A cloud-edge collaborative road sign verification and correction method

Abstract

The application discloses a cloud-edge collaborative road sign verification and correction method, and belongs to the technical field of intelligent verification of road signs. In order to solve the problem of giving consideration to edge real-time performance and cloud verification accuracy, the application comprises the following steps: performing sign candidate detection on continuous road images of road inspection at an edge end; calculating a quality score of a sign image, and selecting an optimal frame sign image to perform sign region cropping to obtain a sign region image, and then encapsulating the sign region image into an edge data packet and uploading the edge data packet to a cloud end; performing line-level text extraction and Chinese-English pairing at the cloud end, obtaining structured entries, and performing initial classification; the classification is divided into road place name entries and non-road place name entries; performing field mapping on the classified structured entries to obtain Chinese fields and English fields, and generating standard reference entries; performing field-by-field verification and error positioning on the obtained fields, extracting error types and error positions of error fields; generating correction contents for the error fields corresponding to the standard reference entries; and the cloud end encapsulates the correction contents into a result packet and outputs the result packet to a road maintenance platform.

Description

Technical Field

[0001] This invention belongs to the field of intelligent verification technology for road signs, specifically relating to a cloud-edge collaborative method for verifying and correcting road signs. Background Technology

[0002] The Chinese place names, pinyin transliterations, English translations of common terms, and layout format of road signs directly affect the accuracy of traffic facility maintenance, road information dissemination, and intelligent driving perception. With the development of digitalization in urban road maintenance and automation of inspections, inspection systems not only need to collect sign images from the road environment, but also need to verify whether the sign content itself complies with national standards, and provide corrective results that can be directly used for maintenance when non-compliant content is found.

[0003] Most existing solutions remain at the stage of "detecting signs and recognizing text," lacking a technical chain for verification and error correction tasks. One type of solution only performs character recognition and cannot determine whether pinyin, general terms, ordinal numbers, directional words, and formats are compliant according to standards such as GB5768.2 "Road Traffic Signs and Markings Part 2: Road Traffic Signs" and GB17733 "Place Name Signs." Another type of solution introduces rule matching, but usually performs character comparison directly on the entire line of text, failing to complete the classification mapping between road names and non-road place names, and failing to generate structured error correction results for each field and each error type. At the same time, in the continuous operation scenario of road inspection, if all the original images are directly uploaded to the cloud, it will cause excessive consumption of communication bandwidth; if detection, recognition, verification, and error correction are all deployed at the edge, it will significantly increase the computing power and storage pressure on the vehicle. Summary of the Invention

[0004] The present invention addresses four common problems in current road sign processing technologies: First, the edge acquisition stage lacks a unified quality quantification and optimal frame selection mechanism; second, text recognition results are difficult to directly map to the field structure required by national standards; third, the error correction process often remains at the level of "providing suggestions" in terms of technical effect description, lacking feasible implementation schemes for generating standard reference items and replacing each field; and fourth, the lack of a cloud-edge collaborative architecture makes it impossible to simultaneously consider edge real-time performance and cloud-based verification accuracy. The present invention proposes a cloud-edge collaborative road sign verification and error correction method.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A cloud-edge collaborative method for verifying and correcting road signs includes the following steps:

[0007] S1. Perform sign candidate detection at the edge of the continuous road image of the road inspection to obtain a set of sign detection boxes. Then, perform candidate association based on the temporal proximity and the degree of overlap of the detection boxes to group the continuous detection boxes belonging to the same physical sign into the same candidate window.

[0008] S2. For all sign images in the same candidate window obtained in step S1, calculate the quality score at the edge end, and select the sign image with the quality score that reaches the quality threshold or the one with the highest quality score as the best frame sign image.

[0009] S3. The edge end performs sign area cropping on all the best frame sign images obtained in step S2 to obtain sign area images, and then encapsulates them into edge data packets and uploads them to the cloud;

[0010] The specific implementation method of step S3 includes the following steps:

[0011] S3.1. At the edge end, the sign detection box in the best frame sign image is expanded to obtain the cropped area; after the cropped area is extracted from the best frame sign image, it is uniformly scaled to a preset size and formed into a sign area image Crop using a lossless compression format;

[0012] S3.2. Crop the image of the signage area and set the acquisition time T. cap Geographic coordinates G pos Vehicle orientation H dir Signage Type sign and the best frame quality score Q best They are jointly encapsulated into edge data packets P edge ;

[0013] S3.3. When the network is normal, upload to the cloud in real time; when the network is abnormal, write to the edge cache queue and re-upload in chronological order after the network is restored.

