Multimodal archival file automated classification storage system

Abstract

The application discloses a kind of multi-modal archive file automated classification storage systems, it includes data preprocessing unit, multi-modal information analysis unit, dynamic rule generation unit and distributed storage unit.Unit is optimized by edge detection and affine transformation to document image, and the classification precision is improved by combining text semantics and visual feature extraction, and the classification rule is dynamically updated using the decay window model, and the safety of classification result is ensured by hash generation and cross-chain verification.The application can significantly improve the classification accuracy, processing efficiency and system security, meet the needs of modern archives management for intelligentization and high efficiency.

Description

Technical Field

[0001] This invention belongs to the field of information technology and data processing technology, specifically a multimodal archive file automated classification and storage system. Background Technology

[0002] With the increasing demand for archival management, archival document classification systems have played a vital role in improving archival management efficiency and optimizing storage and retrieval functions. However, existing archival classification systems still have shortcomings in areas such as intelligent processing, multimodal data support, and user experience optimization, which limits their further promotion and application in complex application scenarios.

[0003] A search revealed a patent for an archival classification system for document organization, with publication number CN114996365B and publication date May 24, 2024. This patent, through the design of an archival classification system, combines feature extraction, keyword matching, and cloud database verification to achieve preliminary screening and automatic classification of documents. It also introduces a classification timing module to avoid program backlog. However, this technical solution primarily relies on text analysis and keyword matching, exhibiting limitations in its adaptive processing capabilities for multimodal data such as images and audio. Furthermore, its cloud verification mechanism may face efficiency bottlenecks when dealing with large-scale document classification, failing to adequately address storage expansion and real-time issues, thus impacting the overall system performance.

[0004] A search revealed a document classification system for archival management, with publication number CN117453982B and publication date June 21, 2024. This patent improves the efficiency of user document retrieval by generating a key information tree diagram and providing links to relevant documents, and introduces a user feedback mechanism to optimize system functions. However, this technical solution has shortcomings in dynamic cross-modal data association and updating, making it difficult to achieve in-depth mining and intelligent analysis of archival content. Furthermore, its static entity relationship modeling method has limited performance in adapting to the diversity and complexity of archival data, especially when processing unstructured data, which may lead to a decrease in the accuracy of classification results.

[0005] The aforementioned problems indicate that existing archival document classification systems still have room for improvement in areas such as multimodal data processing, storage optimization, real-time performance assurance, and cross-modal dynamic correlation updates. Therefore, this invention provides a novel archival document classification system that aims to improve the accuracy and efficiency of archival classification by introducing intelligent multimodal analysis, closed-loop verification mechanisms, and dynamic knowledge graph update functions, thereby meeting the demands of modern archival management for efficient and intelligent systems. Summary of the Invention

[0006] This invention relates to the field of archival management technology, specifically providing a multimodal automated classification and storage system for archival documents. This system addresses the shortcomings of existing archival management systems in areas such as intelligent processing, multimodal data support, real-time performance assurance, and cross-modal dynamic correlation updates by introducing innovative technologies in data processing, feature extraction, dynamic classification, and reliable evidence preservation.

[0007] To address the aforementioned technical problems, this invention employs the following technical solution: an automated classification and storage system for multimodal archival documents, comprising a data preprocessing unit, a multimodal information parsing unit, a dynamic rule generation unit, and a distributed evidence storage unit. The data preprocessing unit performs format standardization and quality optimization on the input archival documents; the multimodal information parsing unit extracts key semantic and visual features from various types of data; the dynamic rule generation unit generates and adjusts classification rules based on historical data and real-time input; and the distributed evidence storage unit records classification results through a decentralized storage mechanism and supports cross-chain verification operations.

[0008] Preferably, the data preprocessing unit includes a document calibration module and an adaptive binarization module; the document calibration module extracts the document boundary contour using an edge detection algorithm, calculates the minimum bounding rectangle to determine the document tilt angle θ, and corrects the document based on the affine transformation matrix; the adaptive binarization module uses the formula:

[0009] The size of the local threshold block is dynamically calculated to achieve image optimization processing, where W and H represent the width and height of the input image, respectively, in pixels; I(i,j) represents the gray value of the image at position (i,j); and the offset C = 8 is used to suppress noise interference.

