A method, system, device and storage medium for vehicle nameplate recognition

Abstract

This invention discloses a vehicle nameplate recognition method, system, device, and storage medium. The method includes: dividing the input image into multi-scale gridded partitions and extracting features; establishing a regional feature map to determine the imaging attribute category; performing differentiated image enhancement and fusing the images to generate a preprocessed image; constructing a multi-hypothesis verification model based on the geometric attributes and layout relationship of the vehicle identification number for rotation correction; constructing a perspective transformation model using text line geometric constraints; correcting spatial deformation through adaptive block perspective transformation and affine transformation; outputting structured information using a mechanism of collaborative decision-making between general recognition and a dedicated engine; and finally performing image feature generation, duplication verification, semantic consistency verification, and blockchain notarization. This invention effectively solves the problems of poor image quality, difficult geometric deformation correction, and low recognition accuracy of vehicle nameplates in complex imaging environments, achieving end-to-end high-precision recognition from low-quality images to reliable structured information.

Description

Technical Field

[0001] This application relates to the field of computer vision and optical character recognition technology, specifically to a vehicle nameplate recognition method, system, device, and storage medium. Background Technology

[0002] As the core carrier of vehicle identity, the vehicle identification plate records detailed information such as the Vehicle Identification Number (VIN), engine number, vehicle model, manufacturer, production date, and key technical parameters. In numerous industrial and commercial scenarios, including intelligent manufacturing, vehicle transactions, vehicle management, annual vehicle inspections, and parts traceability, the automated and high-precision collection and identification of identification plate information has significant application value and urgent market demand.

[0003] However, in practical applications, traditional optical character recognition (OCR) solutions often fall short of actual requirements when dealing with vehicle nameplate images, mainly due to the following long-standing technical bottlenecks that have not been effectively resolved:

[0004] The imaging environment is complex: vehicle nameplates are usually installed in narrow and poorly lit locations such as engine compartments, resulting in images with severe uneven lighting, local shadows, metallic surface reflections, and low overall contrast. These factors significantly reduce the distinguishability of text areas from the background, directly affecting the accuracy of subsequent text detection and recognition.

[0005] Image orientation determination is difficult: Due to the limited shooting angle, the captured nameplate images are prone to rotation at any angle, or even inversion. Traditional OCR systems usually rely on a general orientation classifier to determine the image orientation, but in special cases such as when the nameplate image is vertical, inverted, or tilted at a large angle, this classifier is prone to failure, causing the entire recognition process to fail due to orientation errors.

[0006] Challenges in Spatial Deformation Correction: Due to the shooting angle and the potential curved surface installation of the nameplate itself, text in images often exhibits complex perspective distortion and nonlinear surface deformation. Existing technologies (such as line detection and correction methods based on Hough transform) struggle to accurately model and effectively correct this non-global, irregular spatial distortion, leading to character shape distortion and recognition difficulties.

[0007] Diverse text layouts and difficulty in locating key information: The design of nameplates from different manufacturers and car models lacks a unified standard, and the placement of key information (such as VIN codes) is inconsistent. A single, rigid OCR processing strategy is insufficient to adaptively locate and recognize this key text information scattered across different layouts.

[0008] Most existing technical solutions focus on local optimization of the aforementioned single problem, such as enhancing image contrast through histogram equalization or using Hough transform for line-based rotation correction. These methods have significant shortcomings: first, they lack overall adaptability to complex imaging conditions, resulting in a rigid processing flow; second, they fail to achieve end-to-end deep integration of image preprocessing, geometric correction, and semantic information recognition. This leads to significant deficiencies in the final recognition rate, robustness, and practicality.

[0009] In view of this, the present invention proposes a vehicle nameplate recognition method, system, device and storage medium, which can achieve end-to-end high-precision recognition from low-quality images to structured information. Summary of the Invention

[0010] To address the problems of poor adaptability to imaging environments, inaccurate geometric deformation correction, reliance on a single classifier for orientation determination, and difficulty in locating key information in complex layouts in existing technologies, this invention provides a vehicle nameplate recognition method, system, device, and storage medium to solve the aforementioned technical deficiencies.

[0011] In a first aspect, the present invention proposes a vehicle nameplate recognition method, comprising the following steps:

[0012] S1. Divide the input image into multi-scale grids, extract the brightness, contrast and texture features of each grid region, establish a regional feature map based on the extracted features, determine the imaging attribute category of each region according to the regional feature map, perform differential image enhancement according to the imaging attribute category, and generate a preprocessed image through adaptive weight fusion.

[0013] S2. Use a text detection algorithm to obtain the position information of all text in the preprocessed image, locate the vehicle identification number from the detected text, construct a multi-hypothesis verification model based on the geometric properties of the vehicle identification number and its layout relationship with adjacent text to calculate the global deflection angle of the image, and perform rotation correction on the preprocessed image according to the global deflection angle to generate a rotation-corrected image.

[0014] S3. Construct a perspective transformation model based on the geometric constraints of text lines in the rotation-corrected image, use adaptive block perspective transformation to correct plane tilt, restore the horizontal state of the text through affine transformation, and generate a geometrically corrected image.

[0015] S4. A dual-engine collaborative decision-making mechanism is used to identify geometrically corrected images and output structured information.

[0016] S5. For the geometrically corrected image that has been identified, generate an image feature vector and perform a repeatability check. For images that have not been repeated after verification, verify the semantic consistency between their image content and the structured information. Generate a combined hash value for the geometrically corrected image and the structured information that have passed the verification and register it in the blockchain. Finally, associate the geometrically corrected image, the structured information and the blockchain transaction number and store them in the database.

[0017] Preferably, step S1 specifically includes the following sub-steps:

[0018] S11. Perform multi-scale grid partitioning on the input image, dividing the input image into multiple grid units;

[0019] S12. Extract the mean brightness, standard deviation of contrast, and gray-level co-occurrence matrix texture features of each grid cell;

[0020] S13. Integrate the features of all grid cells to generate a regional feature map that represents the global and local characteristics of the image;

[0021] S14. Based on the regional characteristic map, determine the imaging attribute category of each grid region, where the imaging attribute categories include: shadow interference area, metallic reflection area, normal text area and transition area;

[0022] S15. Based on the imaging attribute category, perform corresponding image enhancement processing on different categories of regions;

[0023] S16. Calculate the fusion weights between adjacent regions through distance transformation, and perform adaptive weight fusion on the processed regions to generate a preprocessed image.

