Electronic archive management method and system

Abstract

The invention relates to the technical field of electronic archive management, in particular to an electronic archive management method and system, and the method comprises the steps: obtaining a paper archive image, and carrying out the character recognition of the paper archive image based on an image processing technology, so as to obtain a first target text; and receiving electronic files uploaded by a plurality of preset data sources every a preset sampling period, converting the electronic files into a second target text, and preprocessing the first target text and the second target text to obtain a target text. Through the image processing technology and the character recognition technology, the paper archives are converted into the electronic texts, the efficiency of converting the archives from the paper archives into the electronic archives is greatly improved, the workload of manual input is reduced, the electronic archives from multiple data sources can be periodically and automatically received and processed by setting the preset sampling period, and the user experience is improved. Therefore, manual intervention can be avoided, and the real-time performance and comprehensiveness of archive management are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic archive management, and in particular to an electronic archive management method and system. Background Art

[0002] Electronic archives refer to archival materials preserved and managed through digital means. They exist in the form of electronic files and can be stored, managed, retrieved, transmitted and backed up through computer systems, storage devices or networks.

[0003] Traditional methods usually rely on manual input and organization of paper files, which is not only time-consuming and labor-intensive, but also prone to errors. The manual input process takes a long time and may lead to information errors due to human negligence; and traditional methods usually adopt a fixed storage method and lack a mechanism for dynamically allocating storage resources. The storage location of files is often static and cannot be optimized according to the importance of files, resulting in low resource utilization efficiency; and traditional methods lack an intelligent file importance assessment mechanism. The storage and management of documents often rely on manual judgment, and there are no clear standards to ensure that important files are stored and protected first. Therefore, key files may not be managed or backed up in a timely manner, increasing the risk of data loss. Summary of the invention The technical problem to be solved by the present invention is to overcome the disadvantages of the above-mentioned prior art and provide a method and system for managing electronic archives.

[0004] The technical solution adopted to solve the above technical problems is: a method for managing electronic archives, comprising: Acquire a paper archive image, and perform text recognition on the paper archive image based on image processing technology to obtain a first target text; receiving electronic files uploaded by a plurality of preset data sources at every preset sampling period, converting the electronic files into a second target text, and preprocessing the first target text and the second target text to obtain a target text; Performing a word segmentation operation on the target text to obtain a word sequence corresponding to the target text, and classifying the target text based on the word sequence to obtain a text type label corresponding to the target text; Acquiring the importance of the target text based on the text type label corresponding to the target text, and matching the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text; The target text is stored in the matching text storage node.

[0005] Preferably, performing text recognition on the paper archive image based on image processing technology to obtain the first target text includes: Performing image enhancement on the paper archive image to obtain a paper archive enhanced image; Performing text box detection on the paper archive enhanced image based on the trained first object detection model to obtain a plurality of first text boxes corresponding to the paper archive enhanced image, and dividing the plurality of first text boxes into columns to obtain a text box column set; Performing text box detection on the text box column set based on the trained second object detection model to obtain a plurality of second text boxes corresponding to the text box column set, and dividing the plurality of second text boxes into rows to obtain a text box row set; Performing text recognition on the text box row set based on a trained text recognition model to obtain a text recognition result corresponding to the text box row set; The text recognition results are sorted to obtain a first target text.

[0006] Preferably, the text recognition model includes a local feature extraction module, a global feature extraction module, a weighted aggregation module and a classification module, the local feature extraction module is used to extract local image features of the text box row set, the global feature extraction module is used to extract global image features of the text box row set, the weighted aggregation module is used to weightedly aggregate the local image features and the global image features to obtain aggregated image features, and the classification module is used to classify the aggregated image features to obtain the text recognition results corresponding to the text box row set.

[0007] Preferably, the local feature extraction module uses deep separable convolution to extract image features, the global feature extraction module includes a window multi-head attention layer, a sliding window multi-head attention layer and an MLP layer, each MLP layer uses a normalization layer before being connected to the window multi-head attention layer and the sliding window multi-head attention layer, and each layer uses a residual connection, the weighted aggregation module weightedly combines the local image features and the global image features through GCN to obtain aggregated image features, and the classification module classifies the aggregated image features through a fully connected layer and a Softmax function to obtain the text recognition results corresponding to the text box row set.