[0014] S4. The cloud performs line-level text extraction and Chinese-English pairing on the signage area image in the received edge data packets to obtain structured entries;

[0015] S5. The cloud performs initial classification of structured entries into road place name entries and non-road place name entries; then, the classified structured entries are mapped to Chinese fields and English fields to generate standard reference items;

[0016] The specific implementation method of step S5 includes the following steps:

[0017] S5.1. For each structured entry, first perform an initial classification based on the sign type, the general terminology dictionary in the Chinese field, and the location attribute dictionary, dividing the entry into road name entries or non-road place name entries;

[0018] S5.2. For road name entries, split the Chinese name into a proper name field N. spec and the general term field T gen For non-road place name entries, the Chinese name is split into a proper name field N. spec and landmark attribute field A geo Complete the field mapping;

[0019] S5.3. Generate English standard reference items. These items are determined by a rule function and a standard lexicon, and the expression is:

[0020] E std1 =Py(N spec )+Map(T gen E std2 =Lex(N cn E std3 =Py(N spec )+Map(A geo )

[0021] Among them, E std1 For the English standard reference items used for road name entries, E std2 E is a standard reference for non-road place name entries that have official fixed English names. std3 For English standard reference items used for non-road place name entries without official fixed English names, Py(·) represents the proper name transliteration function implemented according to national standards; Map(·) represents the mapping function from common terms or landmark attributes to standard English words; Lex(·) represents the retrieval function of official proper names or conventional translation databases; N cn Indicates the full Chinese name;

[0022] S6. The cloud performs field-by-field validation and error location on the fields mapped in step S5, extracting the error type and error location of the error fields; then, it generates error correction content for the standard reference items corresponding to the error fields.

[0023] S7. The cloud generates a unique sign identifier based on the geographic coordinates, acquisition time, sign type, and sign area image in the edge data packet from step S3. The unique sign identifier, Chinese fields, English fields, standard reference items, error type, error location, correction content, acquisition time, geographic coordinates, and quality score are then uniformly packaged into a result package P. out The data is then output to the road maintenance platform.

[0024] Furthermore, in step S1, a lightweight road sign detection model trained on the road sign dataset using YOLO11n is used to detect candidate signs. The retained sign categories include directional signs, place name signs, prohibitory signs, and speed limit signs.

[0025] Furthermore, in step S2, the quality score is calculated using three indicators: sharpness, brightness fit, and occlusion fit.

[0026] Furthermore, the specific implementation method of step S4 includes the following steps:

[0027] S4.1. Cloud receives edge data packets P edge Then, the visual language model, enhanced and fine-tuned through the road sign content recognition task, is first used to perform line-level text extraction on the sign area image, outputting the line frame coordinates, text content, and language label for each text line, forming a line record set L={l i};

[0028] S4.2. For the row records obtained in step S4.1, complete the matching using rule scores, as shown in the expression:

[0029] P i,j =α y ·Y i,j +α x ·X i,j +α t ·T i,j , and α y +α x +α t =1

[0030] Among them, P i,j Y represents the pairing score between the i-th Chinese line and the j-th English line; i,j X represents the vertical alignment between the i-th Chinese line and the j-th English line; i,j T represents the horizontal proximity between the i-th Chinese line and the j-th English line; i,j This represents the complementarity of language and structure between the i-th Chinese line and the j-th English line. It is set to 1 when the i-th line is Chinese and the j-th line is English, and the field types of the two lines correspond; otherwise, it is set to 0. y α x α t These are the weighting coefficients for the three scores; when P i,j Not lower than the pairing threshold τ p When the i-th Chinese line is paired with the j-th English line, they are considered as the same structured entry.

[0031] Y i,j =max(0,1-|y i -y j | / H s )

[0032] X i,j =max(0,1-|x i -x j | / Ws )

[0033] Among them, y i y j x represents the y-coordinate of the center point of the i-th Chinese line and the j-th English line, respectively; i x j H represents the x-coordinate of the center point of the i-th Chinese line and the j-th English line, respectively; s and W s These represent the height and width of the current signage area, respectively.

