AI intelligent single auditing-based graphic character recognition method and system

Abstract

The application discloses a graphomotoric character recognition method and system based on AI intelligent single auditing, and relates to the field of character recognition.The method comprises the following steps: calculating image features by image pixel intensity and local noise, and generating self-adaptive enhanced images; identifying effective areas in combination with regional weight, local expansion index and geometric distortion; performing time sequence feature analysis on candidate text areas to generate character sequences; correcting character confidence values and extracting key fields; calculating nonlinear verification to form overall single verification values; and generating self-adaptive optimization signals according to the verification values and feedback errors, and dynamically updating recognition models and field extraction models.Through self-adaptive image enhancement, candidate area identification, character confidence correction, key field extraction and logical verification, high-precision graphomotoric character recognition is realized, and dynamic optimization and self-adaptive updating of the model are supported.

Description

Technical Field

[0001] This invention relates to the field of character recognition, specifically to a method and system for recognizing graphic characters based on AI-powered intelligent order review. Background Technology

[0002] With the widespread adoption of electronic office systems and digital document processing, businesses generate a large amount of document data in their daily operations, including invoices, contracts, and expense reports. These documents typically contain various information types, such as images, tables, text, and symbols. Traditional manual document review methods suffer from low efficiency, error-proneness, and high costs, while existing automated recognition methods still face technical bottlenecks when processing complex documents, including low recognition rates, inaccurate field extraction, insufficient logical validation, and the inability of models to adapt and adjust.

[0003] In existing technologies, image and text character recognition mainly relies on Optical Character Recognition (OCR) technology, which extracts characters through single image features. However, simple OCR methods are prone to character recognition errors when processing low-quality images, documents with noise interference, image distortion, or diverse fonts, leading to inaccurate field extraction results. Existing methods typically lack full utilization of the context and linguistic information of the recognition results, cannot effectively correct recognition confidence, and struggle to extract key fields and perform logical consistency checks.

[0004] Furthermore, existing systems cannot dynamically and adaptively optimize the recognition and field extraction models during the document review process, resulting in inconsistent performance when faced with documents from different sources or in diverse formats. Especially in large-scale, diverse document environments, existing methods cannot adjust model parameters promptly based on review results, thus limiting recognition accuracy and extraction precision.

[0005] Therefore, there is an urgent need for an AI-based image and text character recognition method and system that can achieve high-precision processing in stages such as image feature enhancement, candidate region recognition, character sequence recognition, confidence correction, key field extraction, and logical verification, and can adaptively optimize the recognition model and field extraction model through overall order review feedback to improve order review efficiency and accuracy. Summary of the Invention

[0006] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a method and system for image and text character recognition based on AI intelligent order review, so as to solve the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a graphic character recognition method based on AI-powered intelligent order review, comprising:

[0008] Image features are calculated by analyzing the pixel intensity and local noise perturbation of the document image to be reviewed. Based on these image features, an adaptive enhanced image is generated by combining the enhancement amplitude factor, nonlinear exponent, and interference suppression coefficient.

[0009] Based on adaptive image enhancement and region weights, local expansion index and geometric distortion metric function, the region response values ​​of candidate text, table and symbol regions are generated to identify the effective regions to be processed;

[0010] A nonlinear recognition representation is calculated by using the temporal feature response of candidate text regions and cross-time interference measurement, and a character sequence recognition result is generated based on this nonlinear representation.

[0011] A corrected character confidence value is generated based on the character sequence recognition results, language model weights, candidate word matching coefficients, and context consistency errors.

[0012] Key field values ​​are generated using the corrected character confidence value, field nonlinear expansion index, redundancy metric function, and extraction ratio coefficient.

[0013] The non-linear verification representation is calculated based on the key field values ​​and the logical deviation measurement between fields, and the overall audit verification value is formed based on the non-linear verification representation.

[0014] By generating adaptive optimization signals for the model through overall order verification values ​​and feedback error metrics, the identification model and field extraction model are dynamically updated.

[0015] The present invention is further configured such that the step of calculating image features from the image of the document to be reviewed and generating an adaptive enhanced image includes:

[0016] The pixel intensity of the document image to be reviewed is normalized to obtain a normalized image with a uniform numerical range.

[0017] Local noise response is calculated based on normalized images, and local noise perturbation is generated by the difference amplitude between each pixel in the image and its neighboring pixels;

[0018] Image features are calculated by combining normalized image pixel intensity and local noise perturbation, and the influence of local noise on image features is adjusted by the noise suppression coefficient.

[0019] Image features are processed by combining a nonlinear enhancement function with an enhancement amplitude factor, a nonlinear exponent, and an interference suppression coefficient to form an adaptive enhanced image.

[0020] The present invention is further configured such that the generation of regional response values ​​for candidate text, tables, and symbol regions, and the identification of valid regions to be processed, include:

[0021] The adaptively enhanced image is divided into several candidate regions, each of which consists of a set of consecutive pixels.

[0022] Region weights are generated based on the pixel intensity distribution of the candidate region, local expansion indexes are generated based on the local gradient changes of the candidate region, and geometric distortion measures are generated based on the geometric shape offset of the candidate region.

[0023] The regional response value is generated through a nonlinear combination of regional weights, local spread index, and geometric distortion measure.

[0024] Based on the comparison between the regional response value and the set threshold, candidate regions that meet the conditions are filtered to form candidate text regions, candidate table regions, and candidate symbol regions;

[0025] The candidate regions obtained from the screening are determined as valid regions to be processed.

[0026] The present invention is further configured such that the calculation of the nonlinear recognition representation and the generation of character sequence recognition results based on the nonlinear representation include:

[0027] The candidate text regions are expanded in the order from left to right and from top to bottom in the image to form a temporal sequence;

[0028] The temporal feature response is calculated based on the pixel features of adjacent regions in the temporal sequence. The features of the neighboring regions are weighted and superimposed by the convolution kernel weights to obtain the feature values ​​at each time step.

[0029] Based on the difference between the feature values ​​at each time point and the feature values ​​at other times in the entire sequence range, a cross-time-time interference metric is generated by combining the position difference attenuation weight.

