A digital archive platform based on the digitization of paper documents

Abstract

This invention provides a digital archive platform based on the digitization of paper documents, belonging to the field of data processing technology. The platform includes: a scanning and acquisition module, an intelligent image preprocessing module, a deep learning OCR recognition module, a semantic understanding and classification module, an automatic metadata extraction module, an encrypted storage and retrieval module, and a version management and access control module. This invention improves the accuracy of text recognition in complex scenarios (such as deformed or blurred documents) by utilizing surface fitting correction technology in the deep learning OCR recognition module, significantly reducing the cost of manual proofreading.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a digital archive platform based on the digitization of paper documents. Background Technology

[0002] With the deepening of digital transformation, the digitization of paper documents has become a core requirement in the field of archival management. Traditional paper-based archival management methods have many drawbacks: on the one hand, paper archives occupy a large amount of physical storage space and are easily affected by environmental factors (such as humidity and pests), leading to information loss or damage; on the other hand, manual retrieval is extremely inefficient, making it difficult to quickly locate and retrieve the required documents, and thus failing to meet the needs of modern enterprises and institutions for efficient information management.

[0003] In the field of digitizing paper documents, early technologies mainly relied on simple scanning and storage, lacking the ability to optimize image quality and perform in-depth text processing. Traditional OCR technology has low accuracy when dealing with deformed or blurred documents, making it difficult to handle complex real-world scenarios; semantic analysis and classification mostly use rule-based methods, which lack flexibility and adaptability, and cannot handle diverse document types. In terms of security, traditional encryption methods are disconnected from document content, key management lacks dynamism, and there is a risk of data leakage; access control is also mostly based on static rules, making it difficult to meet the fine-grained access needs of different users in different scenarios.

[0004] In addition, existing digital archive systems often have independent functional modules, lacking effective data linkage and collaborative processing, and cannot form a complete digital archive management loop, resulting in low overall efficiency and difficulty in meeting the needs of large-scale, high-frequency document processing. Summary of the Invention

[0005] This invention provides a digital archive platform based on the digitization of paper documents to solve one or more of the problems mentioned above.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A digital archive platform based on the digitization of paper documents includes:

[0008] The scanning and acquisition module is used to convert paper documents into digital images and perform geometric correction to obtain corrected images.

[0009] The intelligent image preprocessing module is connected to the scanning acquisition module and is used to perform noise filtering, binarization and enhancement processing on the corrected image to obtain a preprocessed image.

[0010] A deep learning OCR recognition module is connected to the intelligent image preprocessing module and is used to extract text sequences from the preprocessed image;

[0011] The semantic understanding and classification module is connected to the deep learning OCR recognition module and is used to analyze the semantics of the text based on the text sequence and classify the document to obtain category labels and entity information;

[0012] The metadata automatic extraction module, connected to the semantic understanding and classification module, is used to extract structured metadata from text sequences, category labels, and entity information.

[0013] The encrypted storage and retrieval module is connected to the automatic metadata extraction module and is used to encrypt and store structured metadata and build a retrieval index to obtain ciphertext, retrieval index and structured metadata;

[0014] The version management and access control module is connected to the encrypted storage and retrieval module and is used to manage document versions and control user access permissions based on ciphertext, retrieval index and structured metadata.

[0015] In this specification, the deep learning OCR recognition module includes:

[0016] The surface fitting and correction unit is used to construct a parametric surface model to perform geometric correction on the preprocessed image and output the corrected image features.

[0017] A multi-scale feature extraction unit, connected to the surface fitting and correction unit, is used to extract local and global features of the image based on the corrected image features and output a feature vector sequence.

[0018] The temporal modeling unit, connected to the multi-scale feature extraction unit, is used to perform temporal modeling on the feature vector sequence and output a predicted distribution.

[0019] The CTC decoding unit, connected to the time-series modeling unit, is used to convert the predicted distribution into a text sequence.

[0020] In this specification, the semantic understanding and classification module includes:

[0021] BERT encoding units are used to perform semantic encoding on the text sequence and output a context representation vector;

[0022] The surface classifier unit, connected to the BERT encoding unit, is used to construct a surface classification model with local weighted regression based on the context representation vector, classify documents, and output category labels;

[0023] An entity recognition unit, connected to the BERT encoding unit, is used to identify named entities from a text sequence and output entity information.

[0024] In this specification, the automatic metadata extraction module includes:

[0025] The template matching unit is used to select the corresponding metadata extraction template based on the category labels output by the semantic understanding and classification module.

[0026] The information extraction unit, connected to the template matching unit, is used to extract metadata by combining text sequences and entity information through regular expressions and semantic reasoning.

[0027] The metadata verification unit, connected to the information extraction unit, is used to perform format standardization and integrity verification on the extracted metadata and output structured metadata.

[0028] In this specification, the encrypted storage and retrieval module includes:

[0029] The feature generation unit is used to extract document feature vectors from structured metadata;

[0030] The surface key generation unit, connected to the feature generation unit, is used to construct a polynomial key surface based on the document feature vector, and to generate an encryption key by selecting sampling points based on the key surface;

[0031] The data encryption unit, connected to the surface key generation unit, is used to encrypt document data using an encryption key and output ciphertext.

[0032] The index building unit, connected to the feature generation unit, is used to build a retrieval index based on document feature vectors.

[0033] In this specification, the surface fitting correction unit specifically includes:

[0034] The feature point detection subunit is used to detect text edge feature points in the preprocessed image;

[0035] The surface parameter calculation subunit is connected to the corner detection subunit and is used to calculate NURBS surface parameters based on text edge feature points.

[0036] The image correction subunit is connected to the surface parameter calculation subunit and is used to perform geometric transformations on the preprocessed image based on the surface parameters, and output the corrected image features.

[0037] In this specification, the surface classifier unit specifically includes:

[0038] The feature mapping subunit is used to map the context representation vector to the feature space;

[0039] The local weighted regression calculation subunit is connected to the feature mapping subunit and is used to construct a surface classification model with local weighted regression in the feature space.

[0040] The category prediction subunit, connected to the local weighted regression calculation subunit, is used to calculate the probability of a document belonging to each category based on the surface classification model of local weighted regression, and output the category label.

[0041] In this specification, the surface key generation unit specifically includes:

[0042] The polynomial generating subunit is used to generate a polynomial key surface based on the document feature vector.