[0010] Preferably, the multimodal information parsing unit includes a text semantic extraction submodule and a visual feature extraction submodule; the text semantic extraction submodule is based on an improved deep learning model, capturing contextual dependencies through a 12-layer Transformer encoder, and outputting a 768-dimensional semantic vector; the visual feature extraction submodule employs a channel attention mechanism to enhance the response of important features, and its feature recalibration formula is as follows:

[0011] Where F in Represents the input feature map; F avg W1 and W2 represent the feature vectors after global average pooling; W1 and W2 are trainable parameter matrices; σ represents the Sigmoid activation function; and δ represents the ReLU activation function.

[0012] Preferably, the dynamic rule generation unit includes a hybrid feature modeling module and a rule update module; the hybrid feature modeling module uses the formula:

[0013]

[0014] α v =1-α t

[0015] Dynamically assign weight coefficients to text semantic features and visual features, where w t and w v represents the weight parameters for text and visual features, respectively; TF-IDF represents the term frequency-inverse document frequency value of text features; Clarity represents the image sharpness score; the rule update module updates the rule support using a decaying window model, with the formula being...

[0016] Support new =0.85·Support old +0.15·Support current

[0017] When Support new When the value is less than 0.005, the corresponding rule is automatically eliminated.

[0018] Preferably, the distributed evidence storage unit includes a hash generation module, a consensus writing module, and a cross-chain verification module; the hash generation module generates a composite hash value by bit-level concatenation of file content, category tags, and timestamps, using the following formula:

[0019] HASH=SHA3-256(FileContent||ClassificationLabel||Timestamp)

[0020] The consensus writing module writes classification tags to the consortium blockchain based on the Kafka consensus mechanism, with a block time of 2 seconds and a block size of 8MB. The cross-chain verification module compares the hash values ​​of classification tags on different chains through smart contract logic and returns a verification status code, where 0 indicates validity and 1 indicates tampering.

[0021] Preferably, the digital signature of the distributed evidence storage unit adopts the national cryptographic SM2 algorithm, and the private key is stored in a hardware security module that conforms to the FIPS140-2 Level 3 standard. The signature verification time does not exceed 15 milliseconds.

[0022] Preferably, the document calibration module uses the Canny operator to implement the edge detection algorithm. After Gaussian filtering and denoising the image, the gradient magnitude and direction are calculated to determine the bounding rectangle boundary of the document. The document tilt angle θ is obtained by calculating the rotation angle of the minimum bounding rectangle. The translation amount of the affine transformation matrix is ​​determined by the position of the document center point, and the correction error is controlled within 3 pixels.

[0023] Preferably, the input text data of the text semantic extraction submodule is preprocessed and the word segmentation length is limited to ≤512; the input feature map of the visual feature extraction submodule is generated into a feature vector after global average pooling, and W1 and W2 in the feature recalibration formula are trainable parameter matrices.

[0024] Preferably, when the cross-chain verification module compares the classification tag hash values ​​on different chains through smart contract logic, if a single bit change is detected, it returns verification status code 1 to indicate data tampering.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] First, the document calibration module and the adaptive binarization module are used to optimize the processing of low-quality document images. The edge detection and affine transformation technologies are used to significantly improve the OCR recognition performance. The correction error is controlled within 3 pixels, and the system is highly adaptable and stable.

[0027] Second, the multimodal information parsing unit combines the advantages of text semantic extraction and visual feature extraction, and enhances the responsiveness of key features through the channel attention mechanism, improving the classification accuracy to over 89.3%.

[0028] Third, the dynamic rule generation unit realizes the dynamic updating of classification rules through the decay window model and conflict detection mechanism, avoiding the interference of old rules on new data. The rule base update efficiency reaches 12,000 rules / second, and the misclassification rate is reduced by 42%.

