Digital import and export intelligent auditing method based on CV-NLP fusion algorithm

Abstract

The application relates to the technical field of intelligent auditing, and discloses a digital import and export intelligent auditing method based on a CV-NLP fusion algorithm, which comprises the following steps: performing image coordinate standardization processing on an image sequence of an import and export document package, generating a set of mark candidate blocks containing continuous and reliable weights by using semantic probability and prior attributes of a layout, calculating a layout level complexity parameter representing the fragmentation degree of the hierarchical structure in the document package according to the normalized area scale distribution of the block set, generating a multi-level resolution sampling weight distribution based on the parameter, and performing multi-scale identification and probability fusion on the candidate blocks by using the weight determined by the distribution, calculating a character structure fragmentation measurement value representing the structural fragmentation bias and a region positioning volatility measurement value representing the dispersion degree of the cross-page layout landing point in combination with the layout geometric attributes, and finally constructing a comprehensive risk assessment model to output an intelligent auditing result quantitatively representing the document level identity observability risk.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent auditing, more specifically, it relates to a digital import and export intelligent auditing method based on a CV-NLP fusion algorithm. BACKGROUND

[0002] In the field of international import and export logistics, consolidation is a common form of cargo transportation. Its operation process is to gather multiple invoices of goods from different shippers into the same container at the container freight station, and to unpack and distribute them at the destination port. In order to accurately distinguish different goods mixed in the same container, marks & numbers have become the key identification information for identifying the identity of goods at the piece level.

[0003] Corresponding to this operation process, the consolidation business involves a complex set of documents, usually including Master B / L covering full container, House B / L corresponding to each invoice of goods, and Consolidated Manifest detailing the loading conditions. In the digital auditing scenario, intelligent systems need to automatically process the above document package to verify the authenticity and compliance of the goods information.

[0004] The prior art usually adopts the technical route of optical character recognition plus rule comparison when processing such import and export documents. That is, first, use optical character recognition technology to extract text information from various document images, and then compare the contents of the key fields on each document through pre-set logical rules to determine whether they are consistent. However, in the consolidation scenario, there are significant multi-scale hierarchical differences in the layout of the documents:

[0005] In the Consolidated Manifest, in order to list dozens of invoices of goods in a page or a few pages, the marks information is usually compressed in dense table rows, presenting as high-density small-scale graphic text blocks;

[0006] In the House B / L, since each document only describes one invoice of goods, the marks information usually occupies an independent wide column, presenting as low-density large-scale graphic text blocks;

[0007] In the packing list or label page, the marks information may present as medium-scale.

[0008] The existing OCR and layout analysis algorithm usually adopts a unified scale standard or a fixed resolution strategy to process all images in the same single certificate package. When facing the cascade changes of the layout scale in the LCL scenario, the existing technology is often difficult to adapt. For example, it is acceptable when using high resolution to process large-scale fields, but when processing dense line items in LCL manifest, it is easy to produce fragmentation cutting errors (cutting a mark into multiple fragments) or coalescence errors (identifying the mark of two adjacent lines as one line) due to too small character spacing.

[0009] This recognition deviation caused by scale effect makes the system unable to construct a stable and unified piece-level identity representation in the machine dimension. Even if the subsequent comparison rules are perfect, due to the structural distortion and fragmentation of the recognition data itself, the final intelligent audit result often has a high false positive rate or false negative rate, which is difficult to meet the actual needs of high-precision automatic audit. SUMMARY

[0010] The present application provides a digital import and export intelligent audit method based on CV-NLP fusion algorithm, which solves the technical problems raised in the background art.

[0011] The present application provides a digital import and export intelligent audit method based on CV-NLP fusion algorithm, which includes:

[0012] The image sequence of the import and export certificate package is subjected to image coordinate standardization processing, and a set of mark candidate blocks containing continuous and reliable weights is generated through joint positioning of semantic probability and layout prior attributes;

[0013] According to the normalized area scale distribution of each block in the set of mark candidate blocks, a layout level complexity parameter configured to represent the fragmentation degree of the hierarchical structure of the certificate package is calculated;

[0014] Based on the layout level complexity parameter, a multi-level resolution sampling weight distribution is generated, and the set of mark candidate blocks is subjected to multi-scale recognition and probability fusion using the weight determined by the distribution;

[0015] Based on the multi-scale recognition result and the layout geometric attribute, a character structure fragmentation metric value configured to represent the structural fragmentation bias is calculated, and a region positioning volatility metric value configured to represent the dispersion degree of the cross-page layout landing point is calculated;

[0016] An integrated risk assessment model containing the layout level complexity parameter, the character structure fragmentation metric value and the region positioning volatility metric value is constructed, and an intelligent audit result quantitatively representing the piece-level identity observability risk is output through the integrated risk assessment model.

[0017] The beneficial effects of the present application include: by constructing a layout level complexity parameter representing the degree of fragmentation of the layout level, driving multi-level resolution sampling and probability fusion, effectively overcoming the character structure fragmentation and recombination bias problem caused by the size difference of single certificate layout in the case of packing; at the same time, by introducing a scale and shape coupled measurement mechanism and an adaptive comprehensive risk assessment model, the identity observability risk of the piece level is quantified and evaluated, without cross-single certificate consistency comparison, the robustness and accuracy of complex import and export single certificate package in automatic auditing are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the digital import and export intelligent auditing method based on the CV-NLP fusion algorithm of the present application;

[0019] Figure 2 is a specific implementation schematic diagram of the present application. DETAILED DESCRIPTION

[0020] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed in function and arrangement without departing from the scope of the present specification. Various processes or components can be omitted, substituted, or added according to desired implementations. Additionally, features described with respect to some examples can be combined in other examples.

[0021] As shown in Figure 1 , the digital import and export intelligent auditing method based on the CV-NLP fusion algorithm comprises:

[0022] The image sequence of the import and export single certificate package is subjected to image coordinate standardization processing, and a set of mark head candidate blocks containing continuous credible weights is generated through joint positioning of semantic probability and layout prior attributes;

[0023] According to the normalized area size distribution of each block in the set of mark head candidate blocks, a layout level complexity parameter configured to represent the fragmentation degree of the hierarchical structure in the single certificate package is calculated;

[0024] Based on the layout level complexity parameter, a multi-level resolution sampling weight distribution is generated, and the set of mark head candidate blocks is subjected to multi-scale recognition and probability fusion using the weight determined by the distribution;

[0025] Based on the multi-scale recognition result and the layout geometric attribute, a character structure fragmentation metric value configured to represent the structure fragmentation bias, and a region positioning volatility metric value configured to represent the cross-page layout landing point dispersion degree are respectively calculated;

[0026] A comprehensive risk assessment model is constructed to include the page-level complexity parameter, the character structure fragmentation metric value and the region positioning volatility metric value, and an intelligent audit result quantitatively representing the document level identity observability risk is output through the comprehensive risk assessment model.