[0034] Furthermore, the specific implementation method of step S6 includes the following steps:

[0035] S6.1. Denote the field obtained by mapping in step S5 as R. m The field corresponding to the standard reference item is denoted as S. m Before validation, a unified standardization process is performed, including removing leading and trailing spaces, merging duplicate spaces, unifying capitalization rules, and removing punctuation differences that do not affect meaning, to obtain the standardized fields.

[0036] S6.2. Then, a field-by-field comparison is performed on the normalized fields, using the following expression:

[0037] e m =0, when Norm(R) m )=Norm(S m );

[0038] Otherwise e m =1

[0039] Among them, e m This indicates an error flag for the m-th field; Norm(·) represents the normalization function;

[0040] S6.3. For e m When =1, an error location operation is performed. If the difference only appears in the general term field, it is marked as a general term error; if the difference appears in the proper name pinyin field, it is marked as a pinyin transcription error; if the difference appears in the case, ordinal number, direction word, or connector position, it is marked as a format error; if a required field is missing, it is marked as a missing error.

[0041] The error location is given by comparing the position of the first inconsistent character and the end position of the identification field with the standard reference field;

[0042] When multiple rules are triggered simultaneously for the same field, the visual language big model, which has been enhanced and fine-tuned by the road sign content recognition task, is called to output candidate error category descriptions. The final error type is determined by the comparison results of rule priority and standard reference items.

[0043] The beneficial effects of this invention are:

[0044] The cloud-edge collaborative road sign verification and correction method described in this invention does not directly upload the full original image at the edge. Instead, it first detects sign candidates in consecutive frames and performs quality scoring and best frame selection for the same sign candidate within a short time window. Then, it only uploads the standardized cropped sign area and its spatiotemporal metadata to the cloud. In the cloud, it performs structured parsing of line-level text, language tags, and text box coordinates, and generates corresponding standard reference items through the national standard rule base and terminology database. Finally, it compares the identified fields with the standard reference items field by field and outputs the error type, error location, and correction content.

[0045] The cloud-edge collaborative road sign verification and error correction method described in this invention enables the performance scoring and optimal frame selection of the same sign candidate in consecutive frames, which can improve the consistency and clarity of subsequent verification inputs without uploading the full video.

[0046] The cloud-edge collaborative road sign verification and error correction method described in this invention completes the structured extraction of line-level text, language tags, and text box coordinates, which can reduce the difficulty of adapting traditional OCR to complex layouts and mixed Chinese and English text.

[0047] The cloud-edge collaborative road sign verification and error correction method described in this invention, through a technical chain of "field mapping - standard reference item generation - field-by-field verification - direct replacement error correction", can transform the error correction process from an abstract suggestion into an executable standard content generation process.

[0048] The cloud-edge collaborative road sign verification and error correction method described in this invention involves uploading only edge data packets P. edge Output the result package P in the cloud. out It can balance the real-time performance at the edge, network bandwidth, and cloud verification accuracy. Attached Figure Description

[0049] Figure 1 This is a flowchart of a cloud-edge collaborative road sign verification and error correction method according to the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0051] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0052] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 Detailed explanation is as follows:

[0053] Example 1:

[0054] A cloud-edge collaborative method for verifying and correcting road signs includes the following steps:

[0055] S1. Perform sign candidate detection at the edge of the continuous road image of the road inspection to obtain a set of sign detection boxes. Then, perform candidate association based on the temporal proximity and the degree of overlap of the detection boxes to group the continuous detection boxes belonging to the same physical sign into the same candidate window.

[0056] Furthermore, in step S1, a lightweight road sign detection model trained on the road sign dataset using YOLO11n is used to detect candidate signs. The retained sign categories include directional signs, place name signs, prohibitory signs, and speed limit signs.

[0057] Furthermore, YOLO11n or a similar single-stage detection network is preferably used as the backbone, with the input being a scaled and normalized RGB image and the output being a set of detection boxes B. t The expression is:

[0058] B t ={b i |p i ≥τ d ,c i ∈C sign}

[0059] Among them, B t b represents the set of sign detection boxes retained in frame t; i p represents the i-th detection box; i Indicates the confidence level of the detection box; τ d Indicates the edge detection retention threshold; c i Indicates the detection category; C sign This represents the set of sign categories.

