Automated data review system based on visual scanning

Abstract

The application discloses an automatic data review system based on visual scanning and relates to the technical field of data review.The automatic data review system comprises a seal image generation unit, a binary graph construction unit, a positioning coordinate determination unit and a positioning coordinate optimization unit.The positioning coordinate optimization unit adjusts the positions of all pixel points in the image, so that the CRNN model can be more accurate in the recognition process.The content verification unit is used for image scanning on the region to be reviewed, so that the text information of the seal part in the region can be accurately recognized.After multiple rounds of verification, in order to prevent the missing data in the template library from causing verification failure, manual correction is introduced after the first round of verification, a fitting seal record is additionally generated according to the text information of the manual correction, and verification is initiated again according to the seal record, so that the adaptability of the review process is enhanced to a certain extent.

Description

Technical Field

[0001] This invention relates to the field of data verification technology, specifically to an automated data verification system based on visual scanning. Background Technology

[0002] An automated data verification system based on visual scanning captures the stamped area using high-definition image acquisition equipment, extracts character information using image processing technology, and compares it with standard data. This can quickly identify problems such as blurred characters and misprints, reducing the number of manual verifications and improving the efficiency and accuracy of stamped information verification. Patent application number 202210489379.1 discloses "an image data verification method and system based on a neural network classifier. The method includes: setting an input vector for the image data to be verified, the input vector being a vector composed of area feature values, energy feature values, and region number feature values ​​of the image data to be verified; inputting the input vector into a pre-trained neural network classifier to convert it into an output vector; and determining whether the image data to be verified is a false positive based on the value of the output vector. The neural network classifier is trained using image data samples with known results, and the feature dimensions of the image data samples include area feature values, energy feature values, and region number feature values." This method and system solve the technical problems of low efficiency and high workload in secondary image verification.

[0003] The aforementioned existing technologies have solved problems such as low verification efficiency. However, when the system is running, since the text of the stamped part is not identified separately, this part of the text often cannot be effectively obtained. It still relies on manual verification of the stamped part. At the same time, the standard data cannot be updated in real time during the verification process, which makes the data easily misdetected as abnormal during the verification process, resulting in poor overall system adaptability. Summary of the Invention

[0004] The purpose of this invention is to provide an automated data verification system based on visual scanning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automated data verification system based on visual scanning, comprising a stamp image generation unit, a binary image construction unit, a positioning coordinate determination unit, a positioning coordinate optimization unit, and a content verification unit;

[0006] The steel stamp image generation unit collects multiple real steel stamp image data, stores the real steel stamp image data as samples in a set, and generates multiple plain text image data using the SDXL model. Then, it extracts attribute representations from the samples and plain text image data respectively, and uses the attribute representations to generate new steel stamp image data.

[0007] Binary image construction unit: After extracting the stamp image data, the binary image construction unit transmits it to the CNN model for analysis to obtain the corresponding binary image;

[0008] The positioning coordinate determination unit, after acquiring the binary image, inputs the image data into the CRAFT model for analysis, and uses morphological closing operation to connect discontinuous parts of text strokes in the image, removes noise, and obtains the four corner coordinates of the recognition box corresponding to all text in the image. Based on the four corner coordinates, the tilt angle and actual positioning coordinates of each recognition box are calculated.

[0009] The positioning coordinate optimization unit sets constraint values, counts the tilt angle of each recognition box, and determines that the box is a non-orthogonal text box if the tilt angle of the recognition box is greater than the constraint value, and otherwise determines that the box is an orthogonal text box. The optimized positioning coordinates of the two types of text boxes are determined respectively.

[0010] The content verification unit acquires the original image of the area to be verified, determines the text information, text spacing, and the depth and width of the stamped edge in the image, and then verifies the image. If the verification fails, it performs manual correction, generates a fitted stamped record, and then performs verification again.

