≤ H ≤ 120 pixels. Multiple license plate size constraints: The standard deviation σ of the dimensions of all license plate regions should be less than 20 to avoid false detections.

[0134] According to the size constraint of a single license plate, the aspect ratio and connectivity characteristics of each candidate region are used to screen the license plate region, and a single license plate candidate region set is obtained;

[0135] In this embodiment, size screening is performed on the candidate regions, and only regions that meet the size constraint of a single license plate and whose aspect ratio satisfies 2.5≤R≤5.0 are retained to obtain a single license plate candidate region set.

[0136] Calculate the regional contrast distribution of the candidate region set of a single license plate to obtain regional gradient change data, and use the regional gradient change data to perform regional superpixel segmentation to obtain a single license plate region image set;

[0137] In this embodiment, the regional contrast is calculated based on Laplace transform: C = ∑ i,j |I i,j -I mean |; where I i,j is the pixel value, I mean is the average brightness of the region. Subsequently, the SLIC superpixel segmentation algorithm is used to divide the candidate license plate region into K = 100 superpixel regions to remove background interference, and finally a single license plate region image set is obtained.

[0138] Connected component analysis is performed based on the aspect ratio and connectivity characteristics of each candidate region, and low-density regions are filtered to obtain a preliminary set of multiple license plate candidate regions;

[0139] In this embodiment, connected component analysis is performed to calculate the density concentration of the license plate area: where N c is the number of connected components, A is the area of the region. The threshold D ≥ 0.005 is set to eliminate low-density areas and obtain a preliminary set of candidate regions for multiple license plates.

[0140] Performing license plate area screening on the preliminary multiple license plate candidate area sets according to the multiple license plate size constraints to obtain multiple license plate area sets;

[0141] In this embodiment, size screening is performed on the preliminary multiple license plate regions, only the regions that meet the size constraints of the multiple license plates are retained, and abnormal license plate regions with size standard deviations exceeding 20 are removed.

[0142] Perform superpixel segmentation on multiple license plate region sets and perform region segmentation density equalization to obtain multiple license plate region image sets;

[0143] In this embodiment, SLIC superpixel segmentation (K=150K) is performed on multiple license plate regions, and then the density is adjusted based on K-means clustering to balance the regional segmentation density, making the license plate region clearer.

[0144] The single license plate region image set and multiple license plate candidate region sets are merged to obtain a license plate region image set.

[0145] In this embodiment, a single license plate region image set is combined with multiple license plate region image sets, and redundant license plate regions are removed through similarity detection based on Hausdorff distance: After merging, the final license plate area image set is obtained.

[0146] Optionally, the image diffusion reverse enhancement in step S2 is specifically:

[0147] The local gradient change of the license plate area image set is calculated to obtain the local gradient change rate of the pixel point, and the diffusion factor range is set to [0,1] to extract the image diffusion factor from the local gradient change rate of the pixel point;

[0148] In this embodiment, for the license plate area image set, the Sobel operator is used to calculate the gradient of the image in the horizontal and vertical directions. The specific calculation method is as follows: G x =∑ i,j I i,j ·[10-1], Then calculate the local gradient change rate of each pixel in the image At this time, for each pixel, the local gradient change rate δ i Represents the edge strength of the point. Next, set the diffusion factor range to [0,1] and extract the image diffusion factor from it: That is, the diffusion factor of each point is calculated based on the ratio of the gradient intensity of the pixel point to the maximum value, providing a basis for subsequent image diffusion calculations.

[0149] The reverse diffusion calculation is performed on the license plate area image set according to the image diffusion factor, and the iterative calculation is performed for [50,100] rounds. The diffusion range is set to [0.1,0.5] to obtain the reverse diffusion image set.

[0150] In this embodiment, image reverse diffusion calculation is performed on the license plate area image set based on the local diffusion factor. The main steps of the reverse diffusion algorithm include: first, setting the diffusion factor in the diffusion range of [0.1, 0.5] and initializing the image I0. Then, the image is iteratively updated through the reverse diffusion model: Where α is the learning rate, is the Laplacian operator of the image, representing the degree of local variation in the image. This process is performed iteratively, with a set number of iterations ranging from 50 to 100, gradually adjusting image details and enhancing edge information in the license plate area. Ultimately, a set of inverse diffusion images is obtained, containing the license plate area image after diffusion inversion.