[0024] More preferably, in step S15, performing corresponding image enhancement processing on different categories of regions includes:

[0025] S151. Perform a brightness compensation algorithm based on limited contrast adaptive histogram equalization on the regions identified as shadow interference areas.

[0026] S152. Perform a detail enhancement algorithm based on nonlinear brightness mapping on the regions identified as metallic reflective areas;

[0027] S153. Perform a contrast optimization algorithm based on local statistical characteristics on the regions identified as normal text areas;

[0028] S154. Apply gradient blending technology to areas identified as transition zones to eliminate boundary traces.

[0029] Preferably, step S2 specifically includes the following sub-steps:

[0030] S21. Perform trial rotations on the preprocessed image from the preset candidate angle set, and use a text detection algorithm to obtain text location information for the image at each rotation angle;

[0031] S22. Analyze the orientation angle, aspect ratio, and orientation consistency of the internal character outlines of the vehicle identification number text box;

[0032] S23. Analyze the alignment relationship between the vehicle identification number and adjacent text in the horizontal direction and their relative positions in the vertical direction.

[0033] S24. When the initial rotation angle cannot be verified by the consistency of the character outline orientation, the backtracking mechanism is initiated. Multiple correction hypotheses are constructed by comprehensively utilizing geometric attributes and layout relationships. The optimal global image deflection angle is determined from multiple correction hypotheses through a weighted voting strategy.

[0034] S25. Rotate the preprocessed image according to the global deflection angle to generate a rotated corrected image.

[0035] Preferably, step S3 specifically includes the following sub-steps:

[0036] S31. Use a text detection algorithm to obtain the position, orientation and bounding box information of text lines in the rotation-corrected image, and fit the initial perspective transformation matrix based on the text line bounding boxes.

[0037] S32. Construct an objective function by calculating the straightness, parallelism, and spacing uniformity of the transformed text lines, and use the gradient descent algorithm to optimize the parameters of the perspective transformation matrix in order to minimize the objective function.

[0038] S33. Divide the rotation-corrected image into multiple blocks, calculate the optimal local perspective transformation parameters for each block independently, and ensure the continuity of transformation between blocks through a smooth spline function to complete perspective correction.

[0039] S34. Based on the angular residual of the text line baseline, apply affine transformation for fine-tuning to ensure that the text line reaches a horizontal state and generate a geometrically corrected image.

[0040] Preferably, step S4 specifically includes the following sub-steps:

[0041] S41. Call the general character recognition algorithm to recognize the geometrically corrected image and obtain the first recognition result;

[0042] S42. Verify the validity of the vehicle identification number in the first identification result;

[0043] S43. When the vehicle identification number fails the validity check, a dedicated recognition engine based on the CRAFT text detection model and CRNN recognition model is started to perform secondary recognition and obtain a second recognition result.

[0044] S44. Perform confidence fusion and conflict arbitration on the first and second identification results, and output structured information.

[0045] Preferably, step S5 specifically includes the following sub-steps:

[0046] S51. Extract features from the geometrically corrected image to generate an image feature vector;

[0047] S52. Calculate the similarity between the image feature vector and the existing image feature vectors in the database, and perform a duplication check;

[0048] S53. For non-repeating geometrically corrected images, verify the semantic consistency between their image content and structured information. For geometrically corrected images and structured information that pass the semantic consistency verification, generate a combined hash value.

[0049] S54. Register the combined hash value to the blockchain network to obtain the blockchain transaction number;

[0050] S55. Link the geometrically corrected image, structured information, and blockchain transaction number to the database.

[0051] Secondly, the present invention proposes a vehicle nameplate recognition system for performing any of the above-described vehicle nameplate recognition methods, the system comprising:

[0052] The image regionalization adaptive enhancement processing module is configured to perform multi-scale grid partitioning of the input image, extract the brightness, contrast and texture features of each grid region, establish a regional feature map based on the extracted features, determine the imaging attribute category of each region according to the regional feature map, perform differentiated image enhancement according to the imaging attribute category, and generate a preprocessed image through adaptive weight fusion.

[0053] The rotation correction module is configured to use a text detection algorithm to obtain the position information of all text in the preprocessed image, locate the vehicle identification number from the detected text, construct a multi-hypothesis verification model based on the geometric properties of the vehicle identification number and its layout relationship with adjacent text to calculate the global deflection angle of the image, and perform rotation correction on the preprocessed image according to the global deflection angle to generate a rotation-corrected image.

[0054] The spatial distortion correction module is configured to construct a perspective transformation model based on the geometric constraints of text lines in the rotationally corrected image, use adaptive block perspective transformation to correct plane tilt, restore the horizontal state of the text through affine transformation, and generate a geometrically corrected image.

[0055] The dual-engine collaborative decision-making and structured output module is configured to recognize geometrically corrected images using a dual-engine collaborative decision-making mechanism and output structured information.

[0056] The data annotation and evidence storage module is configured to generate image feature vectors and perform repeatability checks on the geometrically corrected images that have been identified. For images that have not been repeated, the module verifies the semantic consistency between the image content and the structured information. It generates a combined hash value for the verified geometrically corrected images and structured information and registers it on the blockchain. Finally, it associates the geometrically corrected images, structured information and blockchain transaction numbers and stores them in the database.

[0057] Thirdly, the present invention proposes a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of any of the vehicle nameplate recognition methods described above.

[0058] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the vehicle nameplate recognition methods described above.

[0059] Compared with the prior art, the beneficial results of the present invention are as follows:

[0060] (1) By employing multi-scale gridded partitioning and regional characteristic map construction techniques, this invention achieves refined regional analysis of vehicle nameplate images. Compared to traditional global processing methods, this method can accurately identify regions with different imaging attributes, such as shadow interference areas and metallic reflective areas, and implement targeted differentiated enhancement strategies. Combined with adaptive weight fusion technology, it effectively eliminates image quality degradation problems under complex lighting conditions, providing a high-quality image foundation for subsequent recognition processes.

