Intelligent license plate recognition system based on multi-view image acquisition matching

Abstract

The application discloses an intelligent license plate recognition system based on multi-view image acquisition matching and belongs to the technical field of multi-view image acquisition and intelligent matching fusion. The system comprises a data image acquisition module, a license plate detection module, a data fusion module and a feedback module. The image acquisition module is used for synchronously acquiring multi-angle vehicle image data and performing pretreatment. The vehicle detection module is used for license plate region positioning and license plate character recognition, obtains a detection result and evaluates a confidence level. When the confidence level is greater than a confidence threshold, data fusion is performed. The data fusion module is used for fusing multi-view detection results, outputting an optimal license plate number matching result and a global confidence level. The feedback module is used for judging whether an artificial auditing instruction is triggered according to the global confidence level. Through multi-view data fusion and intelligent matching algorithm, the application significantly improves the accuracy and stability of license plate recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an intelligent license plate recognition system based on multi-view image acquisition matching. BACKGROUND

[0002] As a core component of intelligent transportation systems, license plate recognition technology is widely used in parking lot management, traffic monitoring, and law enforcement. Existing license plate recognition systems mainly use a single camera to capture images at a fixed angle, and extract license plate information through image preprocessing, license plate positioning, character segmentation, and character recognition. However, a single camera can only capture images from a fixed angle, and when vehicles pass at an angle or the license plate is tilted, the recognition accuracy decreases significantly. In poor environmental conditions such as insufficient light, backlight, rain, and fog, the quality of images from a single perspective is poor, leading to a high failure rate of recognition. When the license plate is partially obscured, a single angle cannot obtain complete license plate information, and the recognition system may misjudge or fail to recognize. Traditional systems need to wait for the best shooting opportunity, which can cause recognition delays in high-traffic scenarios, affecting traffic efficiency. In addition, the recognition accuracy of a single perspective system fluctuates greatly due to factors such as changes in light, license plate damage, and character blurring, making it difficult to meet the needs of practical applications.

[0003] Therefore, there is an urgent need for an intelligent license plate recognition system based on multi-view image acquisition matching to solve the above problems. SUMMARY

[0004] The present application aims to provide an intelligent license plate recognition system based on multi-view image acquisition matching, which solves the limitations of single-angle recognition through multi-view image acquisition. By utilizing the advantages of multiple cameras at different positions and angles, higher-quality image data can be obtained in complex environmental conditions. According to multi-view information fusion, even if there is an obstruction at a single angle, complete license plate information can be obtained from other angles. At the same time, through parallel processing of multi-view images and intelligent matching algorithms, recognition delays are reduced, and the real-time response capability of the system is improved, thus solving the problems raised in the background technology.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] The intelligent license plate recognition system based on multi-view image acquisition matching includes a data image acquisition module, a license plate detection module, a data fusion module, and a feedback module:

[0007] The image acquisition module is used to synchronously acquire multi-angle vehicle image data and perform preprocessing.

[0008] The vehicle detection module is used for license plate region positioning and license plate character recognition, and obtains detection results and evaluates the confidence level. When the confidence level is greater than the confidence threshold, data fusion is performed.

[0009] a data fusion module for fusing multi-view detection results to output an optimal license plate number matching result and a global confidence level;

[0010] a feedback module for determining whether to trigger a manual review instruction according to the global confidence level.

[0011] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition matching of the application, the image acquisition module comprises an acquisition unit and a preprocessing unit:

[0012] The acquisition unit acquires license plate image data in real time through high-resolution cameras deployed at different angle positions;

[0013] The preprocessing unit is configured to preprocess vehicle image data to obtain processed vehicle image data, including:

[0014] performing geometric correction on the vehicle image data through a perspective transformation matrix;

[0015] optimizing image brightness contrast of the vehicle image data through histogram equalization and Gamma correction;

[0016] removing image noise of the vehicle image data through a bilateral filter while keeping the edges clear;

[0017] enhancing character edge features through a sharpening filter.

