College financial expense report automatic auditing method based on OCR fusion

Abstract

This invention discloses an automatic review method for university financial reimbursement documents integrating OCR, belonging to the field of intelligent recognition and review technology for financial documents. The method includes acquiring original images of reimbursement documents and preprocessing them to obtain standardized images; using OCR to identify structured text containing text and coordinates; and relying on a template knowledge base to complete field alignment and entity extraction to generate a reimbursement data table. A dynamic threshold anomaly detection algorithm based on a multi-dimensional feature space of document types is used to complete logical verification, matching and identifying conflict points with a financial rule base, and generating suggestions through historical decision case similarity review. Finally, an audit report containing audit conclusions, conflict details, and suggestions is generated. This method optimizes the anomaly detection and conflict review mechanism, improving the automation level and accuracy of university financial reimbursement review.

Description

Technical Field

[0001] This invention belongs to the field of intelligent recognition and review technology for financial documents, specifically a method for automatic review of financial reimbursement documents in universities that integrates OCR. Background Technology

[0002] Currently, the review of financial reimbursement documents in universities mostly adopts a processing mode that combines manual verification with conventional OCR recognition. After basic preprocessing of the document images, the text content is extracted through a general OCR engine, simple field correspondence is completed by relying on fixed templates, data anomaly judgment is carried out by static fixed thresholds, and the extracted information is matched with the financial rule base word by word. Disputed content is reviewed and judged manually one by one.

[0003] Conventional OCR recognition only outputs fragmented text, failing to combine coordinate information with a template knowledge base to achieve accurate field alignment and entity extraction, resulting in insufficient accuracy in structured information extraction. Anomaly detection uses a uniform, fixed threshold, failing to construct a multi-dimensional feature space based on document type, thus unable to adapt to the attribute differences of different documents, leading to numerous false positives and false negatives in logical consistency checks. Rule matching can only perform explicit clause comparisons, lacking automated methods for determining ambiguous rule conflicts. Manual review relies on personal experience, and the criteria for judging similar conflicts vary, making it difficult to guarantee the consistency of audit results.

[0004] To address the issue of insufficient adaptability of fixed threshold anomaly detection, a multi-dimensional feature space matching document types needs to be constructed to achieve logical verification of dynamic thresholds. To address the problem of inconsistent standards for manual review of rule conflicts, similarity comparisons based on historical decision-making cases are necessary to form an automated review process and generate corresponding review evidence. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes an automatic verification method for university financial reimbursement documents that integrates OCR, including:

[0007] The original images of the financial reimbursement documents to be reviewed are acquired using an image acquisition device.

[0008] The original image is preprocessed to obtain a standardized document image;

[0009] The standardized document image is subjected to text recognition using an optical character recognition engine to extract structured text information containing text content and coordinate positions;

[0010] Based on a pre-defined document template knowledge base, the structured text information is aligned with fields and entities are extracted to generate a preliminary reimbursement information data table.

[0011] An improved anomaly detection algorithm is used to perform logical consistency verification on the preliminary reimbursement information data table. The improved anomaly detection algorithm performs dynamic threshold determination based on the multi-dimensional feature space of the document type.

[0012] The reimbursement information data table, which has undergone logical consistency verification, is matched against the preset financial audit rule base one by one to identify potential rule conflict points.

[0013] For the identified potential rule conflicts, a similarity review process based on historical decision-making cases is initiated to generate review recommendations;

[0014] Based on the matching results between the review recommendations and the financial audit rule base, a final audit report is generated, which includes the audit conclusions, conflict details, and recommendations.

[0015] Further, the original image is preprocessed to obtain a standardized document image, including:

[0016] Perform a grayscale conversion on the original image to obtain a grayscale image;

[0017] An adaptive threshold segmentation algorithm is applied to the grayscale image to separate the foreground text, stamp area and background area;

[0018] The perspective distortion of the document's main body area caused by the shooting angle in the grayscale image is detected and corrected so that the document's main body area appears as a regular rectangle in the image.

[0019] The corrected image undergoes edge enhancement processing, and its resolution is uniformly adjusted to a preset standard resolution, outputting a standardized document image.

[0020] Furthermore, text recognition is performed on the standardized document image using an optical character recognition engine, including:

[0021] The trained optical character recognition model is invoked to perform multi-scale sliding window scanning on standardized document images;

[0022] Within each scanning window, the optical character recognition model outputs the recognized text character, the corresponding confidence level, and the bounding box coordinates of the text character in the original image.

[0023] The recognition results of all scanning windows are merged, and the lines and paragraphs are reorganized according to the bounding box coordinates of the text characters to form structured text information containing text content, confidence level, and hierarchical positional information consisting of line numbers and paragraph numbers.

[0024] The trained optical character recognition model is constructed as follows:

[0025] Obtain a large-scale dataset of images of financial reimbursement documents from universities, and perform detailed annotation on the text content in the images to form a training sample set;

[0026] A deep learning model comprising a feature extraction network, a sequence modeling layer, and a transcription layer is constructed as the initial optical character recognition model. The feature extraction network uses a convolutional neural network to extract visual features from the input image. The sequence modeling layer uses a bidirectional long short-term memory network to capture the contextual dependencies of the features. The transcription layer uses a connectionist temporal classification decoder to map the sequence features into a text sequence.

[0027] Input the document images from the training sample set into the initial optical character recognition model, and calculate the predicted text sequence output by the model through forward propagation;

[0028] The difference between the predicted text sequence and the labeled text sequence is calculated using the connectionist temporal classification loss function, and the network parameters of the initial optical character recognition model are updated using the backpropagation algorithm.

[0029] The forward propagation and parameter update process is iteratively executed until the text recognition accuracy of the model on the independent validation set converges to a preset threshold, thus obtaining the trained optical character recognition model.

[0030] Furthermore, based on a pre-defined document template knowledge base, field alignment and entity extraction are performed on the structured text information, including:

[0031] Based on the keyword matching results in the structured text information, determine the document type of the document to be reviewed;

[0032] Retrieve the standard template corresponding to the document type from the document template knowledge base. The standard template defines the name, logical position, and expected data type of each standard field.

