Audit data security management system and method based on large model

Abstract

The application discloses an audit data security management system and method based on a large model, relates to the technical field of audit data security management, and specifically comprises the following steps: obtaining a to-be-detected transaction voucher image, and simultaneously collecting standard and abnormal transaction voucher images; performing regional feature recognition on the standard and abnormal images to generate a segmentation frame of a text region; obtaining a first and a second feature image set through partitioning and intercepting by the segmentation frame; obtaining a first pixel feature and a second pixel feature through pixel analysis, and comparing the first pixel feature with the second pixel feature to obtain an abnormal pixel feature; extracting a target pixel feature of the to-be-detected image, comparing the target pixel feature with the abnormal pixel feature, and judging whether the to-be-detected image is an abnormal image; and performing corresponding processing according to the judgment result. The application can effectively identify transaction vouchers that are tampered with by image editing software through pixel-level feature comparison, improve the security of audit data, and prevent fake vouchers from flowing into the audit process.

Description

Technical Field

[0001] This invention relates to the field of audit data security management technology, specifically to an audit data security management system and method based on a large model. Background Technology

[0002] With the widespread adoption of electronic payments, transaction detail screenshots generated by various mobile payment platforms, such as WeChat and Alipay, are gradually becoming important payment vouchers in auditing work. In the process of auditing data security management, existing technologies typically use manual review or optical character recognition to extract key information such as amount, parties involved in the transaction, and transaction time from transaction detail images. However, as static image files, these screenshots are easily tampered with using image editing software or photo retouching tools. The aforementioned methods of manual review or optical character recognition lack the ability to identify tampered information, thus leading to the falsification of audit data. Summary of the Invention

[0003] The purpose of this invention is to provide an audit data security management system and method based on a large model to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for audit data security management based on a large model, the method comprising: The system acquires images of transaction vouchers to be detected, and simultaneously collects several standard transaction voucher images and several abnormal transaction voucher images. Standard transaction voucher images refer to original voucher images generated by a legitimate payment platform, without any image editing, and whose text information matches the actual transaction. Abnormal transaction voucher images refer to transaction voucher images where key information such as amount, object, time, and account number has been altered, replaced, or forged using image editing tools. Regional feature recognition is performed on the standard and abnormal transaction voucher images respectively, generating segmentation boxes corresponding to the text regions in each type of image. The standard transaction certificate image and the abnormal transaction certificate image are partitioned and cropped using the segmentation box to obtain a first feature image set and a second feature image set. The pixels of several feature images in the first feature image set and the second feature image set are traversed respectively; the first pixel features corresponding to the first feature image set and the second pixel features corresponding to the second feature image set are analyzed and obtained; the first pixel features and the second pixel features are compared and analyzed to obtain abnormal pixel features; Extract the target pixel features of the transaction voucher image to be detected, compare the target pixel features with the abnormal pixel features, and determine whether the transaction voucher image to be detected is an abnormal image; The transaction image to be detected is processed accordingly based on the judgment result.

[0005] Furthermore, region feature recognition is performed on the standard transaction voucher image and the abnormal transaction voucher image respectively to generate segmentation boxes corresponding to the text regions in various images, including: The standard transaction certificate image and the abnormal transaction certificate image are converted into binary images respectively; in the binary image, pixels with gray values ​​higher than the overall average gray value of the binary image are labeled with a first label, wherein the first label represents non-text pixels, and pixels with gray values ​​lower than the average gray value are labeled with a second label, wherein the second label represents text pixels; The text pixels in the binary image are traversed row by row and column by column. Adjacent pixels that are both labeled with the second tag are connected to form several independent feature blocks, where each feature block represents an independent character. The leftmost, rightmost, topmost, and bottommost pixels on the boundary of each feature block are extracted. Vertical lines are constructed based on the leftmost and rightmost pixels as the left and right boundaries, respectively. Horizontal lines are constructed based on the topmost and bottommost pixels as the top and bottom boundaries, respectively. A rectangle completely enclosing the feature block is formed by the left, right, top, and bottom boundaries. Connect the upper and lower boundaries of the rectangles of two adjacent feature blocks in the same row to form a segmentation frame; the area enclosed by the segmentation frame is a feature region, where one feature region corresponds to one item of financial data to be extracted.