[0018] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition matching of the application, the vehicle detection module comprises a license plate positioning unit, a character recognition unit and a confidence unit:

[0019] The license plate positioning unit is configured to perform license plate region positioning on the processed vehicle image data through an improved YOLO-v5 network architecture, and output license plate bounding boxes and positioning confidence levels, specifically including:

[0020] inputting the processed vehicle image data through the YOLO network to output license plate bounding boxes and license plate category probabilities ;

[0021] extracting historical bounding box recognition data including predicted bounding box data and corresponding real bounding box data, calculating regression loss of the detection box , and the calculation formula is:

[0022]

[0023]

[0024] wherein, denotes the deviation of the predicted bounding box from the real bounding box, and denote the horizontal coordinates of the predicted bounding box and the real bounding box, respectively, and denote the vertical coordinates of the predicted bounding box and the real bounding box, respectively, and denote the length of the predicted bounding box and the real bounding box, respectively, and denote the height of the predicted bounding box and the real bounding box, respectively, and N denotes the total number of historical bounding box recognition data, denotes the deviation of the predicted bounding box from the real bounding box of the i-th historical bounding box recognition data;

[0025] The average intersection over union is obtained as the IoU overlap degree through the intersection over union of the predicted bounding box and the real bounding box in the historical bounding box recognition data ;

[0026] The localization confidence is calculated according to the regression loss of the detection box, the license plate category probability, and the IoU overlap degree , and the calculation formula is: ; wherein, , and denote localization weights;

[0027] The character recognition unit is configured to perform character recognition on the processed image recognition data according to a preset character recognition network structure, output a character recognition result, and calculate a recognition confidence Cr, and the specific process is as follows:

[0028] The character recognition result is obtained, the classification probability of each character is extracted, the average value of the classification probability of the characters is calculated as the character classification probability, the probability of sequence-to-sequence decoding is taken as the CTC decoding probability according to the sequence-to-sequence decoding, and the sequence length matching is taken as the sequence length matching according to the matching degree of the character recognition length and the standard license plate length.

[0029] The recognition confidence is calculated according to the character classification probability , the CTC decoding probability , and the sequence length matching , and the calculation formula is: ; wherein, , and denote recognition coefficients;

[0030] The credibility unit is configured to perform credibility evaluation on the license plate region positioning result and the character recognition result, and includes:

[0031] The image sharpness score is calculated based on the sharpness evaluation of the Laplacian operator The illumination condition score is calculated by evaluating the uniformity of the image brightness distribution The angle offset score is calculated by the angle deviation between the license plate plane and the camera plane The image quality evaluation confidence Cq is obtained, and the calculation formula is: ; wherein, , and represent the image quality evaluation coefficient;

[0032] The character recognition result is matched with the standard license plate format to obtain the matching degree M_format, the character type is judged according to the character recognition result, and the license plate length is judged according to the license plate boundary box to determine whether the license plate length conforms to the standard, and the format verification confidence Cf is obtained, and the calculation formula is: ; wherein, , and represent the format evaluation coefficient, the character type judgment result, and C represents the license plate length judgment result;

[0033] The confidence of the license plate region positioning result and the character recognition result is evaluated by combining the positioning confidence, the recognition confidence, the image quality evaluation confidence and the format verification confidence, and the evaluation formula is: ; wherein, , , and represent the coefficient;

[0034] When the confidence is greater than the confidence threshold, data fusion is performed; when the confidence is less than or equal to the confidence threshold, an artificial review instruction is triggered, and an artificial review queue is entered.

[0035] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition matching of the application, the data fusion module comprises an initial processing unit, a weight calculation unit, a fusion strategy unit and a threshold adjustment unit:

[0036] The initial processing unit is used to construct a detection result set, standardize the format of the detection results in the detection result set, and normalize the confidence of the detection results, sort the normalized confidence from large to small, and record the confidence level of each view;

[0037] The time stamp of the vehicle image data corresponding to the multi-view detection result is obtained, the detection results exceeding the time window are discarded, and the detection result set is constructed according to the remaining detection results; the format of the detection results in the detection result set is standardized, the character encoding format of the license plate number is unified, and the character position index is unified; the confidence of the detection results in the detection result set is normalized;

[0038] The weight calculation unit is configured to calculate the maximum weight;

[0039] The angle between the view angle and the license plate normal is calculated to obtain the view angle weight, and the calculation formula is:

[0040]

[0041] wherein, the view angle weight of the jth view angle, the normalized confidence of the jth view angle, the angle between the jth view angle and the license plate normal, the normalized confidence of the kth view angle, and M represents the number of view angles;

[0042] The historical accuracy of the detection result of each view angle is obtained, and the maximum weight of each view angle is calculated, and the calculation formula is:

[0043]

[0044] wherein, the historical accuracy of the jth view angle, the maximum weight of the jth view angle;

[0045] The fusion strategy unit is configured to calculate the global confidence of the license plate detection result by character-level weighted voting, sequence-level weighted fusion and confidence weighted average calculation;

[0046] The threshold adjustment unit is configured to dynamically adjust the confidence threshold.