[0027] In a preferred embodiment, the image sequence of the import and export document package is subjected to image coordinate standardization processing, and a set of mark candidate blocks containing continuous credible weights is generated through joint positioning of semantic probability and layout prior attributes, including:

[0028] Define the first page image normalization coordinate transformation :

[0029]

[0030] wherein, is the original pixel coordinate, is the pixel width and height of the first page image, and all block coordinates are mapped to space;

[0031] It should be noted that, is used to identify single-page images in the import and export document package, which is the serial number index of the page. Its core role is to distinguish different document pages and ensure that the coordinate standardization processing can accurately correspond to each page, avoiding confusion of the width, height and coordinate information of different pages. This parameter has no fixed value range, and only needs to be assigned in sequence according to the order of the pages in the document package.

[0032] It should be noted that, refers to the original pixel coordinate of any point on the document page image, wherein corresponds to the horizontal coordinate of the pixel, corresponds to the vertical coordinate of the pixel. Its role is to serve as the original input data for coordinate standardization. The value range is determined by the pixel width and height of the corresponding page.

[0033] It should be noted that, refers to the pixel width of the first page image, i.e. the total number of pixels contained by the page in the horizontal direction. Its role is to normalize the original pixel coordinate and eliminate the measurement deviation caused by the difference in physical size and resolution of different pages. The acquisition method is to obtain the metadata of the first page image through image analysis tools.

[0034] It should be noted that, refers to the pixel height of the first page image, i.e. the total number of pixels contained by the page in the vertical direction. Its role is the same as Correspondingly, it is used to convert the original pixel coordinates Normalization ensures that the coordinates of different pages are in the same order of magnitude space. The acquisition method is the same. Read the image using an image analysis tool Metadata for the page image.

[0035] It should be noted that, Defined as the first The normalized coordinate transformation function for a page image. Its function is to transform the original pixel coordinates... Mapped to The normalized physical coordinate space enables a unified measurement of coordinates for pages of different sizes and resolutions.

[0036] The continuous confidence weight of each candidate block is calculated using the following formula. :

[0037]

[0038] in, This refers to the confidence score of semantic associations extracted based on natural language processing. These are prior values ​​for layout structure extracted based on computer vision. To integrate the weighted coefficients, The activation function is non-linear, ultimately forming the set of candidate marker blocks. , This represents the geometric region information of the k-th candidate block, including the boundary of the candidate block in the normalized coordinate space, used to define the spatial range of the k-th candidate block.

[0039] It should be noted that, Used to identify a single section in the set of candidate sections for shipping marks, it is a section number index, and its function is to distinguish different candidate sections for shipping marks.

[0040] It should be noted that, Refers to the first The semantic association confidence score of each candidate section, with a value ranging from [0,1]. Its purpose is to quantify the degree of semantic alignment between the text content of the candidate section and the marking / part number. It is obtained by performing semantic analysis on the OCR text of the candidate sections using a natural language processing network.

[0041] It should be noted that, Refers to the first The prior values ​​of the layout structure of each candidate section are non-negative real numbers. Their purpose is to comprehensively reflect the degree to which the morphological characteristics of the candidate sections match the typical label fields, including the section's aspect ratio, edge regularity, and whether it is located in the typical label area of ​​the document, providing a basis for the layout dimension of the continuous reliable weights. They are obtained by extracting the morphological characteristics of the candidate sections through a layout analysis network and then calculating them.

[0042] It should be noted that, Semantic relevance confidence Prior values ​​of layout structure The fusion weighting coefficients are all non-negative constants. Their function is to adjust the contribution ratio of semantic features and layout features in the calculation of continuous reliability weights, adapting to the differences in the importance of the two types of features under different document scenarios. Optimal values ​​are obtained through offline training, specifically by using a document dataset containing a large amount of labeled data and performing grid search optimization with the prediction accuracy of the continuous reliability weights as the target. and The value range is [0.1, 5.0].

[0043] It should be noted that, This refers to the logistic function. Its function is to map the numerical values ​​after a linear combination to the interval [0,1], so that the output results have a quantitative meaning of credibility or risk probability, avoiding the difficulties in calculation or interpretation caused by the unbounded range of numerical values.

[0044] It should be noted that, Refers to the first The continuous credibility weight of each candidate mark segment, with a value range of [0,1]. Its function is to quantify the credibility of the segment as a genuine mark area.

[0045] It should be noted that, The set of candidate sections for the shipping mark is expressed as follows: Its function is to integrate all candidate blocks obtained through CV-NLP joint localization and their corresponding continuous confidence weights. Representing the Geometric region information (such as boundary coordinates) of each candidate block. Its continuous credible weight, This represents the total number of sections in the set.

[0046] It should be noted that, Refers to the collection of candidate sections for shipping marks The total number of candidate sections included is a positive integer. Its purpose is to specify the size of the section set. The value is determined by the complexity of the document package's content; the more shipping mark-related areas the document contains and the more complex the layout, the better. The larger the value.

[0047] It should be noted that points to the geometric area information of the first candidate version of the label, including the boundary coordinates of the version in the normalized coordinate space (such as the upper left corner and the lower right corner coordinates). Its role is to define the spatial range of each candidate version. The acquisition method is to obtain the region segmentation of the normalized image through the layout analysis network.

[0048] It should be noted that image rendering is performed on each page of the import and export document package, and various formats of documents (including PDF, scans, pictures, etc.) are uniformly converted into standardized image data.

[0049] It should be noted that first, considering the differences in size and resolution of different document pages, directly using pixel coordinates for block position and area measurement will produce deviations due to different page specifications, therefore, a normalized coordinate transformation is designed, which eliminates the influence of page physical size and resolution by dividing the pixel coordinates by the page width and height, so that the coordinate information of all blocks is in the same order of space. Second, the positioning of the label area needs to consider both the layout form features and the text semantic features, relying only on layout analysis may misjudge the non-label area (such as ordinary table columns, note columns) with similar formats as candidates, and relying only on semantic analysis is difficult to accurately lock the specific area position in the layout, therefore, the combination of the two is designed, the form rationality of the region is captured through the layout structure prior value , the semantic relevance of the content is captured through the semantic association confidence , and double verification of positioning is realized. The introduction of the fusion weighting coefficient and is to adapt to the importance difference of the two types of features in different scenarios, for example, in some documents, the layout features of the label are more prominent, while in some other documents, the semantic features are more critical, through offline training to determine the optimal value of and , the fusion result can be more suitable for the actual scenario. The selection of the nonlinear activation function is the logistic function, because it can smoothly map any real number result after linear weighting to the interval, so that has a reliable quantitative meaning, which not only avoids the loss of candidate block information caused by hard threshold segmentation, but also ensures that the contribution of each candidate block matches its credibility.