[0006] Furthermore, by segmenting the standard transaction certificate image and the abnormal transaction certificate image using the segmentation box, a first feature image set and a second feature image set are obtained, including: According to the feature region boundary enclosed by the segmentation box, the feature region corresponding to the same transaction item in the standard transaction feature image is uniformly cropped to obtain several feature region images. The feature region images corresponding to the same transaction item in all standard transaction voucher images are grouped into a feature subset. In the above steps, the feature regions corresponding to the same transaction item are uniformly extracted and grouped into a feature subset. For example, in the transaction amount item in the WeChat transaction screenshot, the text size of the item is the same. Grouping them into a subset allows for centralized analysis of the uniform features corresponding to this item. The first feature image set is formed by combining several feature subsets corresponding to different transaction entries in the standard transaction certificate image; According to the feature region boundary enclosed by the segmentation box, the feature region corresponding to the same transaction item in the abnormal transaction feature image is uniformly cropped to obtain several feature region images. The feature region images corresponding to the same transaction item in all abnormal transaction voucher images are grouped into a feature subset. The second feature image set is formed by combining several feature subsets corresponding to different transaction entries in the abnormal transaction certificate image.

[0007] Furthermore, the analysis and acquisition of the first pixel features corresponding to the first feature image set and the second pixel features corresponding to the second feature image set includes: For each feature region image in the same feature subset of the first feature image set and the second feature image set, perform a row-by-row and column-by-column pixel traversal to identify the top boundary, bottom boundary, left boundary, and right boundary of all feature blocks in each feature region image; The number of vertical pixel rows between the top and bottom boundaries of each feature block in each feature region image is counted. The average height of the feature block corresponding to the feature region image is obtained by summing the number of vertical pixel rows of all feature blocks in the same feature region image and taking the average value. The number of horizontal pixel columns between the left and right boundaries of each feature block in each feature region image is counted. The average width of the feature block corresponding to the feature region image is obtained by summing the number of horizontal pixel columns of all feature blocks in the same feature region image and taking the average value. The average height of feature blocks in all feature region images within the same feature subset is summarized to obtain the height set corresponding to the feature subset. The values ​​in the height set are arranged in ascending order, and the height value in the middle position of the arrangement is extracted as the height feature of the feature subset. The average width of feature blocks in all feature region images within the same feature subset is summarized to obtain the width set corresponding to the feature subset. The values ​​in the width set are arranged in ascending order, and the width value in the middle position of the arrangement is extracted as the width feature of the feature subset. The height and width features corresponding to each feature subset in the first feature image set are combined to form feature pairs corresponding to each feature subset; the feature pairs corresponding to each feature subset are combined to obtain the first pixel feature corresponding to the first feature image set. The height and width features corresponding to each feature subset in the second feature image set are combined to form feature pairs corresponding to each feature subset. The feature pairs corresponding to each feature subset are then combined to obtain the second pixel feature corresponding to the second feature image set.

[0008] Furthermore, by comparing and analyzing the first pixel feature with the second pixel feature, the abnormal pixel features are obtained as follows: Establish a feature coordinate system with the height feature of the feature subset as the vertical axis and the width feature as the horizontal axis; The feature pairs corresponding to each feature subset in the first pixel feature are marked in the feature coordinate system to form a first feature point set. The feature points in the first feature point set are connected in sequence to enclose the first feature enclosed region. The feature pairs corresponding to each feature subset in the second pixel feature are marked in the feature coordinate system to form a second feature point set. The feature points in the second feature point set are connected in sequence to form a second feature enclosed region. The overlapping areas between the second feature-enclosed region and the first feature-enclosed region are removed, and the non-overlapping areas between the second feature-enclosed region and the first feature-enclosed region are retained. The non-overlapping areas are the abnormal pixel features.