[0043] A sampling point subunit, connected to the polynomial generation subunit, is used to select feature sampling points on the polynomial key surface;

[0044] A key derivation subunit is connected to the sampling point subunit and is used to generate an encryption key based on feature sampling points through the HMAC function.

[0045] The polynomial generation subunit uses a neural network to learn the mapping relationship between document feature vectors and the coefficients of the polynomial key surface.

[0046] In this specification, the index building unit specifically includes:

[0047] The keyword extraction subunit is used to extract keywords from the document feature vector;

[0048] The inverted index generation sub-unit is connected to the keyword extraction sub-unit to build an inverted index from keywords to documents;

[0049] The semantic index enhancement subunit, connected to the inverted index generation subunit, is used to semantically expand the inverted index by combining category labels and entity information to obtain the retrieval index.

[0050] The version management and access control module described in this specification includes:

[0051] The version record unit is used to store different versions of the ciphertext, search index, and structured metadata output by the encryption storage and retrieval module;

[0052] The permission allocation unit, connected to the version record unit, is used to allocate access permissions for each version of the document based on user roles.

[0053] The operation audit unit, connected to the permission allocation unit, is used to record user access to and modification operations on document versions.

[0054] In summary, the present invention has at least the following beneficial effects:

[0055] Highly efficient digital processing capabilities: By using surface fitting and correction technology in the deep learning OCR recognition module, the accuracy of text recognition in complex scenarios (such as deformed or blurred documents) is improved, significantly reducing the cost of manual proofreading.

[0056] Precise Semantic Understanding and Classification: Based on BERT and surface fitting techniques, the semantic understanding and classification module can deeply mine semantic information from text. Compared to traditional rule-based classification methods, it improves the accuracy of automatic classification of document topics and categories, and can handle different types of documents such as contracts, invoices, and reports. The automatic metadata extraction module improves the accuracy of key metadata extraction through template matching and multi-strategy information extraction, providing a solid foundation for accurate document retrieval and efficient management.

[0057] Enhanced data security: The surface key binding encryption technology in the encrypted storage and retrieval module deeply associates key generation with document content features, significantly improving security compared to traditional fixed-key encryption methods. Even if data is leaked, attackers cannot generate the correct key without the corresponding document features. The RBAC model and dynamic permission verification in the version management and access control module achieve fine-grained access control, meeting the permission requirements of different users and scenarios. Meanwhile, the audit log function fully records operational behavior, effectively preventing data leakage risks and ensuring data security and compliance.

[0058] A complete closed-loop management process: Close collaboration between modules, from scanning and data acquisition to final storage and access control, forms a complete digital records management loop. Data flows smoothly between modules, with the output of one module precisely serving the output of the next, avoiding redundant data processing and information gaps, and improving overall processing efficiency. This end-to-end solution not only improves the automation level of records management but also provides enterprises and institutions with an efficient, secure, and intelligent digital records management experience, driving the development of records management towards intelligence and modernization. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a schematic diagram illustrating the composition of the digital archive platform based on the digitization of paper documents involved in this invention.

[0061] Figure 2 This is a schematic diagram of the composition of the deep learning OCR recognition module involved in this invention.

[0062] Figure 3 This is a schematic diagram illustrating the composition of the semantic understanding and classification module involved in this invention.

[0063] Figure 4 This is a schematic diagram illustrating the composition of the encrypted storage and retrieval module involved in this invention. Detailed Implementation

[0064] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0065] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0066] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0067] like Figure 1 As shown, this embodiment provides a digital archive platform based on the digitization of paper documents, including:

[0068] The scanning and acquisition module is used to convert paper documents into digital images and perform geometric correction to obtain corrected images.

[0069] The intelligent image preprocessing module is connected to the scanning acquisition module and is used to perform noise filtering, binarization and enhancement processing on the corrected image to obtain a preprocessed image.

[0070] A deep learning OCR recognition module is connected to the intelligent image preprocessing module and is used to extract text sequences from the preprocessed image;

[0071] The semantic understanding and classification module is connected to the deep learning OCR recognition module and is used to analyze the semantics of the text based on the text sequence and classify the document to obtain category labels and entity information;

[0072] The metadata automatic extraction module, connected to the semantic understanding and classification module, is used to extract structured metadata from text sequences, category labels, and entity information.

[0073] The encrypted storage and retrieval module is connected to the automatic metadata extraction module and is used to encrypt and store structured metadata and build a retrieval index to obtain ciphertext, retrieval index and structured metadata;

[0074] The version management and access control module is connected to the encrypted storage and retrieval module and is used to manage document versions and control user access permissions based on ciphertext, retrieval index and structured metadata.

[0075] In some embodiments, such as Figure 2 As shown, the deep learning OCR recognition module includes:

[0076] The surface fitting and correction unit is used to construct a parametric surface model to perform geometric correction on the preprocessed image and output the corrected image features.

[0077] A multi-scale feature extraction unit, connected to the surface fitting and correction unit, is used to extract local and global features of the image based on the corrected image features and output a feature vector sequence.

[0078] The temporal modeling unit, connected to the multi-scale feature extraction unit, is used to perform temporal modeling on the feature vector sequence and output a predicted distribution.

[0079] The CTC decoding unit, connected to the time-series modeling unit, is used to convert the predicted distribution into a text sequence.

[0080] In some embodiments, such as Figure 3 As shown, the semantic understanding and classification module includes:

[0081] BERT encoding units are used to perform semantic encoding on the text sequence and output a context representation vector;

[0082] The surface classifier unit, connected to the BERT encoding unit, is used to construct a surface classification model with local weighted regression based on the context representation vector, classify documents, and output category labels;

[0083] An entity recognition unit, connected to the BERT encoding unit, is used to identify named entities from a text sequence and output entity information.

[0084] In some embodiments, the metadata automatic extraction module includes:

[0085] The template matching unit is used to select the corresponding metadata extraction template based on the category labels output by the semantic understanding and classification module.

[0086] The information extraction unit, connected to the template matching unit, is used to extract metadata by combining text sequences and entity information through regular expressions and semantic reasoning.

[0087] The metadata verification unit, connected to the information extraction unit, is used to perform format standardization and integrity verification on the extracted metadata and output structured metadata.