[0050] It should be noted that the training samples of the natural language processing network are derived from real document text data sets in the import and export LCL business scenario, covering OCR analysis texts of multiple types of documents such as master bill of lading (MBL), house bill of lading (HBL), LCL manifest and packing list, each sample containing text segments of different sections in the document (such as fixed column text, table row item text, note text), and covering multi-format mark expression in Chinese-English bilingual environment (including mixed code string containing letters, numbers and special separators, and pure letter / number code string), and some noisy text segments (such as character confusion caused by OCR recognition error, local occluded text) are also included to improve network robustness; the sample label includes two types of core labels, one is the semantic association label, that is, for each text segment, the matching result of the text segment and the mark / number semantic (binary label 1 represents matching, 0 represents non-matching) and the corresponding real semantic association confidence value (value range [0, 1], labeled by domain experts according to the degree of fit between the text and the mark semantic) are labeled, and the other is the morphological structure label, that is, for the mark text segment, the segment type sequence (such as letter segment A, number field D, separator segment S) and the length proportion of each segment (proportion relative to the total length of the text segment) are labeled; the loss function uses a multi-task joint loss to simultaneously optimize the semantic recognition and morphological analysis capabilities, wherein the binary cross-entropy loss is used for the semantic matching task;

[0051] (in the formula is the total number of samples, is the semantic association binary label of the th sample, is the semantic association confidence value output by the network corresponding to the sample value), the semantic association confidence regression task uses the mean square error loss (in the formula is the real semantic association confidence value of the th sample), the morphological segment classification task uses the class cross-entropy loss (in the formula is the total number of segments in all samples, is the one-hot label of the th segment belonging to type , and is the segment type probability output by the network), and the final total loss is obtained by weighted summation of the task weight coefficients (in the formula is a non-negative weight coefficient, which is optimized by grid search through an offline validation set to ensure balanced contribution of each task loss).

[0052] In a preferred embodiment, based on the normalized area scale distribution of each section in the candidate label section set, a section hierarchy complexity parameter, configured to characterize the degree of fragmentation of the hierarchical structure within a single document package, is calculated, including:

[0053] Calculate the first Normalized area scale of each section :

[0054]

[0055] in, This represents the area of ​​the k-th candidate section. This represents the total area of ​​the page containing the k-th candidate section;

[0056] It should be noted that, Refers to the first The normalized area scale of each candidate label block, with values ​​ranging from (0,1), serves to eliminate the influence of differences in page sizes, objectively reflecting the relative size of the block within its page, and providing a unified scale measure for calculating page layout complexity parameters. For the first The area of ​​each section This represents the total area of ​​the page containing this section.

[0057] It should be noted that, Refers to the first The area of ​​each candidate label block is calculated in normalized coordinate space. Its function is to serve as a normalized area scale. The molecules provide the size data of the plate itself for scale measurement. The calculation method is based on the geometric region information of the plate. calculate.

[0058] It should be noted that, Refers to the first The total area of ​​the pages containing each candidate label section is always 1 in normalized coordinate space. Its function is to serve as a normalized area scale. The denominator is used to eliminate the impact of differences in page physical dimensions on section scale measurement, ensuring the comparability of section scales across different pages. This is determined by coordinate standardization, as page coordinates are standardized... After the transformation, both the width and height are normalized to 1, so the page area is 1×1=1.

[0059] Constructing about scale Weighted empirical survival statistics :

[0060]

[0061] in, is the continuous credible weight, is an indicator function;

[0062] It should be noted that, denotes the set scale threshold, and the value range is (0, 1]. Its role is to serve as a judgment standard for the weighted empirical survival statistics , and is used to count the sum of the block weights whose normalized area scale is greater than the threshold, to characterize the distribution characteristics of the blocks in different scale intervals. The value can be set according to the common scale range of the single certificate scene.

[0063] It should be noted that, denotes the weighted empirical survival statistics about the scale . Its role is to quantitatively represent the cumulative sum of the credible weights of all the label blocks whose normalized area scale is greater than , and reflects the total sum of the credibility of the blocks above the scale threshold, to provide core statistical data for the regression fitting of the layout level complexity parameter.

[0064] It should be noted that, is an indicator function, and its rule is: when the condition in the parentheses is true, the function value is 1; when the condition is not true, the function value is 0. Its role is to filter out the blocks that meet the scale condition in the calculation of the weighted empirical survival statistics, and only sum the continuous credible weights of these blocks, to ensure that the statistics can accurately reflect the block characteristics of the target scale interval.

[0065] Select a number of scale sampling points , and solve the slope of the following regression equation in the logarithmic domain by the weighted least squares method:

[0066]

[0067] Take the solved as the layout level complexity parameter, , and the index of the scale sampling point.

[0068] It should be noted that, denotes the selected scale sampling point, , and the value range is (0, 1]. Its role is to provide multiple discrete scale data points for the linear regression fitting in the logarithmic domain, and to construct a regression equation by the values corresponding to the multiple sampling points to solve the layout level complexity parameter , to ensure the robustness of the fitting result. The preferred sampling method is logarithmic uniform sampling, and 5-20 sampling points are selected in the common scale interval.

[0069] It should be noted that, The page layout complexity parameter is a positive real number and is a linear regression equation in the logarithmic field. The absolute value of the slope. Its function is to quantitatively characterize the degree of fragmentation of the hierarchical structure within a single document package. The larger the value, the higher the proportion of small-scale markings and the more fragmented the document structure. The smaller the value, the higher the proportion of large-scale logo sections and the more regular the structure. The value was obtained by using weighted least squares. and The result was obtained by linear regression fitting.

[0070] It should be noted that, This refers to the regression constant term in a logarithmic linear regression equation, which is a real number. Its function is to fit the logarithmic data. and The linear relationship between them ensures that the regression equation can accurately describe the association between the two, and its value is automatically determined by the sampling point data and the fitting algorithm.

[0071] It should be noted that, This serves as the index for the scale sampling points, distinguishing different sampling points and ensuring that the data for each sampling point can be accurately matched during regression fitting.

[0072] It should be noted that, firstly, considering the differences in physical size and resolution among different document pages, directly comparing the actual area of ​​the sections would be affected by page specifications and would not objectively reflect the relative size and distribution characteristics of the sections within the page. Therefore, a normalized area scale is obtained by comparing the area of ​​the section with the total area of ​​the page. First, it eliminates measurement bias caused by differences in page specifications, making the section scales of different pages comparable. Second, the credibility of each section in the candidate section set varies, and some low-credibility sections may be pseudo-candidate regions caused by noise interference. If all sections are directly weighted equally, the statistical results will be distorted. Therefore, continuous credibility weights are introduced. By weighting the empirical survival statistics, higher-confidence segments are given greater weight in the statistics, while the influence of lower-confidence segments is weakened, thereby increasing the statistical value. The accuracy of representing the actual regional scale distribution of shipping marks. Furthermore, based on the structural characteristics of LCL (Less than Container Load) documents, the scale distribution of their shipping marks exhibits typical multi-scale hierarchical characteristics. This characteristic statistically manifests as a power-law distribution, i.e. This nonlinear relationship can be transformed into a linear regression equation through logarithmic transformation. At this point, the regression slope It can directly quantify the steepness of the power-law distribution. The greater the value, the higher the proportion of small-scale label blocks, and the more serious the fragmentation of the hierarchical structure of the document package. Conversely, it indicates that the proportion of large-scale label blocks is higher, and the structure is more regular. Through this derivation process, the quantitative mapping from block size distribution to hierarchical structure fragmentation is realized.

