Multi-modal pdf content purification and intelligent proofreading system and method based on visual perception enhancement

Abstract

The present application belongs to the technical field of digital publication processing, and provides a multi-modal PDF content purification and intelligent proofreading system and method based on visual perception enhancement, which comprises a visual perception module, a semantic understanding module and an intelligent decision module; the visual perception module eliminates physical layer interference by generating multi-resolution images, extracting fine structure features and locating invisible characters; the semantic understanding module constructs a knowledge graph and a DST-Transformer model to realize character recognition, semantic coding and image-text association repair; the intelligent decision module outputs a purified document through weighted confidence calculation, dynamic threshold adjustment and cross-platform verification; the method corresponds to the three module steps, realizes full-process processing from physical layer analysis to logical layer reconstruction and then to intelligent output, improves the PDF content purification accuracy and cross-platform compatibility, and is suitable for electronic document deep purification and proofreading.

Description

Technical Field

[0001] This invention belongs to the field of digital publication processing technology, specifically a multimodal PDF content purification and intelligent review system and method based on enhanced visual perception. Background Technology

[0002] Existing PDF processing technologies face multiple technical bottlenecks in the digital document circulation, editing, and archiving processes, severely restricting the efficient utilization of electronic documents. Specifically: Limited parsing capabilities and poor compatibility with proprietary formats: Traditional PDF parsing engines rely on open encoding standards such as PDF 1.7 and cannot recognize proprietary format tags added by commercial software such as Adobe Acrobat and Foxit PhantomPDF, resulting in structural breaks, garbled text, or missing elements during format restoration.

[0003] Rigid noise tolerance mechanism leads to damage to text integrity: Character stream-based OCR technology uses a fixed tolerance threshold for control characters such as spaces and tabs. In scenarios such as tilted scans or blurry fonts, it is easy to misjudge effective delimiters as redundant noise or retain extra whitespace characters, resulting in damage to text continuity.

[0004] Semantic association is missing, and the accuracy of format reconstruction is low: Existing algorithms focus on processing single characters or short sentences and do not build a global semantic network for the document, making it difficult to identify format dependencies across paragraphs.

[0005] Poor cross-platform compatibility and inconsistent formatting after editing: Converted documents are prone to problems such as font replacement, misaligned table lines, and chaotic text and image layout in editors such as Word and WPS.

[0006] To this end, the present invention provides a multimodal PDF content purification and intelligent review system and method based on enhanced visual perception. Summary of the Invention

[0007] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0008] The technical solution adopted by this invention to solve its technical problem is: In a first aspect, the present invention provides a multimodal PDF content cleanup and intelligent review system based on enhanced visual perception, comprising: The visual perception module scans the original PDF to generate a base layer image, uses an edge detection algorithm to identify the main text region in the base layer image, dynamically generates a multi-resolution image sequence for the main text region using bilinear interpolation, inputs the multi-resolution image sequence into a deep convolutional network to output a fine structure feature map, develops a text region detection model based on OpenCV, scans the multi-resolution image sequence, and locates invisible characters. Semantic understanding module: Establish a knowledge graph, construct a DST-Transformer model, input text region features into the DST-Transformer model and combine with the knowledge graph to generate structural units, use a convolutional recurrent neural network to perform character recognition on the main text region, generate character-level text sequences, extract visual features from the main text region, perform semantic encoding on the character-level text sequences to generate semantic vectors, calculate the similarity between visual features and semantic vectors to generate image-text associations, and correct errors in image-text associations; Intelligent decision-making module: It calculates the confidence score by weighting the character-level text sequence recognition results, image-text association, and semantic vector results. If the confidence score is unreliable, it enters the dynamic threshold adjustment stage and performs cross-platform verification on the text content after dynamic threshold adjustment.

[0009] As a further solution of the present invention, the specific process of generating the base layer image by scanning the original PDF is as follows: A grayscale CCD scanner is used to scan the electronic or paper document corresponding to the original PDF at a resolution of 72 dpi to generate a base layer image.

[0010] As a further solution of the present invention, the specific process of identifying the main text region in the base layer image using the edge detection algorithm is as follows: Using ROI automatic recognition technology, the effective area of ​​document content in the base layer image is located, and edge blank margins are removed by cropping with a cropping accuracy controlled within ±0.5mm. The text main body area in the base layer image is identified based on the Canny edge detection algorithm; the text main body area includes: body text, tables, and charts.

