A method and system for compliance document review based on text recognition and understanding

Abstract

The application relates to the technical field of image processing, and discloses a compliance document examination method and system based on text recognition and understanding, which comprises the following steps: collecting a contract document to be examined, and converting the contract document to obtain a contract image to be examined; performing document type acquisition and examination on the contract image to be examined based on a knowledge base; obtaining a plurality of single signature areas and a plurality of connected domains in the single signature areas; acquiring the width fluctuation degree and the gray scale fluctuation degree of the single signature areas; obtaining the handwriting duplication probability of the contract image to be examined; decomposing the single signature areas through RPCA to obtain a text area and a text noise area; obtaining a repair area through an image repair algorithm; analyzing the similar relationship of the gray scale noise performance to obtain the noise duplication probability of the contract image to be examined; acquiring the comprehensive duplication probability of the contract image to be examined; and performing compliance document examination on the contract image to be examined. The application aims to solve the problem that secondary printing as a highly concealed tampering method affects the security of compliance document examination.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a compliance document review method and system based on text recognition and understanding. Background Technology

[0002] Current compliance reviews suffer from fundamental flaws when dealing with highly sophisticated document tampering methods like "reprinting." Existing methods analyze documents by converting them into linear text streams, completely stripping away crucial physical attributes such as the spatial structure, layout features, and visual presentation of the text on the original page. This approach prevents the system from perceiving and identifying numerous traces of tampering in reprinted documents, such as subtle adjustments to character spacing and minor deviations in margin settings. Because the system can only acquire and process the text content itself, once the altered text is accurately identified, this physical evidence revealing the document's forgery is completely filtered and ignored at the outset of the processing. This makes it difficult for existing review systems to effectively identify meticulously crafted reprinted forgeries, leaving insurmountable security vulnerabilities for highly deceptive contract fraud. Summary of the Invention

[0003] This invention provides a compliance document review method and system based on text recognition and understanding to address the problem that existing methods of reprinting can be used as a highly covert means of tampering, affecting the security of compliance document review. The specific technical solution adopted is as follows:

[0004] This invention proposes a compliance document review method based on text recognition and understanding, which includes the following steps:

[0005] Collect the contract documents to be reviewed and convert them into images of the contracts to be reviewed;

[0006] The document type of the contract image under review is obtained and examined based on a knowledge base; the handwritten part of the signature area in the contract image under review is extracted, and several single signature areas and several connected components therein are obtained; the width fluctuation degree of the single signature area is obtained based on the difference in the shape and width of different connected components in the same single signature area; the grayscale fluctuation degree of the single signature area is obtained based on the difference in the grayscale appearance of different connected components in the same single signature area; the handwriting copying probability of the contract image under review is obtained by combining the width fluctuation degree and the grayscale fluctuation degree.

[0007] The low-rank text region and sparse text noise region are obtained by RPCA decomposition of the single signature region; the background region of the single signature region is obtained based on the distribution of the single signature region and other neighboring single signature regions, and the text region is repaired based on the background region by the image inpainting algorithm to obtain the repaired region; the similarity between the gray-scale noise performance of the repaired region and the text noise region is analyzed to obtain the noise copying probability of the contract image to be examined.

[0008] Based on the handwriting copying probability and noise copying probability of the contract image to be reviewed, the comprehensive copying probability of the contract image to be reviewed is obtained, and then compliance document review and graded response of the contract image to be reviewed are carried out.

[0009] Optionally, the specific method for extracting the handwritten portion of the signature area in the contract image to be reviewed and obtaining several single signature areas and several connected components therein includes:

[0010] The signature area in the contract image to be reviewed is extracted. The handwritten part in the signature area is obtained through knowledge base and semantic segmentation technology. The handwritten part is segmented into characters, and each character corresponds to a region, which is recorded as a single signature region.

[0011] For any single signature region, several connected components of that single signature region are obtained through connected component analysis.

[0012] Optionally, the method for obtaining the width fluctuation of the single signature area is as follows:

[0013] For any single signature region, the skeleton of each connected component is obtained through a morphological algorithm. For any connected component, the width of each pixel on the skeleton of the connected component is obtained along the skeleton of the connected component based on the perpendicular line to the pixel on the skeleton, and is used as the width set of the connected component. The variance of the width set of all connected components in the single signature region is calculated and used as the width variance of the single signature region.

[0014] Obtain the width variance of all single signature areas, perform linear normalization on all width variances, and use the result as the width fluctuation degree of each single signature area.