[0073] In a preferred embodiment, generating a multi-level resolution sampling weight distribution based on the page-level complexity parameter comprises:

[0074] Defining a target attention scale value determined by the page-level complexity parameter :

[0075]

[0076] wherein, is a mapping coefficient, is the page-level complexity parameter;

[0077] It should be noted that, is a target attention scale value determined by the page-level complexity parameter . Its role is to map the abstract complexity parameter to a specific scale value, realizing the reverse correlation between complexity and target attention scale, i.e. the greater (the more fragmented the structure), the smaller.

[0078] It should be noted that, is a basic coefficient for target attention scale mapping, which is a real number. Its role is to set the basic range of the target attention scale, adjust the overall numerical level of , and ensure that it is within a reasonable scale interval, adapting to the common scale range of single document blocks. The preferred value is obtained through offline training, combining the scale distribution characteristics of a large amount of single document data, and is usually in the range of [0.1, 2.0] after tuning.

[0079] It should be noted that, is an adjustment coefficient for target attention scale mapping, which is a positive real number. Its role is to adjust the influence intensity of the page-level complexity parameter on the target attention scale . The greater , the more significant the change caused by . The preferred value is also determined through offline training, combining the scale response needs of different complexity documents, and is in the range of [0.5, 3.0] after tuning.

[0080] Establishing an image pyramid and defining the equivalent area scale of the first ​;

[0081] It should be noted that, refers to the level index of the multi-level resolution pyramid, which is a non-negative integer, that is, Its role is to distinguish different resolution levels of the pyramid, The smaller the value, the lower the resolution of the corresponding level image; The larger the value, the higher the resolution, which provides different resolution image data for multi-scale recognition.

[0082] It should be noted that, refers to the equivalent area scale of the layer of the multi-level resolution pyramid, which satisfies the proportional relationship of Its role is to quantitatively represent the scale level corresponding to the resolution of the layer image, and to provide a basis for calculating the matching degree of the level and the target scale of interest For example, , the reference scale is ; , the reference scale is 1 / 4 of the reference scale, and so on. The higher the level, the smaller the equivalent area scale.

[0083] The weight of the layer in the multi-level resolution sampling weight distribution is calculated according to the following formula: :

[0084]

[0085] Wherein, is the bandwidth parameter, is the total number of pyramid layers.

[0086] It should be noted that, refers to the bandwidth parameter of the Gaussian kernel function, which is a positive real number. Its role is to control the smoothness of the multi-scale sampling weight distribution, The smaller the value, the more concentrated the weight is on the level that best matches the target scale of interest; The larger the value, the more uniform the weight distribution, balancing the needs of local optimization and global coverage. The preferred value range is [0.1, 1.0], which can be adjusted according to the scale distribution complexity of the single certificate.

[0087] It should be noted that, refers to the weight of the layer in the multi-level resolution sampling weight distribution, which is in the range of [0, 1], and the sum of all level weights is 1. Its role is to allocate the contribution proportion of each resolution level in multi-scale recognition, and the higher the matching degree of the level and the target scale of interest , ​The greater, ensure the identification of resources to adapt to the current certificate complexity level tilt.

[0088] It should be noted that, The total number of layers of the multi-level resolution pyramid, which is a non-negative integer. Its role is to define the total number of resolution levels of the pyramid, determine the resolution coverage of multi-scale identification, the more layers, the more accurate the resolution gradient, the better it can adapt to different scale mark blocks, but at the same time it will increase the amount of calculation. The preferred value range is 3-8 layers.

[0089] It should be noted that, The auxiliary index of the multi-level resolution pyramid level, the value is . Its role is to traverse all levels to complete the normalized summation when calculating the layer weight, to ensure that the weight distribution meets the constraint that the sum is 1.

[0090] It should be noted that, first of all, considering that the mark block of the LCLB certificate has significant multi-scale differences, a single resolution cannot meet the identification needs of large-scale fields and small-scale rows, so a multi-level resolution pyramid is constructed to provide corresponding processing resolution levels for different scale mark blocks, among which the equivalent area scale The proportion is set based on the geometric relationship between image resolution and area, to ensure the regularity of scale change between levels. Secondly, the page level complexity parameter directly reflects the fragmentation degree of the certificate structure, The greater the proportion of small-scale mark blocks, the more it needs to rely on high-resolution levels for fine identification, so the is mapped to the target scale of interest using an exponential decay function, which uses the monotonic decreasing property of the exponential function to achieve accurate association between complexity and target scale, to set the basic scale range, to adjust the influence strength of on the target scale, so that the mapping relationship adapts to the scale distribution characteristics of different certificates. Furthermore, the introduction of the Gaussian kernel function is to quantify the degree of fit between each resolution level and the target scale of interest. The logarithmic transformation can convert the multiplicative difference of the scale into an additive difference, which is more consistent with the distance measurement logic in numerical calculation. The bandwidth parameter is used to control the smoothness of the weight distribution, The smaller the weight is more concentrated in the level that best matches the target scale, The greater the weight distribution is more uniform, thereby balancing the needs of local optimization and global coverage. Finally, the normalization process converts the matching degree of each level into a proportion form, ensuring that the weight distribution meets the constraint that the sum is 1, so that the high-matching-degree level plays a greater role in the identification process, achieving adaptive allocation of resolution resources.

[0091] In a preferred embodiment, the weights determined by the multi-level resolution sampling weight distribution are used to perform multi-scale identification and probability fusion on the candidate mark block set, including:

[0092] Get the The section in the pyramid Layer, First The character position belongs to the character recognition probability ;

[0093] It should be noted that, This is a sequence index for character positions, used to distinguish different character positions within a string, providing a positional index for calculating and fusing the probability distribution of character sequences. The value range is determined based on the specific scenario, and the sampling point sequence number... The value ranges from 1 to the total number of sampling points, and the character position number. The value can range from 1 to the maximum length of the string.

[0094] It should be noted that, The first character in the character sequence The character at position 1 ,in This is a character table. Its purpose is to serve as the object of character recognition probability calculation, specifying the possible character types at each position and providing a foundation for constructing the probability distribution of character sequences. Character table It covers common letters (uppercase and lowercase), numbers, and special separators (such as -, / , *, etc.) used in import and export documents.