[0030] By combining the temporal feature response with cross-time interference measurement in a nonlinear manner, a recognition representation that can characterize the temporal pattern of a character is generated.

[0031] By performing sequence decoding on the recognition representation, a character sequence recognition result is formed.

[0032] The present invention is further configured such that generating the corrected character confidence value includes:

[0033] The character sequence recognition result is represented as the initial recognition probability of each character;

[0034] By combining the language model weights, the probability correction value of each character under its context is calculated;

[0035] For each character, a set of candidate words is retrieved, and the candidate word matching coefficient is calculated, which represents the degree of matching between the character and the candidate word;

[0036] Context consistency error is generated based on the difference between the joint probability and independent probability of a character within a local context, combined with positional weights.

[0037] The initial character recognition probability, language model probability correction value, candidate word matching coefficient, and context consistency error are combined in a non-linear manner to generate the corrected character confidence value.

[0038] The present invention is further configured such that generating key field values ​​using the corrected character confidence values ​​includes:

[0039] The corrected character confidence values ​​are combined with the text sequence to identify candidate character subsequences for each field, forming a candidate field set.

[0040] The character confidence values ​​of candidate fields are non-linearly expanded to enhance the distinguishability of low-confidence characters in field values, and a field non-linear expansion index is generated.

[0041] By combining the redundancy information of candidate fields in the document, the values ​​of non-linear extended fields are corrected through a redundancy measurement function to generate redundancy correction field values;

[0042] Based on the document length and field importance, the extraction ratio coefficient is set, and the redundant correction field values ​​are weighted to generate key field values;

[0043] Map key field values ​​to predefined field names or field locations to form a complete set of key fields.

[0044] The present invention is further configured such that the calculation of the nonlinear verification representation based on the key field values ​​to form the overall audit verification value includes:

[0045] Organize the set of key field values ​​into a key field vector;

[0046] Calculate the logical deviation measure between each field based on the key field vector to represent the deviation of the expected relationship between the fields;

[0047] The key field values ​​and the logical deviation measures between fields are combined in a non-linear manner to generate a verification representation that indicates the internal consistency of the audit.

[0048] The nonlinear verification representation is normalized to form the overall audit verification value.

[0049] The present invention is further configured such that the step of generating the model adaptive optimization signal through the overall order review verification value and feedback error metric includes:

[0050] The overall order review verification value is compared with the preset ideal order review verification standard, and the feedback error metric is calculated.

[0051] The feedback error metric is nonlinearly combined with the optimization signal adjustment coefficient to generate an adaptive optimization signal for the recognition model and the field extraction model.

[0052] The optimized signal is normalized so that the signal value falls within the adjustable range;

[0053] The update direction and magnitude of the recognition model and field extraction model are determined based on the value and intensity of the optimized signal, providing a benchmark for adaptive adjustment of the model.

[0054] The present invention is further configured such that the dynamic updating of the recognition model and the field extraction model includes:

[0055] The adaptive optimization signal is applied to the update of the recognition model parameters, and the weights of the recognition model are adjusted by parameter increment;

[0056] The adaptive optimization signal is applied to adjust the parameters of the field extraction model, and the field extraction rules or weights are corrected according to the proportional coefficient.

[0057] The update process of the identification model and field extraction model is iterated, and the parameters are continuously adjusted based on the feedback error metric and the overall order verification value until the error converges or the preset iteration round is reached.

[0058] The updated recognition model and field extraction model will be used as the benchmark model for the next round of document processing to ensure that the model dynamically adapts to changes in document features during the document review task.

[0059] This invention also provides an AI-based intelligent order review system for image and text character recognition, the system comprising:

[0060] Image feature enhancement module: Calculates image features by analyzing pixel intensity and local noise perturbation in the image of the document to be reviewed, and generates an adaptive enhanced image based on the image features combined with enhancement amplitude factor, nonlinear exponent and interference suppression coefficient;

[0061] Region Candidate Generation and Response Module: Generates region response values ​​for candidate text, table, and symbol regions based on adaptive enhanced images and region weights, local expansion index, and geometric distortion metric function, and identifies valid regions to be processed;

[0062] Character sequence recognition module: Calculates a nonlinear recognition representation based on the temporal feature response of candidate text regions and cross-time interference measurement, and generates character sequence recognition results based on this nonlinear representation;

[0063] Character confidence value correction module: Generates corrected character confidence values ​​based on character sequence recognition results, language model weights, candidate word matching coefficients, and context consistency errors;

[0064] Key field generation module: Generates key field values ​​using corrected character confidence values, field non-linear expansion index, redundancy measurement function, and extraction ratio coefficient;

[0065] The document review and verification module calculates a non-linear verification representation based on key field values ​​and logical deviation measurements between fields, and forms an overall document review and verification value based on the non-linear verification representation.

[0066] Adaptive optimization and update module: Generates adaptive optimization signals for the model by using overall order verification values ​​and feedback error metrics, and dynamically updates the identification model and field extraction model.

[0067] This invention provides an AI-based intelligent document review method and system for image and text character recognition. The method calculates image features based on pixel intensity and local noise perturbation of the document image to be reviewed. Based on these image features, it generates an adaptive enhanced image using an enhancement amplitude factor, a nonlinear exponent, and an interference suppression coefficient. Based on the adaptive enhanced image and region weights, a local expansion exponent, and a geometric distortion metric function, it generates region response values ​​for candidate text, table, and symbol regions, identifying the effective regions to be processed. It calculates a nonlinear recognition representation using the temporal feature response of the candidate text regions and cross-time interference metrics, generating character sequence recognition results based on this nonlinear representation. Based on the character sequence recognition results, language model weights, candidate word matching coefficients, and contextual consistency error, it generates corrected character confidence values. It generates key field values ​​using the corrected character confidence values, field nonlinear expansion exponent, redundancy metric function, and extraction ratio coefficient. It calculates a nonlinear verification representation based on the key field values ​​and inter-field logical deviation metric, forming an overall document review verification value. Finally, it generates an adaptive optimization signal for the model using the overall document review verification value and feedback error metric, dynamically updating the recognition model and field extraction model. The resulting benefits include:

[0068] Improve the feature extraction capability and character recognition accuracy of document images: By performing pixel intensity normalization, local noise perturbation calculation and adaptive enhancement processing on the document images to be reviewed, and combining regional weights, local expansion index and geometric distortion measurement to generate candidate regions, accurate recognition of text, table and symbol regions can be achieved, thereby improving the accuracy and completeness of character sequence recognition;

[0069] Improve the reliability of field extraction and logical verification: Generate key fields by correcting character confidence values, non-linear extended field values ​​and redundant measurement functions, and calculate non-linear verification representation based on logical deviations between fields to form an overall document review verification value. This enables accurate extraction of key fields of documents and verification of document review logic consistency, reducing the error recognition rate and field extraction error.