[0009] Furthermore, determining whether the image of the transaction certificate to be detected is an abnormal image includes: Feature pairs corresponding to each feature subset in the target pixel features of the image to be detected are extracted and labeled in the feature coordinate system to form a target point set. The target points in the target point set are connected in sequence to form a target enclosing region. The target enclosing region is the pixel feature to be detected in the transaction certificate image to be detected. Determine whether there is an overlapping area between the pixel feature to be detected and the abnormal pixel feature. If there is an overlapping area between the pixel feature to be detected and the abnormal pixel feature, the transaction voucher image to be detected is an abnormal image; if there is no overlapping area between the pixel feature to be detected and the abnormal pixel feature, the transaction voucher image to be detected is normal.

[0010] Furthermore, based on the judgment result, the corresponding processing of the transaction image to be detected includes: If the transaction voucher image to be detected is an abnormal image, then the feature entries corresponding to the feature pairs included in the region of the transaction voucher image to be detected that overlaps with the abnormal pixel features are extracted, and the information corresponding to the feature entries is output as warning information. If the transaction voucher image to be detected is a normal image, then the text in each target feature image of the transaction voucher image to be detected is extracted to obtain a set of text content corresponding to each target feature image; the content of each set of text content is extracted sequentially according to the parsing target to obtain the parsing content of the parsing target; wherein, the parsing target includes transaction object, transaction address, transaction category, transaction date and transaction amount; the parsing target and the corresponding parsing content are output in pairs sequentially to obtain the analysis result of the transaction voucher image to be detected.

[0011] Furthermore, to better implement the above methods, an audit data security management system based on a large model is also provided. This management system includes: an image acquisition module, a region feature recognition module, a partitioning and cropping module, a pixel feature analysis module, a comparison analysis module, an anomaly judgment module, and a processing module. The image acquisition module is used to acquire images of transaction vouchers to be detected, and simultaneously collects several standard transaction voucher images and several abnormal transaction voucher images. The region feature recognition module is used to perform region feature recognition on the standard transaction voucher image and the abnormal transaction voucher image respectively, and generate segmentation boxes corresponding to the text regions in various images. The partitioning and cropping module is used to partition and crop the standard transaction certificate image and the abnormal transaction certificate image using the segmentation box to obtain a first feature image set and a second feature image set. The pixel feature analysis module is used to traverse the pixels of several feature images in the first feature image set and the second feature image set respectively, and analyze and obtain the first pixel feature corresponding to the first feature image set and the second pixel feature corresponding to the second feature image set. The comparison analysis module is used to compare and analyze the first pixel feature with the second pixel feature to obtain abnormal pixel features; The anomaly detection module is used to extract the target pixel features of the transaction voucher image to be detected, compare the target pixel features with the abnormal pixel features, and determine whether the transaction voucher image to be detected is an abnormal image. The processing module is used to perform corresponding processing on the transaction image to be detected based on the judgment result.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention performs regional feature recognition and pixel traversal on standard and abnormal transaction voucher images, extracts feature pairs composed of height and width features, and establishes a feature coordinate system for comparison to obtain abnormal pixel features. This enables accurate identification of transaction voucher images that have been tampered with by image editing software, solving the problem of existing technologies lacking the ability to identify tampered information.

[0013] 2. This invention determines the overlapping area between the target pixel features and abnormal pixel features of the transaction voucher image to be detected. If there is overlap, it is determined to be an abnormal image and an early warning message is output. If it is normal, the text content is automatically extracted and the analysis results are output according to the parsing target such as transaction object, address, category, date, amount, etc., thereby realizing the automation and intelligence of audit data security management.

[0014] 3. This invention uses a segmentation box to uniformly extract and classify the feature regions corresponding to the same transaction item into feature subsets, and then constructs pixel features based on feature subsets. This can eliminate interference caused by layout differences between different images, and improve the robustness of feature comparison and the accuracy of anomaly detection. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method of the audit data security management system and method based on a large model according to the present invention. Figure 2 This is a schematic diagram illustrating an application scenario of the audit data security management system and method based on a large model according to the present invention. Detailed Implementation

[0016] 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.