[0015] Optionally, the grayscale fluctuation level of the single signature area can be obtained using the following method:

[0016] For any connected component in any single signature region, the mode of the grayscale values ​​of the pixels in the connected component is taken as the grayscale representative value of the connected component. The variance of the grayscale representative values ​​of all connected components in the single signature region is calculated and taken as the grayscale variance of the single signature region.

[0017] Obtain the grayscale variance of all single-signature regions and perform linear normalization. The result is used as the grayscale fluctuation level of each single-signature region.

[0018] Optionally, the specific method for obtaining the handwriting copy probability of the contract image to be examined includes:

[0019] The mean of the width fluctuation of any single signature area and the handwriting fluctuation is taken as the handwriting fluctuation of that single signature area; the mean and variance of the handwriting fluctuation of all single signature areas in the handwriting part are calculated respectively, and the product of the mean and the variance and the result of inverse proportional normalization are taken as the handwriting copy probability of the contract image to be examined.

[0020] Optionally, the specific method for obtaining low-rank text regions and sparse text noise regions from a single signature region through RPCA decomposition includes:

[0021] For any single signature area, obtain the minimum bounding rectangle of the single signature area, and take the area in the minimum bounding rectangle as the handwritten text area of ​​the single signature area. Obtain the low-rank image and sparse image of the handwritten text area through RPCA decomposition.

[0022] The obtained low-rank image is used as the text region of the single signature area, and the obtained sparse image is used as the text noise region of the single signature area.

[0023] Optionally, the method of obtaining the background area of ​​a single signature area based on its distribution with neighboring single signature areas, and then repairing the text area based on the background area using an image inpainting algorithm to obtain the repaired area, includes the following specific methods:

[0024] For any single signature area, obtain the maximum bounding rectangle. Based on the minimum bounding rectangle, expand outward to obtain the maximum bounding rectangle. During the expansion process, stop expanding when it encounters other single signature areas. Subtract the minimum bounding rectangle from the maximum bounding rectangle of the single signature area to obtain the area, which is used as the background area of ​​the single signature area.

[0025] Using the text area of ​​the signature area as the area to be repaired and the background area as the source area, the repair algorithm is run to obtain the repaired area.

[0026] Optionally, the specific method for obtaining the noise copying probability of the contract image to be examined includes:

[0027] For any single signature area, its text noise area and repair area are vectorized respectively to obtain the text noise area vector and the repair area vector.

[0028] The p-value of the noise region vector and the repair region vector of the single signature area are obtained by KS test and used as the noise similarity of the single signature area;

[0029] The mean of the noise similarity of all single-signature regions is used as the noise copy probability of the contract image to be examined.

[0030] Optionally, the comprehensive copy probability of the contract image to be examined is obtained by the following method:

[0031] The average of the handwriting copy probability and the noise copy probability of the contract image under review is used as the overall copy probability of the contract image under review.

[0032] The present invention also proposes a compliance document review system based on text recognition and understanding. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the above method.

[0033] The beneficial effects of this invention are as follows: After the review system collects and receives contract documents, it converts them to PDF format and performs standardized type acquisition and corresponding standard review based on computer vision. Furthermore, considering the hidden tampering forms that exist in secondary copying, the invention extracts handwritten text during the review process to obtain a single signature area. Considering the differences in the thickness and density of handwritten characters, which do not consistently exhibit similar thickness and density, the invention analyzes the width and grayscale differences of each connected region in the single signature area to obtain the degree of width and grayscale fluctuation, thereby obtaining the probability of handwriting copying to reflect the copying possibility based on the thickness and density of the handwriting. Furthermore, considering that secondary copying will cause the handwritten text and background to be actually on the same layer, their toner noise will have a similar distribution relationship, the invention further analyzes the single signature area... RPCA decomposition is performed to extract text regions and text noise regions. The low-rank and sparse images obtained by RPCA can reflect the original text and noise parts in a single signature area. Then, an image inpainting algorithm is used to inpaint the image based on the background area surrounding the single signature area, so as to intelligently generate noise in the text region through background noise. By analyzing the noise distribution of the inpainted area and the text noise region, consistency analysis is performed in vector form to obtain noise similarity. The greater the similarity, the more similar the toner noise distribution of the text part is to the toner noise distribution of the background part, and the greater the corresponding noise copying probability. By combining the handwriting copying probability and the noise copying probability, a comprehensive noise probability is obtained and a graded response judgment is made. Thus, based on the traditional review system, a second copy review of contract documents is further realized, ultimately improving the security and robustness of compliance document review. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of a compliance document review method based on text recognition and understanding, provided as an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of toner noise. Detailed Implementation

[0037] 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.