[0088] In some embodiments, such as Figure 4 As shown, the encrypted storage and retrieval module includes:

[0089] The feature generation unit is used to extract document feature vectors from structured metadata;

[0090] The surface key generation unit, connected to the feature generation unit, is used to construct a polynomial key surface based on the document feature vector, and to generate an encryption key by selecting sampling points based on the key surface;

[0091] The data encryption unit, connected to the surface key generation unit, is used to encrypt document data using an encryption key and output ciphertext.

[0092] The index building unit, connected to the feature generation unit, is used to build a retrieval index based on document feature vectors.

[0093] In some embodiments, the surface fitting correction unit specifically includes:

[0094] The feature point detection subunit is used to detect text edge feature points in the preprocessed image;

[0095] The surface parameter calculation subunit is connected to the corner detection subunit and is used to calculate NURBS surface parameters based on text edge feature points.

[0096] The image correction subunit is connected to the surface parameter calculation subunit and is used to perform geometric transformations on the preprocessed image based on the surface parameters, and output the corrected image features.

[0097] In some embodiments, the surface classifier unit specifically includes:

[0098] The feature mapping subunit is used to map the context representation vector to the feature space;

[0099] The local weighted regression calculation subunit is connected to the feature mapping subunit and is used to construct a surface classification model with local weighted regression in the feature space.

[0100] The category prediction subunit, connected to the local weighted regression calculation subunit, is used to calculate the probability of a document belonging to each category based on the surface classification model of local weighted regression, and output the category label.

[0101] In some embodiments, the surface key generation unit specifically includes:

[0102] The polynomial generation subunit is used to generate a polynomial key surface based on the document feature vector;

[0103] A sampling point subunit, connected to the polynomial generation subunit, is used to select feature sampling points on the polynomial key surface;

[0104] A key derivation subunit is connected to the sampling point subunit and is used to generate an encryption key based on feature sampling points through the HMAC function.

[0105] The polynomial generation subunit uses a neural network to learn the mapping relationship between document feature vectors and the coefficients of the polynomial key surface.

[0106] In some embodiments, the index building unit specifically includes:

[0107] The keyword extraction subunit is used to extract keywords from the document feature vector;

[0108] The inverted index generation sub-unit is connected to the keyword extraction sub-unit to build an inverted index from keywords to documents;

[0109] The semantic index enhancement subunit, connected to the inverted index generation subunit, is used to semantically expand the inverted index by combining category labels and entity information to obtain the retrieval index.

[0110] In some embodiments, the version management and access control module includes:

[0111] The version record unit is used to store different versions of the ciphertext, search index, and structured metadata output by the encryption storage and retrieval module;

[0112] The permission allocation unit, connected to the version record unit, is used to allocate access permissions for each version of the document based on user roles.

[0113] The operation audit unit, connected to the permission allocation unit, is used to record user access to and modification operations on document versions.

[0114] The technical concept of this invention is as follows:

[0115] 1. Scanning and Acquisition Module

[0116] Function: As the starting point for digitizing paper documents, the scanning and acquisition module is responsible for quickly and efficiently converting paper documents into digital images, while performing basic geometric corrections on the images to provide high-quality data for subsequent processing. This module supports scanning various document types (single pages, booklets, different sizes) and features automatic detection and parameter adaptation capabilities. Original paper documents include contracts, reports, invoices, and other paper-based media.

[0117] Processing flow:

[0118] Multi-mode scanning adaptation:

[0119] Dynamic resolution adjustment: The system automatically selects the scanning resolution based on the document type. For example, for engineering drawings containing detailed diagrams, it automatically switches to 600 DPI; while ordinary office documents use 300 DPI to balance quality and storage costs. Users can also manually set the resolution to meet specific needs.

[0120] Document type recognition: By detecting physical features such as document thickness and material through sensors, and combining them with existing edge detection algorithms to analyze the document outline, the document type can be determined.

[0121] Color mode selection: Automatically selects color, grayscale, or black and white mode based on the document content. For documents containing color charts or photos, color mode is used; for plain text documents, black and white mode is preferred to reduce file size.

[0122] Real-time image correction:

[0123] Tilt detection and correction: The Canny edge detection algorithm is used to extract the document's edge contours, and Hough transform is used to detect straight lines and calculate the document's tilt angle. For example, if a 5° tilt is detected, an affine transform is applied to rotate and correct the image, ensuring that the text lines are horizontally aligned and preventing character distortion in subsequent OCR recognition.

[0124] Perspective Correction: For perspective distortion caused by scanning angle tilt, a four-point perspective transformation algorithm is used. Users can manually mark the four corner points of the document, and the system automatically calculates the transformation matrix to convert irregular images into rectangles, restoring the original shape of the document.

[0125] Intelligent optimization of scanning parameters:

[0126] Exposure compensation: By analyzing the image's grayscale histogram, it determines whether there is overexposure or underexposure. If the histogram is concentrated on the left, it indicates that the image is too dark, and the exposure value is automatically increased; if it is concentrated on the right, the exposure value is decreased to ensure that the contrast between text and background is appropriate.

[0127] Moiré removal: To address moiré interference generated during the scanning of printed materials, Fourier transform is used to convert the image to the frequency domain, identify and filter out high-frequency interference signals, and then inverse Fourier transform is used to restore the image, improving clarity.

[0128] The final corrected image is obtained.

[0129] Technical details:

[0130] On the hardware side, it adopts an industrial-grade sheet-fed scanner (such as the Fujitsu fi-7900), supports A3 size, has a scanning speed of up to 200 pages per minute, and is equipped with ultrasonic double-sheet detection function to avoid image overlap problems caused by scanning multiple pages at the same time.

[0131] On the software side, it integrates TWAIN (Image Device Interface Standard) and WIA (Windows Image Acquisition) drivers, is compatible with mainstream scanning devices on the market, and provides a unified scanning control interface.

[0132] 2. Intelligent Image Preprocessing Module

[0133] Function: Performs deep optimization on the corrected image output by the scanning acquisition module, eliminating noise, enhancing text contrast, and unifying image format, providing ideal input for the deep learning OCR recognition module and significantly improving text recognition accuracy.

[0134] Processing flow:

[0135] Multi-stage noise filtering:

[0136] Gaussian noise suppression: To address Gaussian blurring caused by sensor thermal noise during scanning, Gaussian filters (such as 3×3 or 5×5 kernels) are applied for convolution operations to smooth the image and reduce random pixel interference.

