A chemical single line diagram tube support information extraction method

Abstract

The application discloses a chemical single-line diagram pipe support information extraction method, which first analyzes the layout of the chemical single-line diagram, generates and screens target candidate regions based on multiple source prior characteristics such as text hot areas, lead wire endpoints and graphic element densities; long straight line background suppression is performed in the candidate regions to weaken pipeline interference, and a multi-scale feature pyramid network is combined to perform high-precision detection of small symbols; subsequently, candidate triplets of associated symbols, characters and line items are constructed; a multi-modal large model is introduced to perform deep semantic judgment, OCR error correction and attribute completion on the triplets, and output support objects with comprehensive confidence; finally, the chemical field rule library and knowledge graph are used for verification, and a self-adaptive fallback mechanism is triggered when the verification fails or the confidence is insufficient. The whole-process closed loop from knowledge recognition to cognitive understanding to expert verification is realized, and the integrity and accuracy of the pipe support information extraction are significantly improved.

Description

Technical Field

[0001] This application relates to the field of engineering construction technology, specifically to a method for extracting information from a single-line diagram of a chemical engineering pipe support. Background Technology

[0002] In chemical and petrochemical engineering projects, isometric drawings are core design documents used to guide pipeline prefabrication, installation, and material statistics. These drawings not only contain pipeline routes, fittings, and valve information, but also a wealth of critical engineering data, expressed in a combination of symbols and text, including pipe support numbers, specifications, elevations, and material properties. The accurate extraction of pipe support information directly impacts material procurement quantities, budget accuracy, and construction schedule planning.

[0003] In current engineering practice, pipe support statistics mainly rely on manual reading and summarizing of individual single-line diagrams. Due to the high information density, dense isometric lines, small support symbols that easily overlap with pipelines, and differences in design standards in chemical engineering single-line diagrams, manual identification is labor-intensive, time-consuming, and prone to omissions, misreadings, and duplicate statistics. Any deviation in the statistical data will directly affect the quantity of materials procured, the accuracy of budgets, and the construction organization, thereby increasing project risks.

[0004] While existing technologies have proposed drawing parsing methods based on OCR or object detection, they generally suffer from the following technical shortcomings: (1) The problem of low accuracy in identifying pipe supports in complex chemical single-line diagrams; Due to the dense information in chemical single-line diagrams, pipelines, isometric lines, component lines and text annotations intersect and overlap, and the pipe support symbols are extremely small and easily overlap with isometric lines, resulting in problems of missed detection, false detection and missing fields in existing OCR or single image detection methods, making it difficult to reliably identify pipe support symbols and their related information.

[0005] (2) The problem of the inability to form a complete support object due to the cross-modal dispersion of support information; the support information is usually scattered in different locations on the drawing. Existing technologies can generally only identify some of the information and lack a reliable cross-modal association mechanism. It is impossible to automatically bind symbols, annotations and list items, thus making it difficult to generate a complete support structured attribute table.

[0006] (3) The lack of professional verification and error correction makes it difficult to deliver the results to the project. Most existing methods are limited to the identification level and lack rule verification and constraint reasoning error correction mechanisms based on common sense in chemical engineering. This results in errors such as numbering, specifications, elevation, and model in the identification results, which are difficult to detect and correct automatically. A lot of manual review is still required, and they cannot be directly used in the project delivery process such as budgeting, procurement and construction management. Summary of the Invention

[0007] This application provides a method for extracting pipe support information from a single-line diagram in the chemical industry, which can solve the technical problem of low accuracy in identifying pipe supports in complex single-line diagrams in the prior art.

[0008] In a first aspect, embodiments of this application provide a method for extracting information from a single-line diagram of a chemical engineering pipe support, comprising: Perform layout analysis on the chemical engineering single-line diagram to obtain row and item information in the material table area and image and annotation text information in the drawing area; Initial candidate regions are generated in the drawing area based on preset prior features, and then target candidate regions are selected through confidence evaluation. Background interference suppression and multi-scale target detection are performed on the target candidate regions to output the tube support symbol recognition results. The tube support symbol recognition results include symbol category, bounding box coordinates and detection confidence.

[0009] Preferably, the following steps are also included: Using the tube support symbol recognition result as the anchor point, and based on the lead line direction and spatial topology of the image of the corresponding target candidate region, the corresponding candidate text set is identified from the labeled text information. At the same time, the candidate line item set corresponding to the candidate text set is retrieved from the line item information. The tube support symbol recognition result, the candidate text set, and the candidate line item set are combined to generate multiple candidate triplets of the target candidate region. Semantic determination is performed on multiple candidate triples to output the comprehensive confidence score of each candidate triple; candidate triples with a comprehensive confidence score greater than a first preset threshold are used as target scaffold objects for the target candidate region; Multiple target scaffold objects are validated using a rule base and knowledge graph in the chemical engineering field to output the final target scaffold object.