[0008] Preferably, classifying the target text based on the word sequence to obtain a text type label corresponding to the target text includes: Acquire contextual semantic features and word embedding vectors of the word sequence based on a trained pre-trained model; Calculating the similarity between each word in the word sequence based on the word embedding vector of the word sequence to obtain the similarity between each word in the word sequence, if the similarity is greater than a preset similarity threshold, it indicates that there is a similar relationship between the two words; Constructing a text directed graph based on the word sequence and the similarity relationship between the words, and transferring and aggregating the text directed graph based on a graph convolution operation to obtain text hidden features of the word sequence; The contextual semantic features of the word sequence and the text hidden features are fused to obtain fused features, and the fused features are classified to obtain a text type label corresponding to the target text.

[0009] Preferably, the pre-trained model adopts BERT, and the text directed graph includes multiple nodes and multiple directed edges, wherein the nodes correspond to words in the word sequence, and the directed edges correspond to the similarity relationship between the two words.

[0010] Preferably, the text directed graph is transferred and aggregated based on a graph convolution operation to obtain text hidden features of the word sequence, including: The text directed graph is embedded based on the graph convolution operation to obtain a node representation, wherein the formula of the node representation is as follows: ; in, Represents the first Hidden representation of layer nodes, represents the activation function, Indicates that the column represents the type The adjacency matrix of the text directed graph is Indicates that the column represents the type The first part of the text directed graph Hidden representation of layer nodes, Indicates that the column represents the type The first part of the text directed graph The weight of the layer nodes, Represents different types of nodes; The type-level attention scores of the adjacent nodes are calculated based on the node representation of each node in the text directed graph, wherein the calculation formula of the type-level attention scores is as follows: ; in, represents the type-level attention scores of neighboring nodes, Indicates that the column represents the type The attention vector of Represents a splicing operation, represents the hidden representation of the current node, Indicates that the column represents the type The embedding representation of the neighboring nodes of , represents the adjacent nodes, represents the embedded representation of the directed edges between the current node and the adjacent nodes, Represents the hidden representation of adjacent nodes; The type-level attention weight of the adjacent node is obtained based on the type-level attention score of the adjacent node, wherein the calculation formula of the type-level attention weight is as follows: ; in, represents the type-level attention weight; The node-level attention score of the adjacent node is calculated based on the type-level attention weight of the adjacent node, wherein the calculation formula of the node-level attention score is as follows: ; in, represents the type-level attention weights of neighboring nodes, represents the attention vector; The dual-level attention mechanism of type-level and node-level attention is fused to obtain the text hidden features of the word sequence.

[0011] Preferably, matching the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text includes: Acquire behavior scores of the plurality of text storage nodes based on the behavior transactions of the plurality of text storage nodes; Comparing the behavior scores of the plurality of text storage nodes with a preset behavior score threshold, and if the behavior score of the text storage node is less than the preset behavior score threshold, eliminating the text storage node from the plurality of text storage nodes; sorting the remaining text storage nodes in the plurality of text storage nodes based on the behavior scores to obtain a sequential set of candidate text storage nodes; Dividing the candidate text storage node sequence set into a preset number of candidate text storage node sequence subsets, and matching the target text to the candidate text storage node sequence subsets based on the importance of the target text; Based on the remaining amount of computing resources of each candidate text storage node in the sequential subset of candidate text storage nodes, the target text is matched to the candidate text storage node with the largest remaining amount of computing resources.

[0012] Preferably, the calculation formula of the behavior score is as follows: ; in, Indicates the first of multiple text storage nodes The behavior scores of text storage nodes, Indicates the first of multiple text storage nodes The transaction sequence in which the text storage nodes participate, Indicates the first of multiple text storage nodes Whether the transaction in which the text storage node participates is legal, if legal, then , otherwise, , Indicates the first of multiple text storage nodes Whether the transaction in which the text storage node participates is illegal. If illegal, then , otherwise, , Indicates the preset penalty coefficient.

[0013] The technical solution adopted to solve the above technical problem is: an electronic archive management system, which is applicable to the above electronic archive management method, including: A text recognition unit, the text recognition unit is used to obtain a paper file image, and perform text recognition on the paper file image based on image processing technology to obtain a first target text; A text processing unit, the text processing unit is used to receive electronic files uploaded by multiple preset data sources every preset sampling period, convert the electronic documents into a second target text, and pre-process the first target text and the second target text to obtain a target text; A text classification unit, the text classification unit is used to perform a word segmentation operation on the target text to obtain a word sequence corresponding to the target text, and classify the target text based on the word sequence to obtain a text type label corresponding to the target text; A storage matching unit, the storage matching unit is used to obtain the importance of the target text based on the text type label corresponding to the target text, and match the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text; A text storage unit, wherein the text storage unit is used to store the target text in the matching text storage node.