[0137] Salt and pepper noise removal: A median filtering algorithm is used to select a neighborhood (e.g., a 5×5 window) centered on each pixel and replace the pixel value with the median of its neighboring pixels. For black and white binary images, this effectively eliminates isolated white or black noise.

[0138] Noise removal and scratch repair: Connected component analysis is used to identify and remove noise regions with an area smaller than a threshold; for thin scratches, morphological opening operations (erosion + dilation) are used for repair to maintain the continuity of text strokes.

[0139] Adaptive binarization:

[0140] Global Thresholding (Otsu): For documents with uniform lighting, the Otsu algorithm calculates the inter-class variance of the image's grayscale histogram, automatically determines the optimal threshold, and converts the image into a binary black and white image. For example, for plain text documents, the Otsu algorithm can effectively separate text from the background.

[0141] Local thresholding (Niblack): For images with uneven illumination (such as shadows during scanning), the image is divided into multiple sub-blocks, and a local threshold is calculated for each sub-block to achieve adaptive binarization and avoid text loss or background residue caused by global thresholding.

[0142] Text enhancement and sharpening:

[0143] Contrast Enhancement: Histogram equalization is applied to expand the range of pixel value distribution, making the grayscale difference between text and background more obvious. For low-contrast documents (such as photocopies), CLAHE (Limited Contrast Adaptive Histogram Equalization) is used to enhance contrast in local areas while preventing over-enhancement that could amplify noise.

[0144] Edge sharpening: Using the Laplacian operator or the Unsharp Mask algorithm to enhance text edges and highlight stroke details. For example, for blurry handwritten text, sharpening can make stroke boundaries clearer, facilitating OCR recognition.

[0145] Image normalization:

[0146] Size adjustment: Scaling the images to a specified resolution (e.g., 200×200dpi) ensures consistent image size across different source documents, reducing the differences in OCR model input.

[0147] Format conversion: Convert images to standard formats supported by the OCR model (such as PNG, TIFF), remove unnecessary metadata, reduce file size, and improve processing efficiency.

[0148] The final preprocessed image is obtained.

[0149] Technical Details: The image processing algorithm is implemented using the OpenCV library and accelerated by CUDA technology, supporting GPU parallel computing and significantly improving the speed of large-scale image preprocessing. An automatic parameter tuning mechanism is designed to dynamically select the most suitable denoising intensity, binarization threshold, and other parameters by analyzing image features (such as grayscale distribution and texture complexity), without requiring manual intervention.

[0150] 3. Deep Learning OCR Recognition Module

[0151] In the process of digitizing paper documents, scanned images often exhibit text distortion due to paper wrinkles, scanning device tilt, and other issues, affecting recognition accuracy. The deep learning OCR recognition module effectively solves this problem by innovatively integrating surface fitting technology with an improved CRNN model.

[0152] Improved CRNN model construction:

[0153] First, clarify the definitions of the relevant symbols: The input is a scanned image of a paper document, with a height of H, a width of W, and a number of channels of TD. Generally, for color images, TD = 3 (RGB three channels), and for grayscale images, TD = 1. It is the image after surface fitting correction; P = {p1, p2, ..., p...} nLet} be the set of feature points detected from image I, and p i =(x i y i , z i () represents the coordinates of each feature point in three-dimensional space; these points are typically selected at key locations such as text edges and corners; f CNN (·) is the CNN feature extraction function; Let be the hidden state of the RNN at time t; L represents the predicted character distribution at time t, where L is the size of the character set. For example, in Chinese character recognition, L may cover thousands of commonly used Chinese characters, punctuation marks, and other characters.

[0154] The model building process is as follows:

[0155] Surface fitting correction unit:

[0156] The feature point detection subunit is used to detect text edge feature points in the preprocessed image: For the input scanned image I, the first step is to use a corner detection algorithm (such as Harris corner detection). This algorithm determines the corner position by calculating the eigenvalues ​​of the autocorrelation matrix based on the local gray-level changes of the image, thereby extracting the feature point set P.

[0157] The surface parameter calculation subunit is used to calculate NURBS surface parameters based on text edge feature points. Next, the Non-Uniform Rational B-Spline (NURBS) surface fitting technique is used to construct the surface function.

[0158]

[0159] Where, N i，k (u) and N j，l (v) are B-spline basis functions, which determine the local shape and smoothness of the surface; d ij These are control points, and their positions directly affect the overall shape of the surface; w ij , where is the weight, used to adjust the degree of influence of the control points on the surface; u, v∈[0,1] are the parameter coordinates.

[0160] The surface parameters are solved by minimizing the error function:

[0161] Where p k For feature points, λΩ(d) ij ) represents the smoothing regularization term, and λ is the regularization parameter.

[0162] The image correction subunit performs geometric transformations on the preprocessed image based on surface parameters, outputting the corrected image features: To ensure the constructed surface accurately fits the actual deformation of the document, the least squares method is used, by continuously adjusting d... ij and wij Minimize the error between the surface and the feature points, and finally output the corrected image. To restore distorted text to its normal form.

[0163] Multi-scale feature extraction unit: extracts the corrected image Inputting data into a CNN network typically involves multiple convolutional layers, pooling layers, and activation layers. Convolutional layers slide different kernels across the image to extract local features such as edges and textures. Pooling layers downsample the feature maps, reducing data volume while preserving key features. Activation layers introduce non-linearity into the network, enhancing its expressive power. After multiple layers of processing, a feature sequence is extracted. Where F is the feature sequence and T1 is the time length, these feature sequences contain key information about the text in the image.

[0164] Temporal Modeling Unit: To process the temporal information in the feature sequence, a bidirectional LSTM (BiLSTM) is used to model the feature sequence. BiLSTM consists of two LSTMs operating in opposite directions, enabling simultaneous processing of the sequence from both forward and backward directions, thus better capturing contextual information. At time t, through h... t =BiLSTM(f t ,h t-1 Calculate the hidden state h t , where f t h is the feature vector at the current time. t-1 This is the hidden state from the previous time step. Then, the predicted character distribution o is calculated using a fully connected layer and a Softmax function. t =Softmax(W h h t +b), where W h b are parameters of the fully connected layer, o t This represents the probability of each character appearing at time t.

[0165] CTC Decoding Unit: Since the OCR recognition output is a character sequence and the alignment between the input image and the output text is unclear, the Connectionist Temporal Classification (CTC) loss function is used to calculate the difference between the prediction and the true label. Where P(y|o) is the conditional probability of the true label y given the output o. By minimizing this loss function, the model parameters are continuously adjusted so that the model can recognize text more accurately.

