A license plate recognition multi-result fusion method

Abstract

The application discloses a license plate recognition multi-result fusion method, which comprises the following steps: selecting and cropping license plate character pictures meeting preset conditions from an original data set to determine corresponding representative pictures; obtaining HOG feature vectors of different characters in the pictures according to the representative pictures, and then determining character similarity between different characters; recognizing a to-be-recognized picture by using several recognition algorithms to obtain whole plate recognition results of the to-be-recognized picture by each algorithm; screening a voteable picture from the to-be-recognized picture according to a weighted voting rule; calculating weighted voting weights corresponding to the several whole plate recognition results, and then determining a voting result; calculating a confidence degree of the voting result to complete license plate recognition multi-result fusion voting. The application has high recognition accuracy, rich application scenarios, can quantize the credibility of the voting result, and can be widely applied to the technical field of computers.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method for fusing multiple results in license plate recognition. Background Technology

[0002] As an important application of target recognition technology, license plate recognition aims to locate the coordinates of vehicle license plates within an image and further extract and recognize the license plate characters. License plate recognition is a crucial step in vehicle information identification; high-precision detection and recognition of license plate information in images is of great significance in areas such as motor vehicle management, checkpoint system construction, and accountability for traffic violations.

[0003] In real-world traffic scenarios, license plate images are often captured in complex environments; lighting intensity, shooting angle, and shooting distance all vary. Using a single recognition algorithm to identify a license plate image results in a discrepancy between the algorithm's performance in ideal experimental conditions and its actual performance. Furthermore, different recognition algorithms often have varying degrees of accuracy when identifying the same license plate image. Therefore, utilizing multiple recognition algorithms to identify the same license plate image and integrating and filtering the results can effectively combine the advantages of multiple algorithms, thereby improving the accuracy of license plate recognition.

[0004] Existing methods for fusing multiple recognition results can be divided into two types: voting strategy methods and Naive Bayes classifier methods. Voting strategy methods can be further divided into hard voting and soft voting. Hard voting treats each recognition result as a vote, selecting the result with the most votes; that is, choosing the result that appears most frequently from multiple recognition results as the fusion result. Soft voting is similar in principle to hard voting, but selects the recognition result with the highest sum of probabilities among all voting results as the fusion result. The probability of each recognition result is given by the recognition algorithm. The Naive Bayes classifier method mainly utilizes Bayes' theorem and the assumption of conditional independence of attributes as a probabilistic classification algorithm. It finds the most likely recognition result class through three steps: finding and classifying feature attributes, training the classifier, and applying the classifier.

[0005] The existing multi-result fusion voting method for license plate recognition has the following drawbacks: (1) It lacks consideration of character similarity during the fusion voting process, which leads to frequent confusion between similar characters; (2) During weighted voting, the voting weights of different methods for license plate recognition results are the same, or the weight of the same method remains unchanged when participating in different sample voting. This results in the voting method having a single application scenario and being unable to handle license plate recognition results in multiple situations; (3) It only gives the final fusion voting result and does not quantify the reliability of the result. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a license plate recognition multi-result fusion method with high recognition accuracy and wide application scenarios, which can quantify the credibility of voting results.

[0007] One aspect of this invention provides a method for fusing multiple results in license plate recognition, comprising:

[0008] Select and crop license plate character images that meet preset conditions from the original dataset to determine the corresponding representative images;

[0009] Based on the representative image, obtain the HOG feature vectors of different characters in the image, and then determine the character similarity between different characters;

[0010] Several recognition algorithms are used to recognize the image to be recognized, and the recognition result of each algorithm for the image to be recognized is obtained. Votable images are then selected from the images to be recognized according to the weighted voting rules.

[0011] Calculate the weighted voting weights corresponding to several whole-card recognition results, and then determine the voting result;

[0012] Calculate the confidence level of the voting results to complete the multi-result fusion voting for license plate recognition.

[0013] Optionally, the step of selecting and cropping license plate character images that meet preset conditions from the original dataset to determine the corresponding representative images includes:

[0014] Select license plate images from the original dataset, and crop out license plate character images that include 24 English characters, 10 numeric characters and 41 Chinese characters from the license plate images;

[0015] Select a representative image corresponding to each license plate character from the license plate character images.