[0014] The beneficial effects of the present invention are as follows: (1) The present invention converts paper archives into electronic texts through image processing technology and text recognition technology, which greatly improves the efficiency of converting archives from paper to electronic archives and reduces the workload of manual input. By setting a preset sampling period, the present invention can automatically receive and process electronic archives from multiple data sources on a regular basis, thereby avoiding manual intervention and ensuring the real-time and comprehensiveness of archive management. By matching appropriate storage nodes, the system can ensure that documents stored in different nodes are effectively managed and reduce the risk of data loss. (2) After the present invention performs text recognition on the paper archive image, it pre-processes the first target text and the second target text, thereby eliminating the source file in the conversion process. The possible errors or inconsistencies make the final target text more accurate and reliable, and the text is processed based on word segmentation technology and the type of text is identified through classification, which can help automatically mark and classify archives, thereby reducing errors in manual classification and improving the accuracy and consistency of archive classification; (3) The present invention automatically evaluates the importance of the text based on its type label. This intelligent evaluation method can make the archive management system more flexible, determine the storage strategy based on the importance of the document, ensure that key documents are stored and managed preferentially, and match the importance of the target text with the characteristics of the preset storage node. It can dynamically allocate storage resources according to the priority of the document, optimize the utilization of storage space and the efficiency of data retrieval. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic flow chart of the steps of an overall method in an embodiment of the present invention; Figure 2 A schematic diagram of the system architecture of an overall system in an embodiment of the present invention.

[0016] Figure numerals: 1. Character recognition unit; 2. Text processing unit; 3. Text classification unit; 4. Storage matching unit; 5. Text storage unit. DETAILED DESCRIPTION

[0017] Embodiment 1, as Figure 1 As shown, the present invention proposes a method for managing electronic archives, comprising: S1, obtaining a paper archive image, and performing text recognition on the paper archive image based on image processing technology to obtain a first target text; S2, receiving electronic files uploaded by a plurality of preset data sources at every preset sampling period, converting the electronic files into a second target text, and preprocessing the first target text and the second target text to obtain a target text; S3, performing a word segmentation operation on the target text to obtain a word sequence corresponding to the target text, and classifying the target text based on the word sequence to obtain a text type label corresponding to the target text; S4, obtaining the importance of the target text based on the text type label corresponding to the target text, and matching the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text; S5. Store the target text into a matching text storage node.

[0018] In the present invention, paper archive image refers to the conversion of paper documents into digital image format by scanning or taking photos; electronic archive refers to archive files stored in electronic file formats (such as PDF, Word documents, Excel tables, etc.), which are usually directly generated digital documents rather than scanned images; text preprocessing refers to the operation of cleaning and normalizing the original text before further analysis; word segmentation is a very important step in text preprocessing, especially for languages ​​without space separation such as Chinese. The word segmentation operation is to divide continuous text strings into individual words; word sequence refers to a series of ordered word sets obtained by word segmentation. Each word usually carries certain semantic information. By analyzing the word sequence, the meaning of the text can be understood. meaning and structure; text classification is a task in natural language processing, which aims to classify the text into one or more predefined categories based on its content, for example, classify it into news, technology, finance and other categories based on the content of the article; text type label is the result of text classification, which is the category label of the target text, for example, the target text may be classified into types such as contracts, reports, invoices, etc., and each category corresponds to a label; text storage node refers to the physical or virtual storage location where text data is stored, each node may have different storage policies, permissions and security, etc. According to the importance of the text, the target text will be assigned to the corresponding storage node; text storage refers to storing text data (such as target text) in a storage medium, usually in a database, file system or cloud storage.

[0019] Embodiment 2, a method for managing electronic archives proposed by the present invention, compared with embodiment 1, further comprises: performing text recognition on a paper archive image based on image processing technology to obtain a first target text, including: A1. Performing image enhancement on the paper archive image to obtain a paper archive enhanced image; A2. Performing text box detection on the paper archive enhanced image based on the trained first object detection model to obtain a plurality of first text boxes corresponding to the paper archive enhanced image, and dividing the plurality of first text boxes by columns to obtain a text box column set; A3. Performing text box detection on the text box column set based on the trained second object detection model to obtain a plurality of second text boxes corresponding to the text box column set, and dividing the plurality of second text boxes into rows to obtain a text box row set; A4. Performing text recognition on the text box row set based on the trained text recognition model to obtain a text recognition result corresponding to the text box row set; A5. Sort the text recognition results to obtain the first target text.