[0166] The model training process is as follows:

[0167] The training process relies on a large number of labeled image datasets. Among them I i For the scanned image, y iThese correspond to text labels. To enhance the model's generalization ability, data augmentation operations are performed on the images before training, including rotation, scaling, and distortion, to simulate various deformations that may occur in real-world scenes. During the training phase, the Adam optimizer is used to update the model parameters θ by minimizing the CTC loss function. Meanwhile, during backpropagation, not only are the parameters of the CRNN model updated, but the control points and weights of the NURBS surface are also optimized, so that the surface fitting can better serve text recognition and further improve the accuracy of the model.

[0168] The model application process is as follows:

[0169] In practical applications, the input scanned image I first passes through a surface fitting correction layer to correct text distortion in the image, resulting in... Then The data is fed into a CNN for feature extraction, then into an RNN for sequence modeling and prediction. Finally, CTC decoding is used to convert the predicted character distribution into the final text T, which will be used as output and passed to the subsequent semantic understanding and classification modules to provide a foundation for the semantic analysis of the document.

[0170] 4. Semantic Understanding and Classification Module

[0171] The main task of the semantic understanding and classification module is to perform in-depth analysis of the text recognized by OCR, understand its semantic information, and classify the documents into the corresponding categories. This module, through an improved BERT+surface fitting model, can more accurately capture the semantic features and category distribution of the text.

[0172] Improved BERT+ surface fitting model construction:

[0173] Clarify the meaning of the symbols: T represents the text sequence output by the OCR recognition module; It is the word vector of the i-th word in the text sequence, which maps the word to a d-dimensional vector space to represent its semantic information; Z represents the [CLS] tag vector output by the BERT model, where D is the hidden layer dimension. The [CLS] tag output incorporates the semantic information of the entire text; Z = {z1, z2, ..., z...} n} is from h [CLS] The resulting set of feature vectors; f surf (Z) is the surface fitting function; c represents the document category label; p(c|Z) is the probability that the text belongs to category c.

[0174] The model is constructed as follows:

[0175] BERT encoding unit: performs word segmentation on the text sequence T output by OCR, converting each word into a corresponding word vector e. i This forms a word vector sequence [e1, e2, ..., e] m The word vector sequence is then input into a pre-trained BERT model. Based on the Transformer architecture, the BERT model utilizes multi-head self-attention to simultaneously focus on information from different locations within the text, capturing long-distance dependencies and complex semantic structures. After processing through multiple Transformer layers, the hidden state [h1, h2, ..., h] of each word is output. m Take the output h corresponding to the [CLS] marker. [CLS] =h1, as the overall semantic representation of the entire text, this vector contains the key semantic information of the text.

[0176] Surface classifier unit:

[0177] The feature mapping subunit is used to map the context representation vector to the feature space; specifically, to better analyze the distribution of text semantic features, h... [CLS] It can be mapped to a set of feature vectors Z through linear transformation or other feature engineering methods.

[0178] The locally weighted regression computation unit is used to construct a locally weighted regression surface classification model in the feature space; specifically, it uses a locally weighted regression surface fitting method to fit the feature vectors. For each point z in the feature space, the predicted value... pass Calculate, where the weight w i (z) by kernel function Let z ∈ Z, and σ be the bandwidth parameter of the kernel function, which controls the rate of weight decay. This kernel function makes feature points z closer to the prediction point z more likely to be affected. i The greater the weight of the value, the greater its influence on the predicted value, thus constructing a fitting surface that can reflect the distribution of text semantic features.

[0179] The category prediction subunit is used to calculate the probability of a document belonging to each category based on a surface classification model with local weighted regression, and outputs the category label; specifically, it calculates the probability of a document belonging to each category based on the surface fitting result f. surf (Z) is input to the fully connected layer, which passes through the weight matrix Wf and the bias vector b. f The input is linearly transformed, and then the output is converted into a probability distribution of the document belonging to each category using the Softmax function: p(c|Z) = Softmax(W). f f surf (Z)+b f ), through learning to adjust W f and b fThese parameters enable the model to accurately output the probability that a document belongs to different categories based on the features of the fitted surface.

[0180] Model training process:

[0181] The training data uses a labeled text dataset. Where T i It is the text sequence after OCR recognition, c i Here, M represents the corresponding document category label, and M is the number of training samples. The training process consists of two main stages:

[0182] Pre-trained BERT fine-tuning: First, load the pre-trained BERT model parameters θ. BERT These parameters are obtained through pre-training on a large-scale corpus and possess powerful semantic representation capabilities. Then, they are applied to the labeled dataset for the current document classification task. Fine-tuning was performed on the text sequence T. i Input the BERT model and compute the output labeled [CLS]. By minimizing the cross-entropy loss function To update the BERT model parameters, where As an indicator function, when c = c i The value is 1 if the condition is met, and 0 otherwise. The BERT model parameters θ are continuously adjusted using the backpropagation algorithm. BERT This allows it to adapt to the semantic features of a specific document classification task.

[0183] Surface fitting parameter learning: After fine-tuning the BERT model, fix the BERT parameters and focus on learning the surface fitting function f. surf The parameters of (Z) are obtained by minimizing the classification loss function. To learn about surface parameters, among which Let θ be the cross-entropy loss function. surf Let z be the set of parameters for the surface fitting function. i It is by The transformed feature vector. During training, the surface parameters are continuously adjusted to ensure that the fitted surface better reflects the relationship between text semantic features and categories, thereby further improving the accuracy of document classification.

[0184] Model application process:

[0185] In practical applications, the text T output by the OCR recognition module is first segmented and converted into word vectors according to the training process, and then input into the fine-tuned BERT model to obtain the output h of the [CLS] tag. [CLS] This serves as the overall semantic representation of the text. Next, h... [CLS]Converted into a feature vector z, and fitted by a trained surface fitting function f surf (z) Calculate the output of the fitted surface. Input this output into the classifier to calculate the probability distribution p(c|z) of the document belonging to each category.

[0186] Entity recognition unit: Used to identify named entities from a text sequence and output entity information.

[0187] Finally, the category with the highest probability is selected as the prediction result. Entity information, text T, and category labels Together, they are output and passed to the encrypted storage and retrieval module to provide the necessary information for the encrypted storage of documents.