[0070] Supports adaptive optimization and dynamic updates of the model: Adaptive optimization signals are generated by the overall document review verification value and feedback error measurement, and the recognition model and field extraction model are dynamically iterated and updated. This enables the model to be continuously optimized in diverse document environments, improves system processing efficiency and adaptability, and ensures stable recognition and extraction performance in documents of different formats and qualities.

[0071] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0072] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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. In the drawings:

[0073] Figure 1 A flowchart illustrating an AI-based intelligent order review method for recognizing graphic characters, as shown in an exemplary embodiment of the present invention;

[0074] Figure 2 This is a schematic diagram illustrating the structure of an AI-based intelligent document review image and text character recognition system, which is an exemplary embodiment of the present invention. Detailed Implementation

[0075] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0076] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0077] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0078] Example 1:

[0079] AI-based intelligent order review methods for image and text character recognition, such as Figure 1 As shown, it includes:

[0080] Image features are calculated by analyzing the pixel intensity and local noise perturbation of the document image to be reviewed. Based on these image features, an adaptive enhanced image is generated by combining the enhancement amplitude factor, nonlinear exponent, and interference suppression coefficient.

[0081] Based on adaptive image enhancement and region weights, local expansion index and geometric distortion metric function, the region response values ​​of candidate text, table and symbol regions are generated to identify the effective regions to be processed;

[0082] A nonlinear recognition representation is calculated by using the temporal feature response of candidate text regions and cross-time interference measurement, and a character sequence recognition result is generated based on this nonlinear representation.

[0083] A corrected character confidence value is generated based on the character sequence recognition results, language model weights, candidate word matching coefficients, and context consistency errors.

[0084] Key field values ​​are generated using the corrected character confidence value, field nonlinear expansion index, redundancy metric function, and extraction ratio coefficient.

[0085] The non-linear verification representation is calculated based on the key field values ​​and the logical deviation measurement between fields, and the overall audit verification value is formed based on the non-linear verification representation.

[0086] By generating adaptive optimization signals for the model through overall order verification values ​​and feedback error metrics, the identification model and field extraction model are dynamically updated.

[0087] The present invention is further configured such that the step of calculating image features from the image of the document to be reviewed and generating an adaptive enhanced image includes:

[0088] The pixel intensity of the document image to be reviewed is normalized to obtain a normalized image with a uniform numerical range. Specifically, the pixel intensity range of the document image to be reviewed is unified by scaling the gray values ​​of different pixels in the original image according to the maximum and minimum values, so that the values ​​of all pixels are mapped to a uniform standard range, thereby eliminating the problem of numerical inconsistency between different images due to differences in acquisition conditions, and obtaining a normalized image.

[0089] The local noise response is calculated based on a normalized image. Local noise perturbation is generated by measuring the difference between each pixel and its neighboring pixels. Specifically, for each pixel in the normalized image, its neighboring pixel set is extracted, the grayscale difference between the current pixel and its neighbors is calculated, and the average of these differences is used to characterize the local perturbation intensity of that pixel. If the difference value is large, it indicates the presence of noise or texture fluctuations in the region; if the difference value is small, it indicates that the region is smooth and stable, generating a local noise perturbation value for each pixel.

[0090] Image features are calculated by combining normalized image pixel intensity and local noise perturbation. The impact of local noise on image features is adjusted using a noise suppression coefficient. Specifically, a noise suppression coefficient is set based on the overall noise level of the image to adjust the local noise perturbation value of each pixel. The noise suppression coefficient can be determined as follows: Analyze the distribution of local noise perturbation values ​​across the entire image, and take the median or percentile value as a reference. Map the reference to the range of noise suppression coefficient values, for example, 0.1 to 0.9. The higher the noise reference, the larger the coefficient value, indicating stronger suppression; the lower the noise reference, the smaller the coefficient value, preserving more details. The local noise perturbation value of each pixel is adjusted in conjunction with the noise suppression coefficient: the original local noise perturbation value is multiplied or non-linearly scaled with the noise suppression coefficient to reduce the perturbation value in high-noise areas and keep the perturbation value in low-noise areas unchanged or slightly enhanced. This reduces the impact of noise areas on the enhancement result while preserving details in key areas such as text and tables. The normalized pixel intensity is combined with the adjusted local noise perturbation value to form the image features of each pixel.

[0091] An adaptive enhanced image is generated by processing image features using a nonlinear enhancement function combined with an enhancement amplitude factor, a nonlinear exponent, and an interference suppression coefficient. Specifically, a nonlinear enhancement function is introduced based on the image features to highlight key areas and suppress interference areas. The enhancement amplitude factor is set according to the image contrast level. When the overall contrast is low, a larger value is used to increase the difference between bright and dark areas; when the contrast is high, a smaller value is used to avoid over-enhancement. The nonlinear exponent is set according to the sharpness of character edges. When the edges are blurred, a larger value is used to enhance the edge transition areas; when the edges are sharp, a smaller value is used to maintain a natural transition. The interference suppression coefficient is determined by analyzing the distribution relationship between the background and foreground areas in the image. When there are many background noise areas, a larger value is used to reduce the contribution of the background in the enhancement process; when the background is relatively clean, a smaller value is used. Through the combined effect of these three parameters, a pixel-by-pixel nonlinear mapping is performed on the image features, highlighting character outlines, table lines, and symbol boundaries, while suppressing background textures and interference components, thus generating an adaptive enhanced image.