[0020] In this embodiment, image enhancement is a process of improving image quality through a series of processing methods (such as contrast adjustment, noise removal, sharpening, etc.). For paper archival images, the purpose of image enhancement is to make text and details clearer to facilitate subsequent text recognition and detection. For example, during scanning or shooting, the image may be affected by blur, uneven lighting or noise. Image enhancement technology can improve recognition accuracy by optimizing the visual effect of the image; target detection is a task in computer vision that aims to locate and identify target objects or areas from an image. In this process, the purpose of target detection is to identify text boxes in the image. The target detection model usually marks all areas in the image that may contain text, i.e., text boxes; the first target detection model is used to detect possible text box areas in the enhanced paper archival image. It will identify the possible locations of all texts in the image and output a series of "text box" coordinates; the second target detection model is used to detect possible text box areas in the enhanced paper archival image. It will identify the possible locations of all texts in the image and output a series of "text box" coordinates; the second target detection model is used to detect possible text box areas in the enhanced paper archival image. The detection model is used to further detect the detected text box column set. It identifies each line of text in the text box column and further divides the specific position of each line of text; the text box column set refers to the result of grouping multiple text boxes by column (vertical) obtained by the first target detection model; in the further detection process of the text box row set, the second target detection model divides the text in the text box column by row (horizontally) to form a text box row set; the text results after text recognition are usually output in a certain order, however, there may be confusion in the order of text and character recognition errors during the recognition process. Therefore, the recognition results need to be sorted to ensure the correct order of the text. The sorting process may include re-sorting the recognized text in the order of rows and columns according to the spatial position of the text box (from top to bottom, from left to right).

[0021] In an optional embodiment, the text recognition model includes a local feature extraction module, a global feature extraction module, a weighted aggregation module and a classification module. The local feature extraction module is used to extract local image features of a text box row set, the global feature extraction module is used to extract global image features of a text box row set, the weighted aggregation module is used to weightedly aggregate local image features and global image features to obtain aggregated image features, and the classification module is used to classify the aggregated image features to obtain text recognition results corresponding to the text box row set.

[0022] In an optional embodiment, the local feature extraction module uses deep separable convolution to extract image features, the global feature extraction module includes a window multi-head attention layer, a sliding window multi-head attention layer and an MLP layer, each MLP layer uses a normalization layer before being connected to the window multi-head attention layer and the sliding window multi-head attention layer, and each layer uses a residual connection, the weighted aggregation module weightedly combines local image features and global image features through GCN to obtain aggregated image features, and the classification module classifies the aggregated image features through a fully connected layer and a Softmax function to obtain the text recognition results corresponding to the text box row set.

[0023] It should be noted that the function of the local feature extraction module is to extract features of local areas from the image, especially for each line of text in the text box line set. Local features usually refer to detailed information in the image, such as local attributes such as strokes and glyphs of text. Local feature extraction is performed through convolutional neural networks (CNN) or other deep learning technologies. The purpose is to capture the local structural features of each character or character group; the purpose of the global feature extraction module is to extract global features, that is, information in a larger range, from the entire image or the entire line of text box area. Global features contain a wider range of contextual information in the image, not only limited to the features of a single character, but also including the arrangement of characters, line spacing, text layout, etc. Global features help to identify in more complex situations, especially when the glyphs are similar or ambiguous, the overall context can provide more clues.

[0024] In an optional embodiment, the target text is classified based on the word sequence to obtain a text type label corresponding to the target text, including: B1. Obtain the contextual semantic features and word embedding vectors of word sequences based on the trained pre-trained model; B2. Calculate the similarity between each word in the word sequence based on the word embedding vector of the word sequence to obtain the similarity between each word in the word sequence. If the similarity is greater than a preset similarity threshold, it indicates that there is a similar relationship between the two words. B3. Construct a text directed graph based on the word sequence and the similarity relationship between words, and transfer the aggregated text directed graph based on the graph convolution operation to obtain the text hidden features of the word sequence; B4. Fuse the contextual semantic features of the word sequence and the text hidden features to obtain fused features, and classify the fused features to obtain the text type label corresponding to the target text.