[0060] S2. For all sign images in the same candidate window obtained in step S1, calculate the quality score at the edge end, and select the sign image with the quality score that reaches the quality threshold or the one with the highest quality score as the best frame sign image.

[0061] Furthermore, in step S2, the quality score is calculated using three indicators: sharpness, brightness fit, and occlusion fit.

[0062] Furthermore, the expression for the quality fraction is:

[0063] Q k =w s ·S k +w b ·B k +w o ·O k ,and w s +w b +w o =1S k =min(1,L k / L0)B k =max(0,1-|μ k -μ0| / Δμ)O k =max(0,1-A occ,k / A box,k )k * =argmax 1≤k≤N_{r}+1 Q k

[0064] Among them, Q k S represents the overall quality score of the k-th frame; k Indicates the sharpness score; B k Indicates brightness adaptation score; O k Indicates the occlusion adaptation score; w s w b w o L represents the weighting coefficients of the three scores; k L0 represents the Laplacian variance of the signage region in the k-th frame; L0 represents the sharpness normalization baseline; μ k The average brightness of the sign area in the k-th frame is represented by μ0; the center value of the target brightness is represented by μ0; Δμ represents the allowable brightness deviation; A occ,k A represents the area of ​​the occluded region in the k-th frame; box,k N represents the area of ​​the signage area. r Indicates the number of additional resampling or continuous sampling frames; k * This indicates the sequence number of the frame with the highest quality.

[0065] When Q k Reaching the quality threshold τ qWhen the frame at the edge is below τ, the corresponding frame is directly taken as the best frame; when the quality score of all frames in the window is below τ, the frame at the edge is taken as the best frame. q At the same time, the system still selects the highest quality frame as the upload object, and attaches a low quality mark, so that the cloud can output a manual review prompt in the subsequent results.

[0066] S3. The edge end performs sign area cropping on all the best frame sign images obtained in step S2 to obtain sign area images, and then encapsulates them into edge data packets and uploads them to the cloud;

[0067] The specific implementation method of step S3 includes the following steps:

[0068] S3.1. At the edge end, the sign detection box in the best frame sign image is expanded to obtain the cropped area; after the cropped area is extracted from the best frame sign image, it is uniformly scaled to a preset size and formed into a sign area image Crop using a lossless compression format;

[0069] S3.2. Crop the image of the signage area and set the acquisition time T. cap Geographic coordinates G pos Vehicle orientation H dir Signage Type sign and the best frame quality score Q best They are jointly encapsulated into edge data packets P edge ;

[0070] Furthermore, the Crop is obtained from the original image after boundary expansion of the optimal frame sign detection box output in step S1; the acquisition time T cap Geographic coordinates G pos and the vehicle facing H dir All metadata is taken from the synchronous positioning time corresponding to the sampling time when the best frame was generated; Sign Type sign The category label of the best frame detection box taken from step S1; the best frame quality score Q best The scoring results are taken from step S2.

[0071] S3.3. When the network is normal, upload to the cloud in real time; when the network is abnormal, write to the edge cache queue and re-upload in chronological order after the network is restored.

[0072] S4. The cloud performs line-level text extraction and Chinese-English pairing on the signage area image in the received edge data packets to obtain structured entries;

[0073] Furthermore, the specific implementation method of step S4 includes the following steps:

[0074] S4.1. Cloud receives edge data packets P edgeThen, the visual language model, enhanced and fine-tuned through the road sign content recognition task, is first used to perform line-level text extraction on the sign area image, outputting the line frame coordinates, text content, and language label for each text line, forming a line record set L={l i};

[0075] Furthermore, the visual language model enhanced and fine-tuned for the road sign content recognition task preferably adopts Qwen2.5-VL-7B or a similar level image-text joint understanding model; the visual language model enhanced and fine-tuned for the road sign content recognition task provides candidate segmentation results when there is ambiguity in row segmentation;

[0076] S4.2. For the row records obtained in step S4.1, complete the matching using rule scores, as shown in the expression:

[0077] P i,j =α y ·Y i,j +α x ·X i,j +α t ·T i,j , and α y +α x +α t =1

[0078] Among them, P i,j Y represents the pairing score between the i-th Chinese line and the j-th English line; i,j X represents the vertical alignment between the i-th Chinese line and the j-th English line; i,j T represents the horizontal proximity between the i-th Chinese line and the j-th English line; i,j This represents the complementarity of language and structure between the i-th Chinese line and the j-th English line. It is set to 1 when the i-th line is Chinese and the j-th line is English, and the field types of the two lines correspond; otherwise, it is set to 0. y α x α t These are the weighting coefficients for the three scores; when P i,j Not lower than the pairing threshold τ p When the i-th Chinese line is paired with the j-th English line, they are considered as the same structured entry.