[0017] Example 1: As Figure 1 As shown, the present invention provides a technical solution, a method for audit data security management based on a large model, the management method comprising: Acquire the image of the transaction voucher to be detected, and simultaneously acquire several standard transaction voucher images and several abnormal transaction voucher images; perform region feature recognition on the standard transaction voucher images and abnormal transaction voucher images respectively, and generate segmentation boxes corresponding to the text regions in various images; The standard transaction certificate image and the abnormal transaction certificate image are partitioned and cropped using the segmentation box to obtain a first feature image set and a second feature image set. The pixels of several feature images in the first feature image set and the second feature image set are traversed respectively; the first pixel features corresponding to the first feature image set and the second pixel features corresponding to the second feature image set are analyzed and obtained; the first pixel features and the second pixel features are compared and analyzed to obtain abnormal pixel features; Extract the target pixel features of the transaction voucher image to be detected, compare the target pixel features with the abnormal pixel features, and determine whether the transaction voucher image to be detected is an abnormal image; The transaction image to be detected is processed accordingly based on the judgment result; Specifically, the standard transaction voucher image and the abnormal transaction voucher image are respectively subjected to region feature recognition to generate segmentation boxes corresponding to text regions in various images, including: The standard transaction certificate image and the abnormal transaction certificate image are converted into binary images respectively; in the binary image, pixels with gray values ​​higher than the overall average gray value of the binary image are labeled with a first label, and pixels with gray values ​​lower than the average gray value are labeled with a second label. The text pixels in the binary image are traversed row by row and column by column. Pixels that are adjacent in position and both have the second label are connected to form several independent feature blocks. The leftmost, rightmost, topmost, and bottommost pixels on the boundary of each feature block are extracted. A vertical line is constructed based on the leftmost and rightmost pixels as the left and right boundaries, respectively. A horizontal line is constructed based on the topmost and bottommost pixels as the top and bottom boundaries, respectively. A rectangle completely enclosing the feature block is formed by the left, right, top, and bottom boundaries. Connect the upper and lower boundaries of the rectangles of two adjacent feature blocks in the same row to form a segmentation frame; the area enclosed by the segmentation frame is the feature region. In an embodiment of the present invention, a WeChat payment transaction screenshot is used as the transaction voucher image to be detected. Simultaneously, 100 unmodified original transaction screenshots are collected as standard transaction screenshots, and 100 abnormal transaction screenshots with modified amounts or transaction objects are collected using image editing software. The standard and abnormal transaction screenshots are converted into binary images. Pixels with grayscale values ​​lower than the average grayscale value are marked as text pixels, and pixels with grayscale values ​​higher than the average grayscale value are marked as non-text pixels. The process is repeated row by row and column by column, connecting adjacent text pixels into independent character blocks. The leftmost, rightmost, topmost, and bottommost pixels of each character block are extracted to construct a rectangle. The upper and lower boundaries of adjacent rectangles in the same row are connected to form a segmentation box covering the entire row of text. This segmentation box corresponds to a transaction item; for example, a rectangle corresponds to a transaction item with a transaction amount of 100.00 yuan. Specifically, the standard transaction certificate image and the abnormal transaction certificate image are partitioned and cropped using the segmentation box to obtain a first feature image set and a second feature image set, including: According to the feature region boundary enclosed by the segmentation box, the feature region corresponding to the same transaction item in the standard transaction feature image is uniformly cropped to obtain several feature region images. The feature region images corresponding to the same transaction item in all standard transaction voucher images are grouped into a feature subset. The first feature image set is formed by combining several feature subsets corresponding to different transaction entries in the standard transaction certificate image; According to the feature region boundary enclosed by the segmentation box, the feature region corresponding to the same transaction item in the abnormal transaction feature image is uniformly cropped to obtain several feature region images. The feature region images corresponding to the same transaction item in all abnormal transaction voucher images are grouped into a feature subset. The second feature image set is formed by combining several feature subsets corresponding to different transaction entries in the abnormal transaction voucher image; In an embodiment of the present invention, according to the segmentation box boundary, the region image corresponding to each transaction item is extracted from all standard transaction screenshots, and the region object corresponding to each type of transaction item is grouped into a feature subset; wherein, the transaction item includes but is not limited to the transaction time, transaction counterparty, transaction amount, etc., and the various types of transaction items are uniformly combined to obtain the first feature image set; similarly, the abnormal transaction screenshots are partitioned and extracted to obtain the second feature image set; The analysis and acquisition of the first pixel features corresponding to the first feature image set and the second pixel features corresponding to the second feature image set includes: For each feature region image in the same feature subset of the first feature image set and the second feature image set, perform a row-by-row and column-by-column pixel traversal to identify the top boundary, bottom boundary, left boundary, and right boundary of all feature blocks in each feature region image; The number of vertical pixel rows between the top and bottom boundaries of each feature block in each feature region image is counted. The average height of the feature block corresponding to the feature region image is obtained by summing the number of vertical pixel rows of all feature blocks in the same feature region image and taking the average value. The number of horizontal pixel columns between the left and right boundaries of each feature block in each feature region image is counted. The average width of the feature block corresponding to the feature region image is obtained by summing the number of horizontal pixel columns of all feature blocks in the same feature region image and taking the average value. The average height of feature blocks in all feature region images within the same feature subset is summarized to obtain the height set corresponding to the feature subset. The values ​​in the height set are arranged in ascending order, and the height value in the middle position of the arrangement is extracted as the height feature of the feature subset. The average width of feature blocks in all feature region images within the same feature subset is summarized to obtain the width set corresponding to the feature subset. The values ​​in the width set are arranged