[0061] (2) This invention innovatively adopts a layout analysis method based on vehicle identification numbers (VINs) and combines it with a multi-hypothesis verification model, thus abandoning the easily failing direction classifier in traditional OCR systems. By analyzing the geometric attributes of the VINs and their layout relationship with adjacent text, multiple correction hypotheses are constructed and the optimal rotation angle is determined through a weighted voting strategy, which significantly improves the accuracy of direction judgment at any tilt angle.

[0062] (3) The perspective transformation model based on text line geometric constraints, combined with an adaptive block processing strategy, can effectively address perspective distortion and surface deformation in vehicle nameplate images. By optimizing transformation parameters through gradient descent algorithm and using smooth spline function to ensure the continuity of transformation between blocks, accurate correction of nonlinear deformation is achieved, providing a standardized text layout for character recognition.

[0063] (4) The mechanism of collaborative decision-making between a general character recognition algorithm and a dedicated recognition engine ensures both recognition efficiency and the accuracy of key information (such as vehicle identification numbers). Through confidence fusion and conflict arbitration strategies, the limitations of a single recognition engine are effectively avoided, and the recognition success rate in complex scenarios is significantly improved.

[0064] (5) A labeling mechanism based on image fingerprint generation and blockchain registration technology was introduced, establishing a complete image traceability system. Semantic consistency verification ensures the matching degree between the recognition results and the image content, and the immutability of blockchain technology ensures data security, providing reliable data support for vehicle information management.

[0065] (6) This invention establishes a complete processing pipeline, organically integrating image preprocessing, geometric correction, text recognition, and data management to form an end-to-end solution. The collaborative work among the various technical modules ensures the stable operation of the system in complex industrial environments, significantly improving the practicality and reliability of vehicle nameplate recognition.

[0066] In summary, this invention effectively solves key technical challenges in vehicle nameplate recognition through multi-level technological innovation, achieving significant progress in image quality enhancement, geometric deformation correction, text recognition accuracy, and data management reliability, providing strong technical support for applications such as vehicle management and intelligent manufacturing. Attached Figure Description

[0067] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments, taken with reference to the accompanying drawings:

[0068] Figure 1 This is a schematic diagram of the overall process of a vehicle nameplate recognition method according to the present invention;

[0069] Figure 2 This is a schematic diagram of the refinement process of image regionalization adaptive enhancement processing (S1) in one embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram of the multi-dimensional rotation correction (S2) process in one embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram illustrating the principle of spatial distortion correction (S3) guided by text line distribution in one embodiment of the present invention;

[0072] Figure 5 This is a schematic diagram illustrating the principle of dual-engine collaborative decision-making and structured output (S4) in one embodiment of the present invention;

[0073] Figure 6This is a flowchart of the process of identifying and annotating the image and constructing the database (S5) in one embodiment of the present invention;

[0074] Figure 7 This is a schematic diagram of the vehicle nameplate recognition system of the present invention. Detailed Implementation

[0075] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0076] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0077] Figure 1 This diagram illustrates the overall flow of a vehicle nameplate recognition method according to the present invention. It is specifically applicable to solving the problem of high-precision recognition of vehicle nameplates under complex conditions, tilted shooting, and curved surface deformation scenarios. Figure 1 As shown, the method includes the following steps:

[0078] S1. Divide the input image into multi-scale grid partitions, extract the brightness, contrast, and texture features of each grid region, establish a regional feature map based on the extracted features, determine the imaging attribute category of each region according to the regional feature map, perform differential image enhancement according to the imaging attribute category, and generate a preprocessed image through adaptive weight fusion. Specifically, this includes the following sub-steps:

[0079] S11. Perform multi-scale grid partitioning on the input image, dividing the input image into multiple grid units;

[0080] S12. Extract the mean brightness, standard deviation of contrast, and gray-level co-occurrence matrix texture features of each grid cell;

[0081] S13. Integrate the features of all grid cells to generate a regional feature map that represents the global and local characteristics of the image;

[0082] S14. Based on the regional characteristic map, determine the imaging attribute category of each grid region, where the imaging attribute categories include: shadow interference area, metallic reflection area, normal text area and transition area;

[0083] S15. Based on the imaging attribute category, perform corresponding image enhancement processing on different categories of regions, including: S151. Perform a brightness compensation algorithm based on limited contrast adaptive histogram equalization on regions identified as shadow interference areas; S152. Perform a detail enhancement algorithm based on nonlinear brightness mapping on regions identified as metallic reflection areas; S153. Perform a contrast optimization algorithm based on local statistical characteristics on regions identified as normal text areas; S154. Use gradient blending technology to eliminate processing boundary traces on regions identified as transition areas.

[0084] S16. Calculate the fusion weights between adjacent regions through distance transformation, and perform adaptive weight fusion on the processed regions to generate a preprocessed image.

[0085] Figure 2 This diagram illustrates the thinning process of image regionalization adaptive enhancement processing (S1) in one embodiment of the present invention, as shown below. Figure 2 As shown, the specific implementation process of this step is as follows:

[0086] The input is the original color image I orig ∈R H×W×3 To reduce computational complexity, the image is first scaled according to a preset maximum side length constraint. Then, the image is converted to a grayscale image I and the LAB color space, and the L channel is meshed. Image I is divided into G×G grid cells, with grid indices r∈{1,...,R}(R=G). 2 Let R be the set of pixels in the r-th grid. r Here, image I is a function that maps pixel coordinates Ω to the gray value range [0, 255], and the size of Ω is |Ω| = H × W.

[0087] For each grid cell, the following features are extracted to determine the region attributes:

[0088] Luminance mean μ r It is used to measure the average brightness of a region, and its expression is:

[0089]

[0090] Where I(p) is the intensity value of pixel p (the mean of the grayscale or color channel). Indicates the grid R r Sum the intensity values ​​of all pixels within the range, |R r | represents the number of pixels in the grid, and the average brightness is obtained by averaging.

[0091] Contrast standard deviation σ r This is used to characterize the contrast or texture intensity within a region, and its expression is:

[0092]

[0093] This value reflects the degree of dispersion of pixel intensity relative to the mean; a larger value indicates higher regional contrast or richer texture.