[0047] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition and matching, the fusion strategy unit comprises:

[0048] The detection result is subjected to character-level weighted voting to determine the character position, and the character-level weighted voting comprises:

[0049] For each character position of the license plate, a weighted vote container is prepared for all possible characters, and the initial vote number is zero; the detection result of each camera is traversed: if a certain character is recognized at the current position in the view angle, the maximum weight of the view angle is added to the vote number corresponding to the character; after the voting of all view angles is completed, the weighted vote numbers of the characters are compared, and the character with the largest vote number is selected as the final recognition result of the position; the above process is repeated until the voting of all character positions is completed and the character position is determined;

[0050] The detection result is subjected to sequence weighted fusion, and the sequence weighted fusion comprises:

[0051] The whole license plate string recognized from each view angle is taken as a candidate sequence, for the same candidate sequence, the maximum weight of the view angle and the reliability of the view angle are multiplied to obtain a product, and the product is added to the score of the sequence; in all candidate sequences, the sequence with the highest score is determined as the optimal license plate number matching result;

[0052] The reliability of each view angle is multiplied by the maximum weight of the view angle to obtain a weighted average value as the global confidence of the license plate detection result.

[0053] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition matching, the threshold adjusting unit comprises:

[0054] According to the ratio of the number of effective view angles participating in fusion to the maximum number of view angles supported by the system as a view coverage factor;

[0055] The image quality evaluation confidence of all view angles is obtained, and the comprehensive quality factor is obtained by weighted summation according to the maximum weight of the corresponding view angle;

[0056] The preset basic threshold is multiplied by the view coverage factor and the comprehensive quality factor to obtain the adjusted confidence threshold.

[0057] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition matching, the feedback module triggers the manual review instruction when the global confidence is less than the confidence threshold.

[0058] As a preferred scheme of the intelligent license plate recognition system based on multi-view image acquisition matching, the system further comprises a synchronous control module for uniformly managing the working state of all cameras and synchronizing the time stamps of all cameras; when a vehicle enters the detection area, all cameras are triggered to shoot at the same time.

[0059] Compared with the prior art, the intelligent license plate recognition system based on multi-view image acquisition matching has the following beneficial effects: through multi-view image acquisition and intelligent fusion algorithm, the system recognition accuracy is improved, stable high-precision recognition effect can be maintained in various complex environments, recognition error caused by a single view angle is effectively reduced, and the reliability of the system is significantly improved. The system can maintain high recognition performance under insufficient light, backlight, rain and fog and other adverse weather conditions through the configuration of multiple cameras at different positions and angles, can improve the tolerance range of the system to the inclination angle of the license plate, effectively process the recognition demand when the vehicle passes at various angles, has a wider application range and stronger practicality. Through intelligent system management and adaptive adjustment, the demand for manual intervention and maintenance is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification, illustrate embodiments of the application, and together with the description serve to explain the application, and do not limit the application.

[0061] Figure 1 Fig. 1 is a structural schematic diagram of an intelligent license plate recognition system based on multi-view image acquisition matching according to the present application. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0063] Please refer to Figure 1 In the first embodiment, an intelligent license plate recognition system based on multi-view image acquisition matching is provided, which comprises a data image acquisition module, a license plate detection module, a data fusion module and a feedback module.

[0064] The image acquisition module is used for synchronously acquiring multi-angle vehicle image data and performing preprocessing.

[0065] The vehicle detection module is used for license plate region positioning and license plate character recognition, obtaining a detection result and evaluating a confidence level; when the confidence level is greater than a confidence threshold, data fusion is performed.

[0066] The data fusion module is used for fusing multi-view detection results, outputting an optimal license plate number matching result and a global confidence level.

[0067] The feedback module is used for judging whether to trigger a manual review instruction according to the global confidence level.