[0029] Fourth, the distributed evidence storage unit ensures the security and immutability of the classification results through hash generation and cross-chain verification protocols, with tamper detection sensitivity reaching the level of single-bit changes and traceability efficiency improved by 6 times.

[0030] In summary, this invention achieves comprehensive breakthroughs in classification accuracy, processing efficiency, system security, and scalability through the synergistic effect of the data preprocessing unit, multimodal information parsing unit, dynamic rule generation unit, and distributed evidence storage unit, thus meeting the demands of modern archives management for efficient and intelligent systems. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0032] Figure 2 This is a flowchart illustrating the data preprocessing unit.

[0033] Figure 3 Functional block diagram of the multimodal information parsing unit;

[0034] Figure 4 This is a flowchart of the operation of the distributed evidence storage unit.

[0035] The attached diagram is labeled as follows: 1. Data preprocessing unit; 2. Multimodal information parsing unit; 3. Dynamic rule generation unit; 4. Distributed evidence storage unit; 5. Document calibration module; 6. Adaptive binarization module; 7. Text semantic extraction submodule; 8. Visual feature extraction submodule; 9. Hash generation module; 10. Consensus writing module; 11. Cross-chain verification module. Detailed Implementation

[0036] This invention provides a multimodal automated classification and storage system for archival documents, comprising multiple modules and components. These modules and components work together through specific connections, positional relationships, and cooperative relationships to achieve efficient classification and evidence preservation of archival documents. The specific implementation methods of this system are described in detail below.

[0037] Multimodal archive document automated classification and storage system Figure 1 This diagram illustrates the overall architecture of the system, showing the logical relationships and data flow directions between the data preprocessing unit 1, the multimodal information parsing unit 2, the dynamic rule generation unit 3, and the distributed evidence storage unit 4. Each unit is interconnected through data interfaces, forming a complete closed-loop processing flow to ensure automated operation throughout the entire process from file input to classified storage.

[0038] Data preprocessing unit 1 consists of document calibration module 5 and adaptive binarization module 6, and its working process is as follows: Figure 2 As shown. The document calibration module 5 receives the input archive image and first extracts the document boundary contour using an edge detection algorithm. The edge detection algorithm is based on the Canny operator. After Gaussian filtering to remove noise from the image, the gradient magnitude and direction are calculated to determine the bounding rectangle boundary of the document. The document tilt angle θ is obtained by calculating the rotation angle of the minimum bounding rectangle, and then the document is corrected based on the affine transformation matrix. The translation amount of the affine transformation matrix is ​​determined by the position of the document center point. The corrected document is aligned with the image center, and the error is controlled within 3 pixels.

[0039] The adaptive binarization module 6 uses a local threshold block size formula to optimize the corrected image.

[0040]

[0041] W and H represent the width and height of the input image, respectively, in pixels; I(i,j) represents the grayscale value of the image at position (i,j); the offset C = 8 is used to suppress noise interference. Through the above steps, the quality of the document image is significantly improved, laying the foundation for subsequent feature extraction.

[0042] The multimodal information parsing unit 2 includes a text semantic extraction submodule 7 and a visual feature extraction submodule 8, and its internal structure is as follows: Figure 3 As shown. Text semantic extraction submodule 7 is based on an improved deep learning model, capturing contextual dependencies through a 12-layer Transformer encoder. After preprocessing, the input text data has its word segmentation length limited to ≤512, and the encoder outputs a 768-dimensional semantic vector. Visual feature extraction submodule 8 employs a channel attention mechanism to enhance the response to important features; its feature recalibration formula is:

[0043]

[0044] Where F in Represents the input feature map; F avg W1 and W2 represent the feature vector after global average pooling; W1 and W2 are trainable parameter matrices; σ represents the Sigmoid activation function; δ represents the ReLU activation function. The text semantic extraction submodule 7 and the visual feature extraction submodule 8 are connected via a data pipeline, passing the extracted semantic vectors and visual features to the dynamic rule generation unit 3. The dynamic rule generation unit 3 consists of a hybrid feature modeling module and a rule update module. The hybrid feature modeling module fuses text semantic features and visual features using a dynamic weight allocation formula. The hybrid feature modeling module uses the following formula:

[0045]

[0046] α v =1-α t

[0047] The weight coefficients of text semantic features and visual features are dynamically assigned, where α t and α v The weighting coefficients representing textual and visual features, w t and w v These represent the weight parameters; TF-IDF represents the term frequency-inverse document frequency value of the text features; Clarity represents the image sharpness score; the rule update module updates the rule support using a decaying window model, with the formula being...