[0016] Optionally, the step of obtaining HOG feature vectors of different characters in the image based on the representative image, and then determining the character similarity between different characters, includes:

[0017] After obtaining the representative image, the representative image is converted to grayscale and corrected for gamma.

[0018] Calculate the gradient direction of each pixel in a representative image and count the frequency of different gradient directions within the region to which each pixel belongs;

[0019] The histograms in each gradient direction are normalized, and finally the normalized histograms are combined into feature vectors according to certain rules.

[0020] After obtaining the feature vectors of different characters, the similarity between each pair of characters is calculated.

[0021] Optionally, after obtaining the feature vectors of different characters, calculating the similarity between pairs of characters includes:

[0022] The similarity between feature vectors of different characters is calculated using cosine distance, and the similarity between feature vectors is regarded as the character similarity between the corresponding characters of the feature vectors.

[0023] Optionally, the step of employing several recognition algorithms to recognize the image to be recognized, obtaining the overall recognition result of each algorithm for the image to be recognized, and selecting voteable images from the images to be recognized according to a weighted voting rule includes:

[0024] If all the identification algorithms participating in the voting produce identification results for the image to be identified and the results are of the same length, then the image to be identified is confirmed as a "voteable image", and the result with the highest weighted probability among all identification results is the voting result.

[0025] If at least one of the identification algorithms participating in the voting fails to identify the image to be identified, or if the lengths of all identification results are inconsistent, then the image to be identified will be confirmed as an "unvoting image".

[0026] Optionally, before the step of using several recognition algorithms to recognize the image to be recognized, obtaining the weighted voting weights corresponding to each whole-card recognition result, and obtaining the weighted voting result based on the weighted voting weights, the method further includes:

[0027] Several recognition algorithms are used to recognize the image to be recognized. From the recognition results of all recognition algorithms, the remaining whole card recognition results after removing the same results are formed into a whole card recognition result set C. For each whole card recognition result in the whole card recognition result set C, the algorithm whose recognition result is the same as that whole card recognition result is used as the source algorithm of that whole card recognition result.

[0028] Select any two different whole card recognition results from the whole card recognition result set C. The two different whole card recognition results have multiple pairs of characters with the same position but different results, which are called difference characters.

[0029] The process employs several recognition algorithms to identify the image to be identified, obtaining a weighted voting weight corresponding to each brand recognition result. The weighted voting result obtained based on the weighted voting weight includes:

[0030] For a character recognition result a k Search for the character recognition result a in the whole card recognition result set C. k Collect all the differing characters and form a set D of differing characters; calculate the character recognition result a. kThe character similarity is calculated between the recognition result of each character in the difference character set D and the character recognition result a, and then the character recognition result a is calculated based on Bayes' theorem and the character similarity. k The probability of accurate identification;

[0031] For any whole card recognition result in the whole card recognition result set C, the probability of accurately recognizing each different character in the whole card recognition result is multiplied to obtain the probability of the whole card recognition result being correctly recognized.

[0032] For any whole card recognition result in the whole card recognition result set C, the probability of correctly recognizing the whole card recognition result is multiplied by the number of source algorithms for the whole card recognition result to obtain the weighted voting weight of the whole card recognition result;

[0033] The weighted voting result is the whole card identification result with the highest weight in the weighted voting.

[0034] Optionally, calculating the confidence level of the voting result includes:

[0035] Before the recognition algorithm starts, 1000 candidate images are randomly selected, and the confidence score of the recognition result of each recognition algorithm for the candidate images is output.

[0036] The data distribution range of the confidence scores of the recognition results of each recognition algorithm is statistically analyzed, and the maximum and minimum values ​​are determined.

[0037] For the image to be identified, the confidence of the recognition algorithm for the image to be identified is obtained, and the confidence is normalized to obtain the normalized confidence. Then, the weighted confidence of each recognition algorithm for the image to be identified is calculated.

[0038] The weighted confidence level is used as the probability that the recognition algorithm correctly identifies the image to be recognized.