[0188] 5. Automatic Metadata Extraction Module

[0189] Functionality: This module automatically extracts key metadata from the text, category, and entity information output by the semantic understanding and classification module, constructing structured descriptions to provide a foundation for document retrieval, management, and analysis. It supports metadata extraction from multiple domains and templates, offering high flexibility and scalability.

[0190] Processing flow:

[0191] Template matching unit:

[0192] Category-driven template selection: Based on document category Retrieve the corresponding metadata extraction template from the template library. For example, when When loading the contract template, which defines required metadata fields such as Party A, Party B, contract amount, and signing date; if Then select an invoice template, which includes fields such as invoice number, invoice date, amount, and seller.

[0193] Hybrid Template Strategy: For complex documents (such as comprehensive documents containing contract terms and financial data), the system supports the combined use of multiple templates. The system uses semantic analysis to determine the category of different paragraphs, applies the corresponding templates to extract metadata, and then integrates the results.

[0194] Information extraction unit:

[0195] Regular expression matching:

[0196] Date extraction: Use regular expressions to match various date formats, such as \d{4}-\d{2}-\d{2} (ISO8601 format), \d{4} year \d{1,2} month \d{1,2} day (Chinese date), [A-Za-z]{3}\d{1,2}, \d{4} (English date). For example, extract "2023-10-15" from the text "The contract was signed on October 15, 2023".

[0197] Number extraction: For fields with fixed formats such as invoice numbers and contract numbers, use regular expressions to match. For example, an invoice number is usually a combination of "numbers + letters", and it can be accurately extracted by \d{8}[A-Z]{2}.

[0198] Enhanced named entity recognition (NER):

[0199] Person name / Organization name extraction: Use a pre-trained NER model to identify relevant entities from text T and entity information. For example, extract the organization name "XX Technology Co., Ltd." and the person name "Zhang San" from "Party A: XX Technology Co., Ltd., Person in charge: Zhang San". <00​​​​​​​​​​​​​​​​​​​​​​Technical details: A rule engine (such as Drools) is used to implement template matching and logical judgment, supporting the dynamic addition and modification of template rules. Combined with knowledge graph technology, the relationships between existing metadata are utilized to assist in inferring missing information. For example, given a company's contract, other contracts of the same company can be linked through the knowledge graph to infer similar signing dates or cooperation terms.

[0208] 6. Encrypted storage and retrieval module

[0209] The encrypted storage and retrieval module is responsible for securely encrypting and storing the semantically understood and categorized document data, and quickly and accurately retrieving relevant data when needed by the user. This module achieves a deep correlation between data encryption and document features through a surface key binding encryption model, improving data security and retrieval efficiency.

[0210] Curved surface key binding encryption model:

[0211] The relevant symbols are defined as follows: D represents the text output by the semantic understanding module and its category labels; D is the... The serialized original data block; S is the encryption key, a vector of length k; φ(D) is the data feature extraction function; S K (u, v) is the key surface; E K (D) is the encryption function; C represents the ciphertext; HMAC(m, s) is the hash message authentication code function; a ij These are the polynomial surface fitting coefficients.

[0212] The model building process is as follows:

[0213] Feature generation unit: generates the text T and its category labels output by the semantic understanding and classification module. The data is serialized into a single data block D. Then, using the data feature extraction function φ(D), key information representing the data characteristics is extracted from the data block D to generate a feature vector x = [x1, x2, ..., x...]. n The feature extraction process may include operations such as text segmentation, keyword extraction, and category label encoding. The extracted features cover the semantic information and category attributes of the text, and these features will serve as an important basis for generating the key surface.

[0214]

[0215] TF-IDF(T) is a commonly used weighting technique in information retrieval and text mining, namely Term Frequency-Inverse Document Frequency. It measures the importance of a word to a document by calculating the frequency of each word in a text T (Term Frequency, TF) and combining this with the word's rarity within the entire document set (Inverse Document Frequency, IDF). For example, in a document about digitized archives, the word "archives" may appear frequently and be common in such document sets, while a word like "archives classification" may appear less frequently but is uncommon in the document set, thus having a higher TF-IDF value and better reflecting the characteristics of that document.

[0216] Document category tags Perform one-hot encoding. Assume document categories include "contract," "official document," and "report," etc. If the category is "contract", it is encoded as [1, 0, 0] (assuming there are only these three categories). This encoding method can convert the category label into a vector form that is easy for computers to process, which facilitates the calculation and processing of subsequent models.

[0217] EntityType(E): Here, E represents the entity information in the text T, and EntityType(E) indicates the extraction of the type of these entities. For example, entities such as names of people, places, and times mentioned in the text can be identified and their features extracted to further enrich the feature representation of the data.

[0218] Surface key generation unit:

[0219] The polynomial generation subunit is used to generate a polynomial key surface based on the document feature vector; specifically, it constructs the key surface S based on the extracted feature vector x using polynomial surface fitting techniques. K (u, v). Its expression is: Where p and q are the highest powers of the polynomial in the u and v directions, respectively, and the coefficient a ij The values ​​of the key surface are generated by the data feature vector x through specific algorithmic rules. For example, the feature vector x can be input into a linear transformation matrix or a neural network, outputting a set of coefficient values ​​to determine the shape of the key surface. Different data features will generate surfaces of different shapes, ensuring the uniqueness of the key surface corresponding to each data block and guaranteeing encryption security. In other words, the polynomial generator subunit uses a neural network to learn the mapping relationship between the document feature vector and the coefficients of the polynomial key surface.

[0220] The sampling point subunit is used to select feature sampling points on the polynomial key surface; specifically, on the generated key surface SK After (u, v), specific sampling points (u0, v0) are selected from the surface according to preset rules. The sampling points can be selected by center sampling, random sampling, etc.

[0221] The key derivation subunit is used to generate an encryption key based on feature sampling points using the HMAC function; specifically, it uses the Hash Message Authentication Code (HMAC) function to extract the surface value S of the sampling points. K (u0, v0) is combined with the preset salt value salt to generate the encryption key K, i.e., K = HMAC(S K (u0, v0), salt). This approach makes key generation highly dependent on data characteristics and surface properties. Even if the same data is stored at different times, different keys will be generated due to the randomness of surface sampling points or other factors, further enhancing the security of encryption.