[0038] Please see Figure 1 The diagram illustrates a flowchart of a compliance document review method based on text recognition and understanding, provided by an embodiment of the present invention. The method includes the following steps:

[0039] Step S001: Collect the contract document to be reviewed and convert it into an image of the contract to be reviewed.

[0040] The purpose of this embodiment is to conduct compliance review of contract documents. After the traditional multi-standard review of contract documents through computer vision, it is necessary to perform secondary copying and identification of the contract documents to prevent tamperers from adding content through secondary copying to form document content that is difficult for computers to identify. Therefore, it is necessary to first collect the contract documents to be reviewed and convert them into image format in PDF form, and preprocess them to obtain the image of the contract to be reviewed.

[0041] Specifically, contract documents enter the review system through various channels, including direct user uploads, API integration with existing enterprise business systems (such as CRM and ERP), or automatic retrieval via specific email addresses and shared directories. This enables the review system to collect contract documents to be reviewed. Upon receiving contract documents, the review system automatically records key source data such as contract type and business department. Then, the review system converts all contract documents of various formats into high-resolution PDF files as initial contract images. The initial contract images are then processed through grayscale conversion and contrast enhancement (preprocessing techniques in image processing, which will not be elaborated in this embodiment) to obtain the image of the contract to be reviewed.

[0042] It should be noted that the tamperer may first photocopy the original contract, then add content to the photocopy and photocopy it again, making the altered contract easily recognizable by the computer system, making it difficult to detect the tampering. In such cases, it is necessary to find the handwritten signature on the contract. Since this signature is handwritten, it contains many subtle and unique features that are difficult to replicate. If the current contract has been photocopied twice, the traces of this "secondary photocopy tampering" will be more obvious in the handwritten signature.

[0043] Step S002: Obtain and review the document type of the contract image to be reviewed based on the knowledge base; extract the handwritten part of the signature area in the contract image to be reviewed, and obtain several single signature areas and several connected components therein; obtain the width fluctuation degree of the single signature area based on the difference in the shape and width of different connected components in the same single signature area; obtain the grayscale fluctuation degree of the single signature area based on the difference in the grayscale of different connected components in the same single signature area; combine the width fluctuation degree and the grayscale fluctuation degree to obtain the handwriting copy probability of the contract image to be reviewed.

[0044] Preferably, in one embodiment of the present invention, the method for obtaining and reviewing the document type of the contract image to be reviewed based on a knowledge base includes:

[0045] The contract image under review is processed using OCR technology for text recognition, simultaneously recording the visual attributes and spatial information of the text, which are then used as text information and image information, respectively. The text information is a structured JSON array containing the content, font, font size, bold status, and precise bounding box coordinates (bbox) for each text block. The image information consists of the contract image to be reviewed, marked with several words, and the page number; while the image information is the image of the contract to be reviewed, marked with several words.

[0046] Furthermore, based on textual information, the NLP text classification model is used to identify the contract document type (e.g., procurement contract, labor contract, technology agreement) of the contract images under review. Through collaboration between legal experts and domain engineers, the review points for each type of contract document are structured and entered into a database, forming a knowledge base. This knowledge base includes a rule base, a template library, style templates, and a query interface. The rule base stores information in the form of lists and logical rules, with review criteria including a list of essential clauses, a list of prohibited words, and logical verification rules. The template library stores the text content of standard contracts and marks key paragraphs; these standard contract texts are reviewed during the review process. Style templates define standard layout specifications using structured data formats such as JSON; after determining the contract document type, their standard formatting specifications (e.g., title font, signature area, etc.) are reviewed. Simultaneously, the review system provides a query interface. After determining the contract document type, by inputting the contract document type, the review system performs a database query, obtaining all review criteria for the corresponding contract document type with one click and outputting a risk list, detailing the content, risk type, risk score, and location of each non-compliant text fragment.

[0047] It should be further explained that the original signature has clear changes in pressure at the beginning, end, and turning points of the strokes, which are reflected in the thickness and density of the ink. However, after a second copy, the signature becomes a "flat picture" with a fixed thickness. All the three-dimensional information such as the brushstrokes and pauses caused by pen pressure is lost. Therefore, the possibility of a second copy can be identified by analyzing the rate of change in the stroke width.