[0079] Y i,j =max(0,1-|y i -y j | / H s )

[0080] X i,j =max(0,1-|x i -x j | / W s )

[0081] Among them, y i y j x represents the y-coordinate of the center point of the i-th Chinese line and the j-th English line, respectively; i x j H represents the x-coordinate of the center point of the i-th Chinese line and the j-th English line, respectively; s and W s These represent the height and width of the current signage area, respectively.

[0082] S5. The cloud performs initial classification of structured entries into road place name entries and non-road place name entries; then, the classified structured entries are mapped to Chinese fields and English fields to generate standard reference items;

[0083] Furthermore, the specific implementation method of step S5 includes the following steps:

[0084] S5.1. For each structured entry, first perform an initial classification based on the sign type, the general terminology dictionary in the Chinese field, and the location attribute dictionary, dividing the entry into road name entries or non-road place name entries;

[0085] Furthermore, when dictionary rules cannot uniquely determine the name type or field boundaries, the cloud-based large model M is then invoked. l Output candidate category labels and field segmentation suggestions, and adopt them only if the suggestions are consistent with the rule base constraints.

[0086] S5.2. For road name entries, split the Chinese name into a proper name field N. spec and the general term field T gen For non-road place name entries, the Chinese name is split into a proper name field N. spec and landmark attribute field A geo Complete the field mapping;

[0087] S5.3. Generate English standard reference items. These items are determined by a rule function and a standard lexicon, and the expression is:

[0088] E std1 =Py(N spec )+Map(T gen E std2 =Lex(N cn E std3 =Py(N spec )+Map(A geo )

[0089] Among them, E std1 For the English standard reference items used for road name entries, E std2E is a standard reference for non-road place name entries that have official fixed English names. std3 For English standard reference items used for non-road place name entries without official fixed English names, Py(·) represents the proper name transliteration function implemented according to national standards; Map(·) represents the mapping function from common terms or landmark attributes to standard English words; Lex(·) represents the retrieval function of official proper names or conventional translation databases; N cn This indicates the complete Chinese name. In other words, road names are primarily generated using the method of "property name pinyin + standard translation of general terminology" to create a standard reference item; for non-road place names, if a fixed English name already exists in the official thesaurus, that fixed translation is directly used; if no fixed translation exists, the method of "property name pinyin + standard translation of landmark attribute" is used to generate a standard reference item.

[0090] S6. The cloud performs field-by-field validation and error location on the fields mapped in step S5, extracting the error type and error location of the error fields; then, it generates error correction content for the standard reference items corresponding to the error fields.

[0091] Furthermore, the specific implementation method of step S6 includes the following steps:

[0092] S6.1. Denote the field obtained by mapping in step S5 as R. m The field corresponding to the standard reference item is denoted as S. m Before validation, a unified standardization process is performed, including removing leading and trailing spaces, merging duplicate spaces, unifying capitalization rules, and removing punctuation differences that do not affect meaning, to obtain the standardized fields.

[0093] S6.2. Then, a field-by-field comparison is performed on the normalized fields, using the following expression:

[0094] e m =0, when Norm(R) m )=Norm(S m );

[0095] Otherwise e m =1

[0096] Among them, e m This indicates an error flag for the m-th field; Norm(·) represents the normalization function;

[0097] S6.3. For e m When =1, an error location operation is performed. If the difference only appears in the general term field, it is marked as a general term error; if the difference appears in the proper name pinyin field, it is marked as a pinyin transcription error; if the difference appears in the case, ordinal number, direction word, or connector position, it is marked as a format error; if a required field is missing, it is marked as a missing error.

[0098] The error location is given by comparing the position of the first inconsistent character and the end position of the identification field with the standard reference field;

[0099] When multiple rules are triggered simultaneously for the same field, the visual language big model, which has been enhanced and fine-tuned by the road sign content recognition task, is called to output candidate error category descriptions. The final error type is determined by the comparison results of rule priority and standard reference items.