[0033] By matching text blocks in structured text information with standard fields in standard templates based on the hierarchical and logical positions of the text blocks, field alignment is achieved.

[0034] From each aligned text block, based on the expected data type, specific numerical or text entities are extracted using regular expressions and named entity recognition technology, and then populated into the corresponding fields of the reimbursement information data table.

[0035] Furthermore, the step of using an improved anomaly detection algorithm to perform logical consistency verification on the preliminary reimbursement information data table includes:

[0036] Based on the document type, the corresponding feature space model is loaded from the multi-dimensional feature space model library. The feature space model defines multiple feature dimensions and their normal value ranges for describing the key attributes of the corresponding category of documents.

[0037] Based on the entities in the preliminary reimbursement information data table, calculate their actual feature values ​​in each feature dimension of the feature space model.

[0038] For each feature dimension, the actual feature value is compared with the historical normal range of the feature dimension, and the deviation is calculated.

[0039] Based on the preset correlation between feature dimensions, the judgment threshold of each feature dimension is dynamically adjusted. The dynamically adjusted judgment threshold is weighted and corrected in the same or opposite direction based on the strength of the feature correlation.

[0040] If the actual feature value of a certain feature dimension exceeds the dynamically adjusted judgment threshold, the data corresponding to the feature dimension is marked as a logical anomaly, and all logical anomalies are summarized into a logical consistency verification report.

[0041] Furthermore, the improved anomaly detection algorithm performs dynamic threshold determination based on the multi-dimensional feature space of the document type, and its working principle includes:

[0042] A multi-dimensional feature space is constructed for each document type. The multi-dimensional feature space is spanned by a series of orthogonal or non-orthogonal feature vectors, and each feature vector represents an independent business or logical feature dimension.

[0043] Extract the feature values ​​of similar documents across various feature dimensions from the historical database of audited documents to form a historical feature value distribution;

[0044] By using the kernel density estimation algorithm, the historical feature value distribution of each feature dimension is modeled to obtain the probability density function of the feature dimension;

[0045] The interval boundary of the probability density function at a preset confidence level is used as the benchmark judgment threshold for the feature dimension.

[0046] During runtime, when performing a joint judgment on multiple feature dimensions of a document, the system searches for other feature dimensions that are strongly correlated with the currently judged feature dimension based on a preset feature dimension association diagram.

[0047] Read the current actual feature value of the strongly correlated feature dimension. If the actual feature value deviates from its respective benchmark judgment threshold by more than the preset correlation influence factor, then shrink or expand the benchmark judgment threshold of the current judged feature dimension according to the preset rules to obtain the dynamically adjusted judgment threshold.

[0048] Furthermore, the step of matching the reimbursement information data table, which has undergone logical consistency verification, with the preset financial audit rule base item by item includes:

[0049] From the financial audit rule base, load all audit rules that match the document type and funding source project, wherein the audit rules include conditional expressions and violation conclusions;

[0050] Iterate through each field and its value in the expense reimbursement information data table;

[0051] Substitute the value of each field into the conditional expression of the relevant review rules for logical calculation;

[0052] If the conditional expression evaluates to true, the corresponding violation conclusion of the audit rule is triggered, and the rule number of the violation, the triggered field and its value are recorded as potential rule conflict points.

[0053] Furthermore, for the identified potential rule conflict points, a similarity review process based on historical decision cases is initiated, including:

[0054] A verification query vector is constructed using the rule number, field name, and value involved in potential rule conflict points as key information.

[0055] In the historical decision case database, retrieve historical cases that are similar to the review query vector in terms of rules and fields;

[0056] Calculate the overall similarity between the current case and each retrieved historical case in terms of specific numerical values ​​and contextual information;

[0057] Select historical cases with a comprehensive similarity exceeding a set threshold, extract the final decision results and additional annotations of human reviewers in the historical cases, and form review suggestions.

[0058] Furthermore, the generation of review recommendations includes:

[0059] The final decision results of the extracted historical decision cases are statistically analyzed, and the frequency of occurrence of different decision results is calculated.

[0060] Use the most frequently occurring decision outcomes as the basis for review recommendations;

[0061] If there are historical cases where different decision outcomes have similar frequencies and there are decision-making conflicts, the additional annotations of the conflict cases will be further analyzed to extract the key decision-making considerations. These considerations will then be added as supplementary explanations after the basic review recommendations to form a complete review recommendation.

[0062] Furthermore, the generation of the final audit report, which includes audit conclusions, conflict details, and recommendations, includes:

[0063] Summarize all logical anomalies in the logical consistency verification report, as well as all potential rule conflicts identified in the financial audit rule matching and their corresponding review recommendations;

[0064] If both logical anomalies and potential rule conflicts are empty, a conclusion of approval is generated.

[0065] If there are logical anomalies or potential rule conflicts, the review conclusion will be generated based on the type and severity of the review suggestion, including "suggest resubmitting after modification", "suggest approval with explanation", or "review not approved".

[0066] The audit conclusions, a detailed list of logical anomalies, a list of potential rule conflicts, corresponding review suggestions, and the original key voucher image area index are integrated, formatted, and output as the final audit report document.

[0067] Compared with the prior art, the beneficial effects of the present invention are:

[0068] By constructing a multi-dimensional feature space based on document type and employing dynamic thresholds for anomaly detection, this approach can replace traditional fixed-threshold logical verification methods. The dynamic thresholds can adaptively adjust according to document category and corresponding feature dimensions. Logical consistency verification aligns with the data attributes of different documents, eliminating the mismatch between fixed thresholds and diverse document features, and reducing the probability of false positives and false negatives in anomaly detection. After field alignment and entity extraction, structured text retains complete field attributes and location information. The multi-dimensional feature space can fully integrate various data parameters, resulting in more comprehensive anomaly detection dimensions and a higher degree of consistency between verification results and the actual logic of the document. This improves the accuracy and adaptability of data anomaly identification.