[0095] It should be noted that, This refers to a complete character table containing all character types that may appear in the shipping marks on import and export documents. Its purpose is to define the scope of character recognition, ensuring that the calculation of the probability distribution of character sequences covers all possible characters and avoiding recognition errors due to missing characters. The construction of the character table should refer to the commonly used character specifications for international logistics documents, typically including 26 English letters (uppercase and lowercase), 10 Arabic numerals, and common special separators, totaling 50-100 characters.

[0096] It should be noted that, Refers to the first The section in the pyramid Layer, First The character position belongs to the character The recognition probability is in the range of [0,1], and for the same , , ,all The sum of the probabilities is 1. Its function is to quantify the reliability of recognizing a character at that level and position. It is obtained by using a character recognition model (such as CNN) to analyze the first... The images are obtained by recognizing the blocks of the layer.

[0097] The training samples for the CNN model are derived from real document text datasets in import and export LCL (Less than Container Load) business, covering OCR parsing texts of various documents such as master bills of lading and sub-bills of lading, including bilingual (Chinese and English) multi-format shipping marks and noisy text fragments. Sample labels include semantic association labels (binary matching labels and true semantic association confidence values) and morphological structure labels (fragment type sequence and length proportion). The loss function employs a multi-task joint loss, including binary cross-entropy loss for semantic matching, mean squared error loss for semantic association confidence regression, and category cross-entropy loss for morphological fragment classification, which are weighted and summed to obtain the total loss. When the model outputs the recognition probability of the k-th segment belonging to character c_t at the ℓ-th level of the pyramid and the t-th character position, the segment image at that level is first input into the CNN. After local feature extraction by convolutional layers and dimensionality reduction by pooling layers, it is mapped to the character table dimension through fully connected layers, and then normalized by the softmax function to obtain the probability distribution of each character at that position.

[0098] Utilizing the weights in the multi-level resolution sampling weight distribution Multiplicative fusion of probabilities from each layer yields a non-normalized fused probability. :

[0099]

[0100] It should be noted that, Refers to the first The first section The non-normalized fusion probability of each character position. Its function is to integrate the recognition probabilities of each resolution level, and to amplify the consistency confidence of each level through multiplicative fusion, while suppressing the divergence confidence.

[0101] right Normalization yields the final probability distribution The final string is obtained by decoding using the maximum a posteriori probability. :

[0102]

[0103] It should be noted that, Refers to the first The first section The normalized final probability distribution of each character position takes values ​​in the range [0,1]. Its function is to transform the non-normalized fusion probability into a standard distribution that conforms to probability theory, ensuring that the sum of the probabilities of each character is 1.

[0104] Need to explain, The first The final label string obtained after multi-scale recognition and decoding of the The role is as the corresponding piece-level identity information of the block. The acquisition method is to decode

[0105] Need to explain, The candidate set of the string, containing all possible character sequence combinations. Its role is as the candidate space of maximum a posteriori decoding, and the decoding algorithm selects the character sequence with the maximum probability product from the set as the final string The size of the candidate set is determined by the size of the character table And the maximum length of the string .

[0106] Need to explain, The maximum length of the character sequence, which is a positive integer. Its role is to define the length range of the string, ensure that character recognition and probability fusion can cover the longest possible label string, and avoid character omission due to insufficient length. The preferred value is set according to the common length of import and export documents, usually 10-50 characters.

[0107] Need to explain, first of all, considering the differences in recognition ability of different resolution levels for label blocks, high-resolution levels can capture fine character features of small-scale blocks, and low-resolution levels can avoid excessive segmentation of characters in large-scale blocks, so character recognition is performed at each level and the probability distribution is output, which can fully retain effective recognition information at different scales. Secondly, the choice of multiplicative fusion instead of additive fusion is because multiplicative operation can amplify the consistent confidence of each level in recognizing the same character, while suppressing the divergent confidence, so that the fusion result can better highlight the characters recognized by most levels. The weight As an index, it can realize the nonlinear adjustment of the weight to the probability distribution, and the higher the weight of the level, the more dominant its probability distribution in the fusion, so that the fusion result accurately matches the level complexity characteristics of the document package. The introduction of normalization processing is to convert the non-standard probability distribution after multiplicative fusion into a standard distribution that meets the probability theory, eliminating the interference caused by the difference in probability values of different blocks and different levels, and providing a unified input format for the decoding algorithm. Finally, the maximum a posteriori decoding is used because this algorithm can directly select the character combination with the highest probability at each character position to form a string, which can maximize the reduction of the cumulative impact of character recognition errors while ensuring decoding efficiency, and ensure that the final output label string accurately reflects the piece-level identity information.

[0108] In a preferred embodiment, based on the multi-scale recognition results and page layout geometric attributes, a character structure fragmentation metric configured to characterize structural fragmentation bias is calculated, including:

[0109] Statistical recognition of strings Number of fragments With total length Calculate the morphological complexity :

[0110]

[0111] It should be noted that, Refers to the first Individual header string The number of fragments after decomposition is a positive integer. Its purpose is to quantify the degree of fragmentation of the string; the more fragments, the more fragmented the string's structure, providing basic data for morphological complexity calculations. The decomposition rule is to... Divide into letter segments, number segments, and separator segments.

[0112] It should be noted that, Refers to the first Individual header string The total length, i.e., the number of characters contained in the string, is a positive integer. Its purpose is to serve as the denominator in morphological complexity calculations, reducing the number of segments... Normalization eliminates the impact of string length differences on fragmentation measurement, ensuring that the morphological complexity of strings of different lengths is comparable.

[0113] It should be noted that, Refers to the first The morphological complexity of the header string, with a value ranging from (0,1), is used to quantify the degree of fragmentation in the internal structure of the string. The larger the value, the more fragments there are within a unit length, and the more fragmented the string is. The smaller the value, the more regular the string structure, providing core data for calculating the fragmentation metric of character structure.

[0114] Based on the page layout hierarchy complexity parameter Calculate the expected scale morphological equilibrium value :

[0115]

[0116] in, This is a preset constant;

[0117] It should be noted that, Refers to the complexity parameter based on page layout level The calculated expected scale-form balance value. Its purpose is to characterize the proper balance between scale and form under the current level of document complexity, serving as a benchmark for judging the degree of fragmentation in character structure. The larger, The larger the value, the better it can adapt to string formats with higher degrees of fragmentation.

[0118] It should be noted that, This refers to the fundamental constant of the expected scale morphological equilibrium value, and is a real number. Its function is to set the basic level of the equilibrium value, ensuring... The values ​​are within a reasonable range and are suitable for common scale-morphology coupling relationships. The optimal values ​​are obtained through offline training and, combined with the scale-morphology correlation features of a large amount of single-document data, the optimized range is [-2.0, 0.5].

[0119] It should be noted that, This refers to the adjustment constant for the desired scale-based morphological balance value, and is a positive real number. Its function is to adjust the page layout's hierarchical complexity parameter. right The extent of the impact The larger, Change-induced The more significant the change, the more accurately the balance value can be adapted to documents of varying complexity. The preferred value range is [0.3, 1.5], determined through offline training and optimization.