[0068] Specifically, the image acquisition module comprises an acquisition unit and a preprocessing unit.

[0069] The acquisition unit acquires license plate image data in real time through high-resolution cameras arranged at different angle positions, and contains 3-5 high-resolution cameras arranged at different angle positions.

[0070] The camera layout adopts a triangular distribution scheme.

[0071] The main camera: front shooting, 2.5-3.0 meters from the ground, 15-20 degrees of pitch angle;

[0072] The left camera: left 30-degree angle shooting, 2.2-2.8 meters from the ground;

[0073] Right camera: 30-degree angle shooting to the right, 2.2-2.8 meters from the ground;

[0074] Auxiliary camera: supplementary angle from the back, 1.8-2.5 meters from the ground;

[0075] The preprocessing unit is used for preprocessing the vehicle image data to obtain processed vehicle image data, comprising:

[0076] The vehicle image data is geometrically corrected by a perspective transformation matrix;

[0077] The image brightness contrast of the vehicle image data is optimized by histogram equalization and Gamma correction;

[0078] The image noise of the vehicle image data is removed by a bilateral filter while keeping the edge clear;

[0079] The character edge features are enhanced by a sharpening filter.

[0080] Specifically, the vehicle detection module includes a license plate positioning unit, a character recognition unit, and a trusted unit:

[0081] The license plate positioning unit is used to position the license plate region of the processed vehicle image data by using an improved YOLO-v5 network architecture, and output the license plate bounding box and positioning confidence, specifically comprising:

[0082] The processed vehicle image data is input by the YOLO network, and the license plate bounding box and the license plate category probability are output ;

[0083] The historical bounding box recognition data, including the predicted bounding box data and the corresponding real bounding box data, is extracted, and the regression loss of the detection frame is calculated , the calculation formula is:

[0084]

[0085]

[0086] wherein, represents the deviation of the predicted bounding box and the real bounding box, and respectively represent the horizontal coordinates of the predicted bounding box and the real bounding box, and respectively represent the vertical coordinates of the predicted bounding box and the real bounding box, and respectively represent the length of the predicted bounding box and the real bounding box, and respectively represent the height of the predicted bounding box and the real bounding box, and N represents the total number of historical bounding box recognition data. the deviation of the predicted bounding box representing the i-th historical bounding box recognition data from the true bounding box;

[0087] obtain the average intersection-over-union as the IoU overlap degree through the intersection-over-union of the predicted bounding box and the true bounding box in the historical bounding box recognition data ;

[0088] calculate the positioning confidence according to the regression loss of the detection frame, the license plate category probability and the IoU overlap degree , the calculation formula is: ; wherein, , and represent the positioning coefficient;

[0089] The character recognition unit is configured to perform character recognition on the processed image recognition data according to a preset character recognition network structure, output a character recognition result, and calculate a recognition confidence Cr, and the specific process is as follows:

[0090] obtain the character recognition result, extract the classification probability of each character, calculate the average value of the classification probability of the characters as the character classification probability, obtain the CTC decoding probability according to the probability of sequence-to-sequence decoding, and obtain the sequence length matching according to the matching degree of the character recognition length and the standard license plate length;

[0091] calculate the recognition confidence according to the character classification probability , the CTC decoding probability and the sequence length matching , and the calculation formula is: ; wherein, , and represent the recognition coefficient;

[0092] The confidence unit is configured to perform confidence evaluation on the license plate region positioning result and the character recognition result, and includes:

[0093] calculate the image sharpness score based on the sharpness evaluation of the Laplacian operator , calculate the lighting condition score through the uniformity evaluation of the image brightness distribution , calculate the angle offset score through the angle deviation between the license plate plane and the camera plane , and obtain the image quality evaluation confidence Cq, and the calculation formula is: ; wherein, , and represent the image quality evaluation coefficient;

[0094] The character recognition result is matched with a standard license plate format to obtain a matching degree M_format, the character type is judged according to the character recognition result, and the license plate length is judged according to the license plate boundary box to obtain a format verification confidence Cf, and the calculation formula is: ; wherein, , and represent a format evaluation coefficient, represents a character type judgment result, and C represents a license plate length judgment result;

[0095] The confidence of the license plate region positioning result and the character recognition result is evaluated in combination with the positioning confidence, the recognition confidence, the image quality evaluation confidence and the format verification confidence, and the evaluation formula is: wherein, , , and represent coefficients;

[0096] When the confidence is greater than a confidence threshold, data fusion is performed; and when the confidence is less than or equal to the confidence threshold, an artificial review instruction is triggered to enter an artificial review queue.