[0069] After identifying potential rule conflicts through rule base matching, similarity reviews are performed using historical decision cases, automatically generating corresponding review suggestions. Ambiguous and borderline conflicts can be quantitatively compared using historical data from similar decisions, reducing subjective differences introduced by manual review. The criteria for judging similar conflicts remain consistent, improving the stability of review results. The review process is directly integrated with the rule matching stage, eliminating the need for separate manual intervention in dispute resolution and enhancing the continuity of the review process. Historical case similarity matching can refine conflict categories, providing objective reference points for conflict determination, improving the standardization and consistency of review judgments, and optimizing the automation and reliability of the overall review process. Attached Figure Description

[0070] Figure 1 This is a state diagram of the automatic verification method for university financial reimbursement documents that integrates OCR as described in this invention.

[0071] Figure 2 A flowchart for preprocessing raw images to obtain standardized document images;

[0072] Figure 3 This is a flowchart for field alignment and entity extraction based on a preset document template knowledge base. Detailed Implementation

[0073] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0074] See Figure 1 The present invention provides an automatic review method for university financial reimbursement documents integrating OCR. The overall implementation follows this process: First, an original digital image of the financial reimbursement document to be reviewed is acquired using an image acquisition device, such as a document scanner or high-speed scanner. Second, preprocessing is performed on the original image to eliminate noise, distortion, and other factors, resulting in a standardized document image. Third, a pre-trained optical character recognition engine is used to perform text recognition on the standardized document image. This process not only extracts the text content from the image but also captures the coordinate position of each character or text block in the image, thereby outputting structured text information containing both text content and coordinate positions. Fourth, based on a preset document template knowledge base, the structured text information is matched with standard templates in the knowledge base to complete field alignment and extract specific entity values ​​from the aligned text blocks, generating a preliminary reimbursement information data table. Fifth, an improved anomaly detection algorithm is used to perform logical consistency verification on the preliminary reimbursement information data table. This algorithm can dynamically threshold the numerical values ​​and logical relationships in the data table based on the multi-dimensional feature space of different document types to detect internal contradictions or anomalies. After verification, the reimbursement information data table, which has passed logical consistency verification, is matched against a pre-set financial audit rule base to identify potential rule conflicts between the document information and established financial regulations. For each identified potential rule conflict, the system initiates a similarity review process based on historical decision cases. This process searches for manual review records in similar scenarios from the historical case base and generates review suggestions accordingly. Finally, the system integrates the results of logical consistency verification, rule matching, and review suggestions to generate a final audit report that includes a clear audit conclusion, detailed conflict information, and suggested modifications or handling.

[0075] In one embodiment of the present invention, a detailed description is provided of how to preprocess the original image to obtain a standardized document image. See also... Figure 2The original image is converted to grayscale from color image space to grayscale image space, resulting in a grayscale image containing only brightness information. An adaptive threshold segmentation algorithm is then applied to this grayscale image. This algorithm dynamically calculates the binarization threshold based on the local brightness characteristics of different regions of the image, effectively separating the foreground text and stamp areas from the background. Perspective distortion of the document's main body area caused by the shooting angle is detected and corrected. The four vertices of the document are located through corner detection, and perspective transformation is used to map them into regular rectangles, ensuring the document's main body area is presented frontally in the image. Edge enhancement processing is then applied to the perspective-corrected image to sharpen the edge features of text and lines, and the resolution is uniformly adjusted to a preset standard resolution, such as 300 DPI. The final output is a standardized document image that is conducive to subsequent processing by the optical character recognition engine.

[0076] In practice, the original image acquired by the image acquisition device is a color digital image, such as a JPEG image containing a university travel expense reimbursement form, with a resolution of 1920 pixels multiplied by 1080 pixels. In the implementation, the original image is converted to grayscale using a weighted average method, where the red component has a weight of 0.299, the green component a weight of 0.587, and the blue component a weight of 0.114. The resulting grayscale image has a grayscale value between 0 and 255 for each pixel. An adaptive thresholding algorithm is then applied to the grayscale image. This algorithm uses a local window to calculate the threshold to separate the foreground text and stamp areas from the background area. The formula for the adaptive thresholding algorithm can be expressed as:

[0077]

[0078] in: It is a pixel. Adaptive threshold at the location, Based on pixels The average grayscale value of pixels within a local window centered on the target. It is a constant bias. In some embodiments, for each pixel in a grayscale image, its grayscale value is compared with an adaptive threshold. The image is compared, and if the grayscale value is below an adaptive threshold, the pixel is classified as foreground; otherwise, it is classified as background, thus generating a binary image. In some embodiments, perspective distortion of the document's main body area caused by the shooting angle is detected and corrected in the grayscale image. A corner detection algorithm is used to locate the coordinates of the four vertices of the document in the original image; for example, the coordinates of the top left corner vertex are... The coordinates of the top right vertex are The coordinates of the lower left vertex are The coordinates of the bottom right vertex are Map these four vertices to the four corner points of the target rectangle, and set the coordinates of the corner points of the target rectangle as follows: , , , ,in and It uses the preset standard width and height of the document, and applies a perspective transformation matrix to transform the document area in the original image into a regular rectangular area.

[0079] Optionally, edge enhancement processing is performed on the corrected image using a Laplacian operator convolution kernel, such as a 3x3 kernel with element values ​​[0,1,0;1,-4,1;0,1,0]. This convolution operation emphasizes edge features in the image. Optionally, the resolution of the edge-enhanced image is uniformly adjusted to a preset standard resolution of 300 DPI, resulting in a standardized document image as the output. This can be understood as the local window size in the adaptive thresholding algorithm being set to 31 pixels by 31 pixels, with a constant bias. Set to 10. This is understandable, as the corner detection algorithm uses the Shi-Tomasi corner detection method in perspective distortion correction, with the minimum corner quality set to 0.01 and the minimum corner distance set to 10 pixels.