[0039] The theoretical minimum probability of correctness of the voting result obtained after weighted voting is used as the confidence level of the weighted voting result.

[0040] Another aspect of this invention provides a license plate recognition multi-result fusion device, comprising:

[0041] The first module is used to select and crop license plate character images that meet preset conditions from the original dataset in order to determine the corresponding representative images;

[0042] The second module is used to obtain the HOG feature vectors of different characters in the image based on the representative image, and then determine the character similarity between different characters;

[0043] The third module is used to identify the image to be identified using several recognition algorithms, obtain the overall recognition result of each algorithm for the image to be identified, and select the images to be voted from the images to be identified according to the weighted voting rules.

[0044] The fourth module is used to calculate the weighted voting weights corresponding to several whole-card recognition results, and then determine the voting results;

[0045] The fifth module is used to calculate the confidence level of the voting results and complete the multi-result fusion voting for license plate recognition.

[0046] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0047] The memory is used to store programs;

[0048] The processor executes the program to implement the method described above.

[0049] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the methods described above.

[0050] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0051] In embodiments of this invention, license plate character images that meet preset conditions are selected and cropped from the original dataset to determine representative images. HOG feature vectors of different characters in the representative images are obtained to determine the character similarity between different characters. Several recognition algorithms are used to recognize the images to be identified, obtaining the whole-plate recognition result for each algorithm. Votable images are selected from the images to be identified according to a weighted voting rule. The weighted voting weights corresponding to the several whole-plate recognition results are calculated to determine the voting result. The confidence level of the voting result is calculated to complete the multi-result fusion voting for license plate recognition. This invention has high recognition accuracy, rich application scenarios, and can quantify the credibility of the voting results. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a license plate recognition multi-result fusion method provided in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] To address the problems of existing technologies, this invention provides a multi-result fusion method for license plate recognition. First, feature vectors of character images are extracted using the Histogram of Oriented Gradients (HOG) algorithm, and the similarity between characters is measured using a distance metric. Then, Bayes' theorem is used to derive the error probability (i.e., the character prior probability) of a particular license plate recognition result based on the character similarity. Next, voting weights for the license plate recognition results are calculated based on the character prior probabilities, and the recognition result with the highest weighted average probability is selected as the voting result. Finally, the confidence level of the voting result is derived and calculated based on the character prior probabilities to quantify the credibility of the voting result.

[0056] This invention studies and implements a multi-result fusion method for license plate recognition. The method consists of two parts: (1) an offline character similarity calculation part, where only one offline calculation is needed during the entire fusion voting process; and (2) an online weighted voting and voting result confidence calculation part, where an online calculation is needed for each license plate image during the fusion voting process. The method flow is as follows: Figure 1 As shown.

[0057] 1. Offline section

[0058] The offline part of this method includes the following three steps: (1) Select enough representative license plate images from the dataset to ensure that all English characters (24), numeric characters (10), and Chinese characters (41, including 31 abbreviations of provinces, autonomous regions, and municipalities and 10 abbreviations of special license plates) that appear in the rules of vehicle license plate composition can be cropped. After cropping, select one representative image corresponding to each character. (2) After obtaining the representative images, first perform image grayscale conversion and Gamma correction on the images, then calculate the gradient direction of each pixel in the preprocessed image, and count the frequency of different gradient directions in the region to which each pixel belongs. Then normalize the histogram in each gradient direction, and finally combine the normalized histograms into HOG feature vectors according to certain rules. (3) After obtaining the feature vectors of different characters, calculate the similarity between characters pairwise. This method uses cosine distance to measure the similarity between two feature vectors. For two different characters, their HOG feature vectors are feature vectors E and F, respectively. i F is the component of the feature vector E in each dimension. i S(E,F) represents the components of the feature vector F in each dimension, and is calculated using the following formula:

[0059]

[0060] In subsequent calculations, we consider the similarity S(E, F) between feature vectors as the character similarity between the two characters corresponding to feature vectors E and F.