[0094] Based on the extracted mean brightness and standard deviation of contrast, the imaging attribute category of each grid region is determined according to the following rules:

[0095] is_shadow(r) = [μ r <τ1∧σ r <τ2],

[0096] is_reflection(r) = [μ r >τ3∧σ r <τ4],

[0097] is_Lowcontrast(r)=[σ r <τ5](3)

[0098] [·] Indicator function, returns 1 if the logical condition is true, otherwise returns 0. Here, τ1 to τ5 are thresholds set according to the actual scene, used to distinguish between shadow interference areas (is_shadow(r)), metallic reflection areas (is_reflection(r)), normal text areas, and transition areas.

[0099] In the adaptive image regionalization process, the transition region refers to the boundary region between different attribute regions, characterized by significant changes in brightness or contrast gradients. The low-contrast region (is_Lowcontrast(r)) is an imaging attribute category independently defined based on the contrast standard deviation feature. This category primarily addresses issues of blurred texture and unclear details within the region. It is worth noting that the low-contrast region may overlap with other categories defined based on brightness features (such as shadow interference regions or metallic reflection regions), meaning that the same grid region may simultaneously satisfy the conditions of low contrast and other categories. The processing of the low-contrast region is integrated into the overall image enhancement strategy: the system applies a dedicated contrast optimization algorithm (e.g., contrast enhancement based on local statistical characteristics or histogram adjustment) according to the region category, aiming to improve the distinguishability of text from the background in this region and solve the problem of text recognition difficulties caused by insufficient contrast. This processing method runs parallel to enhancement operations for shadow and reflection regions, achieving targeted enhancement through adaptive parameter configuration, ultimately improving the overall visual quality and recognition robustness of the image.

[0100] Differentiated image enhancement processing is applied to regions with different imaging attribute categories:

[0101] In the target brightness (used to align the brightness of normal regions to a uniform target), if there exists a set N of normal regions, then:

[0102]

[0103] Where clip(x,a,b) is the clipping function, restricting the value to the interval [a,b]; B min and B max These are the lower and upper limits of the target brightness to ensure that the image is neither too dark nor too bright.

[0104] Affine correction for regional brightness (for each pixel p∈R) r (Application), the enhanced pixel value is:

[0105] I' r (p) = clip(α) r ·I(p)+β r ,0,255),β r =B t -α r μ r (5)

[0107] Where α r This is the area brightness gain coefficient, which controls the brightness scaling factor. β r It is the area brightness offset, which aligns the average brightness of the area to the target value B. t In practical implementation, the system independently configures a differentiated brightness gain coefficient α for each identified imaging attribute category (including shadow interference area, metallic reflection area, normal text area, transition area, and low contrast area). r and brightness offset β r Parameters. For example, a high gain coefficient is used for shadow interference areas to increase brightness, and a low gain coefficient is used for metallic reflective areas to suppress overexposure, ensuring that the brightness characteristics of various areas can be adaptively aligned to the target value B. t This allows for precise enhancement.

[0108] Subsequently, a smooth transition fusion is performed on the region boundaries based on the distance transform. First, a binary mask for the region is generated according to the identified imaging attribute category. Here, p represents the pixel coordinates in the image, and M(p)∈{0,1} (the processing region is 1) is a binary function. When pixel p belongs to the processing region to be enhanced, M(p)=1; otherwise, M(p)=0. Then, the distance transform D(p) of the mask M is calculated and normalized:

[0109]

[0110] Where D(p) represents the Euclidean distance from pixel p to the nearest non-processed region boundary (i.e., the nearest pixel where M(p) = 0). dist_transform is the standard distance transform operator. n (p) is the normalized distance value of pixel p, ranging from [0,1]. q D(q) represents the maximum value of the distance transformation value D(q) of all pixels in the entire image.

[0111] The final fusion formula is:

[0112] I fused (p)=D n (p)·I' r (p)+(1-D n (p))·I orig (7)

[0113] Among them, I' r (p) represents the pixel value after region enhancement, I orig This represents the pixel values ​​of the original image. The formula utilizes the normalized distance D. n (p) serves as the weight, enabling the processing of the region within (D) n (p) close to 1, tending to use enhanced results) to the region boundary (D n (p) Smooth transitions (close to 0, tending to preserve the original image) are achieved. Larger distances indicate closer proximity to the interior of the region, resulting in higher weights and a preference for using processed pixels; smooth blending at boundaries eliminates splicing artifacts. Finally, protective sharpening and partitioned denoising are applied to text regions, defining the image sharpness index as the variance of the Laplacian operator. Used to determine whether enhancement / sharpening needs to be applied. Based on S(I) orig The empirical formula for calculating the sharpening factor γ is as follows:

[0114]

[0115] Among them, C(I) orig ) represents the contrast index of the original image, calculated as the standard deviation of grayscale values; ρ is the adjustment parameter of the region brightness gain coefficient, derived from the imaging attribute category discrimination in step S15; 1 ρ∈(0.1,0.5)This describes a triggering logic based on text spatial density for recognizing image content structure. When the proportion of text regions in an image is within a medium density range, the system determines that it belongs to the image class that can be enhanced and enables additional sharpening gain. The value of ρ is 1 when it is between 0.1 and 0.5, and 0 otherwise. The threshold range of the density interval can be preset according to the image quality requirements of the actual application scenario. If the image is blurry, sharpening is adjusted according to sharpness; if the image is sharp but has low contrast, sharpening is adjusted according to contrast. To avoid overprocessing, enhancement verification logic based on quality evaluation is introduced (if the sharpness / contrast is not significantly improved after enhancement, the original image is reverted).

[0116] Continue to refer to Figure 1 The vehicle nameplate recognition method proposed in this invention further includes the following steps:

[0117] S2. Use a text detection algorithm to obtain the positional information of all text in the preprocessed image, locate the vehicle identification number (VIN) from the detected text, construct a multi-hypothesis verification model based on the geometric attributes of the VIN and its layout relationship with adjacent text to calculate the global image rotation angle, and perform rotation correction on the preprocessed image according to the global rotation angle to generate a rotation-corrected image. Specifically, this includes the following sub-steps:

[0118] S21. Perform trial rotations on the preprocessed image from the preset candidate angle set, and use a text detection algorithm to obtain text location information for the image at each rotation angle;

[0119] S22. Analyze the orientation angle, aspect ratio, and orientation consistency of the internal character outlines of the vehicle identification number text box;

[0120] S23. Analyze the alignment relationship between the vehicle identification number and adjacent text in the horizontal direction and their relative positions in the vertical direction.