[0151] Perform edge detection on the reverse diffusion image set, set the regional contrast stretching factor to [1.2.2.0] to calculate the local edge strength, and perform regional contrast nonlinear stretching based on the local edge strength results to obtain the diffusion reverse enhancement image set;

[0152] In this embodiment, the Canny edge detection algorithm is used for the reverse diffusion image set, and the high and low thresholds T are set. low =50 and T high =150, calculate the local edge strength of each pixel. After calculating the edge strength S, use the contrast stretching method to enhance it: S new =α·(SS min )+S min α is the contrast stretching factor, with a setting range of [1.2.2.0]. By adjusting the α value, nonlinear stretching of the local contrast of the image is achieved. This method enhances the image details and highlights the edge information of the license plate, ultimately obtaining a set of diffusion-inverse enhanced images.

[0153] The license plate area image set and the diffusion inverse enhanced image set are adaptively fused. The scale range is set to [3, 5, 7] to calculate the multi-scale clarity score. The image set fusion weight is adjusted according to the multi-scale clarity score result to obtain the license plate area enhanced image set.

[0154] In this embodiment, an adaptive fusion process is performed between the license plate area image set and the diffusion inverse enhancement image set. First, the scale range during fusion is set to [3, 5, 7], and a multi-scale clarity scoring method is used to evaluate the clarity of the image: C score (I)=∑ i,j |I i,j -I mean |; where I mean The image clarity scores at different scales are calculated to reflect the structural information and texture features of the image. Subsequently, the fusion weights of the two image sets are automatically adjusted based on the multi-scale clarity score results. During the fusion process, for each pixel I fusion Computation I fusion =w1·I original +w2·I enhanced Where w1 and w2 are adaptively adjusted fusion weights based on the clarity score. The weights determine the contribution ratio between the original image and the enhanced image. The fused image set further highlights the details and edge information of the license plate area, resulting in an enhanced license plate image set with higher readability and accuracy.

[0155] Optionally, step S3 specifically includes:

[0156] Step S31: performing Sobel operator edge detection on the license plate area enhanced image set, extracting character outline features, and performing morphological closing operation on the character outline features to divide the connected regions and obtain a connected region image set;

[0157] In this embodiment, when edge detection is performed on the license plate area enhanced image set, the Sobel operator is used to extract the edges in the image. The Sobel operator obtains the changes in the horizontal and vertical directions of each pixel by calculating the gradient change of the image. The edge information in the image is obtained by calculating the gradient amplitude. Next, morphological operations are performed on the extracted character contour features. First, a closing operation (i.e., expansion followed by corrosion) is used to fill the edge breaks and small holes to enhance the character contours. The structural element of the closing operation is a 3x3 rectangular kernel. Finally, the character areas in the image are divided through connected region analysis to obtain a connected region image set. At this point, each connected region in the image represents a potential character region.

[0158] Step S32: calculating the aspect ratio, area and rectangle fitting of the connected regions according to the connected region image set, and filtering the character candidate region set according to the calculation results of the aspect ratio, area and rectangle fitting of the connected regions;

[0159] In this embodiment, for each connected region in the connected region image set, the aspect ratio, area, and rectangle fit of the region are first calculated. The aspect ratio is obtained by calculating the ratio of the width W and height H of the region's bounding box: The area is the number of pixels in the region. Next, the minimum enclosing rectangle is used for fitting and the rectangle fit is calculated, usually by calculating the area ratio of the region to the fitted rectangle: Based on these features, we filter by setting a threshold. The aspect ratio is generally set between [2, 5]. Regions with a rectangular fit greater than 0.8 are considered candidate character regions. Ultimately, the selected region set is the character candidate region set.