[0011] Preferably, the specific process for extracting attribute representations from the sample and plain text image data in the stamp image generation unit is as follows:

[0012] S101. After randomly selecting a sample from the set, input it into the CNN model for analysis to obtain feature maps of different sizes. ,in The included channels are , Indicates the total number of feature maps. express The total number of channels contained within;

[0013] S102, Statistics middle Channels and The first in the passage Pixel value at position ,according to Calculate The channel is in Common intensity value at location ,in Indicates the position number;

[0014] S103. After determining the intensity values ​​at different locations, sum them up to obtain... and The correlation coefficient between channels is calculated repeatedly until it is obtained. The correlation coefficient between each channel and other channels;

[0015] S104. Calculate the correlation coefficient between each channel in all feature maps, set the corresponding weight for each feature map, calculate the comprehensive correlation coefficient between each channel according to the weight and correlation coefficient of all feature maps, analyze it, and obtain the texture attribute representation of the current sample.

[0016] S105. After obtaining the texture attribute representation of all samples, extract the text attribute representation of the plain text image data.

[0017] Preferably, the specific process of generating new stamp image data using attribute representation in the stamp image generation unit is as follows:

[0018] S201. Arbitrarily select a single sample and plain text image data, determine the statistical measures corresponding to the texture attribute representation of the sample and the text attribute representation of the plain text image data, align the two statistical measures, and use a statistical analysis algorithm to obtain the fused attribute representation. Generate the corresponding stamp image data based on the fused attribute representation. The statistical analysis algorithm is specifically as follows:

[0019] ;

[0020] in, This represents the attribute representation after fusion. The first and second statistics represent the textual attribute characteristics. The first and second statistics represent the texture attribute characteristics. Represents the attributes of text. Represents texture attributes;

[0021] S202. Repeat the operation until each plain text image data has a corresponding stamped image data for each sample. Perform image blurring, noise addition, and geometric transformation operations on the real stamped image data and the generated stamped image data respectively to obtain the processed image data. After adding the corresponding text labels to all image data, store them in the first dataset.

[0022] Preferably, the specific process of obtaining the corresponding binary map in the binary map construction unit is as follows:

[0023] S301. Obtain feature maps of different sizes corresponding to the image data, fuse all feature maps, and then use the convolutional layers and activation functions in the CNN model to generate the corresponding probability map and threshold map.

[0024] S302, Reading probability map, threshold map, first Pixel value of each pixel and ,use and Calculate the corresponding difference ,in , Indicates the pixel index;

[0025] S303, if Then the first The pixel value of each pixel is set to 1, and the pixel value of the other pixel is set to 0. This process is repeated until the pixel values ​​of all pixels are determined. Then, a binary image is constructed based on the current pixel values.

[0026] Preferably, the specific process by which the positioning coordinate determination unit calculates the tilt angle and actual positioning coordinates of each recognition frame based on the four corner coordinates is as follows:

[0027] S401, Based on the coordinates of the four corners of the first recognition box , , and The tilt angle of the recognition box was analyzed. ,in The coordinates of the four corners are, in order, the coordinates of the top left corner, the top right corner, the bottom left corner, and the bottom right corner;

[0028] S402, according to , , and Calculate the initial first positioning coordinates of the current recognition box. and initial second positioning coordinates ,in , , , ;

[0029] S403, according to the initial first positioning coordinates Initial second positioning coordinates Within the binary image, perform downward and upward searches respectively to determine the pixel that is closest to the initial first positioning coordinates and has a pixel value of 1. Use its location as the actual first positioning coordinates. Also, analyze the pixel that is closest to the initial second positioning coordinates and has a pixel value of 1, and use its location as the actual second positioning coordinates.

[0030] S404. Repeat the operation until the tilt angle, actual first positioning coordinates and actual second positioning coordinates of all recognition boxes are calculated.

[0031] Preferably, the specific process for determining the optimized positioning coordinates of the two types of text boxes in the positioning coordinate optimization unit is as follows:

[0032] S501. After determining the actual first positioning coordinates and actual second positioning coordinates of all non-positive text boxes, calculate the difference between the actual first positioning coordinates and the actual second positioning coordinates in each text box, take the maximum difference as the spacing value, use the DBSCAN clustering algorithm to group the ordinates in all first positioning coordinates, and calculate the average number of ordinates in each group. If the difference between the average numbers of the groups is less than the spacing value, then the current two groups are merged into one group; otherwise, no operation is performed.