[0061] 2. Online section:

[0062] After obtaining the similarity between characters of different license plates offline, this invention calculates the weight of different algorithms' recognition results in a weighted voting process based on character similarity. Then, the recognition results of different algorithms are weighted and voted on to obtain a weighted voting result. Finally, the confidence level of the weighted voting result is calculated.

[0063] 2-1, Data Filtering:

[0064] Before calculating the weights, it is necessary to filter out the images eligible for voting according to the weighted voting rules before proceeding with the next weight calculation. The data filtering rules are set as follows: For the image sample data to be identified, (1) if all license plate recognition algorithms participating in the voting have recognition results for the data and the result length is the same, then the data is a "voteable image", and the result with the highest weighted probability among all recognition results is the voting result; (2) if at least one of the participating algorithms has no recognition result for the data or the length of all recognition results (i.e. the number of characters in the result) is inconsistent, then the data is a "non-voteable image".

[0065] 2-2, Calculate the weights:

[0066] For the image to be identified, assume that several recognition algorithms are used for identification. The remaining complete card recognition results after removing duplicate results from all the algorithm's recognition results form the complete card recognition result set C, R. i Let $S$ be the i-th recognition result in set $C$. For each whole-card recognition result in set $C$, the algorithm whose recognition result is the same as that whole-card recognition result is used as the source algorithm for that whole-card recognition result. Let $S$ be the set of the whole-card recognition results. i The i-th element in} represents R i There are a total of S i The algorithm uses a single source. It selects any two distinct whole-card recognition results from set C. The pairs of characters that are in the same position but have different results between these two distinct whole-card recognition results are called difference characters. R i Together with other recognition results in set C, there are M pairs of differing characters, R i The kth differing character in is 'a' k Search for the character recognition result a in the whole card recognition result set C. k All the differencing characters, and combine all the differencing characters into a differencing character set D. k For the different characters D, excluding a k The probability of error in the recognition result is taken as the prior probability of the character. According to Bayes' theorem and the prior probability of the character: R i The source algorithm for a k The probability of correctly identifying is:

[0067]

[0068] Among them, A k B k a k ,b k The corresponding eigenvector; P i (b k ) is R i The source algorithm for character b k The accuracy of recognition; For R i The source algorithm for character a k The error rate is defined as follows: L is the set of feature vectors corresponding to the characters in the set of differing characters D; G is the feature vector element in set L.

[0069] R i The probability of correctly identifying is:

[0070]

[0071] By incorporating prior character information into the voting weights, this method introduces character similarity as a voting criterion in weighted voting, thereby reducing the possibility of character recognition confusion.

[0072] 2-3, Weighted Voting:

[0073] Recognition result R i The weighted voting weights are:

[0074] V i =S i P(R i )

[0075] The weighted voting result is the license plate recognition result with the highest weight in the weighted voting. Therefore, the final license plate recognition fusion result H after weighted voting is:

[0076]

[0077] 2-4, Confidence level of voting results:

[0078] After obtaining the voting results, the confidence level of the voting results can be calculated. The confidence level of the voting results can be calculated from the confidence levels of the recognition results of the multiple license plate recognition algorithms involved in the voting. Calculating the confidence level of the license plate voting results mainly includes several steps: confidence level normalization and weighted confidence level calculation.

[0079] In the confidence normalization step, the confidence distribution of the target recognition results obtained by different license plate recognition algorithms is normalized to make the confidence of the recognition results of different methods comparable. The normalization method used in this invention is linear normalization, also known as max-min normalization. Specifically, the following operations are performed for algorithm i: (1) Before the recognition algorithm starts, 1000 images are randomly selected, and the confidence of the recognition results of each recognition algorithm for the above images is output. (2) The distribution range of the confidence data of the recognition results of algorithm i is statistically analyzed, and the maximum and minimum values ​​are set to U and U respectively. max U min (3) For the image to be identified, obtain the confidence level D of algorithm i for that sample. i Using formula (4), the following transformation is performed to obtain the normalized confidence level D. i ′:

[0080]

[0081] For the image to be identified, the weight of the recognition result of algorithm i is w. i The confidence level of the normalized recognition result is U. i ′, obtain the weighted reliability T of algorithm i for this sample. i for:

[0082] T i =U i 'w i

[0083] In subsequent calculations, this paper considers the weighted confidence as the probability that algorithm i correctly identifies the image to be identified.