[0120] The character structure fragmentation metric is calculated using the following formula. :

[0121]

[0122] in, For continuous reliable weights, For normalized area scale, This is the morphological adjustment coefficient.

[0123] It should be noted that, This refers to the morphological adjustment coefficient, which is a positive real number. Its function is to calibrate the morphological complexity. Logarithmic value of normalized area scale The difference in dimensions ensures that the summation of the two values ​​accurately reflects the coupling relationship between scale and form, avoiding fragmented measurement bias caused by inconsistent dimensions. The preferred value range is [0.5, 2.0], which can be adjusted according to the morphological distribution characteristics of the document.

[0124] It should be noted that, This refers to a non-negative real number that measures character structure fragmentation. Its function is to quantify the scale-driven character structure fragmentation bias. The larger the value, the more severe the fragmentation of the string and the greater the deviation from the expected scale-based balance value, thus providing a basis for the fragmentation dimension in comprehensive risk assessment.

[0125] It should be noted that, firstly, considering that fragmentation of the shipping mark string in LCL documents is directly manifested as an increased frequency of alternation between letters, numbers, and delimiters, simply counting the number or length of fragments cannot reflect the degree of fragmentation alone. Therefore, the ratio is calculated by combining both. This allows morphological complexity to stably quantify the structural fragmentation features within the string, avoiding biases caused by single-dimensional measurement. Secondly, the layout-level complexity parameter... It directly reflects the degree of fragmentation in the hierarchical structure of the documents. The larger the value, the higher the proportion of small-scale label sections, and the corresponding string morphology complexity should also show a corresponding trend. Therefore, a linear formula is designed. ,pass Set a basic balance level. adjust The magnitude of the impact on the balance value allows the desired scale morphological balance value to adapt to the hierarchical structural characteristics of the document, ensuring the relevance of the measurement. Furthermore, the normalized area scale... The logarithm can transform the multiplicative difference of scale into an additive difference, along with the weighted morphological complexity. Achieving adaptation in the numerical dimension, morphological adjustment coefficient This is used to calibrate the dimensional difference between morphological complexity and the logarithmic value of scale, ensuring that the sum of the two accurately reflects the coupling relationship between scale and morphology. The square operation of the difference highlights the degree of deviation from the desired equilibrium value while avoiding the cancellation of positive and negative deviations, and ensures continuous and reliable weights. The introduction of this method aims to ensure that deviations from high-confidence blocks dominate the final measurement, weakening the interference from low-confidence pseudo-candidate blocks. The weighted average processing further integrates information from all valid blocks, making... It can comprehensively and objectively quantify the scale-driven character structure fragmentation bias.

[0126] In a preferred embodiment, based on the multi-scale recognition results and page layout geometric attributes, a regional positioning volatility metric value configured to characterize the dispersion of cross-page layout landing points is calculated, including:

[0127] Construct the first Geometric anchor vectors of each section :

[0128]

[0129] in, For the normalized center coordinates of the section, This refers to the main axis direction of the section.

[0130] It should be noted that, the first candidate label block is a normalized center coordinate of the first candidate label block, and the value range is [0, 1]. Its role is to represent the center position of the block in the horizontal direction of the page, and to provide the core data of the position dimension for the geometric anchor point vector.

[0131] It should be noted that, the first candidate label block is a normalized center coordinate of the first candidate label block, and the value range is [0, 1]. Its role is to represent the center position of the block in the horizontal direction of the page, and to provide the core data of the position dimension for the geometric anchor point vector. It should be noted that,

[0132] the first candidate label block is a normalized center coordinate of the first candidate label block, and the value range is [0, 1]. Its role is to represent the center position of the block in the horizontal direction of the page, and to provide the core data of the position dimension for the geometric anchor point vector. It should be noted that, the first candidate label block is a normalized center coordinate of the first candidate label block, and the value range is [0, 1]. Its role is to represent the center position of the block in the horizontal direction of the page, and to provide the core data of the position dimension for the geometric anchor point vector.

[0133] It should be noted that, the first candidate label block is a normalized center coordinate of the first candidate label block, and the value range is [0, 1]. Its role is to represent the center position of the block in the horizontal direction of the page, and to provide the core data of the position dimension for the geometric anchor point vector.

[0134] The weighted covariance matrix of the geometric anchor point distribution is calculated as

[0135]

[0136] wherein, is a continuous credible weight, is the weighted mean of the geometric anchor point vector;

[0137] It should be noted that, the first candidate label block is a normalized center coordinate of the first candidate label block, and the value range is [0, 1]. Its role is to represent the center position of the block in the horizontal direction of the page, and to provide the core data of the position dimension for the geometric anchor point vector.

[0138] ​denotes the weighted covariance matrix of the geometric anchor point distribution, which is a 4x4 symmetric matrix. Its role is to quantitatively represent the dispersion degree of all high-confidence block geometric anchor point vectors, and the elements in the matrix reflect the correlation dispersion characteristics between different dimensions.

[0139] calculating the area positioning fluctuation metric value :

[0140]

[0141] wherein, denotes the trace of the matrix.

[0142] It should be noted that, denotes the trace of the weighted covariance matrix , that is, the sum of the main diagonal elements of the matrix. Its role is to comprehensively extract the dispersion information of the geometric anchor point vector in all dimensions, convert the multi-dimensional dispersion characteristics of the 4x4 covariance matrix into a single scalar, avoid the one-sidedness of single-dimensional measurement, and provide the core calculation result for the area positioning fluctuation metric value. The calculation method is to sum the diagonal elements of the covariance matrix in turn.

[0143] It should be noted that, denotes the area positioning fluctuation metric value, which is a non-negative real number. Its role is to quantitatively represent the dispersion degree of the landing point of the header block in the cross-page layout, the greater the value, the more dispersed the center position and the main axis direction distribution of the block, and the poorer the positioning stability; the smaller the value, the more concentrated the block distribution, and the more stable the positioning, which provides a basis for geometric distribution dimension for comprehensive risk assessment.

[0144] It should be noted that, first of all, the dispersion characteristics of the header block are not only reflected in the plane position, but also involve the difference in the main axis direction. Only the coordinates cannot completely describe the dispersion of the landing point shape, so the center coordinates and the main axis direction are combined to construct the geometric anchor point vector , wherein and are introduced to convert the periodic parameter of angle into a linearly calculable value, avoid the measurement deviation caused by angle cycle, and make the direction difference integrated into a unified vector operation system. Secondly, there are low-confidence pseudo-candidate areas in the header candidate block set. If ordinary mean and covariance are directly used for calculation, the dispersion degree will be distorted due to the interference of these areas, so the continuous confidence weight is introduced to weight the mean and covariance matrix, so that high-confidence blocks contribute more weight in the statistical process, and the influence of low-confidence blocks is reasonably weakened, ensuring and The weighted covariance matrix can accurately reflect the distribution characteristics of the real mark area Each element of the weighted covariance matrix corresponds to the discrete correlation degree between different dimensions of the anchor point vector. The trace of the matrix, as the sum of the diagonal elements, can comprehensively extract the discrete information of all dimensions, avoid the one-sidedness caused by single dimension measurement, and make it possible to comprehensively and objectively quantify the discrete degree of the mark block landing point in the cross-page layout.