[0011] As a further solution of the present invention, the specific process of dynamically generating multi-resolution image sequences by using bilinear interpolation on the main text region is as follows: For the identified main text area, bilinear interpolation is used to dynamically generate a multi-resolution image sequence in the range of 200-400 dpi, with a resolution adjustment step of 50 dpi, namely 200 dpi, 250 dpi, 300 dpi, 350 dpi, and 400 dpi, to achieve high-definition capture of fine structures such as the edges and table lines of the main text area.

[0012] As a further solution of the present invention, the specific process of inputting the multi-resolution image sequence into the deep convolutional network to output the fine structure feature map is as follows: A deep convolutional network containing 17 residual blocks is constructed. Each residual block consists of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. The network as a whole integrates two attention mechanisms: Channel attention: The importance of each feature channel is weighted by global average pooling, and then the key channel features are enhanced by 1×1 convolution; Spatial attention: By fusing the results of max pooling and average pooling, key spatial regions such as the edges of the main text area and table lines are located through 1×1 convolution; The generated multi-resolution image sequence is input into A-GResNet, and through forward propagation, it outputs fine structural feature maps such as the edges of the main text region, table lines, and character outlines.

[0013] As a further solution of the present invention, the specific process of using a convolutional recurrent neural network to perform character recognition on the main text region and generate a character-level text sequence is as follows: For the text-major region in the image output by the visual perception module, CRNN is used for character recognition to generate a character-level text sequence.

[0014] As a further solution of the present invention, the specific process for calculating the confidence level is as follows: The system collects character-level text sequence recognition results and image-text association results output by the semantic understanding module, and extracts semantic vector results output by the semantic understanding module. The three types of data, namely character-level text sequence recognition results, image-text association results, and semantic vector results, are assigned weights α=0.4, β=0.3, and γ=0.3, respectively. The final confidence of each content unit is calculated by weighted voting.

[0015] As a further solution of the present invention, the specific process of entering the dynamic threshold adjustment stage is as follows: If a content unit has a confidence level of less than 90%, it is considered unreliable and enters the dynamic threshold adjustment stage. Construct a confusion matrix based on historical processing data, statistically analyze the recognition error distribution of different types of content, and determine the basic fault tolerance threshold for each type of content. The similarity between the merged data and the standard format template is calculated: Similarity > 85%: Determined to be correctable, template rules will be automatically applied for repair; Similarity between 70% and 85%: determined to require manual review and annotation; Similarity <70%: judged as high risk, retain the original content, and record the reason for the error; For content that triggers manual review and labeling or is marked as high-risk, a review work order is generated, which includes the original data, automatic recognition results, error analysis and correction options. It supports online manual editing and synchronous updates of the final results.

[0016] As a further solution of the present invention, the specific process of performing cross-platform verification is as follows: Develop a core module for format conversion, supporting the conversion of cleaned structured data into two mainstream formats: DOCX format: Preserves editability, ensuring that the hierarchical structure and style of text, tables, and images can be modified using Word and WPS editors; PDF / A standard format: meets long-term archiving needs, embeds font and layout information, and avoids display anomalies across devices; Perform a rendering test on the converted document in the editor and check the following: Text style consistency; Table structure integrity; Accuracy of image-text association; For DOCX format files, compress redundant format tags to reduce file size; Optimize image compression ratio for PDF / A format files.

[0017] Secondly, this invention provides a multimodal PDF content cleanup and intelligent review method based on enhanced visual perception, including: S1: Scan the original PDF to generate a base layer image, use an edge detection algorithm to identify the main text region in the base layer image, use bilinear interpolation to dynamically generate a multi-resolution image sequence for the main text region, input the multi-resolution image sequence into a deep convolutional network to output a fine structure feature map, develop a text region detection model based on OpenCV, scan the multi-resolution image sequence, and locate invisible characters. S2: Establish a knowledge graph, construct a DST-Transformer model, input text region features into the DST-Transformer model and combine it with the knowledge graph to generate structural units, use a convolutional recurrent neural network to perform character recognition on the main text region, generate character-level text sequences, extract visual features from the main text region, perform semantic encoding on the character-level text sequences to generate semantic vectors, calculate the similarity between visual features and semantic vectors to generate image-text associations, and correct errors in image-text associations; S3: The confidence score is calculated by weighting the character-level text sequence recognition results, image-text association, and semantic vector results. If the confidence score is unreliable, the dynamic threshold adjustment stage is entered, and the text content after dynamic threshold adjustment is verified across platforms.