[0160] Step S33: performing boundary enhancement on the character candidate region set, and performing superpixel segmentation on the boundary-enhanced character candidate region set to extract character region structural features to obtain a character boundary image set;

[0161] In this embodiment, the character candidate region set is first enhanced by edge enhancement. The high-frequency information of the region boundary is highlighted by using an edge enhancement method based on the Laplace operator. The Laplace operator enhances the edge by calculating the second-order derivative of the pixel point: The enhanced boundaries are segmented into superpixels using the SLIC (Simple Linear Iterative Clustering) algorithm. The SLIC algorithm divides the image into multiple superpixel blocks of uniform size and shape, aiming to reduce computational effort while maintaining pixel consistency within the region. This method generates each superpixel block representing a structural feature in the image, ultimately forming a set of character boundary images that capture the character outlines and regional details.

[0162] Step S34: integrating the geometric features of the character region based on the character boundary image set to obtain the character region outline system features, and performing spatial relationship topological structure modeling based on the character region outline system features to obtain the character region topological structure model;

[0163] In this embodiment, for each character region in the character boundary image set, its geometric features are first extracted, such as the center position of the region, boundary curvature, area, etc. By integrating these geometric features, the contour system features of the character region are obtained. This contour system feature can be expressed in a parameterized form, such as: contour ={x center ,y center ,curvature,area,…}; then, spatial relationships are modeled based on these features. A topological model is used to describe the spatial relationships between character regions, such as adjacency and inclusion, to construct a character region topological structure model. In this model, each character region is considered a node in the topological graph, and the spatial relationships between characters are connected by edges. Topological modeling calculates the geometric distance and relative position between nodes to obtain the topological relationship between each pair of character regions.

[0164] Step S35: performing high-dimensional feature projection on the character region topology structure model to obtain character region topology mapping data, and performing character segmentation based on the character region topology mapping data to obtain a character segmentation matrix.

[0165] In this embodiment, based on the character region topological structure model, methods such as principal component analysis (PCA) are used to perform high-dimensional feature projection. By reducing the dimensionality, a low-dimensional embedding representation of each character region is obtained, namely, the character region topological mapping data. These data represent the relative position and structural characteristics of the character region in the topological space: PCA projected =PCA(F contour ); based on this topological mapping data, character segmentation is performed. The topological mapping data is analyzed using a clustering algorithm (such as K-means or DBSCAN) to identify the boundaries between character regions. Finally, the segmentation results are used to generate a character segmentation matrix that accurately represents the boundaries of each character region, facilitating the subsequent license plate recognition process.

[0166] Optionally, the cross-character semantic compensation in step S5 is specifically:

[0167] According to the statistical distribution characteristics of the character sequence in the character recognition results, the arrangement pattern threshold is set to 0.8 to identify the arrangement pattern between characters and obtain the character arrangement pattern data;

[0168] In this embodiment, the character sequence in the character recognition result is statistically analyzed to extract the arrangement and distribution characteristics of the characters. These characteristics include the distance between characters, the relative positions of characters, and the arrangement patterns of characters in the license plate. By calculating the relative position relationship between characters, a statistical distribution model of the character sequence is constructed. Specifically, the arrangement pattern can be captured by calculating the horizontal and vertical distances between characters, as well as the relative order of characters. Setting the arrangement pattern threshold to 0.8 means that when the similarity between the arrangement pattern between characters and the threshold is higher than 80%, it is regarded as a valid character arrangement pattern. The output of this process is character arrangement pattern data, which represents the arrangement pattern in the character sequence and can be used for subsequent character matching and segmentation.

[0169] The co-occurrence window size is set to 3 to 7 characters, the co-occurrence probability between characters is calculated for the character arrangement pattern data, and based on the co-occurrence probability calculation result, the confidence score range is set to [0, 1] to estimate the character probability combination confidence score;

[0170] In this embodiment, the co-occurrence window size is set to 3 to 7 characters, that is, the co-occurrence probability between them is calculated by considering the continuous arrangement of 3 to 7 characters. By sliding the window on the character arrangement pattern data, the co-occurrence frequency of each character at different positions is counted, and the co-occurrence probability is calculated based on the frequency. The co-occurrence probability formula is as follows: Where count(c1,c2) is the number of times characters c1 and c2 co-occur in the window, and totalcount is the total number of times all character pairs appear. Based on the calculated co-occurrence probability, the confidence score range is set to [0,1], and the confidence of the character probability combination is calculated using the following formula: Where N is the number of characters in the window. This score is used to evaluate the rationality and consistency of the character arrangement, and ultimately generates a character probability combined confidence score for subsequent character semantic matching.