[0097] Specifically, the data fusion module includes an initial processing unit, a weight calculation unit, a fusion strategy unit and a threshold adjustment unit:

[0098] The initial processing unit is configured to construct a detection result set, standardize the format of the detection results in the detection result set, normalize the confidence of the detection results, sort the detection results according to the normalized confidence from large to small, and record the confidence level of each view angle;

[0099] The time stamp of the vehicle image data corresponding to the detection results of multiple views is obtained, the detection results beyond the time window are discarded, and the detection result set is constructed according to the remaining detection results; the format of the detection results in the detection result set is standardized, the character encoding format of the license plate number is unified, and the character position index is unified; the confidence of the detection results in the detection result set is normalized;

[0100] Multi-view data from different cameras are aligned according to timestamps. A 50 ms time window tolerance is set; for each frame of data, the time series of other views are traversed to find one within ±50 ms, which is considered as the observation at the same time; the aligned multi-view data are merged into a synchronized data structure for subsequent processing; using known camera calibration parameters, the detected license plate region coordinates in each view are converted from the pixel coordinate system to the unified world coordinate system: for the pixel-level coordinates of each view, the camera intrinsic parameter matrix and distortion correction are applied first; then, according to the camera pose, such as the rotation matrix and translation vector, the corrected coordinates are projected into the three-dimensional world coordinate system; finally, a set of three-dimensional or planar coordinates represented in the same reference frame is output.

[0101] The weight calculation unit is configured to calculate the maximum weight.

[0102] The weight of each view is calculated according to the angle between the view and the license plate normal, and the calculation formula is:

[0103]

[0104] wherein, represents the view weight of the jth view, represents the normalized confidence of the jth view, represents the angle between the jth view and the license plate normal, represents the normalized confidence of the kth view, and M represents the number of views.

[0105] The historical accuracy of the detection results of each view is obtained, and the maximum weight of each view is calculated, and the calculation formula is:

[0106]

[0107] wherein, represents the historical accuracy of the jth view, represents the maximum weight of the jth view.

[0108] The fusion strategy unit is configured to calculate the global confidence of the license plate detection result by character-level weighted voting, sequence-level weighted fusion, and confidence weighted average.

[0109] The threshold adjustment unit is configured to dynamically adjust the confidence threshold.

[0110] Specifically, the fusion strategy unit:

[0111] The detection result is subjected to character-level weighted voting to determine the character position, including:

[0112] For each character position of the license plate, a weighted vote container is prepared for all possible characters, and the initial vote is zero; traverse the detection results of each camera: if the view angle recognizes a character at the current position, add the maximum weight of the view angle to the vote corresponding to the character; after completing all view angle voting, compare the weighted votes of each character, and the character with the highest vote is selected as the final recognition result of the position; repeat the above process until all character positions complete voting and determine the character position;

[0113] The detection results are sequentially weighted and fused, specifically including:

[0114] The entire license plate string recognized by each view angle is taken as a candidate sequence, and for the same candidate sequence, the maximum weight of the view angle and the view angle reliability product are added to the score of the sequence; among all candidate sequences, the one with the highest score is determined as the optimal license plate number matching result;

[0115] The global confidence of the license plate detection result is obtained by multiplying the reliability of each view angle by its maximum weight and summing them up to obtain a weighted average value.

[0116] Specifically, the threshold adjusting unit comprises:

[0117] According to the ratio of the number of effective view angles currently participating in fusion to the maximum number of view angles supported by the system as a view coverage factor;

[0118] Obtain the image quality evaluation confidence of all view angles, and obtain a comprehensive quality factor by weighted sum according to the maximum weight of the corresponding view angle;

[0119] Multiply the preset basic threshold value by the view coverage factor and the comprehensive quality factor to obtain the adjusted confidence threshold.

[0120] Specifically, the feedback module triggers the manual review instruction when the global confidence is less than the confidence threshold.