[0033] S502. After calculating the average of the ordinates for all groups, set the ordinate value for each group according to the average, and then... The ordinate values ​​of each group The actual first positioning coordinates within this group were analyzed. The corresponding optimized first positioning coordinates ,in , ,according to Determine the corresponding actual second positioning coordinates And calculate the optimized second positioning coordinates corresponding to the actual second positioning coordinates. ,in , , Indicates the group number;

[0034] S503. After determining the actual first positioning coordinates and actual second positioning coordinates of all positive text boxes, use the actual positioning coordinates as the optimized positioning coordinates.

[0035] S504. Construct a position transformation function based on the actual and optimized positioning coordinates of all text boxes, calculate the actual value of each parameter in the position transformation function, and determine the adjusted position coordinates of each pixel in the image data based on the position transformation function.

[0036] S505. Store all adjusted image data and corresponding text labels as samples in the second dataset, and transmit the samples to the CRNN model for analysis to determine the model parameters.

[0037] Preferably, the specific process for determining the text information, text spacing, and the depth and width of the stamped edge in the content verification unit is as follows:

[0038] S601. Obtain the original image of the region to be reviewed, preprocess it, determine the actual and optimized positioning coordinates of all text recognition boxes in the original image, and construct a position transformation function based on the actual and optimized positioning coordinates.

[0039] S602. After adjusting the image data through the position transformation function to obtain new image data, the new image data is transmitted to the CRNN model for analysis, and the text information of all recognition boxes is output.

[0040] S603. After counting all the text information in the area to be reviewed, determine the spacing between each text according to the actual positioning coordinates of the recognition box, and use a laser profilometer to scan the area to be reviewed, and output the depth and width of the steel stamp edge in the area.

[0041] Preferably, the specific process for verifying the image in the content verification unit is as follows:

[0042] S701. Set a threshold and compare the text information, text spacing, and edge depth and width of the area to be reviewed with all steel stamp records in the template library one by one. If the similarity is higher than the threshold, the verification is deemed to pass and the text information of the steel stamp record is output. Otherwise, a text abnormality is indicated and the text information is manually corrected.

[0043] S702. Determine the text attribute representation in the current area to be reviewed based on the corrected text information, analyze the original image of the area to be reviewed using a CNN model to obtain the texture attribute representation in the current area to be reviewed, generate the corresponding stamp image data using the text attribute representation and texture attribute representation, and analyze the fitted stamp record through the stamp image data.

[0044] S703. Calculate the similarity between the area to be reviewed and the fitted steel stamp record. If the similarity is lower than the threshold, the verification is deemed to have failed and an error message is displayed. Otherwise, the verification is deemed to have passed and the steel stamp record is stored in the template library.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. This invention uses a steel stamp image generation unit to fuse and analyze existing real steel stamp image data and plain text image data to obtain new steel stamp image data. This design ensures a sufficient number of samples, facilitating the smooth training of the subsequent recognition model. On the other hand, it provides suitable steel stamp image data for the area to be reviewed after the first round of verification fails, enabling the rapid generation of fitted steel stamp records. Meanwhile, the binary image construction unit introduces probability maps and threshold maps, making the edges of the generated binary images clearer. The positioning coordinate determination unit analyzes the four corner coordinates of the recognition boxes corresponding to all text in the image to obtain initial positioning coordinates, which are then combined with the binary image to make the final actual positioning coordinates more accurate.

[0047] 2. This invention adjusts the position of all pixels in the image through a positioning coordinate optimization unit to ensure that the CRNN model can be more accurate in the recognition process. The content verification unit scans the area to be verified, so that the text information of the stamped part in the area can be accurately identified. After multiple rounds of verification, in order to prevent verification failure due to missing data in the template library, manual correction is introduced after the first round of verification. Based on the manually corrected text information, an additional fitted stamped record is generated. Verification is initiated again based on the stamped record, which enhances the adaptability of the verification process to a certain extent. Attached Figure Description

[0048] Figure 1 A schematic diagram of the overall system flow is provided for embodiments of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1:

[0051] Please see Figure 1 The present invention provides a technical solution: an automated data verification system based on visual scanning, comprising a stamp image generation unit, a binary image construction unit, a positioning coordinate determination unit, a positioning coordinate optimization unit, and a content verification unit;

[0052] The steel stamp image generation unit collects multiple real steel stamp image data, stores the real steel stamp image data as samples in a set, and generates multiple plain text image data using the SDXL model. Then, it extracts attribute representations from the samples and plain text image data respectively, and uses the attribute representations to generate new steel stamp image data.