[0080] In one embodiment of the present invention, a detailed description is provided of how to use an optical character recognition (OCR) engine to perform text recognition on standardized document images. A trained OCR model is invoked to perform multi-scale sliding window scanning on the input standardized document image to capture text regions of different sizes. Within each scanning window, the OCR model outputs the recognized text characters, the confidence score for each character, and the precise bounding box coordinates of the character in the original standardized image. The recognition results from all scanning windows are merged, and the lines and paragraphs are reorganized and sorted according to the bounding box coordinates of the text characters, ultimately forming structured text information containing text content, the confidence score of each character or text line, and hierarchical positional information consisting of line numbers and paragraph numbers. The trained OCR model is constructed by acquiring a large-scale dataset of university financial reimbursement document images and performing fine-grained character-level or line-level annotations on the text content in the images to form a training sample set. A deep learning model comprising a feature extraction network, a sequence modeling layer, and a transcription layer is constructed as the initial optical character recognition model. The feature extraction network uses a deep convolutional neural network to extract visual feature sequences from the input image. The sequence modeling layer uses a bidirectional long short-term memory network to model the feature sequences to capture contextual information. The transcription layer uses a connectionist temporal classification decoder to map the sequence features into the final text sequence. Document images from the training sample set are input into the initial model, and the predicted text sequence output by the model is calculated through forward propagation. The difference between the predicted text sequence and the ground truth labeled text sequence is calculated using a connectionist temporal classification loss function, and the network parameters of the model are updated using a backpropagation algorithm. The above forward propagation and parameter update process is iteratively executed until the text recognition accuracy of the model on an independent validation set converges to a preset threshold. The model obtained at this point is the trained optical character recognition model.

[0081] In practice, a trained optical character recognition (OCR) model is invoked to perform text recognition on standardized document images. These standardized document images are grayscale images with a resolution of 300 DPI. The trained OCR model performs multi-scale sliding window scanning on the input document image. The sizes of the multi-scale sliding windows include 128 pixels x 32 pixels, 256 pixels x 64 pixels, and 512 pixels x 128 pixels, with a sliding step size of 16 pixels for each window size. Within each scanning window, the trained OCR model outputs a recognition result tuple. This tuple contains the sequence of recognized text characters within the scanning window, the confidence score for each character in the sequence, and the bounding box coordinates of the smallest bounding rectangle of the character sequence in the original standardized image. The bounding box coordinates are expressed using the coordinates of the top-left corner. and the coordinates of the bottom right corner The process involves merging the recognition results from all scanning windows and reorganizing lines and paragraphs based on the bounding box coordinates of the text characters. The reorganization process determines elements within the same line based on vertical coordinate overlap and elements within the same paragraph based on horizontal coordinate distance and text semantics. This results in structured text information containing complete text content, the confidence score for each character or line, and hierarchical positional information composed of line and paragraph numbers. In multi-scale sliding window scanning, for a 128-pixel by 32-pixel window, the sliding step size is set to 16 pixels in both the horizontal and vertical directions. Similarly, during line reorganization, if the bounding boxes of two characters overlap by more than 60% vertically, they are considered to belong to the same text line.

[0082] In some embodiments, the construction of the trained optical character recognition model involves a dataset and model structure. A large-scale dataset of university financial reimbursement document images is obtained, containing over 100,000 images covering various types of documents such as travel expense reimbursements, invoices, and purchase contracts. Fine-grained character-level position and content annotations are performed on the text content in the images to form a training sample set. A deep learning model comprising a feature extraction network, a sequence modeling layer, and a transcription layer is constructed as the initial optical character recognition model. The feature extraction network uses a deep residual convolutional neural network, with standardized document images as input and a sequence of feature maps as output. The sequence modeling layer uses a three-layer bidirectional long short-term memory network to perform contextual modeling of the feature sequence. The transcription layer uses a connectionist temporal classification decoder to map the sequence features into a text sequence.

[0083] In some embodiments, model training is completed through iterative optimization. Document images from the training sample set are input into the initial optical character recognition model. The predicted text sequence output by the model is calculated through forward propagation. The difference between the predicted text sequence and the labeled text sequence is calculated using a connectionist temporal classification loss function. The formula for the connectionist temporal classification loss function is as follows:

[0084]

[0085] in: This represents the connectionist temporal classification loss value. Represents the training sample set, To represent a training sample, It is an image of the input document. It is the corresponding labeled text sequence. The model is given input The following output sequence The conditional probability is calculated. The network parameters of the initial optical character recognition model are updated using the backpropagation algorithm and the Adam optimizer. The forward propagation and parameter update process is iteratively executed. Training stops when the text recognition accuracy on the independent validation set converges to 98.5%, resulting in the trained optical character recognition model. Optionally, the backbone of the deep residual convolutional neural network uses a ResNet-34 structure, removing the final fully connected layer. Optionally, the number of hidden units in the bidirectional long short-term memory network is set to 256.

[0086] In one embodiment of the present invention, the process of field alignment and entity extraction based on a preset document template knowledge base is described in detail. See also... Figure 3 Based on the matching results of keywords such as "travel expense reimbursement form" and "purchase invoice" in the structured text information, the specific document type of the document to be reviewed is determined. A standard template uniquely corresponding to this document type is retrieved from the document template knowledge base. This standard template defines the names of the standard fields that should be included in this type of document, their logical arrangement on the document, and the expected data type of each field in a structured manner. Each text block in the structured text information is matched with the standard fields in the standard template based on the hierarchical position information of the text blocks and the logical position of the standard fields, thereby aligning the text content recognized by the image to the correct business field. From each aligned text block, based on the expected data type of the field, specific numerical or text entities are extracted using predefined regular expression patterns and named entity recognition technology. For example, the amount "500" is extracted from "RMB 500 yuan", and the date entity is extracted from "October 1, 2023". These entities are then filled into the corresponding fields of the reimbursement information data table.