[0084] For the image to be identified, Q recognition algorithms are used. After removing duplicate results from all algorithms, N complete identification results remain. A weighted vote is then taken to obtain a vote result of R. The theoretical minimum correct probability of the weighted vote result is used as the confidence level of the weighted vote result. Therefore, the confidence level of the weighted vote result for the image to be identified is:

[0085]

[0086] Where X is the set of indices of algorithms whose identification result is a weighted voting result R. j The recognition result is R j A set of algorithm indices.

[0087] This method uses the minimum voting accuracy rate calculated from the voting results as the confidence level, and uses this to measure the credibility of the voting results, thus solving the problem that previous methods could not quantify the credibility of voting results.

[0088] In summary, the license plate recognition multi-result fusion voting method of the present invention has the following characteristics:

[0089] (1) This invention proposes and implements a multi-result fusion method for license plate recognition.

[0090] (2) The similarity between license plate characters is measured by the cosine distance between the feature vectors of the license plate character images.

[0091] (3) Based on character similarity, Bayes' theorem is used to derive the error probability of different license plate recognition algorithms, which is used as the prior probability of the character.

[0092] (4) The weighted voting weights are calculated based on the prior probabilities of the characters, taking into account the influence of similar characters on the weighted voting results.

[0093] (5) A method for calculating the confidence level of weighted voting results considering character similarity was derived, which quantifies the credibility of the voting results.

[0094] Compared with the prior art, the present invention has the following advantages:

[0095] This invention first calculates the feature vectors of license plate character images, then calculates the distance between different feature vectors as the similarity between different license plate characters, and then calculates the prior probability of the characters based on the character similarity. Next, based on the prior probability of the characters, the voting weight of different license plate recognition results is calculated in a weighted voting process, and the license plate recognition result with the highest weighted probability is selected as the weighted voting result according to the voting rules. Finally, the confidence level of the weighted voting result is derived and calculated, quantifying the credibility of the weighted voting result. By considering the impact of differences between license plate characters on the weighted voting result, this invention proposes and implements a multi-result fusion method for license plate recognition, which can effectively solve the problem of license plate character recognition confusion caused by similar license plate characters.

[0096] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0097] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0098] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0100] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0101] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0102] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0103] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0104] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for fusing multiple results in license plate recognition, characterized in that, include: Select and crop license plate character images that meet preset conditions from the original dataset to determine the corresponding representative images; Based on the representative image, obtain the HOG feature vectors of different characters in the image, and then determine the character similarity between different characters; Several recognition algorithms are used to recognize the image to be recognized, and the recognition result of each algorithm for the image to be recognized is obtained. Votable images are then selected from the images to be recognized according to the weighted voting rules. Calculate the weighted voting weights corresponding to several whole-card recognition results of the voteable image, and obtain the weighted voting result based on the weighted voting weights; Calculate the confidence level of the voting results to complete the multi-result fusion voting for license plate recognition; Before the step of calculating the weighted voting weights corresponding to several whole-card recognition results of the voteable image, and obtaining the weighted voting result based on the weighted voting weights, the method further includes: Several recognition algorithms are used to recognize the image to be recognized. From the recognition results of all recognition algorithms, the remaining whole card recognition results after removing the same results are formed into a whole card recognition result set C. For each whole card recognition result in the whole card recognition result set C, the algorithm whose recognition result is the same as that whole card recognition result is used as the source algorithm of that whole card recognition result. Select any two different whole card recognition results from the whole card recognition result set C. The two different whole card recognition results have multiple pairs of characters with the same position but different results, which are called difference characters. The calculation of the weighted voting weights corresponding to several whole-card recognition results of the voteable image, and the obtaining of the weighted voting result based on the weighted voting weights, includes: For a character recognition result Search for the character recognition result in the whole card recognition result set C. Collect all the differing characters and form a set D of differing characters; calculate the character recognition result. The character similarity is calculated between the recognition result of each character in the set of different characters D and the character recognition result of Bayes' theorem. Then, the character recognition result is calculated based on the character similarity. The probability of accurate identification; For any whole card recognition result in the whole card recognition result set C, the probability of accurately recognizing each different character in the whole card recognition result is multiplied to obtain the probability of the whole card recognition result being correctly recognized. For any whole card recognition result in the whole card recognition result set C, the probability of correctly recognizing the whole card recognition result is multiplied by the number of source algorithms of the whole card recognition result to obtain the weight of the whole card recognition result; The weighted voting result is the whole card identification result with the highest weight in the weighted voting.