[0171] Perform character semantic matching on the character recognition results based on the character probability combined confidence score, filter and mark low-confidence characters, and obtain a character semantic association matrix;

[0172] In this embodiment, character semantic matching is performed based on the confidence score of the character probability combination. For each character combination, its confidence score is compared with the set threshold. If the confidence is lower than the set value (for example, 0.5), the character combination is regarded as a low-confidence combination and needs to be further screened and corrected. The low-confidence character combination is analyzed through the semantic association model between characters to identify possible mismatches between characters and mark them as low-confidence characters. Through this method, possible erroneous character combinations in the license plate recognition process can be effectively identified. Finally, a character semantic association matrix is generated, which contains the similarity and semantic relationship between characters, indicating which character combinations need further correction.

[0173] Combined with the character morphological features, the morphological correction threshold is set to 0.6 to correct the misrecognition of visually similar characters for the low-confidence characters in the character semantic association matrix, and the semantically compensated license plate character vector is obtained.

[0174] In this embodiment, the morphological correction method is used to correct the misidentification of visually similar characters with low confidence characters in combination with the morphological features of the characters. First, the geometric morphological features of the characters (such as aspect ratio, angle, symmetry, etc.) are used to perform similarity analysis. By calculating the Euclidean distance between the morphological features of the characters, other characters that are morphologically similar to the low-confidence characters are identified. The morphological correction threshold is set to 0.6, which means that when the morphological similarity of the characters is greater than 60%, they are considered to be the same characters, and thus misidentification correction is performed. For example, if the character "S" has a high morphological similarity with the character "5", it is regarded as a possible misidentification and is replaced or adjusted. Finally, a semantically compensated license plate character vector is generated, which contains the corrected license plate character information, which helps to improve the accuracy of license plate recognition.

[0175] Optionally, this specification further provides a license plate recognition system based on image technology, which is used to execute the license plate recognition method based on image technology as described above. The license plate recognition system based on image technology includes:

[0176] An environmental compensation module is used to obtain regional sensing data and a camera image set, and to compensate for deformation effects of the camera image set based on the regional sensing data to obtain an environmental compensation image set;

[0177] The license plate area enhancement module is used to perform adaptive license plate area division on the environment compensation image set to obtain a license plate area image set; perform image diffusion reverse enhancement on the license plate area image set to obtain a license plate area enhanced image set;

[0178] The character segmentation module is used to perform character boundary morphology perception based on the license plate area enhanced image set to obtain a character boundary image set; perform high-dimensional topological mapping of the character area based on the character boundary image set to obtain a character segmentation matrix;

[0179] The character recognition module is used to extract character morphological features according to the character segmentation matrix, and perform character recognition on the character morphological features to obtain character recognition results;

[0180] The license plate recognition result verification module is used to perform cross-character semantic compensation on the character recognition results to obtain the semantically compensated license plate character vector, and perform multi-target cross-validation on the adjacent semantically compensated license plate characters to obtain the license plate recognition result.

[0181] Optionally, this specification also provides a computer-readable storage medium having a computer program stored thereon, which implements the license plate recognition method based on image technology as described above when the computer program is executed.

[0182] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0183] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A license plate recognition method based on image technology, characterized in that: The following steps are involved: Step S1: Acquire regional sensing data and a camera image set, and perform deformation effect compensation on the camera image set based on the regional sensing data to obtain an environment-compensated image set; Step S2: performing adaptive license plate area division on the environment compensation image set to obtain a license plate area image set; Perform image diffusion reverse enhancement on the license plate area image set to obtain the license plate area enhanced image set; Step S3: performing character boundary morphology perception based on the license plate region enhanced image set to obtain a character boundary image set; performing high-dimensional topological mapping of the character region based on the character boundary image set to obtain a character segmentation matrix; Step S4: extracting character morphological features according to the character segmentation matrix, and performing character recognition on the character morphological features to obtain a character recognition result; Step S5: Perform cross-character semantic compensation on the character recognition result to obtain a semantically compensated license plate character vector, and perform multi-target cross-validation on the adjacent semantically compensated license plate characters to obtain a license plate recognition result.