[0092] The present invention is further configured such that the generation of regional response values ​​for candidate text, tables, and symbol regions, and the identification of valid regions to be processed, include:

[0093] The adaptively enhanced image is divided into several candidate regions, each consisting of a continuous set of pixels. Specifically, the adaptively enhanced image undergoes preliminary segmentation, dividing the entire image into several candidate regions. Candidate regions are determined by pixel connectivity; that is, adjacent pixels that maintain continuity in grayscale, edge, or orientation features are grouped into the same set. Common methods include connected component analysis or watershed-based region segmentation. The resulting candidate regions correspond to potential text blocks, table cells, or symbol sets, respectively.

[0094] Region weights are generated based on the pixel intensity distribution of candidate regions, local expansion exponents are generated based on the local gradient changes of candidate regions, and geometric distortion measures are generated based on the geometric shape offset of candidate regions. Specifically, region weights are calculated, and the pixel intensity distribution within each candidate region is statistically analyzed to form a grayscale histogram. The histogram is then normalized, and the pixel mean and standard deviation of the region are calculated. The mean reflects the overall brightness level of the region, while the standard deviation reflects the region's contrast and sharpness. Region weights are determined jointly by the mean and standard deviation: regions with high contrast and moderate brightness distribution concentration are assigned higher weights; noisy or background regions are assigned lower weights. Region weights range from zero to one. A local expansion index is generated based on the local gradient changes of candidate regions. Within a candidate region, the local gradient of pixels is calculated, i.e., the directional change in the gray-level difference between adjacent pixels. The gradient distribution of candidate regions is statistically analyzed to obtain an edge density and direction consistency index. The local expansion index is obtained by non-linearly mapping the gradient density; a larger value indicates the presence of numerous continuous edge structures in the region, while a smaller value indicates smoothness within the region. The local expansion index ranges from zero to one. A geometric distortion metric is generated based on the geometric shape offset of candidate regions. Geometric features of candidate regions are extracted, including aspect ratio, the tilt angle of the boundary rectangle, and the curvature of the region's contour. The degree of offset between the region's shape and an ideal regular shape (such as rectangles, straight lines, etc.) is calculated. The greater the offset, the higher the geometric distortion metric value, indicating that the region may be affected by the shooting angle or image distortion. The geometric distortion metric ranges from zero to one.

[0095] Regional response values ​​are generated through a nonlinear combination of regional weights, local expansion exponents, and geometric distortion measures. Specifically, these factors are combined nonlinearly: a larger regional weight results in a higher regional response value; a larger local expansion exponent further increases the regional response value; and a larger geometric distortion measure decreases the regional response value to eliminate false candidate regions caused by distortion. This combination process can be achieved through weighted multiplication or an exponential function to ensure that the response values ​​of strong feature regions are amplified, while the response values ​​of weak feature regions are weakened. The final regional response value for each candidate region is obtained, ranging from zero to one.

[0096] Based on the comparison between the region response value and a set threshold, candidate regions that meet the conditions are selected to form candidate text regions, candidate table regions, and candidate symbol regions. Specifically, the region response value is compared with the set threshold. The threshold is determined through statistical analysis of a large number of sample images. Candidate regions with region response values ​​higher than the threshold are retained, while regions with values ​​lower than the threshold are discarded. The retained candidate regions are classified according to their aspect ratio, edge density, and geometric structure characteristics: regions with a large aspect ratio and uniform horizontal lines inside are marked as candidate table regions; regions with an outline that is approximately rectangular and continuous character edges inside are marked as candidate text regions; regions with regular geometric structures but that do not meet the above conditions are marked as candidate symbol regions.

[0097] The candidate regions obtained from the screening are determined as valid regions to be processed.

[0098] The present invention is further configured such that the calculation of the nonlinear recognition representation and the generation of character sequence recognition results based on the nonlinear representation include:

[0099] Candidate text regions are expanded to form a temporal sequence according to the left-to-right and top-to-bottom order in the image. Specifically, the identified candidate text regions are expanded according to the natural reading order of the image, with a priority order of top-to-bottom and then left-to-right. For example, in a document, the title area at the top is arranged first, followed by the lines of text in the body, and then the characters in the table cells. This expansion method ensures that the sequence order is consistent with the human reading order and avoids misidentification across rows or columns.

[0100] The temporal feature response is calculated based on the pixel features of adjacent regions in the temporal sequence. The features of neighboring regions are weighted and superimposed by the convolution kernel weights to obtain the feature values ​​at each time step. Specifically, in each candidate text region, feature extraction is performed on the region pixels: the region is divided into small blocks, and the edge direction, stroke width and gray level distribution of each block are extracted. Convolution operation is performed on the features between adjacent small blocks. The convolution kernel can be designed as a set of pre-trained weights, which can highlight the line continuity and local changes of the character. The features of adjacent regions are weighted by the convolution kernel weights to obtain the feature value at the current time step. The magnitude of the feature value represents the recognizability of the current character in the sequence. A high value represents clear strokes and complete structure.

[0101] Based on the differences between the feature values ​​at each time point and the feature values ​​at other times throughout the entire sequence, a cross-time-sequence interference metric is generated by combining positional difference attenuation weights. Specifically, the difference magnitude between any two feature values ​​at any given time point is calculated throughout the entire time series. A larger difference indicates a significant difference in the character structures corresponding to the two values; a small difference may indicate duplication or interference. To avoid excessive interference from distant time points, a positional difference attenuation mechanism is introduced: the greater the time interval between the two points, the lower the weight, ensuring that the main interference comes from adjacent or local characters. Combining the difference magnitude with the positional attenuation weights yields the cross-time-sequence interference metric. This metric reflects the degree to which characters are interfered with by adjacent or other characters during the temporal unfolding process.

[0102] By combining temporal feature responses and cross-time interference measures in a nonlinear manner, a recognition representation capable of characterizing temporal patterns is generated; specifically, by combining temporal feature responses and cross-time interference measures...

[0103] If the regional features are clear and the interference metric is low, the recognition representation is enhanced; if the features are blurry or the interference metric is high, the recognition representation is weakened. The combination process uses nonlinear function mapping, such as threshold amplification or exponential compression, to avoid extreme values ​​from distorting the results. The final generated recognition representation is a set of temporal vectors that can fully characterize the character patterns of the candidate regions and their contextual relationships.