[0087] In practice, the document type of the document to be reviewed is determined based on the keyword matching results in the structured text information. The structured text information includes text content identified from the document image and hierarchical position information of text blocks. For example, if keywords such as "travel expense reimbursement form," "traveler," and "start and end time" appear in the structured text information, the document type is determined to be "travel expense reimbursement form." In practice, a standard template corresponding to the "travel expense reimbursement form" document type is retrieved from the document template knowledge base. The document template knowledge base is stored in JSON format, and the standard template defines the name, logical position, and expected data type of each standard field in the "travel expense reimbursement form." Partial definitions of the standard template are shown in Table 1.

[0088] Table 1: Standard Template Field Definitions for Travel Expense Reimbursement Forms

[0089]

[0090] In practice, text blocks in structured text information are matched with standard fields in a standard template to achieve field alignment. The matching process considers both the hierarchical positional information of the text blocks and the semantics of the text content. The formula for calculating the similarity between a text block to be matched and a standard field is as follows:

[0091]

[0092] in: This represents the total similarity score. and It is a weighting coefficient and satisfies , This represents the location similarity calculated based on hierarchical location information. This represents the semantic similarity of keywords calculated based on text content and standard field names. Positional similarity. Semantic similarity is obtained by comparing the predicted row and column numbers of the text blocks with the normalized distances of the logical row and column coordinates defined in the standard fields. The similarity is calculated by taking the cosine similarity between the keywords extracted from the text blocks and the text vectors of the standard field names. For each standard field, the total similarity score is selected from all text blocks. The highest-ranking text block is aligned with it. In some embodiments, the weighting factor... Set to 0.6, weighting coefficient Set to 0.4. In some embodiments, the text vectors are obtained through a pre-trained word embedding model, Word2Vec.

[0093] Optionally, specific entities are extracted from each aligned text block according to the expected data type. For fields with the expected data type of "numerical (RMB)", such as the "transportation fee amount" field, the regular expression pattern "\d+(.\d+)?" is used to extract the numerical entity "500.00" from the aligned text block text "RMB Five Hundred Yuan (500.00)". Optionally, for fields with the expected data type of "date range", such as the "start and end time" field, named entity recognition technology is used to identify and extract the date entities "2023-10-01" and "2023-10-05" from the aligned text block text "From October 1st to October 5th, 2023, a total of 5 days". It can be understood that the named entity recognition model adopts a sequence labeling model based on the BERT architecture and is fine-tuned and trained on a financial text dataset. It can be understood that all extracted entities are filled into the corresponding fields of the reimbursement information data table to form a structured record.

[0094] In one embodiment of the present invention, a detailed implementation method is provided for logical consistency verification using an improved anomaly detection algorithm, and for matching data tables against a financial audit rule base item by item. The logical consistency verification process begins by loading a corresponding feature space model from a multi-dimensional feature space model library based on the document type. This model defines multiple feature dimensions and their normal value ranges for describing the key attributes of such documents. Based on the entities in the initial reimbursement information data table, the actual feature values ​​of the document on each feature dimension of the feature space model are calculated. For each feature dimension, the actual feature value is compared with the historical normal value range of that feature dimension obtained from historical data statistics, and the degree of deviation is calculated. According to the preset correlation between feature dimensions, the judgment thresholds of each feature dimension are dynamically adjusted. For example, when the value of associated feature A is abnormal, the judgment threshold of feature B will be tightened. This dynamic adjustment is based on the strength of the feature correlation, performing weighted corrections in the same or opposite directions.

[0095] If the actual feature value of a certain feature dimension exceeds the dynamically adjusted judgment threshold, the data corresponding to that feature dimension is marked as a logical anomaly, and all logical anomalies are summarized into a logical consistency verification report. The working principle of the dynamic threshold judgment includes: constructing a multi-dimensional feature space for each document type, spanned by a series of orthogonal or non-orthogonal feature vectors, each representing an independent business or logical feature dimension. From the historically reviewed document database, feature values ​​of similar documents on each feature dimension are extracted to form a historical feature value distribution. Using a kernel density estimation algorithm, nonparametric modeling is performed on the historical feature value distribution of each feature dimension to obtain the probability density function of that feature dimension. The interval boundary of the probability density function at a preset confidence level is used as the benchmark judgment threshold for that feature dimension. During runtime, when jointly judging multiple feature dimensions of a document, other feature dimensions that are strongly correlated with the currently judged feature dimension are searched according to a preset feature dimension association graph.

[0096] The system reads the current actual feature values ​​of these strongly correlated feature dimensions. If the deviation of the actual feature value from its respective benchmark judgment threshold exceeds the preset correlation influence factor, the benchmark judgment threshold of the currently judged feature dimension is contracted or expanded according to preset rules to obtain a dynamically adjusted judgment threshold. In the rule matching phase, all audit rules matching the current document type and funding source project are loaded from the financial audit rule base. Each audit rule contains a conditional expression and a violation conclusion. The system iterates through each field and its value in the reimbursement information data table. The value of each field is substituted into the conditional expression of the relevant audit rule for logical calculation. If the conditional expression is true, the violation conclusion corresponding to the audit rule is triggered, and the rule number, the triggered field, and the specific value are recorded as a potential rule conflict point.

[0097] In implementation, an improved anomaly detection algorithm is used to perform logical consistency checks on the initial reimbursement information data table. This table contains structured field information extracted from "Travel Expense Reimbursement Forms." Based on the document type "Travel Expense Reimbursement Form," the corresponding feature space model is loaded from a multi-dimensional feature space model library. This model defines multiple feature dimensions describing the key attributes of the travel expense reimbursement form and their normal value ranges. In practice, actual feature values ​​are calculated based on the entities in the initial reimbursement information data table. For example, the "Number of Days on Business Trip" feature value is calculated from the "Start and End Time" field; the "Total Expenses" feature value is calculated from fields such as "Transportation Amount," "Accommodation Amount," and "Meal Allowance"; and the "Average Amount per Document" feature value is calculated from the "Number of Documents" field. For each feature dimension, the actual feature value is compared with the historical normal value range of that feature dimension, and the deviation is calculated using the following formula:

[0098]

[0099] in: Indicates the degree of deviation. Represents the actual eigenvalues. This represents the center value of the historical normal range. This represents the width of the historical normal value range (i.e., the difference between the maximum and minimum values). The judgment thresholds for each feature dimension are dynamically adjusted based on preset correlations between them. For example, there is a strong positive correlation between "number of business trip days" and "accommodation cost." If the actual value of "number of business trip days" is close to its historical upper limit, the dynamic adjustment algorithm will correspondingly raise the upper limit of the judgment threshold for the "accommodation cost" dimension. If the actual feature value of a certain feature dimension exceeds the dynamically adjusted judgment threshold, the data corresponding to that feature dimension is marked as a logical outlier, and all logical outliers are summarized in a logical consistency verification report.