[0018] The beneficial effects of this invention are as follows: By generating multi-resolution images, extracting features using deep convolutional networks, and detecting implicit format markers, this technology accurately eliminates invisible format interference at the physical layer, improving the accuracy of text feature extraction and addressing the sensitivity of traditional techniques to noise. Leveraging knowledge graphs and the DST-Transformer model to capture long-distance format dependencies, and combining cross-modal semantic alignment technology to achieve image-text association, it can automatically correct structural errors such as abnormal paragraph indentation and disordered heading levels, achieving paragraph-level format reconstruction and compensating for the lack of contextual awareness. By calculating confidence levels through weighted fusion of multi-source data, coupled with a dynamic threshold adjustment strategy, it reduces the risk of over-correction and improves the efficiency of manual review. Supporting DOCX and PDF / A format output, it is compatible with mainstream editors, improving cross-platform display consistency, and resolving the limitations and poor compatibility of parsing. Balancing accuracy and speed, it is suitable for deep cleaning and intelligent review scenarios of various electronic documents. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a system module diagram of the multimodal PDF content purification and intelligent review system based on enhanced visual perception, as described in this invention. Figure 2 This is a flowchart of the steps of the multimodal PDF content purification and intelligent review method based on enhanced visual perception of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] Example 1 like Figure 1 As shown in the embodiment of the present invention, the multimodal PDF content cleansing and intelligent review system based on enhanced visual perception includes: The visual perception module scans the original PDF to generate a base layer image, uses an edge detection algorithm to identify the main text region in the base layer image, dynamically generates a multi-resolution image sequence for the main text region using bilinear interpolation, inputs the multi-resolution image sequence into a deep convolutional network to output a fine structure feature map, develops a text region detection model based on OpenCV, scans the multi-resolution image sequence, and locates invisible characters. Specifically, a grayscale CCD scanner is used to scan the electronic or paper document corresponding to the original PDF at a resolution of 72 dpi to generate a base layer image. By using ROI automatic recognition technology, the effective area of ​​the base layer image document content is located, and the edge blank margins are removed by cropping. The cropping accuracy is controlled within ±0.5mm, which solves the problems of low efficiency and large error of traditional manual cropping. The Canny edge detection algorithm is used to identify the main text region in the base layer image; the main text region includes: body text, tables, charts, etc. For the identified text main area, bilinear interpolation is used to dynamically generate a multi-resolution image sequence in the range of 200-400dpi, with a resolution adjustment step of 50dpi, namely 200dpi, 250dpi, 300dpi, 350dpi, and 400dpi, to achieve high-definition capture of fine structures such as the edges and table lines of the text main area. An adaptive Gamma correction algorithm is applied to the generated multi-resolution image sequence to eliminate color difference and brightness deviation caused by differences in scanning equipment models and uneven illumination intensity, ensuring uniformity of image brightness. PNG-32-bit lossless compression format is used to store multi-resolution image sequences, avoiding the loss of text details caused by lossy compression such as JPEG; The compressed file size should be kept within 1 / 3 of the original data. For example, after processing, the file size of a 200-page document can be compressed from 12.6GB to 4.2GB, reducing storage and transmission costs while maintaining image quality; A deep convolutional network (A-GResNet) containing 17 residual blocks is constructed. Each residual block consists of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. The network as a whole integrates two attention mechanisms: Channel attention: The importance of each feature channel is weighted by global average pooling, and then the key channel features are enhanced by 1×1 convolution; Spatial attention: By fusing the results of max pooling and average pooling, key spatial regions such as the edges of the main text area and table lines are located through 1×1 convolution; The generated multi-resolution image sequence is input into A-GResNet, and through forward propagation calculation, it outputs fine structural feature maps such as the edges of the main text region, table lines, and character outlines. Dropout is used during network training to suppress overfitting and ensure the accuracy of feature extraction, providing high-precision feature support for subsequent format recognition. A text region detection model was developed based on OpenCV, and the MSER (Maximum Stable Extreme Region) algorithm was used to scan multi-resolution image sequences to locate invisible characters. For example, when locating the distribution areas of private format markers, redundant whitespace characters, and control characters, the text region detection model is considered to have completed training when the recall rate reaches ≥99.1%. For the invisible character regions located by the MSER algorithm, morphological operations are used, such as erosion, dilation, and opening operations, to remove noise such as consecutive whitespace characters and misaligned tab characters. By combining connected component analysis, isolated invisible control characters are filtered out, such as proprietary tags added by Adobe Acrobat, while retaining valid characters and structure of the main text area of ​​the document content; The visual perception module simulates the human visual perception characteristics of paper documents through a three-level processing flow of multi-resolution image generation, fine feature extraction, and implicit formatting cleanup. It eliminates invisible formatting interference inserted during the electronic document generation process at the physical layer, such as spaces, misaligned layouts, and private tags. This module integrates dynamic imaging technology, deep convolutional networks, and computer vision algorithms to achieve accurate recognition of the document's fine structure and noise filtering.