[0222] Data encryption unit: Using the generated key K, the original data D is encrypted using the Advanced Encryption Standard (AES-256) encryption algorithm to obtain ciphertext C, i.e., C = E. K (D) = AES-256(D, K). AES-256 is a symmetric encryption algorithm with high encryption strength, which can effectively protect the security of data during storage and transmission.

[0223] Index building blocks: the encrypted ciphertext C and the relevant parameters of the key surface (such as polynomial coefficients a). ij Sampling points (u0, v0, etc.) will be stored in a distributed file system, and an index will be built based on the document feature vector for subsequent retrieval. Specifically:

[0224] The keyword extraction subunit is used to extract keywords from the document feature vector;

[0225] The inverted index generation sub-unit is used to build an inverted index from keywords to documents;

[0226] The semantic indexing enhancement subunit is used to semantically expand the inverted index by combining category labels and entity information to obtain a retrieval index.

[0227] Model application process:

[0228] Storage phase: When the output of the semantic understanding and classification module is received... First, the data is serialized to obtain data block D. Then, the data feature extraction function φ(D) is called to obtain the data feature vector x. Based on the feature vector x, a key surface S is generated using a polynomial surface fitting algorithm. K (u,v), determine the coefficients a of the surface. ij Next, from the key surface S KA sampling point (u0, v0) is selected on (u, v), and the encryption key K is calculated using the HMAC function. The original data D is then encrypted using AES-256 with key K to obtain ciphertext C. Finally, the ciphertext C and the key surface coefficients a are... ij The sampling points (u0, v0) and other relevant metadata (such as document category, encryption time, etc.) are stored in a distributed file system and a detailed index is built to facilitate fast retrieval later.

[0229] Retrieval Phase: When a user submits a retrieval request Q, the system first parses the retrieval request and extracts features to generate a query feature vector x. q This process uses natural language processing technology to analyze search keywords, semantic intent, and other factors to extract key features that represent the search needs.

[0230] Next, based on the query feature vector x q The query surface S is constructed using the same polynomial surface fitting technique as the storage stage. Q (u, v). The method for determining its coefficients is similar to that of the key surface. By inputting the query feature vector into a specific algorithm, corresponding polynomial coefficients are generated, thereby determining the shape of the query surface. Then, the system retrieves the ciphertext C related to the retrieval conditions and the corresponding key surface parameters (coefficients a) from the stored encrypted data. ij Sampling points (u0, v0, etc.) are used to reconstruct the original key surface S based on these parameters. K (u, v).

[0231] To determine the degree of matching between the retrieval request and the original data, the system calculates the query surface S. Q (u, v) and the original key surface S K The matching degree of (u, v) is calculated using the following formula:

[0232]

[0233] Among them, (u i v i Let S be N sampling points uniformly selected within the domain of the surface. This formula measures the similarity between two surfaces by calculating the average of the absolute values ​​of the differences between them at multiple sampling points. The smaller the matching score, the more similar the two surfaces are, meaning the higher the relevance between the retrieval request and the original data. When the matching score exceeds a preset threshold, relevant data is considered retrieved. At this point, the system uses the original key surface S... K Given (u,v) and sampling points (u0,v0), recalculate the encryption key K. Specifically, this involves using the previous HMAC function to... K (u0, v0) and the salt value salt are calculated again to obtain the key K.

[0234] Finally, the ciphertext C is decrypted using key K, and the decrypted data is returned to the user, completing the retrieval operation. The entire retrieval process fully utilizes surface fitting technology, performing deep matching between the retrieval request and the features of the encrypted data. This not only improves retrieval accuracy but also ensures that only requests meeting the criteria can retrieve the corresponding data, further enhancing data security and privacy.

[0235] 7. Version Management and Access Control Module

[0236] Functionality: Enables version management throughout the entire document lifecycle, recording each modification and operation information; simultaneously, based on a role-based access control (RBAC) model, it ensures the security and compliance of documents during storage, retrieval, and use.

[0237] Processing flow:

[0238] Fine-grained version management:

[0239] Version Creation and Updates: When a document is first stored, an initial version is generated, containing encrypted document data, a search index, and metadata, recording information such as creation time and creator. When the document is modified, a new version is automatically created. The system calculates the hash values ​​of the old and new version data using a hash algorithm (such as SHA-256), compares the differences, and only stores the changed parts, reducing storage space usage. For example, if only one paragraph of the document is modified, the new version only records the updated content of that paragraph and the version number.

[0240] Version Retrospection and Comparison: Users can select historical versions by version number or timestamp. The system decrypts the encrypted text of the corresponding version and displays the document content. A visual version comparison function is provided, highlighting added, deleted, and modified content between different versions, allowing users to quickly understand the modification details. For example, new content (green), deleted content (red), and modified content (yellow) are marked with colors.

[0241] Version strategy configuration: Supports custom version retention policies, such as automatically cleaning up expired versions by time (retaining versions from the most recent 30 days) or by quantity (retaining the latest 5 versions) to free up storage space.

[0242] Multi-level access control:

[0243] Role-based access control model implementation:

[0244] Role definition: The system predefines various roles, such as "System Administrator" (with full permissions), "Department Manager" (can read and edit documents in their department), and "Regular Employee" (can only read some public documents), and supports user-defined roles.

[0245] Permission Assignment: Administrators can assign roles to users, and users inherit the permissions of the corresponding roles; they can also set individual permissions for users to achieve more granular control. For example, granting a specific employee temporary access to a confidential document.

[0246] Dynamic permission verification: The system verifies permissions in real time when a user accesses a document. For example, when a user requests to download a document, it checks whether they have download permissions; if they attempt to edit a document, it verifies whether they are the document owner or authorized to edit. For sensitive operations (such as deleting documents or modifying permissions), a two-factor authentication mechanism is enabled, using SMS verification codes or hardware tokens to confirm user identity and prevent abuse of permissions.

[0247] Data classification protection: Different access permissions are automatically assigned based on document category and metadata (such as "confidential" or "public" labels). For example, contract documents marked "confidential" can only be accessed by specific user groups.

[0248] Full audit log:

[0249] Operation Log: Records all user operations, including timestamps, user IDs, operation types (read, edit, delete, share), accessed document IDs and version numbers, request parameters, and other detailed information. For example, it records "User A read document ID-123 version v3 on 2024-01-01 10:00:00".

[0250] Security Analysis: By analyzing audit logs, abnormal operational behavior is detected. For example, if a user frequently attempts to access unauthorized documents within a short period, the system automatically triggers an alert and restricts that user's login access.