[0121] S24. When the initial rotation angle cannot be verified by the consistency of the character outline orientation, the backtracking mechanism is initiated. Multiple correction hypotheses are constructed by comprehensively utilizing geometric attributes and layout relationships. The optimal global image deflection angle is determined from multiple correction hypotheses through a weighted voting strategy.

[0122] S25. Rotate the preprocessed image according to the global deflection angle to generate a rotated corrected image.

[0123] Figure 3 A flowchart illustrating multi-dimensional rotation correction (S2) in one embodiment of the present invention is shown, as follows: Figure 3 As shown, the specific implementation process of this step is as follows:

[0124] Using multiple hypothesis testing as its core, the method first performs trial rotations on a pre-defined set of orthogonal candidate angles A = {a1, a2, ...}. For each candidate angle a in set A, optical character recognition (OCR) technology is used to retrieve a potentially complete VIN code (if recognized, it indicates a minor rotation tilt, allowing for slight correction and direct progression to the next step; otherwise, it proves that more thorough rotation correction is needed).

[0125] I a =R(I fused a)

[0126]

[0127] Where R(·,a) represents a function that rotates the image by angle a, and OCR(·) is optical character recognition processing, returning the text lines detected in the current image. Each line contains a bounding box b, a text string t, and a confidence score c. The recognized text strings are then cleaned. Only retain the acceptable character set for the Vehicle Identification Number (VIN). If a line of text meets the following simple character set rule:

[0128]

[0129] The text line is then counted as a valid candidate. Here, 1[·] is an indicator function, taking a value of 1 if the condition is true, and 0 otherwise. For candidates that meet the conditions, their confidence scores are calculated:

[0130]

[0131] The highest confidence score among all selected text lines at that angle is taken as the score for angle a. Then, the angle with the highest score is selected as the approximate rotation angle.

[0132]

[0133] To improve the robustness of angle estimation, character-level analysis is introduced to assist in judgment when the text box orientation is unreliable. Let the set of character-level contours be k, and the orientation angle of the minimum bounding rectangle corresponding to each character contour be θ. k The character-level average angle estimate is:

[0134]

[0135] This average angle can be used to correct the overall orientation estimate of the text.

[0136] Refine the correction based on the selected angle (based on the text box angle and aspect ratio). Let the aspect ratio of the detected VIN text box be r (width / height, i.e., the ratio of width to height), the text box angle be φ, and define the corresponding aspect ratio threshold r. v and tilt angle threshold φ th The correction angle is then:

[0137]

[0138] If the text box is vertical (aspect ratio less than the threshold), rotate it ±90° to make it horizontal; otherwise, perform a small-angle correction based on the detected text tilt angle. Reverse verification (to prevent false corrections) is performed on the rotated image. Application of correction angle θ corr Image I was obtained corr Run OCR(·) again to check if a valid VIN still exists. If it does not exist, try rotating in reverse. If a valid VIN can be identified after rotating in reverse, then use the result of rotating in reverse.

[0139] Character-level analysis (fitting the minimum bounding rectangle by character contours) is used to estimate the average orientation of character sequences as an alternative when text box angles are unreliable, thereby improving the robustness of angle estimation on noisy or long lines of text.

[0140] Continue to refer to Figure 1 The vehicle nameplate recognition method proposed in this invention further includes the following steps:

[0141] S3. Construct a perspective transformation model based on the geometric constraints of the text lines in the rotation-corrected image, use adaptive block perspective transformation to correct plane tilt, and restore the horizontal state of the text through affine transformation to generate a geometrically corrected image. Step S3 specifically includes the following sub-steps:

[0142] S31. Use a text detection algorithm to obtain the position, orientation and bounding box information of text lines in the rotation-corrected image, and fit the initial perspective transformation matrix based on the text line bounding boxes.

[0143] S32. Construct an objective function by calculating the straightness, parallelism, and spacing uniformity of the transformed text lines, and use the gradient descent algorithm to optimize the parameters of the perspective transformation matrix in order to minimize the objective function.

[0144] S33. Divide the rotation-corrected image into multiple blocks, calculate the optimal local perspective transformation parameters for each block independently, and ensure the continuity of transformation between blocks through a smooth spline function to complete perspective correction.

[0145] S34. Based on the angular residual of the text line baseline, apply affine transformation for fine-tuning to ensure that the text line reaches a horizontal state and generate a geometrically corrected image.

[0146] Figure 4 The diagram illustrates the principle of spatial distortion correction (S3) guided by text line distribution in one embodiment of the present invention. Figure 4 As shown, the specific implementation process of this step is as follows:

[0147] First, on the rotation-corrected image, the positional information of all characters / text boxes is obtained using a text detection algorithm (such as CRAFT or DB-Net), and all corner points are extracted. The convex hull of the point set is calculated and an approximate quadrilateral is fitted. Let the set of all character box corner points from the OCR be... Where N is the number of corner points. Calculate the convex hull set:

[0148]

[0149] It is the set of all convex combinations of points; α i These are the convex combination coefficients, non-negative, and their sum is 1. This yields the convex hull. Afterwards, the number of vertices M may be greater than 4 (e.g., hexagonal, octagonal), so for simplification, it needs to be approximated as a rectangle. The Douglas–Peucker algorithm is used for polygon approximation: given a polyline L (from... (Constructed by the vertex order), let the approximation accuracy threshold be ∈, connect the start and end points of the polyline to form a straight line, and find the point p farthest from this line. max Calculate its vertical distance d max If d max ≤∈ Remove intermediate points and approximate with a straight line; otherwise, keep p. max The process recursively processes two sub-polylines. The final result is a polygon with a small number of vertices and a shape approximating the original convex hull. If the number of vertices is 4, it is directly used as the target region; if it is greater than 4, the minimum bounding rectangle (minAreaRect) is calculated, and its vertices can be obtained using a caliper rotation algorithm.

[0150] The approximated quadrilaterals are sorted by top left p0, top right p1, bottom right p2, and bottom left p3, and then expanded outwards by a certain percentage (e.g., 20%) according to empirical values ​​to cover the edge text. Define each quadrilateral...