2. The license plate recognition method based on image technology according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: acquiring regional sensor data and a camera image set, and performing data preprocessing on the regional sensor data and the camera image set respectively to obtain the regional sensor data to be analyzed and the camera image set to be analyzed; Step S12: classifying the sensor data of the area to be analyzed into sensor types to obtain environmental sensor data and camera sensor data; Step S13: performing illumination contrast compensation on the camera image set to be analyzed according to the environmental sensing data to generate an illumination compensated image set; Step S14: performing camera distortion correction on the illumination compensation image set based on the camera sensing data to obtain a distortion corrected image set; Step S15: Estimate motion blur and deformation errors according to the distortion-corrected image set, and perform inverse deformation compensation based on the estimation results to obtain an environment-compensated image set.

3. The license plate recognition method based on image technology according to claim 2, characterized in that: Step S14 is specifically as follows: Step S141: extracting camera posture features based on camera sensor data corresponding to the camera image set to obtain camera posture feature data; Step S142: obtaining camera calibration data, and correlating the camera calibration data with camera posture feature data to integrate camera imaging parameters to obtain a camera imaging parameter matrix; Step S143: performing a preliminary perspective transformation on the illumination-compensated image set, extracting global feature points and local feature points, and estimating the perspective distortion error based on the feature point matching results of the global feature points and the local feature points to obtain a perspective distortion parameter matrix; Step S144: performing nonlinear optical compensation pixel resampling based on the perspective distortion parameter matrix and the camera imaging parameter matrix, and performing pixel mapping on the illumination compensation image set according to the compensation pixels to obtain a preliminary correction image set; Step S145: performing local gradient direction sharpening on the preliminary corrected image set to obtain a distortion corrected image set.

4. The license plate recognition method based on image technology according to claim 2, characterized in that: Step S15 is specifically as follows: Step S151: performing motion blur detection on the distortion-corrected image set based on the distortion-corrected image set to obtain a blurred area image set; Step S152: performing deconvolution on the blurred area image set to obtain a blurred repaired image set; Step S153: performing deformation error estimation on the blurred inpainted image set, identifying geometric distortions between the blurred inpainted images, and thereby generating a deformation error matrix; Step S154: performing inverse geometric correction of deformation on the blurred inpainted image set according to the deformation error matrix, reconstructing the geometrically distorted region in the blurred inpainted image set, and obtaining a deformation corrected image set; Step S155: performing global color adaptive adjustment on the deformation-corrected image set to obtain an environment-compensated image set.

5. The license plate recognition method based on image technology according to claim 1, characterized in that: The adaptive division of the license plate area described in step S2 is specifically as follows: Perform edge detection on the environmental compensation image set, extract image contour features, and obtain image area contour feature data; The foreground area of the license plate image is screened according to the image area contour feature data to obtain a preliminary license plate candidate area set; Perform morphological filtering on the preliminary license plate candidate region set and calculate the aspect ratio and connectivity features of each candidate region; Based on the preset license plate size rules, set the size constraints of a single license plate and multiple license plates according to the license plate size rules; According to the size constraint of a single license plate, the aspect ratio and connectivity characteristics of each candidate region are used to screen the license plate region, and a single license plate candidate region set is obtained; Calculate the regional contrast distribution of the candidate region set of a single license plate to obtain regional gradient change data, and use the regional gradient change data to perform regional superpixel segmentation to obtain a single license plate region image set; Connected component analysis is performed based on the aspect ratio and connectivity characteristics of each candidate region, and low-density regions are filtered to obtain a preliminary set of multiple license plate candidate regions; Performing license plate area screening on the preliminary multiple license plate candidate area sets according to the multiple license plate size constraints to obtain multiple license plate area sets; Perform superpixel segmentation on multiple license plate region sets and perform region segmentation density equalization to obtain multiple license plate region image sets; The single license plate region image set and multiple license plate candidate region sets are merged to obtain a license plate region image set.