[0145] In a preferred embodiment, a comprehensive risk assessment model including the layout hierarchical complexity parameter, the character structure fragmentation measurement value and the area positioning volatility measurement value is constructed, comprising:

[0146] Defining a dimension reduction function :

[0147]

[0148] It should be noted that refers to the dimension reduction function. Its function is to perform smoothing mapping processing on the character structure fragmentation measurement value and the area positioning volatility measurement value , compress the numerical span and regularize the dimension without changing the relative size relationship of the data, so as to avoid the fusion deviation caused by too large difference in magnitude of the two types of measurement values.

[0149] The character structure fragmentation measurement value and the area positioning volatility measurement value are processed by using the dimension reduction function;

[0150] A comprehensive risk linear potential energy is constructed:

[0151]

[0152] Among them, is the layout hierarchical complexity parameter, to is a preset weighting coefficient, is an adaptive penalty cross term.

[0153] It should be noted that refers to the comprehensive risk linear potential energy, which is a real number. Its function is to integrate the layout hierarchical complexity, fragmentation bias, positioning volatility, and the synergistic effect of complexity and fragmentation, to form a unified linear potential energy index. Among them is a reference constant term, describes the direct risk contribution of structural complexity, and To separately characterize the risk contribution of fragmentation and locational volatility, To characterize the synergistic risk contribution of the two.

[0154] It should be noted that, The pre-defined weighting coefficients for the linear potential energy of comprehensive risk are all real numbers, where... Typically, it is a non-negative real number. Its function is to adjust the contribution ratio of each risk factor in the overall potential energy, adapting to the differences in the importance of various risks under different document scenarios. The optimal value is obtained through offline training. Specifically, it is achieved by using a document dataset containing a large number of labeled risk levels and performing grid search or gradient descent optimization with the accuracy of risk assessment as the target. The value is usually in the range of [0.1, 5.0].

[0155] It should be noted that, The intelligent review result refers to a value range within... The probability scalar between these values. Its function is to quantify the observability risk of item-level identity in the LCL (Less than Container Load) scenario. The larger the value, the higher the risk, and the smaller the value, the lower the risk, providing users with intuitive and standardized review conclusions.

[0156] It should be noted that, firstly, the character structure fragmentation metric is taken into account. and regional positioning volatility measure The values ​​may vary considerably, and directly performing a linear combination can lead to fusion bias caused by differences in numerical magnitudes. Meanwhile, the dimensional reduction function... It possesses the characteristics of monotonically increasing and compressing large values ​​while preserving differences for small values. It can map two types of measures to similar numerical ranges without altering the relative size of the data, eliminating interference from inconsistent dimensions and ensuring that each factor has an equal basis for weight adjustment during the fusion process. Secondly, comprehensive risk assessment requires a full integration of structural features and two types of derived risk measures, along with parameters related to page layout complexity. It is a core indicator for describing the structure of documents. and Fragmentation bias and positioning discreteness were quantified separately, and these three factors were incorporated into the formula as independent weighting terms. This allows the formula to capture risk contributions from different dimensions, while the weighting coefficients... to The introduction of this allows for adjusting weights based on the importance of each factor in the actual scenario, enabling the model to adapt to the risk distribution characteristics of different documents. Furthermore, the page layout complexity parameter... With smoothed fragmentation metric There is an inherent connection between them. The larger the value, the more fragmented the document structure. In this case, the impact of fragmentation bias on risk should be amplified accordingly. Therefore, a product cross term should be designed. ,pass Adjusting the strength of this association, achieving adaptive enhancement of risk contribution, enables the model to accurately capture the synergistic effect of structural complexity and fragmentation bias. The preset constant term is used as the reference level for calibrating the linear potential energy of comprehensive risk, ensuring that the numerical range meets the input requirements of nonlinear mapping.

[0157] In a preferred embodiment, the intelligent audit result quantitatively representing the identity observability risk of the piece level is output by the comprehensive risk assessment model, including:

[0158] Using the logistic function to map the linear potential energy of comprehensive risk to the intelligent audit result :

[0159]

[0160] wherein is a scalar with a value range between , used to quantitatively represent the identity observability risk of the piece level caused by multi-scale cascading structure in the consolidation scenario.

[0161] It should be noted that first, considering that the value of the linear potential energy of comprehensive risk may cover the positive and negative infinite interval, the values corresponding to different single certificate packages lack a unified comparison reference, and it is difficult for users to understand and apply directly as an audit result, therefore it needs to be converted into a scalar within a fixed interval through nonlinear mapping. The logistic function is chosen because it has the property of monotonic increasing, which can completely preserve the risk level ordering relationship implied by , and its S-shaped curve can achieve a reasonable characterization of risk, i.e., when tends to negative infinity, tends to 0, corresponding to extremely low risk; when tends to positive infinity, tends to 1, corresponding to extremely high risk; and when is in the middle value range, the change rate is moderate, which can accurately distinguish different degrees of medium risk, avoiding excessive sensitivity or sluggishness of risk assessment. The output interval is limited to because this interval meets the conventional understanding of risk quantification, which is convenient for users to quickly understand the risk level corresponding to the numerical value, and is compatible with the risk scoring system of existing audit systems, without the need for additional scale conversion. At the same time, the numerical form of this interval naturally has the meaning of probability, which is highly consistent with the quantification requirements of the identity observability risk of the piece level.

[0162] As shown in Figure 2 ,Figure 2 The left side first shows the business basis of LCL consolidation: after the multi-ticket goods are loaded in the container freight station to form a container, the corresponding documents are generated, including the master bill of lading covering the whole container, multiple sub-bills of lading corresponding to each ticket, a consolidated manifest and a container manifest detailing the loading situation. All the documents contain the core identification of the goods item level identity, i.e. the mark / item number, and present the multi-scale cascading layout characteristics due to different types of documents. The middle part connects the digital processing link. After the document package is converted into processable document data, the CV-NLP fusion audit mechanism is started. The right side details the audit steps: first, the image coordinate standardization is performed on the image sequence of the import and export document package, the mark candidate block set with continuous and reliable weights is generated through the joint positioning of semantic probability and layout prior attributes, then the layout level complexity parameter representing the fragmentation degree of the document hierarchical structure is calculated according to the normalized area scale distribution of the set, the multi-level resolution sampling weight distribution is generated based on the parameter, and the candidate block is identified and probabilistically fused in multiple scales, subsequently, the character structure fragmentation metric value and the region positioning volatility metric value are calculated combined with the layout geometric attributes, finally, the intelligent audit result quantitatively representing the observable risk of the item level identity is output through the comprehensive risk assessment model integrating the above key parameters , which clearly links the LCL consolidation business scenario, the document system characteristics and the technical link of CV-NLP fusion audit.