[0048] Support new =0.85·Support old +0.15·Support current Support in the formula new Indicates new support. old Indicates old support. current Indicates the current level of support. new When the value is less than 0.005, the corresponding rule is automatically eliminated.

[0049] Distributed evidence storage unit 4 includes a hash generation module 9, a consensus writing module 10, and a cross-chain verification module 11; the hash generation module 9 generates a composite hash value by bit-level concatenation of file content, category tags, and timestamps, using the following formula:

[0050] HASH=SHA3-256(FileContent||ClassificationLabel||Timestamp);

[0051] In the formula, HASH represents a composite hash value, SHA3-256 is the hash algorithm, FileContent represents the file content, ClassificationLabel represents the category label, and Timestamp represents the timestamp. The consensus writing module 10 writes the category label to the consortium blockchain based on the Kafka consensus mechanism, with a block time of 2 seconds and a block size of 8MB. The cross-chain verification module 11 compares the hash values ​​of the category labels on different chains using smart contract logic and returns a verification status code, where 0 indicates validity and 1 indicates tampering.

[0052] The digital signature of the distributed evidence storage unit 4 adopts the national cryptographic SM2 algorithm, and the private key is stored in a hardware security module that conforms to the FIPS 140-2 Level 3 standard. The signature verification time is no more than 15 milliseconds.

[0053] In practical applications, the system's workflow is as follows: First, the archive files are input to the data preprocessing unit 1. The document calibration module 5 corrects the documents using edge detection algorithms and affine transformation matrices. The adaptive binarization module 6 optimizes image quality using a local threshold block size formula. The processed archive data is then transmitted to the multimodal information parsing unit 2. The text semantic extraction submodule 7 and the visual feature extraction submodule 8 extract text semantic features and visual features respectively, and transmit the feature data to the dynamic rule generation unit 3 via a data pipeline. The dynamic rule generation unit 3 fuses multimodal features through a hybrid feature modeling module and generates and adjusts classification rules through a rule update module. Finally, the classification results are transmitted to the distributed evidence storage unit 4. The hash generation module 9 generates a composite hash value, the consensus writing module 10 writes the classification label to the consortium blockchain, and the cross-chain verification module 11 completes cross-chain verification through smart contract logic comparison. The entire system achieves automated classification and storage of archive files through the collaborative efforts of each unit, ensuring a comprehensive improvement in classification accuracy, processing efficiency, system security, and scalability.