[0104] By performing sequence decoding on the recognition representation, a character sequence recognition result is formed. Specifically, the recognition representation is input into the sequence decoder. The decoder can adopt a statistical model based on hidden Markov chains or a learning model based on recurrent neural networks. The decoder outputs the most likely character category at each time step and makes corrections based on the context probability. The decoding results at all time steps are concatenated to obtain a complete character sequence recognition result.

[0105] The present invention is further configured such that generating the corrected character confidence value includes:

[0106] The character sequence recognition results are represented as the initial recognition probability of each character; specifically, the output of the recognition network is represented as the initial recognition probability of each character. For each character at each position, the category with the highest recognition probability is selected as the initial recognition result, and the probability value corresponding to that category is recorded as the initial recognition probability of that character.

[0107] By combining language model weights, the probability correction value of each character under its context is calculated. Specifically, each character is analyzed under the context of its sequence, such as the characters before and after it, word position, and sentence structure. The initial probability is adjusted using language model weights: if the context logic indicates that the character is more likely to appear, its probability is increased; otherwise, its probability is decreased. The language model weights are derived from the training results of a large-scale text corpus and their values ​​are usually between zero and one, representing the enhancement or weakening effect of the context on the probability.

[0108] For each character, a set of candidate words is retrieved, and the matching coefficient of the candidate words is calculated, which represents the degree of matching between the character and the candidate words. Specifically, for each character, a set of candidate words that match it is retrieved, and the degree of matching between each character and the candidate words is calculated, i.e., the matching coefficient of the candidate words. The matching coefficient is quantified based on the position of the character in the candidate words and the consistency of the strokes. The value ranges from zero to one. The higher the value, the more accurate the match. The set of candidate words comes from a pre-established document field dictionary or historical data sample library.

[0109] Context consistency error is generated based on the difference between the joint probability and independent probability of a character within its local context, combined with positional weights. Specifically, for each character within its local context, the difference between the joint probability and independent probability is calculated: the joint probability represents the probability that a character and its neighboring characters will appear simultaneously, while the independent probability represents the probability that a single character will appear. Combining the character's position in the sequence, positional weights are set, giving higher weights to characters closer to core fields or key positions. The difference between the joint probability and independent probability is multiplied by the positional weights to obtain the context consistency error. The error range is usually between zero and one, with a larger value indicating a higher degree of inconsistency in the context.

[0110] The initial character recognition probability, language model probability correction value, candidate word matching coefficient, and context consistency error are combined in a non-linear manner to generate a corrected character confidence value. Specifically, the initial character recognition probability, language model probability correction value, candidate word matching coefficient, and context consistency error are combined in a non-linear manner, such as by non-linear weighting or fusion through a mapping function. The purpose of non-linear combination is to enhance the contribution of reliable features to the confidence value while suppressing the influence of potential interference information. The corrected confidence value generated after combination is still in the range of zero to one, and serves as the final credibility index for each character.

[0111] The present invention is further configured such that generating key field values ​​using the corrected character confidence values ​​includes:

[0112] The corrected character confidence values ​​are combined with the text sequence to identify candidate character subsequences for each field, forming a candidate field set. Specifically, the corrected confidence value of each character is associated with its position in the text sequence to form a character sequence with confidence labels. Based on the document's predefined field templates or rules, the character sequence is divided to identify candidate character subsequences for each field. For example, for the "invoice number" field, a set of consecutive characters that may belong to this field is extracted from the text sequence to form candidate fields. Each candidate field consists of several characters and their corresponding corrected confidence values, forming a candidate field set.

[0113] The confidence values ​​of characters in candidate fields are non-linearly expanded to enhance the discriminative power of low-confidence characters in the field values, generating a field non-linear expansion index. Specifically, the confidence value of each character in the candidate field is non-linearly expanded to increase the discriminative power of low-confidence characters in the field values. Non-linear expansion can use exponential mapping or non-linear gain functions to ensure that high-confidence values ​​remain high and low-confidence values ​​do not completely lose their potential contribution after processing. This generates a field non-linear expansion index for each candidate field, which serves as an important indicator for measuring the confidence distribution within the field.

[0114] By combining the redundant information of candidate fields in the document, the non-linear expansion field value is corrected through a redundancy measurement function to generate a redundancy correction field value. Specifically, for the redundant information of candidate fields in the document, such as the case where the same field appears repeatedly in multiple positions, a redundancy measurement function is calculated. The redundancy measurement function is weighted according to the number of times the candidate field appears, positional consistency and logical relationship between adjacent fields to generate a redundancy correction value. This corrects the non-linear expansion index of the field to obtain the redundancy correction field value, thereby reducing the impact of misidentified or duplicate fields.

[0115] The extraction ratio coefficient is set according to the document length and field importance. The redundancy correction field value is weighted to generate the key field value. Specifically, the extraction ratio coefficient is set according to the document length, field importance and business rules to control the contribution ratio of the redundancy correction field value in the generation of key fields. The redundancy correction field value is weighted to generate the final key field value. The extraction ratio coefficient can be set according to the historical recognition accuracy and importance of the field. The range is 0 to 1. The higher the value, the higher the priority of the field.

[0116] The key field values ​​are mapped to predefined field names or field locations to form a complete set of key fields. Specifically, the generated key field values ​​are matched with the predefined field names or the positions of the fields in the document to form a complete set of key fields. The final output set of key fields includes field names, key field values, and confidence metrics for each field.

[0117] The present invention is further configured such that the calculation of the nonlinear verification representation based on the key field values ​​to form the overall audit verification value includes:

[0118] Organize the set of key field values ​​into a key field vector; specifically, arrange each key field value according to the document's predefined field order, such as invoice number, invoice date, total amount, etc., and combine the arranged key field values ​​into a key field vector. Each field value is accompanied by its confidence information. This vector not only contains the field value, but also the field's position or index information in the document, so as to perform logical relationship verification.