[0151] Side length: W top =||p0-p1||,W bot =||p3-p2||,

[0152] H left =||p0-p3||,H right =||p1-p2|| (16)

[0153] The width W and height H of the target correction area are taken to their maximum values: Construct a target rectangle using the quadrilateral as the source point set, and correct the perspective distortion to a planar orthographic projection by solving for the perspective transformation (homography matrix). Given four sets of corresponding points p i =(x i ,y i ) and target point q i ={x' i ,y' i}(i=1...4), homography matrix H∈R 3×3 Satisfying homogeneity: Where λ i It is a scaling factor. A linear system is constructed using the direct linear transformation (DLT): For each pair of corresponding points, the equation is constructed as follows:

[0154]

[0155] Where h∈R 9 Let H be the column vectors after flattening. Solve this homogeneous least squares problem to reconstruct the homography matrix H.

[0156] To handle local nonlinear surface deformation, adaptive block partitioning and thin-plate spline (TPS) smoothing are used as subsequent thinning steps. The image is divided into grid cells (u,v), and for each cell, the local mapping H is solved based on its local control points. u,v Finally, a weighted interpolation fusion local transformation is applied to pixel position x:

[0157] H(x)=∑ u,v w u,v (x)H u,v ,∑ u,v w u,v (x)=1(18)

[0158] Where w u,v (x) Use bilinear weights or distance-based weights (the closer the distance, the greater the weight). Further, thin-plate splines can be used to smoothly interpolate the control point displacements; the TPS is formalized as follows:

[0159]

[0160] Where (a0,a x ,a y ) represents the affine coefficients, w j For the control point weights, c j Let m be the number of control points and U(r) be the radial basis function. After correction, contrast-limited adaptive histogram equalization (CLAHE) or sharpening is applied again to the text region to ensure character readability.

[0161] Continue to refer to Figure 1 The vehicle nameplate recognition method proposed in this invention further includes the following steps:

[0162] S4. Employ a dual-engine collaborative decision-making mechanism to recognize the geometrically corrected image and output structured information. This includes the following sub-steps:

[0163] S41. Call the general character recognition algorithm to recognize the geometrically corrected image and obtain the first recognition result;

[0164] S42. Verify the validity of the vehicle identification number in the first identification result;

[0165] S43. When the vehicle identification number fails the validity check, a dedicated recognition engine based on the CRAFT text detection model and CRNN recognition model is started to perform secondary recognition and obtain a second recognition result.

[0166] S44. Perform confidence fusion and conflict arbitration on the first and second identification results, and output structured information.

[0167] Figure 5 The diagram illustrates the principle of dual-engine collaborative decision-making and structured output (S4) in one embodiment of the present invention, as shown below. Figure 5 As shown, the specific implementation process of this step is as follows:

[0168] On the geometrically corrected image, a general character recognition (OCR) algorithm is first called for fast full-text recognition to obtain candidate texts and their confidence scores. If the VIN candidate recognized by the general OCR passes the simple character set check and length check and the confidence score meets the threshold, it is directly adopted; otherwise, the high-precision engine returned by the dedicated VIN recognition engine is activated. Let the candidate result set returned by the general OCR engine be G = {(g k ,c g,k The set of candidate results returned by the dedicated engine is S = {(s)}. l ,c s,l )}, where g k s l For candidate strings, c g,k c s,l This corresponds to the confidence level.

[0169] If the Vehicle Identification Number (VIN) candidate identified by the general OCR passes the following verification conditions:

[0170] 1. Character set check: Only includes characters allowed by VIN code (uppercase letters and numbers, excluding easily confused characters such as I, O, Q);

[0171] 2. Length check: The string length is 17 characters;

[0172] 3. Confidence threshold: Confidence level c g,k Not lower than the preset threshold;

[0173] Then the candidate result will be adopted directly.

[0174] If any of the above conditions are not met, a dedicated VIN recognition engine will be activated for secondary recognition.

[0175] The dedicated VIN recognition engine is a deep learning system optimized for vehicle identification number (VIN) recognition tasks. Its processing flow includes:

[0176] 1. Use the CRAFT text detection model to accurately locate text regions in an image;

[0177] 2. A character sequence recognition model based on CRNN (Convolutional Recurrent Neural Network) is used;

[0178] 3. Optimizations were made for the 17-bit fixed length of the VIN code and its specific character set;

[0179] 4. A three-round progressive recognition strategy is adopted: the first round uses a standard detection model for preliminary recognition, the second round re-detects after optimizing the image size, and the third round uses the DB-Net detection model for fine processing of difficult cases.

[0180] If both engines return valid candidate results, they are fused using a weighted confidence strategy. Define the general engine weight w. g With dedicated engine weight w s Typically, w is set s >w g To reflect the priority of dedicated engines, the highest-confidence candidate is selected for each engine:

[0181]

[0182] After merging the candidate sets, the final identification result is selected based on the weighted confidence score:

[0183]

[0184] Where g * , As the best candidate field, the weight function w(x) is assigned a value according to the engine type: if x comes from a general engine, then w(x) = w g If it comes from a dedicated engine, then w(x) = w s .

[0185] To enhance the reliability of the identification results, check digits can be calculated according to the international VIN standard (ISO-3779) to filter out erroneous identification results that do not conform to the rules. The final output is structured information integrating key fields such as Vehicle Identification Number (VIN), vehicle model, engine number, manufacturer information, and production date.

[0186] Continue to refer to Figure 1 The vehicle nameplate recognition method proposed in this invention further includes the following steps:

[0187] S5. For the identified geometrically corrected images, generate image feature vectors and perform repeatability checks. For images that do not repeat after verification, verify the semantic consistency between their image content and structured information. Generate a combined hash value for the verified geometrically corrected images and structured information and register it in the blockchain. Finally, associate the geometrically corrected images, structured information, and blockchain transaction numbers and store them in the database. Specifically, this includes the following sub-steps:

[0188] S51. Extract features from the geometrically corrected image to generate an image feature vector;

[0189] S52. Calculate the similarity between the image feature vector and the existing image feature vectors in the database, and perform a duplication check;

[0190] S53. For non-repeating geometrically corrected images, verify the semantic consistency between their image content and structured information. For geometrically corrected images and structured information that pass the semantic consistency verification, generate a combined hash value.