[0054] To better enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below with reference to a specific application scenario. In the archives management center of a large enterprise, a large number of multimodal archive files containing text, images, and scanned documents need to be processed daily. These archive files come from various sources, including scanned copies of paper documents, spreadsheets, image attachments, and audio recordings. To improve the efficiency of archive classification and storage, the center introduced the automated classification and storage system for multimodal archive files provided by this invention, and implemented automated classification and storage of archives according to the following steps. First, the archive files are input to the data preprocessing unit 1 via a scanner or upload interface. The document calibration module 5 receives the input archive image and uses the Canny operator as an edge detection algorithm to extract the document boundary contour. After Gaussian filtering to remove noise from the image, the gradient magnitude and direction are calculated to determine the minimum bounding rectangle boundary of the document. Subsequently, based on the rotation angle θ of the minimum bounding rectangle, the document is corrected using an affine transformation matrix. The translation amount of the affine transformation matrix is ​​determined by the position of the document center point, ensuring that the corrected document is aligned with the image center, with the error controlled within 3 pixels. Next, the adaptive binarization module 6 optimizes the corrected image using a dynamic local threshold block size formula. In the formula, W and H represent the image width and height in pixels, respectively; I(i,j) represents the grayscale value at position (i,j); and the offset C = 8 is used to suppress noise interference. Through these steps, the quality of the document image is significantly improved, laying the foundation for subsequent feature extraction. Next, the processed archive data is transmitted to the multimodal information parsing unit 2. The text semantic extraction submodule 7, based on an improved deep learning model, captures contextual dependencies through a 12-layer Transformer encoder. After preprocessing, the word segmentation length of the input text data is limited to ≤512, and the encoder outputs a 768-dimensional semantic vector. The visual feature extraction submodule 8 uses a channel attention mechanism to enhance the response of important features. In its feature recalibration formula, F... in F represents the input feature map. avg W1 and W2 represent the feature vector after global average pooling, respectively. W1 and W2 are trainable parameter matrices, σ represents the Sigmoid activation function, and δ represents the ReLU activation function. Through this process, the text semantic extraction submodule 7 and the visual feature extraction submodule 8 extract the semantic and visual features of the archive file, respectively, and pass these feature data to the dynamic rule generation unit 3 via a data pipeline. Subsequently, the dynamic rule generation unit 3 receives feature data from the multimodal information parsing unit 2. The hybrid feature modeling module fuses the text semantic features and visual features through a dynamic weight allocation formula. In the formula, α... t and α v The weighting coefficients representing textual and visual features, w t and w v... new Indicates new support. old Indicates old support. current Indicates the current level of support. When Support... new When the value is less than 0.005, the corresponding rule is automatically eliminated. Through the above operations, the dynamic rule generation unit 3 generates and adjusts the classification rules, and transmits the classification results to the distributed evidence storage unit 4. Finally, the classification results are transmitted to the distributed evidence storage unit 4. The hash generation module 9 generates a composite hash value by concatenating the file content, classification label, and timestamp at the bit level. In the formula, HASH represents the composite hash value, SHA3-256 is the hash algorithm, FileContent represents the file content, ClassificationLabel represents the classification label, and Timestamp represents the timestamp. The consensus writing module 10 writes the classification label to the consortium chain based on the Kafka consensus mechanism, with a block time of 2 seconds and a block size of 8MB. The cross-chain verification module 11 compares the classification label hash values ​​on different chains through smart contract logic and returns a verification status code, where 0 indicates validity and 1 indicates tampering. The digital signature of the distributed evidence storage unit 4 adopts the national cryptographic SM2 algorithm, and the private key is stored in a hardware security module that conforms to the FIPS140-2 Level 3 standard. The signature verification time does not exceed 15 milliseconds. Through the above steps, the classification results are securely stored and support cross-chain verification operations. The entire system achieves automated classification and storage of archival documents through the collaborative efforts of its various units. In the actual application within this enterprise, the system significantly improves the accuracy and efficiency of archival classification while ensuring the security and immutability of the classification results. For example, when processing a batch of scanned documents with complex backgrounds, the document calibration module 5 and the adaptive binarization module 6 effectively improve image quality, resulting in a significant improvement in OCR recognition performance, with correction errors controlled within 3 pixels. Furthermore, the multimodal information parsing unit 2 combines the advantages of text semantic extraction and visual feature extraction, increasing the classification accuracy to over 89.3%. The dynamic rule generation unit 3 achieves dynamic updates of classification rules through a decay window model and conflict detection mechanism, reducing the misclassification rate by 42%. Finally, the distributed evidence storage unit 4 ensures the security of the classification results through hash generation and cross-chain verification protocols, achieving tamper detection sensitivity at the single-bit change level and improving traceability efficiency by 6 times.

[0055] In summary, this invention achieves comprehensive breakthroughs in classification accuracy, processing efficiency, system security, and scalability through the synergistic effect of the data preprocessing unit 1, the multimodal information parsing unit 2, the dynamic rule generation unit 3, and the distributed evidence storage unit 4, thus meeting the demands of modern archives management for efficient and intelligent systems.