[0048] Preferably, in one embodiment of the present invention, the handwritten portion of the signature area in the image of the contract to be reviewed is extracted, and several single signature areas and several connected components therein are obtained. The specific method includes:

[0049] The signature area in the contract image to be reviewed is extracted. Using a knowledge base and semantic segmentation technology, the handwritten portion of the signature area is obtained. The handwritten portion is then segmented into characters, with each character corresponding to a region, which is recorded as a single signature region, resulting in several single signature regions. The signature areas in the contract image are already marked with templates from the knowledge base. A training dataset is constructed based on a large number of contract images in the knowledge base, and the signature areas are manually marked to train a semantic segmentation neural network. The cross-entropy loss function is used to obtain the trained semantic segmentation network. The contract image to be reviewed is input into the trained semantic segmentation network, and the output is the handwritten portion of the signature area, i.e., the text portion of the signature area. The text segmentation of the handwritten portion is performed using existing text extraction techniques in the field of computer vision (such as OCR recognition), which will not be elaborated further in this embodiment.

[0050] Furthermore, for any single signature area, several connected components of the single signature area are obtained through connected component analysis. Each connected component corresponds to a radical or component of the text in the single signature area.

[0051] It should be further explained that by extracting the handwritten part of the signature area and dividing it into single signature areas, and extracting connected components from the single signature areas, a foundation is provided for subsequent analysis of the differences in width and grayscale performance of each connected component.

[0052] Preferably, in one embodiment of the present invention, the method for obtaining the width fluctuation degree of a single signature region based on the difference in the morphological width of different connected regions within the same single signature region includes:

[0053] For any single signature region, the skeleton of each connected component is obtained using a morphological algorithm. Then, for any connected component, the width of each pixel on the skeleton is obtained along the skeleton of that connected component, based on a perpendicular line to the pixels on the skeleton. This width is used as the set of widths of that connected component. The variance of the widths in the width sets of all connected components in the single signature region is calculated and used as the width variance of the single signature region. The width variances of all single signature regions are obtained using the above method. All width variances are then linearly normalized, and the result is used as the width fluctuation degree of each single signature region.

[0054] Preferably, in one embodiment of the present invention, the grayscale fluctuation degree of a single signature region is obtained based on the difference in grayscale performance of different connected regions within the same single signature region, including the following specific method:

[0055] For any connected component in any single signature region, the mode of the grayscale values ​​of the pixels in the connected component is taken as the grayscale representative value of the connected component. The variance of the grayscale representative values ​​of all connected components in the single signature region is calculated and taken as the grayscale variance of the single signature region. The grayscale variance of all single signature regions is obtained by the above method and linearly normalized. The result is taken as the grayscale fluctuation degree of each single signature region.

[0056] It should be noted that the larger the width variance of a single signature area, that is, the greater the variation in the thickness of the handwriting between different strokes in each connected region, the greater the corresponding width variation, and the greater the probability that it is handwritten. The gray value of each connected region reflects the handwriting density of the corresponding stroke. The greater the variation in the gray value between different strokes, the greater the variation in handwriting density, and the greater the degree of gray value variation, the greater the probability that it is handwritten.

[0057] Preferably, in one embodiment of the present invention, the probability of handwriting reproduction of the contract image to be examined is obtained by combining the width fluctuation degree and the grayscale fluctuation degree, and the specific method includes:

[0058] The mean of the width fluctuation of any single signature area and the mean of the handwriting fluctuation are taken as the handwriting fluctuation of that single signature area; the mean and variance of the handwriting fluctuation of all single signature areas in the handwriting portion are calculated respectively, and the product of the mean and the variance and inversely normalized result is taken as the handwriting copy probability of the contract image to be examined; this embodiment adopts The model is used to represent the inverse proportional relationship and for normalization processing. As input to the model, This represents an exponential function with the natural constant as the base. Implementers can set inverse proportional functions and normalization functions according to the actual situation.

[0059] It should be noted that the width fluctuation and handwriting fluctuation reflect the handwriting probability in terms of both thickness and density. By calculating the mean, it is necessary to ensure that both aspects are large, thus ensuring a large handwriting fluctuation to reflect the handwriting probability. Furthermore, the mean and variance of handwriting fluctuation are calculated through multiple signature areas. The larger the mean, the larger the overall handwriting fluctuation. At the same time, it is necessary to ensure that the variance is large to avoid the problem of similar handwriting fluctuation in different signature areas when photocopying, and the lower the probability of handwriting photocopying.