[0100] In some embodiments, the improved anomaly detection algorithm performs dynamic threshold determination based on the multi-dimensional feature space of the document type. A multi-dimensional feature space is constructed for the "travel expense reimbursement form," which consists of multiple feature dimensions such as the number of business trip days, total expense, average amount per receipt, and city consumption level. Ten thousand approved "travel expense reimbursement forms" are extracted from the historical approved document database, and the feature values ​​of each form on each feature dimension are calculated to form a historical feature value distribution. A kernel density estimation algorithm is used to model the historical feature value distribution of each feature dimension to obtain a probability density function. The interval boundary of the probability density function at a 95% confidence level is used as the benchmark determination threshold for the feature dimension. During runtime, when jointly determining multiple feature dimensions of a document, strongly correlated dimensions are found according to a preset feature dimension correlation graph. The feature dimension correlation graph defines the correlation and strength between dimensions. See Table 2 for some relationships.

[0101] Table 2: Relationship between Some Feature Dimensions of Travel Expense Reimbursement Forms

[0102]

[0103] The system reads the current actual feature values ​​of strongly correlated feature dimensions. If the deviation of the actual feature value from its respective baseline judgment threshold exceeds the preset correlation influence factor, the baseline judgment threshold of the currently judged feature dimension is contracted or expanded according to preset rules to obtain a dynamically adjusted judgment threshold. For example, when judging the "total accommodation cost" dimension, the system searches for the "number of business trip days" dimension, which is strongly positively correlated with it. If the actual value of "number of business trip days" exceeds 20% of its baseline judgment threshold upper limit, and the correlation influence factor is set to 0.5, then the upper limit of the judgment threshold for the "total accommodation cost" dimension will be increased by 10% proportionally.

[0104] Optionally, load all audit rules matching the document type and funding source from the financial audit rule base. For example, there are 15 rules matching "Travel Expense Reimbursement Form" with funding source "Vertical Scientific Research Project (National Level)". Iterate through each field and its value in the reimbursement information data table. For example, the value of the field "Transportation Class" is "Economy Class Airplane", and the value of the field "Accommodation Fee Standard" is "500 RMB / day". Substitute the value of each field into the conditional expression of the relevant audit rule for logical calculation. For example, the conditional expression of a rule is: [Transportation Class] == "First Class Airplane" AND [Funding Source] == "Vertical Scientific Research Project (National Level)". If the conditional expression evaluates to true, the corresponding audit rule's violation conclusion is triggered. For example, the violation conclusion is "National-level scientific research projects are prohibited from reimbursing first class airplane tickets". Record the rule number, the triggered field, and its value as potential rule conflict points. It is understandable that the financial audit rule base is stored in the form of database tables, with each rule containing a rule number, applicable document type, applicable project type, conditional expression logical string, and a description of the violation conclusion. It is also understandable that the calculation of the conditional expression is completed through an embedded script engine that parses and executes field values ​​and logical operators.

[0105] In one embodiment of the present invention, a similarity review process based on historical decision-making cases and the generation of a final review report are described in detail. For identified potential rule conflicts, a similarity review process is initiated, constructing a review query vector using the rule number, field name, and value involved in the conflict as key information. Historical decision-making case databases are then searched for historical cases similar to the review query vector in terms of rules and fields.

[0106] Calculate the comprehensive similarity between the current case and each retrieved historical case in terms of specific values ​​and contextual information. Select one or more historical cases whose comprehensive similarity exceeds a set threshold, and extract the final decision results and additional annotations of the human auditors in these cases to form review recommendations. When generating review recommendations, statistically analyze the final decision results of the extracted historical decision cases and calculate the frequency of different decision results. The decision result with the highest frequency is used as the basic review recommendation. If historical cases have similar frequencies of different decision results and conflicting decisions, further analyze the additional annotations of these conflicting cases, extract the key considerations affecting the decision, and add these considerations as supplementary explanations to the basic review recommendations to form a complete review recommendation. When generating the final audit report, summarize all logical anomalies in the logical consistency verification report and all potential rule conflicts identified in the financial audit rule matching, along with their corresponding review recommendations. If both logical anomalies and potential rule conflicts are empty, a conclusion of audit approval is generated.

[0107] If logical anomalies or potential rule conflicts are found, an audit conclusion will be generated based on the type and severity of the review suggestions, including "suggest resubmitting after modification," "suggest approval with explanation," or "audit rejected." The audit conclusions, a detailed list of logical anomalies, a list of potential rule conflicts, review suggestions for each conflict, and an index of the relevant original voucher image areas will be integrated, formatted, and output as a clearly structured final audit report document.

[0108] In implementation, a similarity review process based on historical decision cases is initiated for identified potential rule conflict points. Taking a triggered rule as an example, its rule number is "Rule_2023_Travel_001", the conflicting field is "Transportation Class", and the field value is "Economy Class". In implementation, a review query vector is constructed using the rule number, field name, and value involved in the potential rule conflict point as key information. The review query vector adopts a dictionary structure, for example, {"Rule Number":"Rule_2023_Travel_001", "Field Name":"Transportation Class", "Field Value":"Economy Class"}. In the historical decision case database, historical cases similar to the review query vector in terms of rules and fields are retrieved. The historical decision case database stores all past document records that have undergone manual review and contain clear review conclusions and annotations, along with their triggered rule information. The comprehensive similarity between the current case and each retrieved historical case is calculated in terms of specific values ​​and contextual information. The comprehensive similarity calculation formula is:

[0109]

[0110] in: This represents the overall similarity score. and It is a weighting coefficient and satisfies , Indicates similarity across rule dimensions. This represents the similarity between numerical values ​​and the context dimension. (Rule-based similarity) The value is 1 when the rule numbers are exactly the same, 0.7 when the rule numbers belong to the same parent rule class, and 0 otherwise. Numerical similarity with context dimension. The similarity score is calculated by comparing the matching degree between the document containing the current conflict and historical case documents in multiple contextual fields such as project type, funding amount, and business trip destination. For historical cases that exceed the set threshold, extract the final decision results and additional annotations recorded by human auditors in these historical cases to form review suggestions for potential rule conflicts.