[0163] The above describes the embodiments of the present embodiment, but the present embodiment is not limited to the specific embodiments described above, which are only illustrative and not restrictive. Those skilled in the art can make many forms under the inspiration of the present embodiment, which are all within the protection scope of the present embodiment.

Claims

1. A digital import and export intelligent auditing method based on a CV-NLP fusion algorithm, characterized in that, The method comprises the following steps: image coordinate standardization processing is performed on the image sequence of the import and export document package, and a mark head candidate block set containing continuous confidence weights is generated through joint positioning of semantic probability and layout prior attributes; According to the normalized area scale distribution of each block in the mark head candidate block set, a layout level complexity parameter configured to represent the fragmentation degree of the hierarchical structure in the document package is calculated, including: Calculate the ratio of the area of each block in the mark head candidate block set to the total area of the page where it is located to obtain the normalized area scale; Taking the continuous confidence weight as a weighted item, the sum of the weights of the blocks with normalized area scale greater than a certain threshold is calculated to construct a weighted empirical survival statistic varying with scale; In the logarithmic coordinate system, the relationship between the weighted empirical survival statistic and the normalized area scale is linearly regressed and fitted, and the absolute value of the slope value obtained by fitting is taken as the layout level complexity parameter; Based on the layout level complexity parameter, a multi-level resolution sampling weight distribution is generated, and the weights determined by the distribution are used to perform multi-scale recognition and probability fusion on the mark head candidate block set; Based on the multi-scale recognition result and the layout geometric attribute, a character structure fragmentation metric value configured to represent the structure fragmentation bias and a region positioning volatility metric value configured to represent the dispersion degree of the cross-page layout landing point are calculated respectively; The character structure fragmentation metric value includes: The recognized string is decomposed into a combination sequence of letter segments, number fields and separator segments, the ratio of the number of segments in the string to the total length of the string is calculated to obtain the morphological complexity; The expected scale morphological balance value at the current level is calculated using the layout level complexity parameter; The sum of the logarithmic value of the normalized area scale of each block and the morphological complexity after weighted adjustment is calculated, and the square of the difference between the sum and the expected scale morphological balance value is calculated; The weighted average of the square of the difference is calculated using the continuous confidence weight, and the weighted average result is taken as the character structure fragmentation metric value; The region positioning volatility metric value includes: The normalized center coordinates and the principal axis direction angle of each block in the mark head candidate block set are extracted to construct the geometric anchor vector of each block; Using the continuous confidence weight, the weighted covariance matrix of the geometric anchor vector of all blocks relative to the weighted mean vector is calculated; The trace of the weighted covariance matrix is calculated, and the value of the trace is taken as the region positioning volatility metric value; A comprehensive risk assessment model containing the layout level complexity parameter, the character structure fragmentation metric value and the region positioning volatility metric value is constructed, and an intelligent audit result quantitatively representing the document level identity observability risk is output through the comprehensive risk assessment model.

2. The digital import and export intelligent auditing method based on the CV-NLP fusion algorithm according to claim 1, characterized in that, The method comprises the following steps: image coordinate standardization processing is performed on the image sequence of the import and export document package, and a mark head candidate block set containing continuous confidence weights is generated through joint positioning of semantic probability and layout prior attributes, including: Render each page of documents in the import and export single document package into an image, and map the pixel coordinates in the image to normalized physical coordinates with values between zero and one by dividing the width and height of the image; Extract the layout structure prior of the candidate layout block using a layout analysis network, and extract the semantic association confidence of the candidate layout block using a natural language processing network; Linearly weight and sum the semantic association confidence and the layout structure prior, and input the sum into a nonlinear activation function for mapping to obtain the continuous confidence weight of each candidate layout block, thereby constructing the set of head candidate layout blocks.

3. The digital import and export intelligent auditing method based on the CV-NLP fusion algorithm according to claim 1, characterized in that, Based on the layout level complexity parameter, a multi-level resolution sampling weight distribution is generated, including: A multi-level resolution pyramid of the input image is constructed, and the equivalent area scale of each level of the pyramid is determined; An exponential decay function is used to map the layout level complexity parameter to a target attention scale value, so that the higher the complexity, the smaller the target attention scale value; A Gaussian kernel function is used to calculate the matching degree between the equivalent area scale of each level of the pyramid and the target attention scale value; The matching degrees of all levels are normalized, and the obtained normalized values are used as the weight values of the corresponding levels in the multi-level resolution sampling weight distribution.

4. The digital import and export intelligent auditing method based on the CV-NLP fusion algorithm according to claim 1, characterized in that, Multi-scale recognition and probability fusion of the set of head candidate layout blocks are performed using the weights determined by the multi-level resolution sampling weight distribution, including: For each layout block in the set of head candidate layout blocks, character recognition is performed at each level of the image pyramid, and the character sequence probability distribution of that level is output; The weights of each level in the multi-level resolution sampling weight distribution are used as exponents to multiply and weight the character sequence probability distribution of each level, and a fused non-normalized probability distribution is obtained; The non-normalized probability distribution is normalized, and a sequence decoding algorithm is used to parse the final string result from the normalized probability distribution.

5. The digital import and export intelligent auditing method based on the CV-NLP fusion algorithm according to claim 1, characterized in that, A comprehensive risk assessment model is constructed, including the layout level complexity parameter, the character structure fragmentation metric value, and the region positioning volatility metric value, including: The character structure fragmentation metric value and the region positioning volatility metric value are respectively smoothed using a natural logarithm function to obtain a smoothed fragmentation metric value and a smoothed volatility metric value; A linear weighted combination formula is constructed, and the weighted items of the formula include at least the layout level complexity parameter, the smoothed fragmentation metric value, the smoothed volatility metric value, and the product of the layout level complexity parameter and the smoothed fragmentation metric value; The weighted coefficients are calculated by summing the weighted items, and the sum is used as the comprehensive risk linear potential, thereby completing the construction of the comprehensive risk assessment model.

6. The digital import and export intelligent auditing method based on the CV-NLP fusion algorithm according to claim 5, characterized in that, An intelligent audit result quantitatively representing the identity observability risk of the piece is output by the comprehensive risk assessment model, including: The comprehensive risk linear potential output by the comprehensive risk assessment model is input into a logistic function for nonlinear mapping; The output value of the logistic function is calculated to obtain a probability scalar with a value between zero and one; The probability scalar is taken as the smart audit result, wherein the greater the numerical value of the result indicates the higher the part-level identity observability risk under the current layout level complexity. The probability scalar is taken as the smart audit result, wherein the greater the numerical value of the result indicates the higher the part-level identity observability risk under the current layout level complexity.

Images