[0023] Semantic understanding module: Establish a knowledge graph, construct a DST-Transformer model, input text region features into the DST-Transformer model and combine with the knowledge graph to generate structural units, use a convolutional recurrent neural network to perform character recognition on the main text region to generate character-level text sequences, extract visual features from the main text region, perform semantic encoding on the character-level text sequences to generate semantic vectors, calculate the similarity between visual features and semantic vectors to generate image-text associations, and correct errors in image-text associations; Specifically, based on the formatting specifications of mainstream design software such as Adobe Acrobat and WPS, a knowledge graph containing 2,147 formatting patterns was compiled and established, covering the heading level; Examples include document elements such as first-level headings, second-level headings, paragraph styles such as first-line indentation and line spacing, table structures such as border styles and cell merging rules; Based on the Transformer architecture, a DST-Transformer model is built, which includes a 6-layer encoder and a 6-layer decoder. It captures long-distance format dependencies across paragraphs and pages in the document through a self-attention mechanism. The base layer image, main text region, multi-resolution image sequence, and multi-resolution image sequence output by the visual perception module are integrated into text region features. The text region features are input into the DST-Transformer model, and the document structure units are initially divided by matching them with the format patterns in the knowledge graph. The similarity between input features and format patterns in the knowledge graph is calculated using a dynamic programming algorithm, and the format type is automatically labeled for each structural unit. For special formats that do not match existing patterns, such as custom headers and footers, mark them as pending verification to reserve an interface for subsequent manual review; For the main text region in the image output by the visual perception module, CRNN (Convolutional Recurrent Neural Network) is used for character recognition to generate a character-level text sequence; For non-text elements such as images and charts in the main text area, visual features are extracted using image feature extraction techniques, such as ResNet-50 (50-layer residual network). Using the BERT-base (basic bidirectional Transformer encoder representation model) language model, semantic encoding is performed on character-level text sequences to generate semantic vectors; It should be noted that semantic vectors include word-level semantic vectors and sentence-level semantic vectors; Cross-modal feature alignment technology is used to calculate the cosine similarity between the semantic vector of the main text region and the visual features to establish image-text association; For example, the binding relationship between chart description text and corresponding chart; For ambiguous text-image relationships, such as a text that may correspond to multiple charts, the element with the closest spatial distance on the page is prioritized for matching, based on the location information output by the document structure parser, such as the page distance between the text and the chart. Based on the context prediction capability of the BERT-based language model, character errors in CRNN recognition are corrected; For issues such as incorrect sentence segmentation and missing punctuation, the sentence-end punctuation prediction model (based on bidirectional LSTM) is used to complete the text and improve its semantic coherence. Based on 500 typical template formats, such as academic paper templates, official document templates, and resume templates, we extracted format consistency rules, such as that a first-level heading must be followed by a second-level heading or body paragraph, the line spacing of the body text within the same chapter must be consistent, and the number of columns in a table must match the number of columns in the table header, etc. Rules are stored in the form of regular expressions and style rule libraries, and support dynamic updates; Traverse the structural units output by the document structure parser and perform item-by-item validation according to the rule base: Check whether the heading levels are continuous. For example, check if there is a skipping error where a fourth-level heading appears directly after a first-level heading. Verifying the consistency of styles for similar text types is exemplified by whether the body text size and font are consistent across chapters. Checking the integrity of the table structure is an example of checking for broken cells across pages or missing borders. For format errors detected during validation, a repair solution is automatically generated based on the rule base: For heading errors involving skipping levels, insert a placeholder for the missing intermediate level heading. For texts with inconsistent styles, uniformly adjust them to the mainstream style of the same type of text. For example, change the Song font 10pt of a certain paragraph to the mainstream Song font 12pt of the same chapter. For table structure errors, the page breakage problem is fixed and missing borders are supplemented using cell merging / splitting algorithms; The semantic understanding module constructs a document semantic network through a three-level processing flow of document structure parsing, cross-modal semantic alignment, and format consistency verification, realizing the transformation from physical layer features to logical layer structure. This module integrates the Transformer architecture, language model, and rule verification technology to capture long-distance semantic dependencies in documents, align cross-modal information between text and images, and correct format and logic errors, providing structured semantic data for subsequent intelligent decision-making.