[0191] S54. Register the combined hash value to the blockchain network to obtain the blockchain transaction number;

[0192] S55. Link the geometrically corrected image, structured information, and blockchain transaction number to the database.

[0193] Figure 6 A flowchart of the image annotation and database construction (S5) process in one embodiment of the present invention is shown, as follows: Figure 6 As shown, the specific implementation process of this step is as follows:

[0194] After completing the dual-engine collaborative recognition (S4), the system introduces a joint mechanism of artificial intelligence and blockchain to label (e.g., label "recognized") and store evidence of the recognized images to prevent duplicate recognition and build a trusted database. Input geometrically corrected image I∈R H×w×3 First, normalization is performed, and then the result is fed into a pre-trained feature extraction network f. θ Output feature map Where C represents the number of feature map channels. Image feature vectors are generated using generalized mean pooling (GeM).

[0195]

[0196] Where p is the pooling index (usually set as a learnable parameter or a fixed value such as p = 3), F c,i,j This represents the value of the feature map at channel c and position (i,j); The aggregated feature vector; the L2 norm of the ||v||2 vector.

[0197] In database index collection Searching for cosine similarity:

[0198]

[0199] Where s(I,j) is the current image feature vector f I The feature vector of the j-th image in the database similarity, T dup A duplicate detection threshold is set (usually 0.9-0.95). If the similarity reaches the threshold, it is determined to be a duplicate image, and the existing recognition result is returned directly to avoid duplicate processing.

[0200] For images not identified as duplicates, the semantic consistency between their image content and the recognition results (structured information) is further verified. A visual encoder E is used. v (Mapping image regions to vectors) and text encoder E t (Map the recognition results to vectors), and output the embedding vectors respectively. and The visual encoder can use a pre-trained ResNet model, and the text encoder uses a BERT model. The recognition result is then compared to the image content to calculate semantic similarity.

[0201]

[0202] Where s mm It is the semantic similarity between visual and textual data, T mm As the semantic consistency threshold, when s mm ≥T mm Then the semantic verification passes.

[0203] For each verified image, a unique registration record is constructed. First, the image fingerprint f is processed. I Calculate the SHA-256 hash value for both the recognition result (structured text information output by S4, including fields such as VIN, vehicle model, and production date, stored in normalized JSON format) and the recognition result (structured text information output by S4, including fields such as VIN, vehicle model, and production date, stored in normalized JSON format).

[0204] h I =SHA256(f I ),hRES =SHA256(RES)(25)

[0205] By combining metadata (including timestamps, model versions, processing node IDs, etc.), a combined hash value is generated as a unique identifier:

[0206] R = SHA256(h) I ||h RES ||meta) (26)

[0207] Here, || represents the string concatenation operation. R is registered in the blockchain smart contract to obtain the transaction number tx_id. Finally, the image I, the recognition result RES, the unique identifier R, and the transaction number tx_id are associated (I, RES, R, tx_id) and stored in the database. This ensures that the same image is recognized only once through fingerprint hashing; semantic verification ensures that the recognition result matches the image content; and a high-quality database is established to prevent duplicate recognition and provide reliable data for subsequent training.

[0208] In summary, this invention provides a vehicle nameplate recognition method based on regionalized adaptive processing and layout analysis. This method simulates the human visual cognitive process, constructing a hierarchical, progressively advanced intelligent recognition pipeline, which effectively solves the problems of low recognition rate and poor robustness of vehicle nameplate images in complex environments.

[0209] Further reference Figure 7 As an implementation of the above method, in a second aspect, the present invention provides an embodiment of a structural diagram of a vehicle nameplate recognition system 700, which can be specifically applied to various electronic devices. The vehicle nameplate recognition system 700 includes the following modules:

[0210] The image regionalization adaptive enhancement processing module 710 is configured to perform multi-scale grid partitioning on the input image, extract the brightness, contrast and texture features of each grid region, establish a regional feature map based on the extracted features, determine the imaging attribute category of each region according to the regional feature map, perform differentiated image enhancement according to the imaging attribute category, and generate a preprocessed image through adaptive weight fusion.

[0211] The rotation correction module 720 is configured to use a text detection algorithm to obtain the position information of all text in the preprocessed image, locate the vehicle identification number from the detected text, construct a multi-hypothesis verification model based on the geometric properties of the vehicle identification number and its layout relationship with adjacent text to calculate the global deflection angle of the image, and perform rotation correction on the preprocessed image according to the global deflection angle to generate a rotation-corrected image.

[0212] The spatial distortion correction module 730 is configured to construct a perspective transformation model based on the geometric constraints of text lines in the rotationally corrected image, use adaptive block perspective transformation to correct plane tilt, restore the horizontal state of the text through affine transformation, and generate a geometrically corrected image.

[0213] The dual-engine collaborative decision-making and structured output module 740 is configured to recognize geometrically corrected images using a dual-engine collaborative decision-making mechanism and output structured information.

[0214] The data annotation and evidence storage module 750 is configured to generate image feature vectors and perform repeatability checks on the geometrically corrected images that have been identified. For images that have not been repeated, it verifies the semantic consistency between the image content and the structured information. It generates a combined hash value for the geometrically corrected images and structured information that have passed the verification and registers it in the blockchain. Finally, it associates the geometrically corrected images, structured information and blockchain transaction numbers and stores them in the database.

[0215] Thirdly, the present invention proposes a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the steps of any of the vehicle nameplate recognition methods described above.

[0216] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the vehicle nameplate recognition methods described above.