[0119] The logical deviation metric between fields is calculated based on the key field vector to represent the deviation from the expected relationship between fields. Specifically, the expected relationship between fields is defined according to document templates or business rules. For example, the total amount should be greater than or equal to the sum of individual amounts; the invoice date should be later than the invoice date. For each pair of related fields in the key field vector, the deviation between its actual value and the expected relationship is calculated to generate a logical deviation metric between fields. The deviation metric uses normalized distance or ratio and is weighted by field confidence, so that fields with high deviation values ​​have a greater impact on the validation representation.

[0120] By combining key field values ​​and inter-field logical deviation measures in a non-linear manner, a verification representation representing the internal consistency of the audit is generated. Specifically, by combining key field values ​​and inter-field logical deviation measures in a non-linear manner, such as through exponential mapping, non-linear gain functions, or polynomial mapping, the impact of field deviation on the overall verification is amplified or suppressed. The resulting combination generates a non-linear verification representation that comprehensively reflects the integrity of field values ​​and internal logical consistency. This representation not only reflects the confidence of a single field but also the logical consistency between fields, forming an overall internal consistency measurement index for the audit.

[0121] The nonlinear verification representation is normalized to form an overall document review verification value. Specifically, the generated nonlinear verification representation is normalized so that its value falls within a uniform range, such as 0 to 1. The normalized value is used as the overall document review verification value, reflecting the completeness and logical consistency of the document's key fields. The overall document review verification value can be directly used to determine whether the document processing meets the preset standards.

[0122] The present invention is further configured such that the step of generating the model adaptive optimization signal through the overall order review verification value and feedback error metric includes:

[0123] The overall document review verification value is compared with the preset ideal document review verification standard to calculate the feedback error metric. Specifically, the overall document review verification value generated by the current document processing is compared with the preset ideal document review verification standard. The preset standard can be a document integrity and consistency indicator defined internally by the industry or enterprise, such as a standard value of 1 or 0.95. The difference between the current verification value and the standard value is calculated to form a feedback error metric, reflecting the degree of deviation between the document processing result and the ideal state. The feedback error metric can include positive and negative signs to indicate whether the current model output is too high or too low.

[0124] The feedback error metric and the optimization signal adjustment coefficient are nonlinearly combined to generate an adaptive optimization signal for both the recognition model and the field extraction model. Specifically, the adjustment coefficient is used to control the amplitude of the optimization signal, avoiding updates that are too large or too small. The nonlinear combination method can employ exponential amplification, square gain, or polynomial mapping to amplify larger errors to enhance the optimization signal and suppress smaller errors to stabilize the model. The generated adaptive optimization signal contains both error direction information and reflects the nonlinear effect of the error amplitude and the adjustment coefficient.

[0125] The optimized signal is normalized so that the signal value falls within an adjustable range. Specifically, the generated optimized signal is normalized so that the signal value falls within an adjustable range to avoid over- or under-updating the model. The normalization method can be based on maximum and minimum normalization or scaling according to the absolute value of the signal.

[0126] The update direction and magnitude of the recognition model and field extraction model are determined based on the value and intensity of the optimization signal, providing a benchmark for adaptive adjustment of the model. Specifically, the parameter update direction of the recognition model and field extraction model (increasing or decreasing weights, adjusting rule intensity, etc.) is determined based on the sign and magnitude of the normalized optimization signal. The update magnitude is set according to the signal magnitude. Optimization signals with larger magnitudes lead to larger parameter adjustments, while optimization signals with smaller magnitudes lead to fine-tuning, in order to balance the model's convergence speed and stability. The optimization signal is used as a reference for the next round of document processing to achieve adaptive adjustment of the recognition model and field extraction model.

[0127] The present invention is further configured such that the dynamic updating of the recognition model and the field extraction model includes:

[0128] The adaptive optimization signal is applied to update the parameters of the recognition model, and the weights of the recognition model are adjusted by parameter increment. Specifically, the normalized adaptive optimization signal is matched with the current parameters of the recognition model to form parameter increment. The parameter increment is applied to the weights of the recognition model in a weighted manner to adjust the model’s sensitivity to image features and character recognition probability distribution. The adjustment magnitude is proportional to the value of the optimization signal. A larger magnitude results in a more significant adjustment, while a smaller magnitude results in a fine-tuning of the model.

[0129] The adaptive optimization signal is applied to adjust the parameters of the field extraction model, and the field extraction rules or weights are corrected according to the proportional coefficient. Specifically, the optimization signal is matched with the rules or weights of the field extraction model and corrected according to the set proportional coefficient. The proportional coefficient can be set according to the importance of the field, the length of the field, or the historical recognition accuracy. It is used to distinguish the update sensitivity of different fields. The positive or negative signal indicates the increase or decrease of the extraction rule strength, and the field recognition ability of the model under different document structures is dynamically adjusted.

[0130] The update process of the recognition model and field extraction model is iterated. The parameters are continuously adjusted based on the feedback error metric and the overall review verification value until the error converges or the preset iteration round is reached. Specifically, the updated model is run on the new batch of documents to obtain the new overall review verification value and feedback error metric. The new feedback error is used to calculate the new optimization signal, and the parameters of the recognition model and field extraction model are readjusted. The iterative update process is repeated until the feedback error converges to the preset threshold or the maximum iteration round is reached.

[0131] The updated recognition model and field extraction model will be used as the benchmark model for the next round of document processing to ensure that the model can dynamically adapt to changes in document features during the document review task. Specifically, after the iteration is completed, the parameters of the final updated recognition model and field extraction model will be saved as the initial model for the next batch of document processing. In subsequent document processing, the model will use this benchmark to dynamically adapt to changes in document features and achieve continuous learning and optimization.

[0132] Example 2:

[0133] Please see Figure 2 This exemplary AI-based intelligent order review image and text character recognition system includes:

[0134] Image feature enhancement module: Calculates image features by analyzing pixel intensity and local noise perturbation in the image of the document to be reviewed, and generates an adaptive enhanced image based on the image features combined with enhancement amplitude factor, nonlinear exponent and interference suppression coefficient;

[0135] Region Candidate Generation and Response Module: Generates region response values ​​for candidate text, table, and symbol regions based on adaptive enhanced images and region weights, local expansion index, and geometric distortion metric function, and identifies valid regions to be processed;

[0136] Character sequence recognition module: Calculates a nonlinear recognition representation based on the temporal feature response of candidate text regions and cross-time interference measurement, and generates character sequence recognition results based on this nonlinear representation;

[0137] Character confidence value correction module: Generates corrected character confidence values ​​based on character sequence recognition results, language model weights, candidate word matching coefficients, and context consistency errors;

[0138] Key field generation module: Generates key field values ​​using corrected character confidence values, field non-linear expansion index, redundancy measurement function, and extraction ratio coefficient;

[0139] The document review and verification module calculates a non-linear verification representation based on key field values ​​and logical deviation measurements between fields, and forms an overall document review and verification value based on the non-linear verification representation.