in ascending order, and the width value in the middle position of the arrangement is extracted as the width feature of the feature subset. The height and width features corresponding to each feature subset in the first feature image set are combined to form feature pairs corresponding to each feature subset; the feature pairs corresponding to each feature subset are combined to obtain the first pixel feature corresponding to the first feature image set. The height and width features corresponding to each feature subset in the second feature image set are combined to form feature pairs corresponding to each feature subset. The feature pairs corresponding to each feature subset are then combined to obtain the second pixel feature corresponding to the second feature image set. In an embodiment of the present invention, for example, for each image within the transaction amount subset of the first feature image set, the boundary of each character block is traversed pixel by pixel, and the height and width of each character block are calculated, i.e., the number of vertical pixel rows and the number of horizontal pixel columns. The average height and average width of all character blocks in the image are then taken. The average height of all images in the subset are sorted, and the median is taken as the height feature. Similarly, the width feature is obtained. The height feature and the width feature are combined into a feature pair. The feature pairs of each item are summarized to form the first pixel feature. Similarly, the second feature image set is analyzed to obtain the second pixel feature. Specifically, by comparing and analyzing the first pixel feature with the second pixel feature, the abnormal pixel features obtained include: Establish a feature coordinate system with the height feature of the feature subset as the vertical axis and the width feature as the horizontal axis; The feature pairs corresponding to each feature subset in the first pixel feature are marked in the feature coordinate system to form a first feature point set. The feature points in the first feature point set are connected in sequence to enclose the first feature enclosed region. The feature pairs corresponding to each feature subset in the second pixel feature are marked in the feature coordinate system to form a second feature point set. The feature points in the second feature point set are connected in sequence to enclose the second feature bounding region. The overlapping areas between the second feature-enclosed region and the first feature-enclosed region are removed, and the non-overlapping areas between the second feature-enclosed region and the first feature-enclosed region are retained. The non-overlapping areas are the abnormal pixel features. In an embodiment of the present invention, for example, if the width or height of the transaction amount character in the abnormal image deviates significantly from the normal range, then the area enclosed by each feature point in the abnormal image must have a non-overlapping area with the first enclosing area. This part of the area represents the abnormal pixel feature corresponding to the transaction amount in the abnormal image. The determination of whether the transaction certificate image to be detected is an abnormal image includes: Feature pairs corresponding to each feature subset in the target pixel features of the image to be detected are extracted and labeled in the feature coordinate system to form a target point set. The target points in the target point set are connected in sequence to form a target enclosing region. The target enclosing region is the pixel feature to be detected in the transaction certificate image to be detected. Determine whether there is an overlapping area between the pixel feature to be detected and the abnormal pixel feature. If there is an overlapping area between the pixel feature to be detected and the abnormal pixel feature, then the transaction voucher image to be detected is an abnormal image; if there is no overlapping area between the pixel feature to be detected and the abnormal pixel feature, then the transaction voucher image to be detected is normal. In an embodiment of the present invention, for example, if the size of the transaction amount characters in a WeChat payment transaction screenshot falls within an abnormal area, then the WeChat payment transaction screenshot is determined to be abnormal. The corresponding processing of the transaction image to be detected based on the judgment result includes: If the transaction voucher image to be detected is an abnormal image, then the feature entries corresponding to the feature pairs included in the region of the transaction voucher image to be detected that overlaps with the abnormal pixel features are extracted, and the information corresponding to the feature entries is output as warning information. If the transaction voucher image to be detected is a normal image, then the text in each target feature image of the transaction voucher image to be detected is extracted to obtain a set of text content corresponding to each target feature image; the content of each set of text content is extracted sequentially according to the parsing target to obtain the parsing content of the parsing target; wherein, the parsing target includes transaction object, transaction address, transaction category, transaction date and transaction amount; the parsing target and the corresponding parsing content are output in pairs sequentially to obtain the analysis result of the transaction voucher image to be detected; In an embodiment of the present invention, for example, if the abnormal transaction item is the transaction amount, then the warning message is output as "the transaction amount field is suspected to be abnormal"; if the WeChat payment transaction screenshot is a normal image, then the text in the image is extracted, for example, transaction object: Zhang San; transaction amount: 200.00 yuan; Example 2: Figure 2 As shown, in order to better implement the above method, an audit data security management system based on a large model is also provided. The management system includes: an image acquisition module, a region feature recognition module, a partitioning and cropping module, a pixel feature analysis module, a comparison analysis module, an anomaly judgment module, and a processing module. The image acquisition module is used to acquire images of transaction vouchers to be detected, and simultaneously collects several standard transaction voucher images and several abnormal transaction voucher images. The region feature recognition module is used to perform region feature recognition on the standard transaction voucher image and the abnormal transaction voucher image respectively, and generate segmentation boxes corresponding to the text regions in various images. The partitioning and cropping module is used to partition and crop the standard transaction certificate image and the abnormal transaction certificate image using the segmentation box to obtain a first feature image set and a second feature image set. The pixel feature analysis module is used to traverse the pixels of several feature images in the first feature image set and the second feature image set respectively, and analyze and obtain the first pixel feature corresponding to the first feature image set and the second pixel feature corresponding to the second feature image set. The comparison analysis module is used to compare and analyze the first pixel feature with the second pixel feature to obtain abnormal pixel features; The anomaly detection module is used to extract the target pixel features of the transaction voucher image to be detected, compare the target pixel features with the abnormal pixel features, and determine whether the transaction voucher image to be detected is an abnormal image. The processing module is used to perform corresponding processing on the transaction image to be detected based on the judgment result.