[0100] Furthermore, the rule priority is set as follows: missing errors > pinyin transcription errors > general terminology errors > formatting errors.

[0101] Furthermore, for each field containing an error, the corresponding standard reference field S is directly used. m As correction content C m For fields without errors, C m Set to null. This avoids the problem of uncertainty in "error correction suggestions," ensuring that the error correction content is uniquely determined by the standard reference item, rather than being freely generated by the large model.

[0102] S7. The cloud generates a unique sign identifier based on the geographic coordinates, acquisition time, sign type, and sign area image in the edge data packet from step S3. The unique sign identifier, Chinese fields, English fields, standard reference items, error type, error location, correction content, acquisition time, geographic coordinates, and quality score are then uniformly packaged into a result package P. out The data is then output to the road maintenance platform.

[0103] Furthermore, each entry in the results package contains three sets of fields: original content, standard content, and error correction content, which facilitates the platform in directly generating disease records and rectification suggestions.

[0104] The parameters set in this embodiment are as follows:

[0105] 1. Edge detection preservation threshold τ d It can be set to 0.40–0.60; the candidate window length can be set to 3–6 frames; the quality threshold τ q It can be set to 0.65~0.80;

[0106] 2. The optimal weight for the quality score is w. s =0.50、w b =0.30、w o =0.20;

[0107] 3. Pairing threshold τ p It can be set to 0.60~0.75;

[0108] 4. Edge data packet P edge At least including Crop, Tcap G pos H dir Type sign and Q best Six types of fields;

[0109] 5. Result package P out It should include at least the sign's unique identifier, original field, standard reference field, error type, error location, and correction content.

[0110] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0111] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A cloud-edge collaborative method for verifying and correcting road signs, characterized in that, Includes the following steps: S1. Perform sign candidate detection at the edge of the continuous road image of the road inspection to obtain a set of sign detection boxes. Then, perform candidate association based on the temporal proximity and the degree of overlap of the detection boxes to group the continuous detection boxes belonging to the same physical sign into the same candidate window. S2. For all sign images in the same candidate window obtained in step S1, calculate the quality score at the edge end, and select the sign image with the quality score that reaches the quality threshold or the sign image with the highest quality score as the best frame sign image; S3. The edge end performs sign area cropping on all the best frame sign images obtained in step S2 to obtain sign area images, and then encapsulates them into edge data packets and uploads them to the cloud; The specific implementation method of step S3 includes the following steps: S3.

1. At the edge end, the sign detection box in the best frame sign image is expanded to obtain the cropped area; after the cropped area is extracted from the best frame sign image, it is uniformly scaled to the preset size and formed into a sign area image Crop using a lossless compression format; S3.

2. Crop the image of the signage area and set the acquisition time T. cap Geographic coordinates G pos Vehicle orientation H dir Signage Type sign and the best frame quality score Q best They are encapsulated together as edge data packets P edge ; S3.

3. When the network is normal, upload to the cloud in real time; when the network is abnormal, write to the edge cache queue and re-upload in chronological order after the network is restored. S4. The cloud performs line-level text extraction and Chinese-English pairing on the signage area image in the received edge data packets to obtain structured entries; S5. The cloud performs initial classification of structured entries into road place name entries and non-road place name entries; then, the classified structured entries are mapped to Chinese fields and English fields to generate standard reference items; The specific implementation method of step S5 includes the following steps: S5.

1. For each structured entry, first perform an initial classification based on the sign type, the general terminology dictionary in the Chinese field, and the location attribute dictionary, dividing the entry into road name entries or non-road place name entries; S5.

2. For road name entries, split the Chinese name into a proper name field N. spec and the general term field T gen For non-road place name entries, the Chinese name is split into a proper name field N. spec and landmark attribute field A geo Complete the field mapping; S5.