[0053] Binary image construction unit: After extracting the stamp image data, the binary image construction unit transmits it to the CNN model for analysis to obtain the corresponding binary image;

[0054] The positioning coordinate determination unit, after acquiring the binary image, inputs the image data into the CRAFT model for analysis, and uses morphological closing operation to connect discontinuous parts of text strokes in the image, removes noise, and obtains the four corner coordinates of the recognition box corresponding to all text in the image. Based on the four corner coordinates, the tilt angle and actual positioning coordinates of each recognition box are calculated.

[0055] The positioning coordinate optimization unit sets constraint values, counts the tilt angle of each recognition box, and determines that the box is a non-orthogonal text box if the tilt angle of the recognition box is greater than the constraint value, and otherwise determines that the box is an orthogonal text box. The optimized positioning coordinates of the two types of text boxes are determined respectively.

[0056] The content verification unit acquires the original image of the area to be verified, determines the text information, text spacing, and the depth and width of the stamped edge in the image, and then verifies the image. If the verification fails, it performs manual correction, generates a fitted stamped record, and then performs verification again.

[0057] The specific process of extracting attribute representations from sample and plain text image data in the stamp image generation unit is as follows:

[0058] S101. After randomly selecting a sample from the set, input it into the CNN model for analysis to obtain feature maps of different sizes. ,in The included channels are , Indicates the total number of feature maps. express The total number of channels contained within;

[0059] S102, Statistics middle Channels and The first in the passage Pixel value at position ,according to Calculate The channel is in Common intensity value at location ,in , Indicates the position number;

[0060] S103. After determining the intensity values ​​at different locations, sum them up to obtain... and The correlation coefficient between channels is calculated repeatedly until it is obtained. The correlation coefficient between each channel and other channels;

[0061] S104. Calculate the correlation coefficient between each channel in all feature maps, set the corresponding weight for each feature map, calculate the comprehensive correlation coefficient between each channel according to the weight and correlation coefficient of all feature maps, analyze it, and obtain the texture attribute representation of the current sample.

[0062] S105. After obtaining the texture attribute representation of all samples, extract the text attribute representation of the plain text image data.

[0063] The specific process of generating new steel stamp image data using attribute representation in the steel stamp image generation unit is as follows:

[0064] S201. Arbitrarily select a single sample and plain text image data, determine the statistical measures corresponding to the texture attribute representation of the sample and the text attribute representation of the plain text image data, align the two statistical measures, and use a statistical analysis algorithm to obtain the fused attribute representation. Generate the corresponding stamp image data based on the fused attribute representation. The statistical analysis algorithm is specifically as follows:

[0065] ;

[0066] in, This represents the attribute representation after fusion. The first and second statistics represent the textual attribute characteristics. The first and second statistics represent the texture attribute characteristics. Represents the attributes of text. Represents texture attributes;

[0067] S202. Repeat the operation until each plain text image data has a corresponding stamped image data generated for each sample. Perform image blurring, noise addition, and geometric transformation operations on the real stamped image data and the generated stamped image data respectively to obtain the processed image data. After adding the corresponding text labels to all image data, store them in the first dataset.

[0068] The specific process of obtaining the corresponding binary map in the binary map construction unit is as follows:

[0069] S301. Obtain feature maps of different sizes corresponding to the image data, fuse all feature maps, and then use the convolutional layers and activation functions in the CNN model to generate the corresponding probability map and threshold map.

[0070] S302, Reading probability map, threshold map, first Pixel value of each pixel and ,use and Calculate the corresponding difference ,in , Indicates the pixel index;

[0071] S303, if Then the first The pixel value of each pixel is set to 1, and the pixel value of the other pixel is set to 0. This operation is repeated until the pixel value of all pixels is determined. Then, a binary image is constructed based on the current pixel value.