[0018] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for audit data security management based on a large model, characterized by: The management method includes: Acquire the image of the transaction voucher to be detected, and simultaneously acquire several standard transaction voucher images and several abnormal transaction voucher images; perform region feature recognition on the standard transaction voucher images and abnormal transaction voucher images respectively, and generate segmentation boxes corresponding to the text regions in various images; The standard transaction certificate image and the abnormal transaction certificate image are partitioned and cropped using the segmentation box to obtain a first feature image set and a second feature image set. The pixels of several feature images in the first feature image set and the second feature image set are traversed respectively; the first pixel features corresponding to the first feature image set and the second pixel features corresponding to the second feature image set are analyzed and obtained; the first pixel features and the second pixel features are compared and analyzed to obtain abnormal pixel features; Extract the target pixel features of the transaction voucher image to be detected, compare the target pixel features with the abnormal pixel features, and determine whether the transaction voucher image to be detected is an abnormal image; The transaction image to be detected is processed accordingly based on the judgment result.

2. The audit data security management method based on a large model according to claim 1, characterized in that: The standard transaction voucher image and the abnormal transaction voucher image are respectively subjected to region feature recognition to generate segmentation boxes corresponding to text regions in various images, including: The standard transaction certificate image and the abnormal transaction certificate image are converted into binary images respectively; in the binary image, pixels with gray values ​​higher than the overall average gray value of the binary image are labeled with a first label, and pixels with gray values ​​lower than the average gray value are labeled with a second label. The text pixels in the binary image are traversed row by row and column by column. Pixels that are adjacent in position and both have the second label are connected to form several independent feature blocks. The leftmost, rightmost, topmost, and bottommost pixels on the boundary of each feature block are extracted. A vertical line is constructed based on the leftmost and rightmost pixels as the left and right boundaries, respectively. A horizontal line is constructed based on the topmost and bottommost pixels as the top and bottom boundaries, respectively. A rectangle completely enclosing the feature block is formed by the left, right, top, and bottom boundaries. Connect the upper and lower boundaries of the rectangles of two adjacent feature blocks in the same row to form a segmentation frame; the area enclosed by the segmentation frame is the feature region.