[0121] Specifically, it further comprises a synchronous control module for uniformly managing the working state of all cameras and synchronizing the time stamps of all cameras; when a vehicle enters the detection area, all cameras are triggered to take pictures at the same time.

[0122] It is to be noted that, in the present text, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0123] Finally, it should be noted that the above-mentioned only constitutes preferred embodiments of the present application and is not intended to limit the present application, and although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still make modifications to the technical solutions described in the foregoing embodiments or make equivalent replacements to some of the technical features, all of which shall fall within the scope of protection of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the scope of protection of the present application.

Claims

1. An intelligent license plate recognition system based on multi-view image acquisition matching, characterized in that, The system comprises a data image acquisition module, a license plate detection module, a data fusion module and a feedback module: The image acquisition module is used for synchronously acquiring multi-angle vehicle image data and pre-processing; The vehicle detection module is used for license plate region positioning and license plate character recognition, obtaining a detection result and evaluating a confidence level; when the confidence level is greater than a confidence threshold, data fusion is performed; the vehicle detection module comprises a license plate positioning unit, a character recognition unit and a confidence unit: The license plate positioning unit is used for positioning a license plate region of the processed vehicle image data by using an improved YOLO-v5 network architecture, outputting a license plate bounding box and a positioning confidence level, and specifically comprising: Through the YOLO network, input the processed vehicle image data, and output the license plate boundary box and license plate category probability ; extract the historical bounding box recognition data, including the predicted bounding box data and the corresponding real bounding box data, to calculate the regression loss of the detection frame The calculation formula is: ； ； wherein, represents a deviation of the predicted bounding box from the real bounding box, and respectively represent the horizontal coordinates of the predicted bounding box and the real bounding box, and respectively represent the vertical coordinates of the predicted bounding box and the real bounding box, and respectively represent the length of the predicted bounding box and the real bounding box, and respectively represent the height of the predicted bounding box and the real bounding box, represents a deviation of the predicted bounding box from the real bounding box of the i-th historical bounding box recognition data, and N represents the total number of the historical bounding box recognition data; An average intersection over union is obtained as the IoU overlap degree by identifying the intersection over union of the predicted bounding box and the true bounding box in the historical bounding box recognition data ; According to the regression loss of the detection frame, the license plate category probability and the IoU overlap, a positioning confidence is calculated , the calculation formula is: ; wherein, , and represent the positioning coefficient; The character recognition unit is used for recognizing characters in the processed image recognition data according to a preset character recognition network structure, outputting a character recognition result and calculating a recognition confidence level Cr, and the specific process is as follows: The character recognition result is obtained, the classification probability of each character is extracted, the average value of the classification probability of the characters is calculated as the character classification probability; the probability of sequence-to-sequence decoding is taken as the CTC decoding probability; and the matching degree of the character recognition length and the standard license plate length is taken as the sequence length matching; According to the character classification probability , CTC decoding probability , and sequence length matching The recognition confidence is calculated according to the formula: ; wherein , and denote the recognition coefficient; The confidence unit is used for evaluating the confidence level of the license plate region positioning result and the character recognition result, comprising: The clarity evaluation based on Laplacian operator calculates the image clarity score The lighting condition score is calculated by evaluating the uniformity of the image brightness distribution The angle deviation score is calculated by the angle deviation between the license plate plane and the camera plane The image quality evaluation confidence Cq is obtained, and the calculation formula is: ; wherein , and represent the image quality evaluation coefficient; The character recognition result is matched with a standard license plate format to obtain a matching degree M format. Whether the character type is correct is judged according to the character recognition result, and whether the license plate length conforms to the specification is judged according to the license plate boundary box to obtain a format verification confidence Cf. The calculation formula is: ; wherein, , and represent a format evaluation coefficient, represents a character type judgment result, and C represents a license plate length judgment result. The confidence of the detection result is evaluated in combination with the position confidence, the recognition confidence, the image quality evaluation confidence and the format verification confidence, and the evaluation formula is: ; wherein, , , and represent coefficients; When the confidence level is greater than the confidence threshold, data fusion is performed; when the confidence level is less than or equal to the confidence threshold, an artificial review instruction is triggered, and an artificial review queue is entered; The data fusion module is used for fusing multi-view detection results, outputting an optimal license plate number matching result and a global confidence level; The feedback module is used for judging whether to trigger an artificial review instruction according to the global confidence level.