[0025] It should be noted that pre-trained models refer to models that have been trained on large-scale data sets and can learn rich language features and representations. These models are usually deep learning models that are trained using large amounts of text data to learn semantic information about words, sentences, and context.

[0026] In an optional embodiment, the pre-trained model adopts BERT, and the text directed graph includes multiple nodes and multiple directed edges, wherein the nodes correspond to words in a word sequence, and the directed edges correspond to the similarity relationship between two words.

[0027] It should be noted that BERT is a pre-trained language model based on the Transformer architecture, proposed by Google in 2018. It is an important progress in the field of natural language processing (NLP) and has achieved significant performance improvements in many NLP tasks.

[0028] In an optional embodiment, the aggregated text directed graph is transferred based on the graph convolution operation to obtain the text hidden features of the word sequence, including: C1. Based on the graph convolution operation, the text directed graph is embedded to obtain the node representation, where the formula for the node representation is as follows: ; in, Represents the first Hidden representation of layer nodes, represents the activation function, Indicates that the column represents the type The adjacency matrix of the text directed graph is Indicates that the column represents the type The first part of the text directed graph Hidden representation of layer nodes, Indicates that the column represents the type The first part of the text directed graph The weight of the layer nodes, Represents different types of nodes; C2. Calculate the type-level attention score of the adjacent nodes based on the node representation of each node in the text directed graph. The calculation formula of the type-level attention score is as follows: ; in, represents the type-level attention scores of neighboring nodes, Indicates that the column represents the type The attention vector of Represents a splicing operation, represents the hidden representation of the current node, Indicates that the column represents the type The embedding representation of the neighboring nodes of , represents the adjacent nodes, represents the embedded representation of the directed edges between the current node and the adjacent nodes, Represents the hidden representation of adjacent nodes; C3. Obtain the type-level attention weight of the adjacent nodes based on the type-level attention score of the adjacent nodes, where the calculation formula of the type-level attention weight is as follows: ; in, represents the type-level attention weight; C4. Calculate the node-level attention score of the adjacent nodes based on the type-level attention weights of the adjacent nodes. The calculation formula of the node-level attention score is as follows: ; in, represents the type-level attention weights of neighboring nodes, represents the attention vector; C5. The dual-level attention mechanism of type-level and node-level attention is integrated to obtain the text hidden features of word sequences.

[0029] In an optional embodiment, the target text is matched with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text, including: D1. Obtaining behavior scores of multiple text storage nodes based on the behavior transactions of multiple text storage nodes; D2, comparing the behavior scores of the multiple text storage nodes with a preset behavior score threshold, and if the behavior score of the text storage node is less than the preset behavior score threshold, eliminating the text storage node from the multiple text storage nodes; D3, sorting the remaining text storage nodes in the multiple text storage nodes based on the behavior scores to obtain a sequential set of candidate text storage nodes; D4, dividing the candidate text storage node sequence set into a preset number of candidate text storage node sequence subsets, and matching the target text to the candidate text storage node sequence subsets based on the importance of the target text; D5. Based on the remaining amount of computing resources of each candidate text storage node in the sequential subset of candidate text storage nodes, match the target text to the candidate text storage node with the largest remaining amount of computing resources.

[0030] It should be noted that the candidate text storage node sequence set is a set of text storage nodes sorted according to the behavior scores, and the nodes are arranged in descending order, indicating the order of priority use; the candidate text storage node sequence subset set is a division of the candidate text storage node sequence set, which divides the sorted node set into several subsets; the importance of the target text refers to the priority or processing value of the text in a specific situation; the remaining amount of computing resources refers to the remaining computing power of each text storage node currently available for processing tasks.

[0031] In an optional embodiment, the calculation formula of the behavior score is as follows: ; in, Indicates the first of multiple text storage nodes The behavior scores of text storage nodes, Indicates the first of multiple text storage nodes The transaction sequence in which the text storage nodes participate, Indicates the first of multiple text storage nodes Whether the transaction in which the text storage node participates is legal, if legal, then , otherwise, , Indicates the first of multiple text storage nodes Whether the transaction in which the text storage node participates is illegal. If illegal, then , otherwise, , Indicates the preset penalty coefficient.