[0024] Intelligent decision-making module: It calculates the confidence score by weighting the character-level text sequence recognition results, image-text association, and semantic vector results. If the confidence score is unreliable, it enters the dynamic threshold adjustment stage and performs cross-platform verification on the text content after dynamic threshold adjustment. The system collects character-level text sequence recognition results and image-text association results output by the semantic understanding module, and extracts semantic vector results output by the semantic understanding module. The three types of data—character-level text sequence recognition results, image-text association results, and semantic vector results—are assigned weights (α=0.4, β=0.3, γ=0.3), and the final confidence score of each content unit is calculated through weighted voting. Character-level text sequence (α=0.4): focuses on the recognition accuracy of individual characters, for example, the verification of details such as spaces and punctuation marks; Image-text association (β=0.3): Focuses on the structural integrity of image-text association within text blocks; Semantic vector results (γ=0.3): Focus on contextual semantic consistency, for example, the logical connection between the title and the body text, and the coherence of terminology usage; Content units with a confidence level ≥ 90% after fusion are marked as highly reliable and directly included in the final cleaned text; content units with a confidence level < 90% are judged to have unreliable confidence and enter the dynamic threshold adjustment stage. The specific process of dynamic threshold adjustment is as follows: Construct a confusion matrix based on historical processing data, statistically analyze the recognition error distribution of different types of content, and determine the basic fault tolerance threshold for each type of content. The similarity between the merged data and the standard format template is calculated: Similarity > 85%: Determined to be correctable, template rules will be automatically applied for repair; Similarity between 70% and 85%: determined to require manual review and annotation; Similarity <70%: judged as high risk, retain the original content, and record the reason for the error; For content that triggers manual review and labeling or is marked as high-risk, a review work order is generated, which includes the original data, automatic recognition results, error analysis and correction options. It supports online manual editing and synchronous updates of the final results. Develop a core module for format conversion, supporting the conversion of cleaned structured data into two mainstream formats: DOCX format: Preserves editability, ensuring that the hierarchical structure and style of text, tables, and images can be modified using editors such as Word and WPS; PDF / A standard format: meets the needs of long-term archiving, embeds font and layout information, and avoids display anomalies across devices.

[0025] Perform a rendering test on the converted document in the editor and check the following: Text style consistency; Table structure integrity; Accuracy of image-text association; For DOCX format files, compress redundant format tags to reduce file size; For PDF / A format files, the image compression ratio is optimized to improve transmission efficiency.

[0026] The intelligent decision-making module integrates multi-level data from preceding modules through a three-level process of multi-source evidence fusion, dynamic threshold decision-making, and cross-platform format output. It combines rule-based fault tolerance mechanisms and format adaptation technology to ultimately output high-quality, editable cleaned documents, solving problems such as the single output format and rough processing of low-confidence content in traditional PDF processing.