Claims

1. A multimodal archive document automated classification and storage system, characterized in that... It includes a data preprocessing unit (1), a multimodal information parsing unit (2), a dynamic rule generation unit (3), and a distributed evidence storage unit (4); The data preprocessing unit (1) performs format standardization and quality optimization on the input archive files; The multimodal information parsing unit (2) extracts key semantic features and visual features from various types of data; The dynamic rule generation unit (3) generates and adjusts classification rules based on historical data and real-time input; The distributed evidence storage unit (4) records the classification results through a decentralized storage mechanism and supports cross-chain verification operations; The data preprocessing unit (1) includes a document calibration module (5) and an adaptive binarization module (6); The document calibration module (5) extracts the document boundary contour through an edge detection algorithm, calculates the minimum bounding rectangle to determine the document tilt angle θ, and corrects the document based on the affine transformation matrix; The adaptive binarization module (6) uses the formula: The local threshold block size is dynamically calculated to achieve image optimization processing, where W and H represent the width and height of the input image, respectively, in pixels; I(i,j) represents the gray value of the image at position (i,j); and the offset C=8 is used to suppress noise interference. The dynamic rule generation unit (3) includes a hybrid feature modeling module and a rule update module; The hybrid feature modeling module uses the following formula: The weight coefficients of text semantic features and visual features are dynamically assigned, where and These are the weighting coefficients for textual and visual features, w t and w v These represent the weight parameters; TFIDF represents the term frequency-inverse document frequency value of the text features; Clarity represents the image sharpness score; the rule update module updates the rule support using a decaying window model, with the formula: When Support new When the value is less than 0.005, the corresponding rule is automatically eliminated; The distributed evidence storage unit (4) includes a hash generation module (9), a consensus writing module (10), and a cross-chain verification module (11); the hash generation module (9) generates a composite hash value by bit-level concatenation of file content, category tags, and timestamps, using the following formula: The consensus writing module (10) writes the classification tags into the consortium chain based on the Kafka consensus mechanism. The block time is 2 seconds and the block size is 8MB. The cross-chain verification module (11) compares the classification label hash values ​​on different chains through smart contract logic and returns a verification status code, where 0 indicates validity and 1 indicates tampering. The document calibration module (5) uses the Canny operator to implement the edge detection algorithm. After denoising the image by Gaussian filtering, the gradient magnitude and direction are calculated to determine the bounding rectangle boundary of the document. The document tilt angle θ is obtained by calculating the rotation angle of the minimum bounding rectangle. The translation amount of the affine transformation matrix is ​​determined by the position of the document center point, and the correction error is controlled within 3 pixels.

2. The multimodal archive document automated classification and storage system according to claim 1, characterized in that: The multimodal information parsing unit (2) includes a text semantic extraction submodule (7) and a visual feature extraction submodule (8); the text semantic extraction submodule (7) is based on an improved deep learning model, which captures contextual dependencies through a 12-layer Transformer encoder and outputs a 768-dimensional semantic vector; the visual feature extraction submodule (8) uses a channel attention mechanism to enhance the response of important features, and its feature recalibration formula is: F in Represents the input feature map; F avg W1 and W2 represent the feature vectors after global average pooling; W1 and W2 are trainable parameter matrices. This represents the Sigmoid activation function; This represents the ReLU activation function.

3. The multimodal archive document automated classification and storage system according to claim 1, characterized in that: The digital signature of the distributed evidence storage unit (4) adopts the national cryptographic SM2 algorithm, and the private key is stored in a hardware security module that conforms to the FIPS 140-2 Level 3 standard. The signature verification time is no more than 15 milliseconds.

4. The multimodal archive document automated classification and storage system according to claim 2, characterized in that: After preprocessing, the input text data of the text semantic extraction submodule (7) is limited to word segmentation length of ≤512; the input feature map of the visual feature extraction submodule (8) is generated into a feature vector after global average pooling, and W1 and W2 in the feature recalibration formula are trainable parameter matrices.

5. The multimodal archive document automated classification and storage system according to claim 1, characterized in that: When the cross-chain verification module (11) compares the classification label hash values ​​on different chains through smart contract logic, if a single bit change is detected, it returns verification status code 1 to indicate data tampering.

Images