[0032] It should be noted that the behavior score is a quantitative indicator used to evaluate the performance of a text storage node in the process of executing a task. It reflects the overall health or work efficiency of the node. Transactions in which a text storage node participates include but are not limited to data storage, data reading, data updating, data deletion, data synchronization, data compression, and logging.

[0033] Embodiment three, as Figure 2 As shown, the electronic archive management system proposed by the present invention is applicable to the electronic archive management method described above, and includes: A text recognition unit 1 is used to obtain a paper file image and perform text recognition on the paper file image based on image processing technology to obtain a first target text; The text processing unit 2 is used to receive electronic files uploaded by multiple preset data sources at every preset sampling period, convert the electronic files into a second target text, and pre-process the first target text and the second target text to obtain the target text; The text classification unit 3 is used to perform a word segmentation operation on the target text to obtain a word sequence corresponding to the target text, and classify the target text based on the word sequence to obtain a text type label corresponding to the target text; The storage matching unit 4 is used to obtain the importance of the target text based on the text type label corresponding to the target text, and match the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text; The text storage unit 5 is used to store the target text into a matching text storage node.

[0034] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for managing electronic archives, characterized in that: include: Acquire a paper archive image, and perform text recognition on the paper archive image based on image processing technology to obtain a first target text; receiving electronic files uploaded by a plurality of preset data sources at every preset sampling period, converting the electronic files into a second target text, and preprocessing the first target text and the second target text to obtain a target text; Performing a word segmentation operation on the target text to obtain a word sequence corresponding to the target text, and classifying the target text based on the word sequence to obtain a text type label corresponding to the target text; Acquiring the importance of the target text based on the text type label corresponding to the target text, and matching the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text; The target text is stored in the matching text storage node.

2. The electronic archive management method according to claim 1, characterized in that: Performing text recognition on the paper archive image based on image processing technology to obtain a first target text includes: Performing image enhancement on the paper archive image to obtain a paper archive enhanced image; Performing text box detection on the paper archive enhanced image based on the trained first object detection model to obtain a plurality of first text boxes corresponding to the paper archive enhanced image, and dividing the plurality of first text boxes into columns to obtain a text box column set; Performing text box detection on the text box column set based on the trained second object detection model to obtain a plurality of second text boxes corresponding to the text box column set, and dividing the plurality of second text boxes into rows to obtain a text box row set; Performing text recognition on the text box row set based on a trained text recognition model to obtain a text recognition result corresponding to the text box row set; The text recognition results are sorted to obtain a first target text.

3. The electronic archive management method according to claim 2, characterized in that: The text recognition model includes a local feature extraction module, a global feature extraction module, a weighted aggregation module and a classification module. The local feature extraction module is used to extract local image features of the text box row set, the global feature extraction module is used to extract global image features of the text box row set, the weighted aggregation module is used to weightedly aggregate the local image features and the global image features to obtain aggregated image features, and the classification module is used to classify the aggregated image features to obtain the text recognition results corresponding to the text box row set.

4. The electronic archive management method according to claim 3, characterized in that: The local feature extraction module uses deep separable convolution to extract image features, the global feature extraction module includes a window multi-head attention layer, a sliding window multi-head attention layer and an MLP layer, each of the MLP layers uses a normalization layer before being connected to the window multi-head attention layer and the sliding window multi-head attention layer, and each layer uses a residual connection, the weighted aggregation module weightedly combines the local image features and the global image features through GCN to obtain aggregated image features, and the classification module classifies the aggregated image features through a fully connected layer and a Softmax function to obtain the text recognition results corresponding to the text box row set.

5. The electronic archive management method according to claim 1, characterized in that: Classifying the target text based on the word sequence to obtain a text type label corresponding to the target text includes: Acquire contextual semantic features and word embedding vectors of the word sequence based on a trained pre-trained model; Calculating the similarity between each word in the word sequence based on the word embedding vector of the word sequence to obtain the similarity between each word in the word sequence, if the similarity is greater than a preset similarity threshold, it indicates that there is a similar relationship between the two words; Constructing a text directed graph based on the word sequence and the similarity relationship between the words, and transferring and aggregating the text directed graph based on a graph convolution operation to obtain text hidden features of the word sequence; The contextual semantic features of the word sequence and the text hidden features are fused to obtain fused features, and the fused features are classified to obtain a text type label corresponding to the target text.