[0060] Thus, the probability of copying the handwriting in the image of the contract to be examined is obtained.

[0061] Step S003: Obtain the low-rank text region and sparse text noise region by decomposing the single signature region using RPCA; obtain the background region of the single signature region based on the distribution of the single signature region and other neighboring single signature regions, and repair the text region based on the background region using an image inpainting algorithm to obtain the repaired region; analyze the similarity between the gray-scale noise performance of the repaired region and the text noise region to obtain the noise copying probability of the contract image to be examined.

[0062] It should be noted that because laser copying uses electrostatic adsorption of toner for imaging, the resulting document will have randomly distributed microscopic toner noise (i.e., "copy background gray") in the blank areas. When the first copy is reproduced, this background gray noise is simultaneously solidified into the new document along with the document content (including handwriting). If the handwriting is original, the ink will completely cover the paper fibers, forming opaque handwriting. However, in the second copy, upon magnification, it can be seen that the gray areas of the handwriting strokes are uniformly mixed with fine toner particles consistent with the surrounding blank areas. This "noise penetration" phenomenon proves that the handwriting and background noise belong to the same imaging layer, thus revealing that it is not the original handwriting but a copy formed by the second copy. Figure 2As shown, it illustrates the manifestation of toner noise in an image; that is, if it is a secondary copy, the handwritten text after copying is interfered with by toner noise, resulting in discontinuities at the microscopic level. Therefore, the noise distribution in the noise image obtained after RPCA decomposition is very similar to that of the background noise.

[0063] Preferably, in one embodiment of the present invention, the method for obtaining low-rank text regions and sparse text noise regions from a single signature region through RPCA decomposition includes:

[0064] For any single signature area, obtain the minimum bounding rectangle of the single signature area, and take the area in the minimum bounding rectangle as the handwritten text area of ​​the single signature area. The handwritten text area is decomposed by RPCA to obtain a low-rank image and a sparse image. RPCA decomposition, namely Robust Principal Component Analysis, is an existing technology. The low-rank image obtained is the original noise-free image, and the sparse image is the noisy image. The obtained low-rank image is taken as the text area of ​​the single signature area, and the obtained sparse image is taken as the text noise area of ​​the single signature area.

[0065] It should be further explained that after obtaining the text area and text noise area from the signature area, it is necessary to obtain the maximum bounding rectangle from the signature area. The noise in the other areas of the maximum bounding rectangle, excluding the text area, is the background noise. The noise in the text area is generated by the background noise distribution in the other areas. Then, the generated noise is compared with the text noise. If the similarity is high, it is likely a second copy.

[0066] Preferably, in one embodiment of the present invention, based on the distribution of the single signature area and other adjacent single signature areas, the background area of ​​the single signature area is obtained, and the text area is repaired based on the background area using an image inpainting algorithm to obtain the repaired area. The specific method includes:

[0067] For any single signature area, obtain the maximum bounding rectangle, where the four sides of the maximum and minimum bounding rectangles are parallel and their centers are the same. That is, expand outward from the minimum bounding rectangle to obtain the maximum bounding rectangle. During the expansion process, stop when it encounters other single signature areas. Subtract the minimum bounding rectangle from the maximum bounding rectangle of the single signature area to obtain the background area of ​​the single signature area. Based on the gray value distribution of the pixels in the background area, calculate the mean and standard deviation using a Gaussian model to obtain the probability distribution of the background area, which is used as the probability distribution of noise. The calculation of the probability distribution using a Gaussian model is a prior art and will not be described in detail in this embodiment.

[0068] It should be further explained that in order to preserve the spatial structure of the noise, the text area needs to be repaired using an image inpainting algorithm. The text area is repaired using the background area as the source area to obtain the repaired area. The inpainting algorithm can intelligently fill the text area to be repaired based on the texture and noise pattern of the surrounding background to generate the repaired area.

[0069] Furthermore, based on the background area of ​​the signature area, the Criminisi algorithm (existing technology, which will not be described in detail in this embodiment, is used to repair the noise based on the probability distribution in the background area) is used to select the text area as the area to be repaired and the background area as the source area. The repair algorithm is then run to obtain the repair area. The size of the repair area image is the same as that of the text area and the text noise area.