2. The license plate recognition multi-result fusion method according to claim 1, characterized in that, The step of selecting and cropping license plate character images that meet preset conditions from the original dataset to determine the corresponding representative images includes: Select license plate images from the original dataset, and crop out license plate character images that include 24 English characters, 10 numeric characters and 41 Chinese characters from the license plate images; Select a representative image corresponding to each license plate character from the license plate character images.

3. The license plate recognition multi-result fusion method according to claim 2, characterized in that, The step of obtaining HOG feature vectors of different characters in the representative image, and then determining the character similarity between different characters, includes: After obtaining the representative image, the representative image is converted to grayscale and corrected for gamma. Calculate the gradient direction of each pixel in a representative image and count the frequency of different gradient directions within the region to which each pixel belongs; The histograms in each gradient direction are normalized, and finally the normalized histograms are combined into feature vectors according to certain rules. After obtaining the feature vectors of different characters, the similarity between each pair of characters is calculated.

4. The license plate recognition multi-result fusion method according to claim 3, characterized in that, After obtaining the feature vectors of different characters, the similarity between pairs of characters is calculated, including: The similarity between feature vectors of different characters is calculated using cosine distance, and the similarity between feature vectors is regarded as the character similarity between the corresponding characters of the feature vectors.

5. The license plate recognition multi-result fusion method according to claim 1, characterized in that, The process involves employing several recognition algorithms to identify the image to be identified, obtaining the overall recognition result of each algorithm for the image to be identified, and then selecting voteable images from the images to be identified according to a weighted voting rule, including: If all the identification algorithms participating in the voting produce identification results for the image to be identified and the results are of the same length, then the image to be identified is confirmed as a "voteable image", and the result with the highest weighted probability among all identification results is the voting result. If at least one of the identification algorithms participating in the voting fails to identify the image to be identified, or if the lengths of all identification results are inconsistent, then the image to be identified will be confirmed as an "unvoting image".

6. The license plate recognition multi-result fusion method according to claim 1, characterized in that, The calculation of the confidence level of the voting result includes: Before the recognition algorithm starts, 1000 candidate images are randomly selected, and the confidence level of the recognition result of each recognition algorithm for the candidate images is output. The data distribution range of the confidence scores of the recognition results of each recognition algorithm is statistically analyzed, and the maximum and minimum values ​​are determined. For the image to be identified, the confidence of the recognition algorithm for the image to be identified is obtained, and the confidence is normalized to obtain the normalized confidence. Then, the weighted confidence of each recognition algorithm for the image to be identified is calculated. The weighted confidence level is used as the probability that the recognition algorithm correctly identifies the image to be recognized. The theoretical minimum probability of correctness of the voting result obtained after weighted voting is used as the confidence level of the weighted voting result.

7. An apparatus for implementing the license plate recognition multi-result fusion method as described in any one of claims 1-6, characterized in that, include: The first module is used to select and crop license plate character images that meet preset conditions from the original dataset in order to determine the corresponding representative images; The second module is used to obtain the HOG feature vectors of different characters in the image based on the representative image, and then determine the character similarity between different characters; The third module is used to identify the image to be identified using several recognition algorithms, obtain the overall recognition result of each algorithm for the image to be identified, and select the images to be voted from the images to be identified according to the weighted voting rules. The fourth module is used to calculate the weighted voting weights corresponding to several whole-card recognition results, and then determine the voting results; The fifth module is used to calculate the confidence level of the voting results and complete the multi-result fusion voting for license plate recognition.

8. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 6.

Images