6. The license plate recognition method based on image technology according to claim 1, characterized in that: The image diffusion reverse enhancement described in step S2 is specifically as follows: The local gradient change of the license plate area image set is calculated to obtain the local gradient change rate of the pixel point, and the diffusion factor range is set to [0,1] to extract the image diffusion factor from the local gradient change rate of the pixel point; The reverse diffusion calculation is performed on the license plate area image set according to the image diffusion factor, and the iterative calculation is performed for [50,100] rounds. The diffusion range is set to [0.1,0.5] to obtain the reverse diffusion image set. Perform edge detection on the reverse diffusion image set, set the regional contrast stretching factor to [1.2.2.0] to calculate the local edge strength, and perform regional contrast nonlinear stretching based on the local edge strength results to obtain the diffusion reverse enhancement image set; The license plate area image set and the diffusion inverse enhanced image set are adaptively fused. The scale range is set to [3, 5, 7] to calculate the multi-scale clarity score. The image set fusion weight is adjusted according to the multi-scale clarity score result to obtain the license plate area enhanced image set.

7. The license plate recognition method based on image technology according to claim 1, characterized in that: Step S3 is specifically as follows: Step S31: performing Sobel operator edge detection on the license plate area enhanced image set, extracting character outline features, and performing morphological closing operation on the character outline features to divide the connected regions and obtain a connected region image set; Step S32: calculating the aspect ratio, area and rectangle fitting of the connected regions according to the connected region image set, and filtering the character candidate region set according to the calculation results of the aspect ratio, area and rectangle fitting of the connected regions; Step S33: performing boundary enhancement on the character candidate region set, and performing superpixel segmentation on the boundary-enhanced character candidate region set to extract character region structural features to obtain a character boundary image set; Step S34: integrating the geometric features of the character region based on the character boundary image set to obtain the character region outline system features, and performing spatial relationship topological structure modeling based on the character region outline system features to obtain the character region topological structure model; Step S35: performing high-dimensional feature projection on the character region topology structure model to obtain character region topology mapping data, and performing character segmentation based on the character region topology mapping data to obtain a character segmentation matrix.

8. The license plate recognition method based on image technology according to claim 1, characterized in that: The cross-character semantic compensation described in step S5 is specifically as follows: According to the statistical distribution characteristics of the character sequence in the character recognition results, the arrangement pattern threshold is set to 0.8 to identify the arrangement pattern between characters and obtain the character arrangement pattern data; The co-occurrence window size is set to 3 to 7 characters, the co-occurrence probability between characters is calculated for the character arrangement pattern data, and based on the co-occurrence probability calculation result, the confidence score range is set to [0, 1] to estimate the character probability combination confidence score; Perform character semantic matching on the character recognition results based on the character probability combined confidence score, filter and mark low-confidence characters, and obtain a character semantic association matrix; Combined with the character morphological features, the morphological correction threshold is set to 0.6 to correct the misrecognition of visually similar characters for the low-confidence characters in the character semantic association matrix, and the semantically compensated license plate character vector is obtained.

9. A license plate recognition system based on image technology, characterized in that: For executing the license plate recognition method based on image technology as claimed in claim 1, the license plate recognition system based on image technology comprises: An environmental compensation module is used to obtain regional sensing data and a camera image set, and to compensate for deformation effects of the camera image set based on the regional sensing data to obtain an environmental compensation image set; The license plate area enhancement module is used to perform adaptive license plate area division on the environment compensation image set to obtain a license plate area image set; perform image diffusion reverse enhancement on the license plate area image set to obtain a license plate area enhanced image set; The character segmentation module is used to perform character boundary morphology perception based on the license plate area enhanced image set to obtain a character boundary image set; perform high-dimensional topological mapping of the character area based on the character boundary image set to obtain a character segmentation matrix; The character recognition module is used to extract character morphological features according to the character segmentation matrix, and perform character recognition on the character morphological features to obtain character recognition results; The license plate recognition result verification module is used to perform cross-character semantic compensation on the character recognition results to obtain the semantically compensated license plate character vector, and perform multi-target cross-validation on the adjacent semantically compensated license plate characters to obtain the license plate recognition result.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed, the license plate recognition method based on image technology as described in any one of claims 1 to 8 is implemented.

Images