[0070] Preferably, in one embodiment of the present invention, the method for analyzing the similarity between the grayscale noise performance of the repaired area and the text noise area to obtain the noise copying probability of the contract image to be examined includes:

[0071] For any single signature area, its text noise area and repair area are vectorized separately. In this embodiment, a row-first, column-later format is used, arranging the gray values ​​of the pixels row by row to obtain the text noise area vector and the repair area vector. The text noise area vector and the repair area vector of the single signature area are subjected to the KS test (an existing vector analysis technique, which will not be elaborated in this embodiment) to obtain the p value, which is used as the noise similarity of the single signature area. The mean of the noise similarity of all single signature areas is used as the noise copying probability of the contract image to be examined.

[0072] It should be noted that the greater the vector consistency between the text noise region and the repaired region, the more similar the distribution of text noise and background noise, and the more likely it is that the text and background were formed in the same printing process. This is different from the characteristic that the text and background are not formed in the same process during handwriting, and the probability of it being a copy is higher. On the other hand, the greater the noise similarity of the overall single signature region, that is, the greater the mean value, the greater the possibility that the whole is a copy.

[0073] Thus, the probability of noisy copying of the contract image to be examined is obtained.

[0074] Step S004: Based on the handwriting copy probability and noise copy probability of the contract image to be reviewed, obtain the comprehensive copy probability of the contract image to be reviewed, and then conduct compliance document review and graded response of the contract image to be reviewed.

[0075] It should be noted that both handwriting copying probability and noise copying probability can reflect the possibility of secondary copying of a contract. Therefore, a comprehensive analysis is needed to obtain the overall copying probability in order to analyze the possibility of secondary copying of the contract image under review and to establish a tiered response mechanism to achieve the review of compliant documents.

[0076] Specifically, the average of the handwriting copy probability and the noise copy probability of the contract image under review is used as the comprehensive copy probability of the contract image under review. This embodiment sets five levels: "Safe," "Slightly Abnormal," "Moderate Risk," "Severe Risk," and "Highest Alert," with first to fourth level thresholds described as 0.1, 0.4, 0.7, and 0.9, respectively. If the comprehensive copy probability of the contract image under review is less than the first level threshold, the review system marks it as "Safe," and the subsequent review process proceeds normally. If the comprehensive copy probability is greater than or equal to the first level threshold and less than the second level threshold, it is marked as "Slightly Abnormal," and the subsequent review process proceeds normally. If the overall copying probability is greater than or equal to the second-level threshold but less than the third-level threshold, it is marked as "moderate risk." Reviewers are advised to make a comprehensive judgment based on the contract document corresponding to the image of the contract under review, and then determine whether to proceed with the subsequent review process. If the overall copying probability is greater than or equal to the third-level threshold but less than or equal to the fourth-level threshold, it is marked as "severe risk." The review system will force manual review and prompt for verification of the corresponding original legal documents. If the overall copying probability is greater than the fourth-level threshold, the review system will trigger the highest alert, immediately suspending the relevant business processes for the corresponding contract and automatically reporting to the relevant risk control department to initiate a special investigation. This achieves compliance review of contract documents.

[0077] This concludes the embodiment.

[0078] Another embodiment of the present invention provides a compliance document review system based on text recognition and understanding. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the above method steps S001 to S004.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for compliance document review based on text recognition and understanding, characterized in that, The method comprises the following steps: Collecting a contract document to be examined, and converting to obtain a contract image to be examined; Based on a knowledge base, document type acquisition and examination are performed on the contract image to be examined; a handwritten part of a signature area in the contract image to be examined is extracted, and a plurality of single signature areas and a plurality of connected domains in the single signature areas are obtained; based on the morphological width difference of different connected domains in the same single signature area, the width fluctuation degree of the single signature area is obtained; based on the gray level difference of different connected domains in the same single signature area, the gray level fluctuation degree of the single signature area is obtained; and the width fluctuation degree and the gray level fluctuation degree are comprehensively obtained to obtain a handwriting copy probability of the contract image to be examined; The specific method for obtaining the handwriting copy probability of the contract image to be examined comprises: The width fluctuation degree and the mean value of the handwriting fluctuation degree of any single signature area are taken as the handwriting fluctuation degree of the single signature area; the mean value and the variance of the handwriting fluctuation degree of all single signature areas in the handwritten part are calculated respectively, and the product of the mean value and the variance is normalized in inverse proportion to obtain the handwriting copy probability of the contract image to be examined; The single signature area is decomposed by RPCA to obtain a low-rank text area and a sparse text noise area; based on the distribution of the single signature area and other adjacent single signature areas, a background area of the single signature area is obtained, and the text area is repaired based on the background area by an image repair algorithm to obtain a repaired area; the gray level noise performance similarity relationship between the repaired area and the text noise area is analyzed to obtain a noise copy probability of the contract image to be examined; Based on the handwriting copy probability and the noise copy probability of the contract image to be examined, a comprehensive copy probability of the contract image to be examined is obtained, and then a compliant document examination and a hierarchical response of the contract image to be examined are performed.