[0111] In some embodiments, the historical decision case database establishes a composite index based on rule number and field name to accelerate retrieval. The final decision results of the extracted historical decision cases mainly include three categories: "approved," "approved after modification," and "not approved." The final decision results of the extracted historical decision cases are statistically analyzed to calculate the frequency of different decision results in the set of similar cases. The decision result with the highest frequency is used as the basic review suggestion. For example, if among the five highly similar historical cases retrieved, four cases have a decision result of "approved after modification" and one case has a decision result of "not approved," then the basic review suggestion is "approved after modification."

[0112] In some embodiments, if historical cases with similar frequencies of different decision outcomes and conflicting decision-making outcomes are analyzed, the additional annotations of the conflicting cases will be further analyzed to extract key decision-making considerations. For example, in the annotations for cases that are "approved after modification," "with explanation attached" frequently appears, while in the annotations for cases that are "not approved," "insufficient basis for exceeding the standard" is clearly marked. The key consideration extracted is then "whether sufficient explanation of the reasons for exceeding the standard is provided." This consideration is added as supplementary explanation to the basic review recommendation to constitute a complete review recommendation. Optionally, a weighting coefficient... Set to 0.6, weighting coefficient Set to 0.4 to further emphasize consistency across the rule dimension. Optional, context-dimensional similarity. The calculation employs a weighted scoring method based on exact field value matching and numerical range matching.

[0113] It is understandable that each case record in the historical decision-making case database contains a complete snapshot of the reimbursement information data table, a list of all triggered rules, manual review conclusions, review comments, and processing timestamps. It is also understandable that when calculating the similarity between numerical values ​​and contextual dimensions, the matching of numerical fields such as "funding amount" is determined by comparing whether the values ​​fall within the same range. All logical anomalies in the summary logical consistency verification report, as well as all potential rule conflicts identified in the financial review rule matching and their corresponding review suggestions, are included. Logical anomaly descriptions include, for example, "3 days of business trip, total accommodation fee of 4000 yuan, exceeding the common proportion range," and potential rule conflict descriptions include, for example, "Rule_2023_Travel_001: Transportation class is economy class airfare, violating national project travel standards." If both logical anomalies and potential rule conflicts are empty, a conclusion of approval is generated.

[0114] If logical anomalies or potential rule conflicts exist, an audit conclusion is generated based on the type and severity of the review suggestion. For example, for a review suggestion of "Approved after modification" with a minor conflict, a conclusion of "Approved with explanation" is generated. For a review suggestion of "Review Failed" or a minor conflict, a conclusion of "Review Failed" or "Resubmitted after modification" is generated. The audit conclusion, a detailed list of logical anomalies, a list of potential rule conflicts, review suggestions for each conflict, and the original key voucher image area index are integrated, formatted, and output as the final audit report document. The report document uses a structured text format.

[0115] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An automatic verification method for university financial reimbursement documents integrating OCR, characterized in that, include: The original images of the financial reimbursement documents to be reviewed are acquired using an image acquisition device. The original image is preprocessed to obtain a standardized document image; The standardized document image is subjected to text recognition using an optical character recognition engine to extract structured text information containing text content and coordinate positions; Based on a pre-defined document template knowledge base, the structured text information is aligned with fields and entities are extracted to generate a preliminary reimbursement information data table. An improved anomaly detection algorithm is used to perform logical consistency verification on the preliminary reimbursement information data table. The improved anomaly detection algorithm performs dynamic threshold determination based on the multi-dimensional feature space of the document type. The reimbursement information data table, which has undergone logical consistency verification, is matched against the preset financial audit rule base one by one to identify potential rule conflict points. For the identified potential rule conflicts, a similarity review process based on historical decision-making cases is initiated to generate review recommendations; Based on the matching results between the review recommendations and the financial audit rule base, a final audit report is generated, which includes the audit conclusions, conflict details, and recommendations.

2. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 1, characterized in that, The original image is preprocessed to obtain a standardized document image, including: Perform a grayscale conversion on the original image to obtain a grayscale image; An adaptive threshold segmentation algorithm is applied to the grayscale image to separate the foreground text, stamp area and background area; The perspective distortion of the document's main body area caused by the shooting angle in the grayscale image is detected and corrected so that the document's main body area appears as a regular rectangle in the image. The corrected image undergoes edge enhancement processing, and its resolution is uniformly adjusted to a preset standard resolution, outputting a standardized document image.

3. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 1, characterized in that, Text recognition of the standardized document image is performed using an optical character recognition engine, including: The trained optical character recognition model is invoked to perform multi-scale sliding window scanning on standardized document images; Within each scanning window, the optical character recognition model outputs the recognized text character, the corresponding confidence level, and the bounding box coordinates of the text character in the original image. The recognition results of all scanning windows are merged, and the lines and paragraphs are reorganized according to the bounding box coordinates of the text characters to form structured text information containing text content, confidence level, and hierarchical positional information consisting of line numbers and paragraph numbers. The trained optical character recognition model is constructed as follows: Obtain a large-scale dataset of images of financial reimbursement documents from universities, and perform detailed annotation on the text content in the images to form a training sample set; A deep learning model comprising a feature extraction network, a sequence modeling layer, and a transcription layer is constructed as the initial optical character recognition model. The feature extraction network uses a convolutional neural network to extract visual features from the input image. The sequence modeling layer uses a bidirectional long short-term memory network to capture the contextual dependencies of the features. The transcription layer uses a connectionist temporal classification decoder to map the sequence features into a text sequence. Input the document images from the training sample set into the initial optical character recognition model, and calculate the predicted text sequence output by the model through forward propagation; The difference between the predicted text sequence and the labeled text sequence is calculated using the connectionist temporal classification loss function, and the network parameters of the initial optical character recognition model are updated using the backpropagation algorithm. The forward propagation and parameter update process is iteratively executed until the text recognition accuracy of the model on the independent validation set converges to a preset threshold, thus obtaining the trained optical character recognition model.

4. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 1, characterized in that, Based on a pre-defined document template knowledge base, field alignment and entity extraction are performed on the structured text information, including: Based on the keyword matching results in the structured text information, determine the document type of the document to be reviewed; Retrieve the standard template corresponding to the document type from the document template knowledge base. The standard template defines the name, logical position, and expected data type of each standard field. By matching text blocks in structured text information with standard fields in standard templates based on the hierarchical and logical positions of the text blocks, field alignment is achieved. From each aligned text block, based on the expected data type, specific numerical or text entities are extracted using regular expressions and named entity recognition technology, and then populated into the corresponding fields of the reimbursement information data table.

5. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 1, characterized in that, The step of using an improved anomaly detection algorithm to perform logical consistency verification on the preliminary reimbursement information data table includes: Based on the document type, the corresponding feature space model is loaded from the multi-dimensional feature space model library. The feature space model defines multiple feature dimensions and their normal value ranges for describing the key attributes of the corresponding category of documents. Based on the entities in the preliminary reimbursement information data table, calculate their actual feature values ​​in each feature dimension of the feature space model. For each feature dimension, the actual feature value is compared with the historical normal range of the feature dimension, and the deviation is calculated. Based on the preset correlation between feature dimensions, the judgment threshold of each feature dimension is dynamically adjusted. The dynamically adjusted judgment threshold is weighted and corrected in the same or opposite direction based on the strength of the feature correlation. If the actual feature value of a certain feature dimension exceeds the dynamically adjusted judgment threshold, the data corresponding to the feature dimension is marked as a logical anomaly, and all logical anomalies are summarized into a logical consistency verification report.

6. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 5, characterized in that, The improved anomaly detection algorithm performs dynamic threshold determination based on the multi-dimensional feature space of document types. Its working principle includes: A multi-dimensional feature space is constructed for each document type. The multi-dimensional feature space is spanned by a series of orthogonal or non-orthogonal feature vectors, and each feature vector represents an independent business or logical feature dimension. Extract the feature values ​​of similar documents across various feature dimensions from the historical database of audited documents to form a historical feature value distribution; By using the kernel density estimation algorithm, the historical feature value distribution of each feature dimension is modeled to obtain the probability density function of the feature dimension; The interval boundary of the probability density function at a preset confidence level is used as the benchmark judgment threshold for the feature dimension. During runtime, when performing a joint judgment on multiple feature dimensions of a document, the system searches for other feature dimensions that are strongly correlated with the currently judged feature dimension based on a preset feature dimension association diagram. Read the current actual feature value of the strongly correlated feature dimension. If the actual feature value deviates from its respective benchmark judgment threshold by more than the preset correlation influence factor, then shrink or expand the benchmark judgment threshold of the current judged feature dimension according to the preset rules to obtain the dynamically adjusted judgment threshold.

7. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 1, characterized in that, The step of matching the reimbursement information data table, which has undergone logical consistency verification, with the preset financial audit rule base one by one includes: From the financial audit rule base, load all audit rules that match the document type and funding source project, wherein the audit rules include conditional expressions and violation conclusions; Iterate through each field and its value in the expense reimbursement information data table; Substitute the value of each field into the conditional expression of the relevant review rules for logical calculation; If the conditional expression evaluates to true, the corresponding violation conclusion of the audit rule is triggered, and the rule number of the violation, the triggered field and its value are recorded as potential rule conflict points.

8. The method for automatic verification of university financial reimbursement documents integrating OCR according to claim 1, characterized in that, The process of initiating a similarity review based on historical decision-making cases for the identified potential rule conflict points includes: A verification query vector is constructed using the rule number, field name, and value involved in potential rule conflict points as key information. In the historical decision case database, retrieve historical cases that are similar to the review query vector in terms of rules and fields; Calculate the overall similarity between the current case and each retrieved historical case in terms of specific numerical values ​​and contextual information; Select historical cases with a comprehensive similarity exceeding a set threshold, extract the final decision results and additional annotations of human reviewers in the historical cases, and form review suggestions.

9. The method for automatic verification of university financial reimbursement documents integrating OCR as described in claim 8, characterized in that, The generation of review recommendations includes: The final decision results of the extracted historical decision cases are statistically analyzed, and the frequency of occurrence of different decision results is calculated. Use the most frequently occurring decision outcomes as the basis for review recommendations; If there are historical cases where different decision outcomes have similar frequencies and there are decision-making conflicts, the additional annotations of the conflict cases will be further analyzed to extract the key decision-making considerations. These considerations will then be added as supplementary explanations after the basic review recommendations to form a complete review recommendation.

10. The method for automatic verification of university financial reimbursement documents integrating OCR according to claim 1, characterized in that, The generation of the final audit report, which includes audit conclusions, conflict details, and recommendations, includes: Summarize all logical anomalies in the logical consistency verification report, as well as all potential rule conflicts identified in the financial audit rule matching and their corresponding review recommendations; If both logical anomalies and potential rule conflicts are empty, a conclusion of approval is generated. If there are logical anomalies or potential rule conflicts, the review conclusion will be generated based on the type and severity of the review suggestion, including "suggest resubmitting after modification", "suggest approval with explanation" or "review not approved". The audit conclusions, a detailed list of logical anomalies, a list of potential rule conflicts, corresponding review suggestions, and the original key voucher image area index are integrated, formatted, and output as the final audit report document.

Images