[0027] The technical solution in this embodiment is as follows: The process involves scanning the original PDF to generate a base layer image. An edge detection algorithm is used to identify the main text regions within the base layer image. Bilinear interpolation is then used to dynamically generate multi-resolution image sequences for these regions. These multi-resolution image sequences are input into a deep convolutional network to output fine-structure feature maps. A text region detection model based on OpenCV is developed to scan the multi-resolution image sequences and locate invisible characters. A knowledge graph is established, and a DST-Transformer model is constructed. Text region features are input into the DST-Transformer model, which is then combined with the knowledge graph to generate structural units. A convolutional recurrent neural network is used to recognize characters in the main text regions, generating character-level text sequences. Visual features are extracted from the main text regions, and semantic encoding is performed on the character-level text sequences to generate semantic vectors. The similarity between visual features and semantic vectors is calculated to generate image-text associations, and errors in the image-text associations are corrected. The character-level text sequence recognition results, image-text associations, and semantic vector results are weighted and processed to calculate confidence scores. If the confidence scores are unreliable, a dynamic threshold adjustment stage is initiated. The dynamically thresholded text content is then validated across platforms.

[0028] Example 2 like Figure 2 As shown in Example 1, this invention provides a multimodal PDF content cleansing and intelligent review method based on enhanced visual perception, including: S1: Scan the original PDF to generate a base layer image, use an edge detection algorithm to identify the main text region in the base layer image, use bilinear interpolation to dynamically generate a multi-resolution image sequence for the main text region, input the multi-resolution image sequence into a deep convolutional network to output a fine structure feature map, develop a text region detection model based on OpenCV, scan the multi-resolution image sequence, and locate invisible characters. S2: Establish a knowledge graph, construct a DST-Transformer model, input text region features into the DST-Transformer model and combine it with the knowledge graph to generate structural units, use a convolutional recurrent neural network to perform character recognition on the main text region, generate character-level text sequences, extract visual features from the main text region, perform semantic encoding on the character-level text sequences to generate semantic vectors, calculate the similarity between visual features and semantic vectors to generate image-text associations, and correct errors in image-text associations; S3: The confidence score is calculated by weighting the character-level text sequence recognition results, image-text association, and semantic vector results. If the confidence score is unreliable, the dynamic threshold adjustment stage is entered, and the text content after dynamic threshold adjustment is verified across platforms.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multimodal PDF content cleanup and intelligent review system based on enhanced visual perception, characterized by: include: The visual perception module scans the original PDF to generate a base layer image, uses an edge detection algorithm to identify the main text region in the base layer image, dynamically generates a multi-resolution image sequence for the main text region using bilinear interpolation, inputs the multi-resolution image sequence into a deep convolutional network to output a fine structure feature map, develops a text region detection model based on OpenCV, scans the multi-resolution image sequence, and locates invisible characters. Semantic understanding module: Establish a knowledge graph, construct a DST-Transformer model, input text region features into the DST-Transformer model and combine with the knowledge graph to generate structural units, use a convolutional recurrent neural network to perform character recognition on the main text region, generate character-level text sequences, extract visual features from the main text region, perform semantic encoding on the character-level text sequences to generate semantic vectors, calculate the similarity between visual features and semantic vectors to generate image-text associations, and correct errors in image-text associations; Intelligent decision-making module: It calculates the confidence score by weighting the character-level text sequence recognition results, image-text association, and semantic vector results. If the confidence score is unreliable, it enters the dynamic threshold adjustment stage and performs cross-platform verification on the text content after dynamic threshold adjustment.

2. The multimodal PDF content purification and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process of generating the base layer image by scanning the original PDF is as follows: A grayscale CCD scanner is used to scan the electronic or paper document corresponding to the original PDF at a resolution of 72 dpi to generate a base layer image.

3. The multimodal PDF content purification and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process of using edge detection algorithms to identify the main text region in the base layer image is as follows: Using ROI automatic recognition technology, the effective area of ​​document content in the base layer image is located, and the edge blank margins are cropped and removed. The cropping accuracy is controlled within ±0.5mm. The text main body area in the base layer image is identified based on the Canny edge detection algorithm. The main text area includes: body text, tables, and charts.

4. The multimodal PDF content purification and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process of dynamically generating multi-resolution image sequences by using bilinear interpolation on the main text region is as follows: For the identified main text area, bilinear interpolation is used to dynamically generate a multi-resolution image sequence in the range of 200-400 dpi, with a resolution adjustment step of 50 dpi, namely 200 dpi, 250 dpi, 300 dpi, 350 dpi, and 400 dpi, to achieve high-definition capture of fine structures such as the edges and table lines of the main text area.