[0140] Adaptive optimization and update module: Generates adaptive optimization signals for the model by using overall order verification values ​​and feedback error metrics, and dynamically updates the identification model and field extraction model.

[0141] It should be noted that the AI-based intelligent document review image and text character recognition system and the AI-based intelligent document review image and text character recognition method provided in the above embodiments belong to the same concept. The specific methods of operation of each module and unit have been described in detail in the method embodiments and will not be repeated here. In practical applications, the AI-based intelligent document review image and text character recognition system provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0142] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A graphic character recognition method based on AI-powered intelligent order review, characterized in that: include: Image features are calculated by analyzing the pixel intensity and local noise perturbation of the document image to be reviewed. An adaptive enhanced image is then generated based on these features, combined with an enhancement amplitude factor, a nonlinear exponent, and an interference suppression coefficient. Local noise perturbation is generated by analyzing the difference in amplitude between each pixel and its neighboring pixels. The enhancement amplitude factor is set based on the image contrast level. The nonlinear exponent is set based on the character edge sharpness. The interference suppression coefficient is determined by analyzing the distribution relationship between the background and foreground regions in the image. Through the combined effect of these three parameters, a pixel-by-pixel nonlinear mapping is performed on the image features to generate the adaptive enhanced image. Based on adaptive image enhancement and region weights, local expansion index, and geometric distortion metric function, region response values ​​are generated for candidate text, table, and symbol regions to identify effective regions to be processed. Specifically, a grayscale histogram is formed by statistically analyzing the pixel intensity distribution within each candidate region. The histogram is normalized to calculate the pixel mean and standard deviation for that region, and the region weight is determined jointly based on the mean and standard deviation. Local gradients are calculated for pixels within the candidate regions, and edge density is obtained by statistically analyzing the gradient distribution. The local expansion index is obtained by nonlinearly mapping the gradient density. The geometric distortion metric is generated by calculating the degree of deviation between the region shape and the ideal regular shape. The region response value is generated by nonlinearly combining the above three parameters. Based on the comparison of the region response value with a set threshold, candidate regions that meet the conditions are selected as effective regions to be processed. The nonlinear recognition representation is calculated by using the temporal feature response of the candidate text region and the cross-time interference metric, and the character sequence recognition result is generated based on the nonlinear recognition representation. The cross-time interference metric is generated by calculating the difference magnitude between feature values ​​at any two time points and combining the difference magnitude with the positional attenuation weight. The temporal feature response and the cross-time interference metric are combined in a nonlinear manner to generate the nonlinear recognition representation. A corrected character confidence value is generated based on the character sequence recognition results, language model weights, candidate word matching coefficients, and context consistency errors. Key field values ​​are generated using corrected character confidence values, field nonlinear expansion index, redundancy metric function, and extraction ratio coefficient. Specifically, the field nonlinear expansion index expands low-confidence characters by nonlinearly processing the confidence value of each character within the candidate field. A redundancy metric function is calculated for redundant information appearing in the document for the candidate field. An extraction ratio coefficient is set based on the field's historical recognition accuracy and importance. The field nonlinear expansion index is corrected using the redundancy metric function to generate a redundancy-corrected field value. Finally, the contribution ratio of the redundancy-corrected field value to key field generation is controlled, and a weighted calculation is performed on the redundancy-corrected field value to generate the key field value. A non-linear verification representation is calculated based on key field values ​​and inter-field logical deviation measures. An overall audit verification value is then formed based on this non-linear verification representation. Specifically, each key field value is arranged in the order of predefined fields in the document to generate a key field vector. Based on the document template or business rules, the expected relationship between fields is defined. For each pair of related fields in the key field vector, the deviation between the actual value and the expected relationship is calculated to generate an inter-field logical deviation measure. The key field values ​​and the inter-field logical deviation measure are then combined in a non-linear manner to generate a non-linear verification representation. By generating adaptive optimization signals for the model through overall order verification values ​​and feedback error metrics, the identification model and field extraction model are dynamically updated.

2. The image and text character recognition method based on AI intelligent order review according to claim 1, characterized in that, Calculating image features from the image of the document to be reviewed to generate an adaptively enhanced image includes: The pixel intensity of the document image to be reviewed is normalized to obtain a normalized image with a uniform numerical range. Local noise response is calculated based on normalized images, and local noise perturbation is generated by the difference amplitude between each pixel in the image and its neighboring pixels; Image features are calculated by combining normalized image pixel intensity and local noise perturbation, and the influence of local noise on image features is adjusted by noise suppression coefficient. Specifically, the local noise perturbation value of each pixel is adjusted by combining it with the noise suppression coefficient: the original local noise perturbation value is multiplied or nonlinearly scaled with the noise suppression coefficient, and the normalized pixel intensity is combined with the adjusted local noise perturbation value to form the image features of each pixel. Image features are processed by combining a nonlinear enhancement function with an enhancement amplitude factor, a nonlinear exponent, and an interference suppression coefficient to form an adaptive enhanced image.

3. The image and text character recognition method based on AI intelligent order review according to claim 2, characterized in that, Generate region response values ​​for candidate text, tables, and symbol regions, and identify valid regions to be processed, including: The adaptively enhanced image is divided into several candidate regions, each of which consists of a set of consecutive pixels. Region weights are generated based on the pixel intensity distribution of the candidate region, local expansion indexes are generated based on the local gradient changes of the candidate region, and geometric distortion measures are generated based on the geometric shape offset of the candidate region. The regional response value is generated through a nonlinear combination of regional weights, local spread index, and geometric distortion measure. Based on the comparison between the regional response value and the set threshold, candidate regions that meet the conditions are filtered to form candidate text regions, candidate table regions, and candidate symbol regions; The candidate regions obtained from the screening are determined as valid regions to be processed.