[0251] Compliance support: Audit logs meet compliance requirements (such as the General Data Protection Regulation and the Cybersecurity Classified Protection 2.0 standard), providing enterprises with traceability evidence for data operations and facilitating responses to regulatory inspections.

[0252] Technical details:

[0253] Access control is implemented using the Apache Shiro framework, supporting integration with existing enterprise authentication systems such as LDAP and Active Directory to achieve single sign-on (SSO). Version data is stored in a distributed version control system (such as Git-LFS), leveraging its efficient version management and differential storage features to ensure rapid retrieval and storage of massive document versions.

[0254] The final output is a securely stored collection of documents, supporting version tracking, access control, and audit logs, meeting enterprises' security and compliance needs for full lifecycle document management.

[0255] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0256] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0257] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0258] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0259] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0260] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0261] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0262] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0263] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0264] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A digital archive platform based on the digitization of paper documents, characterized in that, include: The scanning and acquisition module is used to convert paper documents into digital images and perform geometric correction to obtain corrected images. The intelligent image preprocessing module is connected to the scanning acquisition module and is used to perform noise filtering, binarization and enhancement processing on the corrected image to obtain a preprocessed image. A deep learning OCR recognition module is connected to the intelligent image preprocessing module and is used to extract text sequences from the preprocessed image; The semantic understanding and classification module is connected to the deep learning OCR recognition module and is used to analyze the semantics of the text based on the text sequence and classify the document to obtain category labels and entity information; The metadata automatic extraction module, connected to the semantic understanding and classification module, is used to extract structured metadata from text sequences, category labels, and entity information. The encrypted storage and retrieval module is connected to the automatic metadata extraction module and is used to encrypt and store structured metadata and build a retrieval index to obtain ciphertext, retrieval index and structured metadata; The version management and access control module is connected to the encrypted storage and retrieval module and is used to manage document versions and control user access permissions based on ciphertext, retrieval index and structured metadata; The encrypted storage and retrieval module includes: The feature generation unit is used to extract document feature vectors from structured metadata; The surface key generation unit, connected to the feature generation unit, is used to construct a polynomial key surface based on the document feature vector, and to generate an encryption key by selecting sampling points based on the key surface; The data encryption unit, connected to the surface key generation unit, is used to encrypt document data using an encryption key and output ciphertext. The index building unit, connected to the feature generation unit, is used to build a retrieval index based on document feature vectors.

2. The digital archive platform based on the digitization of paper documents according to claim 1, characterized in that, The deep learning OCR recognition module includes: The surface fitting and correction unit is used to construct a parametric surface model to perform geometric correction on the preprocessed image and output the corrected image features. A multi-scale feature extraction unit, connected to the surface fitting and correction unit, is used to extract local and global features of the image based on the corrected image features and output a feature vector sequence. The temporal modeling unit, connected to the multi-scale feature extraction unit, is used to perform temporal modeling on the feature vector sequence and output a predicted distribution. The CTC decoding unit, connected to the time-series modeling unit, is used to convert the predicted distribution into a text sequence.

3. The digital archive platform based on the digitization of paper documents according to claim 2, characterized in that, The semantic understanding and classification module includes: BERT encoding units are used to perform semantic encoding on the text sequence and output a context representation vector; The surface classifier unit, connected to the BERT encoding unit, is used to construct a surface classification model with local weighted regression based on the context representation vector, classify documents, and output category labels; An entity recognition unit, connected to the BERT encoding unit, is used to identify named entities from a text sequence and output entity information.

4. The digital archive platform based on the digitization of paper documents according to claim 3, characterized in that, The automatic metadata extraction module includes: The template matching unit is used to select the corresponding metadata extraction template based on the category labels output by the semantic understanding and classification module. The information extraction unit, connected to the template matching unit, is used to extract metadata by combining text sequences and entity information through regular expressions and semantic reasoning. The metadata verification unit, connected to the information extraction unit, is used to perform format standardization and integrity verification on the extracted metadata and output structured metadata.

5. The digital archive platform based on the digitization of paper documents according to claim 2, characterized in that, The surface fitting correction unit specifically includes: The feature point detection subunit is used to detect text edge feature points in the preprocessed image; The surface parameter calculation subunit is connected to the corner detection subunit and is used to calculate NURBS surface parameters based on text edge feature points. The image correction subunit is connected to the surface parameter calculation subunit and is used to perform geometric transformations on the preprocessed image based on the surface parameters, and output the corrected image features.

6. The digital archive platform based on the digitization of paper documents according to claim 3, characterized in that, The surface classifier unit specifically includes: The feature mapping subunit is used to map the context representation vector to the feature space; The local weighted regression calculation subunit is connected to the feature mapping subunit and is used to construct a surface classification model with local weighted regression in the feature space. The category prediction subunit, connected to the local weighted regression calculation subunit, is used to calculate the probability of a document belonging to each category based on the surface classification model of local weighted regression, and output the category label.

7. The digital archive platform based on the digitization of paper documents according to claim 1, characterized in that, The surface key generation unit specifically includes: The polynomial generation subunit is used to generate a polynomial key surface based on the document feature vector; A sampling point subunit, connected to the polynomial generation subunit, is used to select feature sampling points on the polynomial key surface; A key derivation subunit, connected to the sampling point subunit, is used to generate an encryption key based on feature sampling points using the HMAC function; The polynomial generation subunit uses a neural network to learn the mapping relationship between document feature vectors and the coefficients of the polynomial key surface.

8. The digital archive platform based on the digitization of paper documents according to claim 1, characterized in that, The index construction unit specifically includes: The keyword extraction subunit is used to extract keywords from the document feature vector; The inverted index generation sub-unit is connected to the keyword extraction sub-unit to build an inverted index from keywords to documents; The semantic index enhancement subunit, connected to the inverted index generation subunit, is used to semantically expand the inverted index by combining category labels and entity information to obtain the retrieval index.

9. The digital archive platform based on the digitization of paper documents according to claim 1, characterized in that, The version management and access control module includes: The version record unit is used to store different versions of the ciphertext, search index, and structured metadata output by the encryption storage and retrieval module; The permission allocation unit, connected to the version record unit, is used to allocate access permissions for each version of the document based on user roles. The operation audit unit, connected to the permission allocation unit, is used to record user access to and modification operations on document versions.

Images