2. The method of claim 1, wherein, The specific method for extracting the handwritten part of the signature area in the contract image to be examined and obtaining a plurality of single signature areas and a plurality of connected domains in the single signature areas comprises: The signature area in the contract image to be examined is extracted, the handwritten part in the signature area is obtained by a knowledge base and a semantic segmentation technology, the handwritten part is subjected to text segmentation, each text corresponds to an area, and the area is recorded as a single signature area; For any single signature area, a plurality of connected domains of the single signature area are obtained by connected domain analysis.

3. The method of claim 1, wherein, The specific method for obtaining the width fluctuation degree of the single signature area comprises: For any single signature area, the skeleton of each connected domain is obtained by a morphological algorithm, for any connected domain, the width of each pixel point on the skeleton of the connected domain is obtained based on the perpendicular line perpendicular to the pixel point on the skeleton, as the width set of the connected domain, and the variance of the width in the width set of all connected domains in the single signature area is calculated as the width variance of the single signature area; The width variances of all single signature areas are linearly normalized, and the obtained result is taken as the width fluctuation degree of each single signature area.

4. The method of claim 1, wherein, The specific method for obtaining the gray level fluctuation degree of the single signature area comprises: For any single signature area in any connected domain, the gray value of the pixel point in the connected domain is the mode of the gray value of the connected domain, and the gray representative value of the connected domain is calculated, and the gray variance of all connected domains in the single signature area is calculated as the gray variance of the single signature area. The gray variance of all single signature areas is obtained and linearly normalized, and the result is obtained as the gray fluctuation degree of each single signature area.

5. The method of claim 1, wherein, The low-rank text area and sparse text noise area of the single signature area are obtained by RPCA decomposition, including the following specific methods: For any single signature area, the minimum circumscribed rectangle of the single signature area is obtained, and the region in the minimum circumscribed rectangle is taken as the handwritten character region of the single signature area. The low-rank image and sparse image of the handwritten character region are obtained by RPCA decomposition. The obtained low-rank image is taken as the text region of the single signature area, and the obtained sparse image is taken as the text noise region of the single signature area.

6. The method of claim 5, wherein, The background region of the single signature area is obtained based on the distribution of the single signature area and adjacent other single signature areas, and the text region is repaired based on the background region by the image repair algorithm to obtain a repaired region, including the following specific methods: The maximum circumscribed rectangle of any single signature area is obtained, and the maximum circumscribed rectangle is obtained based on the minimum circumscribed rectangle. During the expansion process, when other single signature areas are contacted, the expansion is stopped. The region obtained by subtracting the minimum circumscribed rectangle from the maximum circumscribed rectangle of the single signature area is taken as the background region of the single signature area. Based on the text region of the single signature area as the to-be-repaired region, the background region as the source region, and the repair algorithm is run to obtain a repaired region.

7. The method of claim 1, wherein, The noise copying probability of the to-be-reviewed contract image is obtained, including the following specific methods: For any single signature area, the text noise region and the repaired region are respectively vectorized to obtain a text noise region vector and a repaired region vector. The p value of the text noise region vector and the repaired region vector of the single signature area is obtained by K-S test, and the noise similarity of the single signature area is obtained. The average of the noise similarities of all single signature areas is taken as the noise copying probability of the to-be-reviewed contract image.

8. The method of claim 1, wherein, The comprehensive copying probability of the to-be-reviewed contract image is obtained, including the following specific methods: The average of the noise copying probability and the handwriting copying probability of the to-be-reviewed contract image is taken as the comprehensive copying probability of the to-be-reviewed contract image.

9. A text recognition and understanding based compliance document review system comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, The processor executes the computer program to realize the steps of the compliance document review method based on text recognition and understanding according to any one of claims 1-8.

Images