3. The audit data security management method based on a large model according to claim 1, characterized in that: The standard transaction certificate image and the abnormal transaction certificate image are partitioned and cropped using the segmentation box to obtain a first feature image set and a second feature image set, including: According to the feature region boundary enclosed by the segmentation box, the feature region corresponding to the same transaction item in the standard transaction feature image is uniformly cropped to obtain several feature region images. The feature region images corresponding to the same transaction item in all standard transaction voucher images are grouped into a feature subset. The first feature image set is formed by combining several feature subsets corresponding to different transaction entries in the standard transaction certificate image; According to the feature region boundary enclosed by the segmentation box, the feature region corresponding to the same transaction item in the abnormal transaction feature image is uniformly cropped to obtain several feature region images. The feature region images corresponding to the same transaction item in all abnormal transaction voucher images are grouped into a feature subset. The second feature image set is formed by combining several feature subsets corresponding to different transaction entries in the abnormal transaction certificate image.

4. The audit data security management method based on a large model according to claim 1, characterized in that: The analysis and acquisition of the first pixel features corresponding to the first feature image set and the second pixel features corresponding to the second feature image set include: For each feature region image in the same feature subset of the first feature image set and the second feature image set, perform a row-by-row and column-by-column pixel traversal to identify the top boundary, bottom boundary, left boundary, and right boundary of all feature blocks in each feature region image; The number of vertical pixel rows between the top and bottom boundaries of each feature block in each feature region image is counted. The average height of the feature block corresponding to the feature region image is obtained by summing the number of vertical pixel rows of all feature blocks in the same feature region image and taking the average value. The number of horizontal pixel columns between the left and right boundaries of each feature block in each feature region image is counted. The average width of the feature block corresponding to the feature region image is obtained by summing the number of horizontal pixel columns of all feature blocks in the same feature region image and taking the average value. The average height of feature blocks in all feature region images within the same feature subset is summarized to obtain the height set corresponding to the feature subset. The values ​​in the height set are arranged in ascending order, and the height value in the middle position of the arrangement is extracted as the height feature of the feature subset. The average width of feature blocks in all feature region images within the same feature subset is summarized to obtain the width set corresponding to the feature subset. The values ​​in the width set are arranged in ascending order, and the width value in the middle position of the arrangement is extracted as the width feature of the feature subset. The height and width features corresponding to each feature subset in the first feature image set are combined to form feature pairs corresponding to each feature subset; the feature pairs corresponding to each feature subset are combined to obtain the first pixel feature corresponding to the first feature image set. The height and width features corresponding to each feature subset in the second feature image set are combined to form feature pairs corresponding to each feature subset. The feature pairs corresponding to each feature subset are then combined to obtain the second pixel feature corresponding to the second feature image set.