[0217] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for recognizing vehicle nameplates, characterized in that, Includes the following steps: S1. Divide the input image into multi-scale grid partitions, extract the brightness, contrast, and texture features of each grid region, establish a regional feature map based on the extracted features, determine the imaging attribute category of each region according to the regional feature map, perform differential image enhancement according to the imaging attribute category, and generate a preprocessed image through adaptive weight fusion. Specifically, this includes the following sub-steps: S11. Perform multi-scale grid partitioning on the input image, dividing the input image into multiple grid units; S12. Extract the mean brightness, standard deviation of contrast, and gray-level co-occurrence matrix texture features of each grid cell; S13. Integrate the features of all grid cells to generate a regional feature map that represents the global and local characteristics of the image; S14. Based on the regional characteristic map, determine the imaging attribute category of each grid region, wherein the imaging attribute category includes: shadow interference area, metallic reflection area, normal text area and transition area; S15. Based on the imaging attribute category, perform corresponding image enhancement processing on regions of different categories; S16. Calculate the fusion weight between adjacent regions by distance transformation, and perform adaptive weight fusion on the processed regions to generate the preprocessed image. In step S15, the corresponding image enhancement processing for different categories of regions includes: S151. Perform a brightness compensation algorithm based on limited contrast adaptive histogram equalization on the regions identified as shadow interference areas. S152. Perform a detail enhancement algorithm based on nonlinear brightness mapping on the regions identified as metallic reflective areas; S153. Perform a contrast optimization algorithm based on local statistical characteristics on the regions identified as normal text areas; S154. Apply gradient blending technology to eliminate boundary traces in areas identified as transition zones; S2. Use a text detection algorithm to obtain the position information of all text in the preprocessed image, locate the vehicle identification number from the detected text, construct a multi-hypothesis verification model based on the geometric attributes of the vehicle identification number and its layout relationship with adjacent text to calculate the global deflection angle of the image, and perform rotation correction on the preprocessed image according to the global deflection angle to generate a rotation-corrected image. S3. Construct a perspective transformation model based on the geometric constraints of the text lines in the rotationally corrected image, use adaptive block perspective transformation to correct the plane tilt, restore the horizontal state of the text through affine transformation, and generate a geometrically corrected image. S4. The geometrically corrected image is identified using a dual-engine collaborative decision-making mechanism, and structured information is output. This includes the following sub-steps: S41. Call a general character recognition algorithm to recognize the geometrically corrected image and obtain a first recognition result; S42. Verify the validity of the vehicle identification number in the first identification result; S43. When the vehicle identification number fails the validity check, a dedicated recognition engine based on the CRAFT text detection model and the CRNN recognition model is started to perform secondary recognition and obtain a second recognition result. S44. Perform confidence fusion and conflict arbitration on the first identification result and the second identification result, and output the structured information; S5. For the geometrically corrected image that has been identified, generate an image feature vector and perform a repeatability check. For images that are not repeated after verification, verify the semantic consistency between their image content and the structured information. Generate a combined hash value for the geometrically corrected image and the structured information that have passed the verification and register it in the blockchain. Finally, associate the geometrically corrected image, the structured information and the blockchain transaction number and store them in the database.

2. The vehicle nameplate recognition method according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S21. Perform a trial rotation on the preprocessed image in the preset candidate angle set, and use a text detection algorithm to obtain text position information for the image under each rotation angle; S22. Analyze the orientation angle, aspect ratio, and orientation consistency of the internal character outline of the vehicle identification number text box; S23. Analyze the alignment relationship between the vehicle identification number and adjacent text in the horizontal direction and their relative positions in the vertical direction; S24. When the initial rotation angle cannot be verified by the consistency of the character outline orientation, the backtracking mechanism is initiated. Multiple correction hypotheses are constructed by comprehensively utilizing the geometric attributes and layout relationships. The optimal global image deflection angle is determined from the multiple correction hypotheses through a weighted voting strategy. S25. Rotate the preprocessed image according to the global deflection angle to generate a rotated corrected image.

3. The vehicle nameplate recognition method according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S31. Use a text detection algorithm to obtain the position, direction and bounding box information of the text lines in the rotation-corrected image, and fit an initial perspective transformation matrix based on the text line bounding boxes; S32. Construct an objective function by calculating the straightness, parallelism, and spacing uniformity of the transformed text lines, and optimize the parameters of the perspective transformation matrix using the gradient descent algorithm to minimize the objective function; S33. Divide the rotation-corrected image into multiple blocks, calculate the optimal local perspective transformation parameters for each block independently, and ensure the continuity of transformation between blocks through a smooth spline function to complete perspective correction. S34. Based on the angular residual of the text line baseline, apply affine transformation for fine-tuning to ensure that the text line reaches a horizontal state and generate a geometrically corrected image.

4. The vehicle nameplate recognition method according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S51. Extract features from the geometrically corrected image to generate an image feature vector; S52. Calculate the similarity between the image feature vector and the existing image feature vectors in the database, and perform a duplication check; S53. For the non-repeating geometrically corrected image, verify the semantic consistency between its image content and the structured information, and generate a combined hash value for the geometrically corrected image that passes the semantic consistency verification and the structured information; S54. Register the combined hash value to the blockchain network to obtain a blockchain transaction number; S55. The geometrically corrected image, the structured information, and the blockchain transaction number are associated and stored in the database.

5. A vehicle nameplate recognition system, used to execute the vehicle nameplate recognition method according to any one of claims 1 to 4, characterized in that, The system includes: The image regionalization adaptive enhancement processing module is configured to perform multi-scale grid partitioning on the input image, extract the brightness, contrast and texture features of each grid region, establish a regional feature map based on the extracted features, determine the imaging attribute category of each region according to the regional feature map, perform differentiated image enhancement according to the imaging attribute category, and generate a preprocessed image through adaptive weight fusion. The rotation correction module is configured to use a text detection algorithm to obtain the position information of all text in the preprocessed image, locate the vehicle identification number from the detected text, construct a multi-hypothesis verification model based on the geometric attributes of the vehicle identification number and its layout relationship with adjacent text to calculate the global deflection angle of the image, and perform rotation correction on the preprocessed image according to the global deflection angle to generate a rotation-corrected image. The spatial distortion correction module is configured to construct a perspective transformation model based on the geometric constraints of the text lines in the rotationally corrected image, use adaptive block perspective transformation to correct plane tilt, restore the horizontal state of the text through affine transformation, and generate a geometrically corrected image. The dual-engine collaborative decision-making and structured output module is configured to recognize the geometrically corrected image using a dual-engine collaborative decision-making mechanism and output structured information. The data annotation and evidence storage module is configured to generate image feature vectors and perform duplicate verification on the geometrically corrected images that have been identified. For images that are not duplicated after verification, the module verifies the semantic consistency between the image content and the structured information. For the geometrically corrected images that pass the verification and the structured information, a combined hash value is generated and registered in the blockchain. Finally, the geometrically corrected images, the structured information and the blockchain transaction number are associated and stored in the database.

6. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the vehicle nameplate recognition method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the vehicle nameplate recognition method as described in any one of claims 1 to 4.

Images