5. The multimodal PDF content cleanup and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process of inputting multi-resolution image sequences into a deep convolutional network to output fine structure feature maps is as follows: A deep convolutional network containing 17 residual blocks is constructed. Each residual block consists of a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function. The network as a whole integrates two attention mechanisms: Channel attention: The importance of each feature channel is weighted by global average pooling, and then the key channel features are enhanced by 1×1 convolution; Spatial attention: By fusing the results of max pooling and average pooling, key spatial regions such as the edges of the main text area and table lines are located through 1×1 convolution; The generated multi-resolution image sequence is input into A-GResNet, and through forward propagation, it outputs fine structural feature maps such as the edges of the main text region, table lines, and character outlines.

6. The multimodal PDF content cleanup and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process of using a convolutional recurrent neural network to perform character recognition in the main text region and generate a character-level text sequence is as follows: For the text-major region in the image output by the visual perception module, CRNN is used for character recognition to generate a character-level text sequence.

7. The multimodal PDF content cleanup and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process for calculating the confidence level is as follows: The system collects character-level text sequence recognition results and image-text association results output by the semantic understanding module, and extracts semantic vector results output by the semantic understanding module. The three types of data, namely character-level text sequence recognition results, image-text association results, and semantic vector results, are assigned weights α=0.4, β=0.3, and γ=0.3, respectively. The final confidence of each content unit is calculated by weighted voting.

8. The multimodal PDF content cleansing and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process of entering the dynamic threshold adjustment stage is as follows: If a content unit has a confidence level of less than 90%, it is considered unreliable and enters the dynamic threshold adjustment stage. Construct a confusion matrix based on historical processing data, statistically analyze the recognition error distribution of different types of content, and determine the basic fault tolerance threshold for each type of content. The similarity between the merged data and the standard format template is calculated: Similarity > 85%: Determined to be correctable, template rules will be automatically applied for repair; Similarity between 70% and 85%: determined to require manual review and annotation; Similarity <70%: judged as high risk, retain the original content, and record the reason for the error; For content that triggers manual review and labeling or is marked as high-risk, a review work order is generated, which includes the original data, automatic recognition results, error analysis and correction options. It supports online manual editing and synchronous updates of the final results.

9. The multimodal PDF content cleanup and intelligent review system based on enhanced visual perception as described in claim 1, characterized in that: The specific process for cross-platform verification is as follows: Develop a core module for format conversion, supporting the conversion of cleaned structured data into two mainstream formats: DOCX format: Preserves editability, ensuring that the hierarchical structure and style of text, tables, and images can be modified using Word and WPS editors; PDF / A standard format: meets long-term archiving needs, embeds font and layout information, and avoids display anomalies across devices; Perform a rendering test on the converted document in the editor and check the following: Text style consistency; Table structure integrity; Accuracy of image-text association; For DOCX format files, compress redundant format tags to reduce file size; Optimize image compression ratio for PDF / A format files.

10. A multimodal PDF content cleanup and intelligent review method based on enhanced visual perception, characterized by: S1: Scan the original PDF to generate a base layer image, use an edge detection algorithm to identify the main text region in the base layer image, use bilinear interpolation to dynamically generate a multi-resolution image sequence for the main text region, input the multi-resolution image sequence into a deep convolutional network to output a fine structure feature map, develop a text region detection model based on OpenCV, scan the multi-resolution image sequence, and locate invisible characters. S2: Establish a knowledge graph, construct a DST-Transformer model, input text region features into the DST-Transformer model and combine it with the knowledge graph to generate structural units, use a convolutional recurrent neural network to perform character recognition on the main text region, generate character-level text sequences, extract visual features from the main text region, perform semantic encoding on the character-level text sequences to generate semantic vectors, calculate the similarity between visual features and semantic vectors to generate image-text associations, and correct errors in image-text associations; S3: The confidence score is calculated by weighting the character-level text sequence recognition results, image-text association, and semantic vector results. If the confidence score is unreliable, the dynamic threshold adjustment stage is entered, and the text content after dynamic threshold adjustment is verified across platforms.

Images