2. The intelligent vehicle license plate recognition system based on multi-view image acquisition matching according to claim 1, characterized in that, The image acquisition module comprises an acquisition unit and a preprocessing unit: The acquisition unit acquires license plate image data in real time by deploying high-resolution cameras at different angle positions; The preprocessing unit is used for preprocessing vehicle image data to obtain processed vehicle image data, comprising: Performing geometric correction on the vehicle image data by using a perspective transformation matrix; Optimizing the image brightness contrast of the vehicle image data by using histogram equalization and Gamma correction; Removing image noise of the vehicle image data by using a bilateral filter while keeping the edges clear; Enhancing the character edge features by using a sharpening filter. 3.The intelligent license plate recognition system based on multi-view image acquisition matching of claim 1, wherein, The data fusion module comprises an initial processing unit, a weight calculation unit, a fusion strategy unit and a threshold adjustment unit: The initial processing unit is used for constructing a detection result set, standardizing the format of the detection results in the detection result set, normalizing the confidence levels corresponding to the detection results, sorting the normalized confidence levels from large to small, and recording the confidence level of each view; The time stamps of the vehicle image data corresponding to the multi-view detection results are obtained, the detection results beyond the time window are discarded, and the detection result set is constructed according to the remaining detection results; The format of the detection results in the detection result set is standardized, the character encoding format of the license plate number is unified, and the character position index is unified; The confidence levels of the detection results in the detection result set are normalized; The weight calculation unit is used for calculating the maximum weight; According to the angle between the view angle and the license plate normal, the view angle weight is calculated, and the calculation formula is: ； wherein, represents the view weight of the jth view, represents the normalized confidence of the jth view, represents the angle between the jth view and the license plate normal, represents the normalized confidence of the kth view, and M represents the number of views. The historical accuracy of the detection result of each view angle is obtained, the maximum weight of each view angle is calculated, and the calculation formula is: ； wherein, represents the historical accuracy rate of the jth view angle, represents the maximum weight of the jth view angle; The fusion strategy unit is used for calculating the global confidence of the license plate detection result by character-level weighted voting, sequence-level weighted fusion and confidence weighted average. The threshold adjusting unit is used for dynamically adjusting the confidence threshold.

4. The intelligent license plate recognition system based on multi-view image acquisition matching according to claim 3, characterized in that, The fusion strategy unit includes: The character-level weighted voting is performed on the detection result to determine the character position, including: For each character position of the license plate, a weighted vote container is prepared for all possible characters, and the initial vote number is zero; the detection result of each camera is traversed: if a character is recognized at the current position in the view angle, the maximum weight of the view angle is added to the vote number corresponding to the character; after all view angles are voted, the weighted vote numbers of each character are compared, and the character with the largest vote number is selected as the final recognition result of the position; the above process is repeated until all character positions complete voting and determine the character position; The sequence weighted fusion is performed on the detection result, including: The whole license plate string recognized by each view angle is taken as a candidate sequence, and for the same candidate sequence, the maximum weight of the view angle and the view angle confidence product are added to the score of the sequence; in all candidate sequences, the one with the highest score is determined as the optimal license plate number matching result; The global confidence of the license plate detection result is obtained by multiplying the confidence of each view angle by its maximum weight and summing them up.

5. The intelligent vehicle license plate recognition system based on multi-view image acquisition matching according to claim 3, characterized in that, The threshold adjusting unit includes: According to the ratio of the number of effective view angles participating in fusion to the maximum number of view angles supported by the system as the view angle coverage factor; The image quality evaluation confidence of all view angles is obtained, and the comprehensive quality factor is obtained by weighted sum according to the maximum weight of the corresponding view angle. The preset basic threshold is multiplied by the view angle coverage factor and the comprehensive quality factor to obtain the adjusted confidence threshold.

6. The intelligent license plate recognition system based on multi-view image acquisition matching of claim 1, wherein, When the global confidence is less than the confidence threshold, the feedback module triggers the manual review instruction.

7. The intelligent license plate recognition system based on multi-view image acquisition matching of claim 1, wherein, The synchronous control module is also included for unified management of the working state of all cameras and synchronization of the time stamps of all cameras; when a vehicle enters the detection area, all cameras are triggered to take pictures at the same time.

Images