4. The image and text character recognition method based on AI intelligent order review according to claim 3, characterized in that, Calculate the nonlinear recognition representation, and generate character sequence recognition results based on the nonlinear representation, including: The candidate text regions are expanded in the order from left to right and from top to bottom in the image to form a temporal sequence; The temporal feature response is calculated based on the pixel features of adjacent regions in the temporal sequence. The features of the neighboring regions are weighted and superimposed by the convolution kernel weights to obtain the feature values ​​at each time step. Based on the difference between the feature values ​​at each time point and the feature values ​​at other times in the entire sequence range, a cross-time-time interference metric is generated by combining the position difference attenuation weight. By combining the temporal feature response with cross-time interference measurement in a nonlinear manner, a recognition representation that can characterize the temporal pattern of a character is generated. By performing sequence decoding on the recognition representation, a character sequence recognition result is formed.

5. The image and text character recognition method based on AI intelligent order review according to claim 1, characterized in that, The generated corrected character confidence values ​​include: The character sequence recognition result is represented as the initial recognition probability of each character; By combining the language model weights, the probability correction value of each character under its context is calculated; For each character, a set of candidate words is retrieved, and the candidate word matching coefficient is calculated, which represents the degree of matching between the character and the candidate word; Context consistency error is generated based on the difference between the joint probability and independent probability of a character within a local context, combined with positional weights. The initial character recognition probability, language model probability correction value, candidate word matching coefficient, and context consistency error are combined in a non-linear manner to generate the corrected character confidence value.

6. The image and text character recognition method based on AI intelligent order review according to claim 1, characterized in that, The key field values ​​generated using the corrected character confidence values ​​include: The corrected character confidence values ​​are combined with the text sequence to identify candidate character subsequences for each field, forming a candidate field set. The confidence values ​​of characters in candidate fields are nonlinearly expanded to enhance the discriminative power of low-confidence characters in the field values, generating a nonlinear expansion index for the field. Specifically, the confidence value of each character in the candidate field is nonlinearly processed to expand low-confidence characters, thereby increasing their discriminative power in the field values. The nonlinear expansion uses exponential mapping or a nonlinear gain function to ensure that high-confidence values ​​remain high and that low-confidence values ​​do not completely lose their potential contribution after processing, generating a nonlinear expansion index for each candidate field. By combining the redundancy information of candidate fields in the document, the nonlinear expansion index of the fields is corrected through a redundancy measurement function, and a redundancy correction field value is generated. Based on the document length and field importance, the extraction ratio coefficient is set, and the redundant correction field values ​​are weighted to generate key field values; Map key field values ​​to predefined field names or field locations to form a complete set of key fields.

7. The image and text character recognition method based on AI intelligent order review according to claim 6, characterized in that, Based on the key field values, a non-linear verification representation is calculated to form the overall audit verification value, which includes: Organize the set of key field values ​​into a key field vector; Calculate the logical deviation measure between each field based on the key field vector to represent the deviation of the expected relationship between the fields; The key field values ​​and the logical deviation measures between fields are combined in a non-linear manner to generate a verification representation that indicates the internal consistency of the audit. The nonlinear verification representation is normalized to form the overall audit verification value.

8. The image and text character recognition method based on AI intelligent order review according to claim 1, characterized in that, The adaptive optimization signal for the model is generated through overall order review verification values ​​and feedback error metrics, including: The overall order review verification value is compared with the preset ideal order review verification standard, and the feedback error metric is calculated. The feedback error metric is nonlinearly combined with the optimization signal adjustment coefficient to generate an adaptive optimization signal for the recognition model and the field extraction model. The optimized signal is normalized so that the signal value falls within the adjustable range; The update direction and magnitude of the recognition model and field extraction model are determined based on the value and intensity of the optimized signal, providing a benchmark for adaptive adjustment of the model.

9. The image and text character recognition method based on AI intelligent order review according to claim 8, characterized in that, Dynamic updates to the recognition model and field extraction model include: The adaptive optimization signal is applied to the update of the recognition model parameters, and the weights of the recognition model are adjusted by parameter increment; The adaptive optimization signal is applied to adjust the parameters of the field extraction model, and the field extraction rules or weights are corrected according to the proportional coefficient. The update process of the identification model and field extraction model is iterated, and the parameters are continuously adjusted based on the feedback error metric and the overall order verification value until the error converges or the preset iteration round is reached. The updated recognition model and field extraction model will be used as the benchmark model for the next round of document processing to ensure that the model dynamically adapts to changes in document features during the document review task.

10. A graphic character recognition system based on AI-powered intelligent document review, used to implement the graphic character recognition method based on AI-powered intelligent document review as described in any one of claims 1-9, characterized in that, include: Image feature enhancement module: Calculates image features by analyzing pixel intensity and local noise perturbation in the image of the document to be reviewed, and generates an adaptive enhanced image based on the image features combined with enhancement amplitude factor, nonlinear exponent and interference suppression coefficient; Region Candidate Generation and Response Module: Generates region response values ​​for candidate text, table, and symbol regions based on adaptive enhanced images and region weights, local expansion index, and geometric distortion metric function, and identifies valid regions to be processed; Character sequence recognition module: Calculates a nonlinear recognition representation based on the temporal feature response of candidate text regions and cross-time interference measurement, and generates character sequence recognition results based on this nonlinear representation; Character confidence value correction module: Generates corrected character confidence values ​​based on character sequence recognition results, language model weights, candidate word matching coefficients, and context consistency errors; Key field generation module: Generates key field values ​​using corrected character confidence values, field non-linear expansion index, redundancy measurement function, and extraction ratio coefficient; The document review and verification module calculates a non-linear verification representation based on key field values ​​and logical deviation measurements between fields, and forms an overall document review and verification value based on the non-linear verification representation. Adaptive optimization and update module: Generates adaptive optimization signals for the model by using overall order verification values ​​and feedback error metrics, and dynamically updates the identification model and field extraction model.

Images