[0072] The specific process by which the positioning coordinate determination unit calculates the tilt angle and actual positioning coordinates of each recognition frame based on the four corner coordinates is as follows:

[0073] S401, Based on the coordinates of the four corners of the first recognition box , , and The tilt angle of the recognition box was analyzed. ,in The coordinates of the four corners are, in order, the coordinates of the top left corner, the top right corner, the bottom left corner, and the bottom right corner;

[0074] S402, according to , , and Calculate the initial first positioning coordinates of the current recognition box. and initial second positioning coordinates ,in , , ,

[0075] ;

[0076] S403, according to the initial first positioning coordinates Initial second positioning coordinates Within the binary image, perform downward and upward searches respectively to determine the pixel that is closest to the initial first positioning coordinates and has a pixel value of 1. Use its location as the actual first positioning coordinates. Also, analyze the pixel that is closest to the initial second positioning coordinates and has a pixel value of 1, and use its location as the actual second positioning coordinates.

[0077] S404. Repeat the operation until the tilt angle, actual first positioning coordinates and actual second positioning coordinates of all recognition boxes are calculated.

[0078] The specific process for determining the optimized positioning coordinates of the two types of text boxes in the positioning coordinate optimization unit is as follows:

[0079] S501. After determining the actual first positioning coordinates and actual second positioning coordinates of all non-positive text boxes, calculate the difference between the actual first positioning coordinates and the actual second positioning coordinates in each text box, take the maximum difference as the spacing value, use the DBSCAN clustering algorithm to group the ordinates in all first positioning coordinates, and calculate the average number of ordinates in each group. If the difference between the average numbers of the groups is less than the spacing value, then the current two groups are merged into one group; otherwise, no operation is performed.

[0080] S502. After calculating the average of the ordinates for all groups, set the ordinate value for each group according to the average, and then... The ordinate values ​​of each group The actual first positioning coordinates within this group were analyzed. The corresponding optimized first positioning coordinates ,in , ,according to Determine the corresponding actual second positioning coordinates And calculate the optimized second positioning coordinates corresponding to the actual second positioning coordinates. ,in , , Indicates the group number;

[0081] S503. After determining the actual first positioning coordinates and actual second positioning coordinates of all positive text boxes, use the actual positioning coordinates as the optimized positioning coordinates.

[0082] S504. Construct a position transformation function based on the actual and optimized positioning coordinates of all text boxes, calculate the actual value of each parameter in the position transformation function, and determine the adjusted position coordinates of each pixel in the image data based on the position transformation function.

[0083] S505. Store all adjusted image data and corresponding text labels as samples in the second dataset, and transmit the samples to the CRNN model for analysis to determine the model parameters.

[0084] The specific process by which the content verification unit determines the text information, text spacing, and the depth and width of the stamped edge in the image is as follows:

[0085] S601. Obtain the original image of the area to be reviewed, preprocess it, determine the actual and optimized positioning coordinates of all text recognition boxes in the original image, and construct a position transformation function based on the actual and optimized positioning coordinates. The position transformation function is specifically as follows:

[0086] ;

[0087] in, This represents the position transformation function. Represents a scalar. This indicates the optimized positioning coordinates. Indicates the actual positioning coordinates. Represents the weight vector. Indicates the value of the control function. Indicates the number of vectors. Indicates a parameter.

[0088] S602. After adjusting the image data through the position transformation function to obtain new image data, the new image data is transmitted to the CRNN model for analysis, and the text information of all recognition boxes is output.

[0089] S603. After counting all the text information in the area to be reviewed, determine the spacing between each text according to the actual positioning coordinates of the recognition box, and use a laser profilometer to scan the area to be reviewed, and output the depth and width of the steel stamp edge in the area.

[0090] The specific process for verifying images in the content verification unit is as follows:

[0091] S701. Set a threshold and compare the text information, text spacing, and edge depth and width of the area to be reviewed with all steel stamp records in the template library one by one. If the similarity is higher than the threshold, the verification is deemed to pass and the text information of the steel stamp record is output. Otherwise, a text abnormality is indicated and the text information is manually corrected.