3. Generate English standard reference items. These items are determined by a rule function and a standard lexicon, and the expression is: E std1 =Py(N spec )+Map(T gen )E std2 =Lex(N cn )E std3 =Py(N spec )+Map(A geo ); Among them, E std1 For the English standard reference items used for road name entries, E std2 E is a standard reference for non-road place name entries that have official fixed English names. std3 For English standard reference items used for non-road place name entries without official fixed English names, Py(·) represents the proper name transliteration function implemented according to national standards; Map(·) represents the mapping function from common terms or landmark attributes to standard English words; Lex(·) represents the retrieval function of official proper names or conventional translation databases; N cn Indicates the full Chinese name; S6. The cloud performs field-by-field validation and error location on the fields mapped in step S5, extracting the error type and error location of the error fields; then, it generates error correction content for the standard reference items corresponding to the error fields. S7. The cloud generates a unique sign identifier based on the geographic coordinates, acquisition time, sign type, and sign area image in the edge data packet from step S3. The unique sign identifier, Chinese fields, English fields, standard reference items, error type, error location, correction content, acquisition time, geographic coordinates, and quality score are then uniformly packaged into a result package P. out The data is then output to the road maintenance platform.

2. The cloud-edge collaborative road sign verification and error correction method according to claim 1, characterized in that, In step S1, a lightweight road sign detection model trained on the road sign dataset using YOLO11n is used to detect candidate signs. The retained sign categories include directional signs, place name signs, prohibitory signs, and speed limit signs.

3. The cloud-edge collaborative road sign verification and error correction method according to claim 2, characterized in that, In step S2, the quality score is calculated using three indicators: sharpness, brightness fit, and occlusion fit.

4. The cloud-edge collaborative road sign verification and error correction method according to claim 3, characterized in that, The specific implementation method of step S4 includes the following steps: S4.

1. Cloud receives edge data packets P edge Then, the visual language model, enhanced and fine-tuned through the road sign content recognition task, is first used to perform line-level text extraction on the sign area image, outputting the line frame coordinates, text content, and language label for each text line, forming a line record set L={l i }; S4.

2. For the row records obtained in step S4.1, complete the matching using rule scores, as shown in the expression: P i,j = α y · Y i,j + α x · X i,j + α t · T i,j and α y + α x + α t = 1; Among them, P i,j Y represents the pairing score between the i-th Chinese line and the j-th English line; i,j X represents the vertical alignment between the i-th Chinese line and the j-th English line; i,j T represents the horizontal proximity between the i-th Chinese line and the j-th English line; i,j This represents the complementarity of language and structure between the i-th Chinese line and the j-th English line. It is set to 1 when the i-th line is Chinese and the j-th line is English, and the field types of the two lines correspond; otherwise, it is set to 0. y α x α t These are the weighting coefficients for the three scores; when P i,j Not lower than the pairing threshold τ p When the i-th Chinese line is paired with the j-th English line, they are combined into the same structured entry. AND i,j =max(0,1-|y i -and j | / H s ); X i,j =max(0,1-|x i -x j | / W s ); Among them, y i y j x represents the y-coordinate of the center point of the i-th Chinese line and the j-th English line, respectively; i x j H represents the x-coordinate of the center point of the i-th Chinese line and the j-th English line, respectively; s and W s These represent the height and width of the current signage area, respectively.

5. The cloud-edge collaborative road sign verification and error correction method according to claim 4, characterized in that, The specific implementation method of step S6 includes the following steps: S6.

1. Denote the field obtained by mapping in step S5 as R. m The field corresponding to the standard reference item is denoted as S. m Before validation, a unified standardization process is performed, including removing leading and trailing spaces, merging duplicate spaces, unifying capitalization rules, and removing punctuation differences that do not affect meaning, to obtain the standardized fields. S6.

2. Then, a field-by-field comparison is performed on the normalized fields, using the following expression: e m =0, when Norm(R) m )=Norm(S m ); Otherwise e m =1; Among them, e m This indicates an error flag for the m-th field; Norm(·) represents the normalization function; S6.

3. For e m When =1, an error location operation is performed. If the difference only appears in the general term field, it is marked as a general term error; if the difference appears in the proper name pinyin field, it is marked as a pinyin transcription error; if the difference appears in the case, ordinal number, direction word, or connector position, it is marked as a format error; if a required field is missing, it is marked as a missing error. The error location is given by comparing the position of the first inconsistent character and the end position of the identification field with the standard reference field; When multiple rules are triggered simultaneously for the same field, the visual language big model, which has been enhanced and fine-tuned by the road sign content recognition task, is called to output candidate error category descriptions. The final error type is determined by the comparison results of rule priority and standard reference items.

Images