5. The audit data security management method based on a large model according to claim 1, characterized in that: By comparing and analyzing the first pixel feature with the second pixel feature, the abnormal pixel features are obtained as follows: Establish a feature coordinate system with the height feature of the feature subset as the vertical axis and the width feature as the horizontal axis; The feature pairs corresponding to each feature subset in the first pixel feature are marked in the feature coordinate system to form a first feature point set. The feature points in the first feature point set are connected in sequence to enclose the first feature enclosed region. The feature pairs corresponding to each feature subset in the second pixel feature are marked in the feature coordinate system to form a second feature point set. The feature points in the second feature point set are connected in sequence to form a second feature enclosed region. The overlapping areas between the second feature-enclosed region and the first feature-enclosed region are removed, and the non-overlapping areas between the second feature-enclosed region and the first feature-enclosed region are retained. The non-overlapping areas are the abnormal pixel features.

6. The audit data security management method based on a large model according to claim 1, characterized in that: Determining whether the image of the transaction certificate to be detected is an abnormal image includes: Feature pairs corresponding to each feature subset in the target pixel features of the image to be detected are extracted and labeled in the feature coordinate system to form a target point set. The target points in the target point set are connected in sequence to form a target enclosing region. The target enclosing region is the pixel feature to be detected in the transaction certificate image to be detected. Determine whether there is an overlapping area between the pixel feature to be detected and the abnormal pixel feature. If there is an overlapping area between the pixel feature to be detected and the abnormal pixel feature, the transaction voucher image to be detected is an abnormal image; if there is no overlapping area between the pixel feature to be detected and the abnormal pixel feature, the transaction voucher image to be detected is normal.

7. The audit data security management method based on a large model according to claim 1, characterized in that: Based on the judgment result, the following processing is performed on the transaction image to be detected: If the transaction voucher image to be detected is an abnormal image, then the feature entries corresponding to the feature pairs included in the region of the transaction voucher image to be detected that overlaps with the abnormal pixel features are extracted, and the information corresponding to the feature entries is output as warning information. If the transaction voucher image to be detected is a normal image, then the text in each target feature image of the transaction voucher image to be detected is extracted to obtain a set of text content corresponding to each target feature image; the content of each set of text content is extracted sequentially according to the parsing target to obtain the parsing content of the parsing target; wherein, the parsing target includes transaction object, transaction address, transaction category, transaction date and transaction amount; the parsing target and the corresponding parsing content are output in pairs sequentially to obtain the analysis result of the transaction voucher image to be detected.

8. A large-scale model-based audit data security management system, used to execute the large-scale model-based audit data security management method according to any one of claims 1-7, characterized in that: The management system includes: an image acquisition module, a region feature recognition module, a partitioning and cropping module, a pixel feature analysis module, a comparison analysis module, an anomaly detection module, and a processing module; The image acquisition module is used to acquire images of transaction vouchers to be detected, and simultaneously collect several standard transaction voucher images and several abnormal transaction voucher images; The region feature recognition module is used to perform region feature recognition on the standard transaction voucher image and the abnormal transaction voucher image respectively, and generate segmentation boxes corresponding to the text regions in various images. The partitioning module is used to partition and extract the standard transaction certificate image and the abnormal transaction certificate image through the segmentation box to obtain a first feature image set and a second feature image set. The pixel feature analysis module is used to traverse the pixels of several feature images in the first feature image set and the second feature image set respectively, and analyze and obtain the first pixel feature corresponding to the first feature image set and the second pixel feature corresponding to the second feature image set. The comparison analysis module is used to compare and analyze the first pixel feature and the second pixel feature to obtain abnormal pixel features; The anomaly detection module is used to extract the target pixel features of the transaction voucher image to be detected, compare the target pixel features with the abnormal pixel features, and determine whether the transaction voucher image to be detected is an abnormal image. The processing module is used to perform corresponding processing on the transaction image to be detected based on the judgment result.

Images