[0092] S702. Determine the text attribute representation in the current area to be reviewed based on the corrected text information, analyze the original image of the area to be reviewed using a CNN model to obtain the texture attribute representation in the current area to be reviewed, generate the corresponding stamp image data using the text attribute representation and texture attribute representation, and analyze the fitted stamp record through the stamp image data.

[0093] S703. Calculate the similarity between the area to be reviewed and the fitted steel stamp record. If the similarity is lower than the threshold, the verification is deemed to have failed and an error message is displayed. Otherwise, the verification is deemed to have passed and the steel stamp record is stored in the template library.

[0094] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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 process, method, article, or apparatus.

[0095] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art 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 appended claims and their equivalents.

Claims

1. An automated data verification system based on visual scanning, comprising a stamp image generation unit, a binary map construction unit, a positioning coordinate determination unit, a positioning coordinate optimization unit, and a content verification unit, characterized in that: The steel stamp image generation unit collects multiple real steel stamp image data, stores the real steel stamp image data as samples in a set, and generates multiple plain text image data using the SDXL model. Then, it extracts attribute representations from the samples and plain text image data respectively, and uses the attribute representations to generate new steel stamp image data. Binary image construction unit: After extracting the stamp image data, the binary image construction unit transmits it to the CNN model for analysis to obtain the corresponding binary image; The positioning coordinate determination unit, after acquiring the binary image, inputs the image data into the CRAFT model for analysis, obtains the four corner coordinates of the recognition boxes corresponding to all text in the image, and calculates the tilt angle and actual positioning coordinates of each recognition box based on the four corner coordinates; The positioning coordinate optimization unit sets constraint values, counts the tilt angle of each recognition box, and determines that the box is a non-orthogonal text box if the tilt angle of the recognition box is greater than the constraint value, and otherwise determines that the box is an orthogonal text box. The optimized positioning coordinates of the two types of text boxes are determined respectively. The content verification unit acquires the original image of the area to be verified, determines the text information, text spacing, and the depth and width of the stamped edge in the image, and then verifies the image. If the verification fails, it performs manual correction, generates a fitted stamped record, and then performs verification again.

2. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process of extracting attribute representations from sample and plain text image data in the stamp image generation unit is as follows: S101. After randomly selecting a sample from the set, input it into the CNN model for analysis to obtain feature maps of different sizes. ,in The included channels are , Indicates the total number of feature maps. express The total number of channels contained within; S102, Statistics middle Channels and The first in the passage Pixel value at position ,according to Calculate The channel is in Common intensity value at location ,in , Indicates the position number; S103. After determining the intensity values ​​at different locations, sum them up to obtain... and The correlation coefficient between channels is calculated repeatedly until it is obtained. The correlation coefficient between each channel and other channels; S104. Calculate the correlation coefficient between each channel in all feature maps, set the corresponding weight for each feature map, calculate the comprehensive correlation coefficient between each channel according to the weight and correlation coefficient of all feature maps, analyze it, and obtain the texture attribute representation of the current sample. S105. After obtaining the texture attribute representation of all samples, extract the text attribute representation of the plain text image data.

3. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process of generating new steel stamp image data using attribute representation in the steel stamp image generation unit is as follows: S201. Arbitrarily select a single sample and plain text image data, determine the statistical quantities corresponding to the texture attribute representation of the sample and the text attribute representation of the plain text image data, align the statistical quantities of the two, use the statistical analysis algorithm to obtain the fused attribute representation, and generate the corresponding stamp image data based on the fused attribute representation. S202. Repeat the operation until each plain text image data has a corresponding stamped image data for each sample. Perform image blurring, noise addition, and geometric transformation operations on the real stamped image data and the generated stamped image data respectively to obtain the processed image data. After adding the corresponding text labels to all image data, store them in the first dataset.

4. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process of obtaining the corresponding binary map in the binary map construction unit is as follows: S301. Obtain feature maps of different sizes corresponding to the image data, fuse all feature maps, and then use the convolutional layers and activation functions in the CNN model to generate the corresponding probability map and threshold map. S302, Reading probability map, threshold map, first Pixel value of each pixel and ,use and Calculate the corresponding difference ,in , Indicates the pixel index; S303, if Then the first The pixel value of each pixel is set to 1, and the pixel value of the other pixel is set to 0. This process is repeated until the pixel values ​​of all pixels are determined. Then, a binary image is constructed based on the current pixel values.

5. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process by which the positioning coordinate determination unit calculates the tilt angle and actual positioning coordinates of each recognition frame based on the four corner coordinates is as follows: S401, Based on the coordinates of the four corners of the first recognition box , , and The tilt angle of the recognition box was analyzed. ,in ; S402, according to , , and Calculate the initial first positioning coordinates of the current recognition box. and initial second positioning coordinates ,in , , , ; S403, according to the initial first positioning coordinates Initial second positioning coordinates Within the binary image, perform downward and upward searches respectively to determine the pixel that is closest to the initial first positioning coordinates and has a pixel value of 1. Use its location as the actual first positioning coordinates. Also, analyze the pixel that is closest to the initial second positioning coordinates and has a pixel value of 1, and use its location as the actual second positioning coordinates. S404. Repeat the operation until the tilt angle, actual first positioning coordinates and actual second positioning coordinates of all recognition boxes are calculated.

6. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process for determining the optimized positioning coordinates of the two types of text boxes in the positioning coordinate optimization unit is as follows: S501. After determining the actual first positioning coordinates and actual second positioning coordinates of all non-positive text boxes, calculate the difference between the actual first positioning coordinates and the actual second positioning coordinates in each text box, take the maximum difference as the spacing value, use the DBSCAN clustering algorithm to group the ordinates in all first positioning coordinates, and calculate the average number of ordinates in each group. If the difference between the average numbers of the groups is less than the spacing value, then the current two groups are merged into one group; otherwise, no operation is performed. S502. After calculating the average of the ordinates for all groups, set the ordinate value for each group according to the average, and then... The ordinate values ​​of each group The actual first positioning coordinates within this group were analyzed. The corresponding optimized first positioning coordinates ,in , ,according to Determine the corresponding actual second positioning coordinates And calculate the optimized second positioning coordinates corresponding to the actual second positioning coordinates. ,in , , Indicates the group number; S503. After determining the actual first positioning coordinates and actual second positioning coordinates of all positive text boxes, use the actual positioning coordinates as the optimized positioning coordinates. S504. Construct a position transformation function based on the actual and optimized positioning coordinates of all text boxes, calculate the actual value of each parameter in the position transformation function, and determine the adjusted position coordinates of each pixel in the image data based on the position transformation function. S505. Store all adjusted image data and corresponding text labels as samples in the second dataset, and transmit the samples to the CRNN model for analysis to determine the model parameters.

7. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process by which the content verification unit determines the text information, text spacing, and the depth and width of the stamped edge in the image is as follows: S601. Obtain the original image of the area to be reviewed, preprocess it, determine the actual and optimized positioning coordinates of all text recognition boxes in the original image, and construct a position transformation function based on the actual and optimized positioning coordinates. S602. After adjusting the image data through the position transformation function to obtain new image data, the new image data is transmitted to the CRNN model for analysis and the text information of all recognition boxes is output. S603. After counting all the text information in the area to be reviewed, determine the spacing between each text according to the actual positioning coordinates of the recognition box, and use a laser profilometer to scan the area to be reviewed, and output the depth and width of the steel stamp edge in the area.

8. The automated data verification system based on visual scanning according to claim 1, characterized in that: The specific process for verifying images in the content verification unit is as follows: S701. Set a threshold and compare the text information, text spacing, and edge depth and width of the area to be reviewed with all steel stamp records in the template library one by one. If the similarity is higher than the threshold, the verification is deemed to pass and the text information of the steel stamp record is output. Otherwise, a text abnormality is indicated and the text information is manually corrected. S702. Determine the text attribute representation in the current area to be reviewed based on the corrected text information, analyze the original image of the area to be reviewed using a CNN model to obtain the texture attribute representation in the current area to be reviewed, generate the corresponding stamp image data using the text attribute representation and texture attribute representation, and analyze the fitted stamp record through the stamp image data. S703. Calculate the similarity between the area to be reviewed and the fitted steel stamp record. If the similarity is lower than the threshold, the verification is deemed to have failed and an error message is displayed. Otherwise, the verification is deemed to have passed and the steel stamp record is stored in the template library.

Images