6. The electronic archive management method according to claim 5, characterized in that: The pre-training model adopts BERT, and the text directed graph includes multiple nodes and multiple directed edges, wherein the nodes correspond to words in the word sequence, and the directed edges correspond to the similarity relationship between the two words.

7. The electronic archive management method according to claim 6, characterized in that: The text directed graph is transferred and aggregated based on a graph convolution operation to obtain text hidden features of the word sequence, including: The text directed graph is embedded based on the graph convolution operation to obtain a node representation, wherein the formula of the node representation is as follows: ； in, Represents the first Hidden representation of layer nodes, represents the activation function, Indicates that the column represents the type The adjacency matrix of the text directed graph is Indicates that the column represents the type The first part of the text directed graph Hidden representation of layer nodes, Indicates that the column represents the type The first part of the text directed graph The weight of the layer nodes, Represents different types of nodes; The type-level attention scores of the adjacent nodes are calculated based on the node representation of each node in the text directed graph, wherein the calculation formula of the type-level attention scores is as follows: ； in, represents the type-level attention scores of neighboring nodes, Indicates that the column represents the type The attention vector of Represents a splicing operation, represents the hidden representation of the current node, Indicates that the column represents the type The embedding representation of the neighboring nodes of , represents the adjacent nodes, represents the embedded representation of the directed edges between the current node and the adjacent nodes, Represents the hidden representation of adjacent nodes; The type-level attention weight of the adjacent node is obtained based on the type-level attention score of the adjacent node, wherein the calculation formula of the type-level attention weight is as follows: ； in, represents the type-level attention weight; The node-level attention score of the adjacent node is calculated based on the type-level attention weight of the adjacent node, wherein the calculation formula of the node-level attention score is as follows: ； in, represents the type-level attention weights of neighboring nodes, represents the attention vector; The dual-level attention mechanism of type-level and node-level attention is fused to obtain the text hidden features of the word sequence.

8. The electronic archive management method according to claim 1, characterized in that: Matching the target text with a plurality of preset text storage nodes based on the importance of the target text to obtain a text storage node matching the target text, including: Acquire behavior scores of the plurality of text storage nodes based on the behavior transactions of the plurality of text storage nodes; Comparing the behavior scores of the plurality of text storage nodes with a preset behavior score threshold, and if the behavior score of the text storage node is less than the preset behavior score threshold, eliminating the text storage node from the plurality of text storage nodes; sorting the remaining text storage nodes in the plurality of text storage nodes based on the behavior scores to obtain a sequential set of candidate text storage nodes; Dividing the candidate text storage node sequence set into a preset number of candidate text storage node sequence subsets, and matching the target text to the candidate text storage node sequence subsets based on the importance of the target text; Based on the remaining amount of computing resources of each candidate text storage node in the sequential subset of candidate text storage nodes, the target text is matched to the candidate text storage node with the largest remaining amount of computing resources.

9. The electronic archive management method according to claim 8, characterized in that: The calculation formula of the behavior score is as follows: ； in, Indicates the first of multiple text storage nodes The behavior scores of text storage nodes, Indicates the first of multiple text storage nodes The transaction sequence in which the text storage nodes participate, Indicates the first of multiple text storage nodes Whether the transaction in which the text storage node participates is legal, if legal, then , otherwise, , Indicates the first of multiple text storage nodes Whether the transaction in which the text storage node participates is illegal. If illegal, then , otherwise, , Indicates the preset penalty coefficient.

10. An electronic archive management system, which is applicable to an electronic archive management method according to any one of claims 1 to 9, characterized in that: include: A text recognition unit (1), the text recognition unit (1) being used to obtain a paper archive image, and to perform text recognition on the paper archive image based on image processing technology to obtain a first target text; A text processing unit (2), the text processing unit (2) being used to receive electronic files uploaded by a plurality of preset data sources at every preset sampling period, convert the electronic files into a second target text, and pre-process the first target text and the second target text to obtain a target text; A text classification unit (3), the text classification unit (3) is used to perform a word segmentation operation on the target text to obtain a word sequence corresponding to the target text, and classify the target text based on the word sequence to obtain a text type label corresponding to the target text; A storage matching unit (4), the storage matching unit (4) being used to obtain the importance of the target text based on the text type label corresponding to the target text, and to match the target text with a plurality of preset text storage nodes based on the importance of the target text, so as to obtain a text storage node matching the target text; A text storage unit (5), wherein the text storage unit (5) is used to store the target text in the matching text storage node.

Images