[0010] Preferably, an initial candidate region is generated in the drawing area based on preset prior features, and then the target candidate region is selected through confidence evaluation, specifically including: The text hotspot, leader endpoint dense area, and primitive density response area of ​​the drawing area are used as candidate windows; The candidate windows are deduplicated and merged to obtain an initial candidate region; the deduplication and merging step includes: when the intersection-union ratio of any two adjacent candidate windows exceeds a second preset threshold θ iou When the time comes, merge the two candidate windows; Calculate the overall score for each initial candidate region. ; The initial candidate regions are sorted from highest to lowest based on their overall scores, and only the top K initial candidate regions are selected as the target candidate regions.

[0011] Preferably, the text hotspot, leader endpoint dense area, and primitive density response area of ​​the drawing area are used as candidate windows, specifically including the following steps: The text block t that conforms to the format of bracket number-elevation-specification in the labeled text information is taken as the text hot area, and then the neighborhood expansion is performed to generate the first type of candidate window; Obtain the dense area of ​​detected lead wire endpoints in the drawing area, and generate a second type of candidate window with each endpoint as the center; Obtain the detected primitive density response area in the drawing area, calculate the density response of the suspected symbol using a sliding window, and take the area corresponding to the maximum density response as the third type of candidate window; The first type of candidate window, the second type of candidate window, and the third type of candidate window are used as candidate windows; The overall score The calculation formula is: Where S_text represents the weighted average of text pattern hit and OCR recognition confidence within the initial candidate region; S_lead represents the consistency between lead endpoint density and direction; S_dens represents the density response; S_shape represents the rapid morphological coarse screening response; and α, β, γ, and η are weights.

[0012] Preferably, background interference suppression and multi-scale target detection are performed on the target candidate region to output the tube support symbol recognition result, specifically including: Within the current target candidate region, detect the set of long straight lines {l n A line mask M_line is generated, and pixel-level attenuation is performed on the image of the current target candidate region to obtain the suppressed image; the attenuation formula is: Where I is the image of the current target candidate region, I′ is the suppressed image, λ is the attenuation coefficient, and ⊙ represents element-wise multiplication; The suppressed image I′ is scaled at different scales and an image pyramid sequence is constructed. This sequence is then input into an object detection model with a pyramid network for inference, outputting a set of original predicted bounding boxes at each scale. During inference, the resolution feature layer of the object detection model is used for feature extraction. An original predicted bounding box includes the symbol category, bounding box coordinates, and detection confidence. The original predicted bounding box set is deduplicated by category using non-maximum suppression or soft non-maximum suppression to remove redundant boxes whose spatial overlap exceeds a set threshold, thus obtaining the retained target boxes. Based on the symbol category, bounding box coordinates, and detection confidence of the retained target box, the tube support symbol recognition result is generated.

[0013] Preferably, if the tube support symbol recognition result and the labeled text information in the current target candidate area meet any of the following triggering conditions, the rollback mechanism is triggered; The triggering conditions include: a first condition, no stent symbol recognition result is detected, but stent number text is detected; and a second condition, the detection confidence of the stent symbol recognition result is lower than a preset threshold for detection confidence. The rollback mechanism includes: Increase the initial candidate region screening K value, expand the candidate window size, adjust the attenuation coefficient λ, or increase the resolution of the resolution feature layer until the tube support symbol recognition result is detected or the maximum number of iterations is reached.

[0014] Preferably, using the tube stent symbol recognition result as the anchor point, and based on the lead-line orientation and spatial topological relationship of the corresponding target candidate region image, the corresponding candidate text set is identified from the labeled text information. Simultaneously, the candidate line item set corresponding to the candidate text set is retrieved from the line item information. Based on the tube stent symbol recognition result, the candidate text set, and the candidate line item set, multiple candidate triples are generated, specifically including: For each tube support symbol recognition result s, calculate its geometric consistency score and semantic consistency score with the neighboring text block t, and then calculate the joint score; retain the top N text blocks according to the joint score and use them as the candidate text set {t}. For each text block in the candidate text set, retrieve the matching row item in the material list row item information, and calculate the field consistency score MatchScore(t,b). Based on the field consistency score, retain the top M candidate row items to form the candidate row item set {b}. The identification result s of each tube support symbol is combined with the elements in the corresponding candidate text set {t} and candidate line item set {b} to generate multiple candidate triples {s,t,b}; Preferably, the geometric consistency score and semantic consistency score between the text block t and its neighboring text block t are calculated, and then the joint score Score(s,t) is calculated, which specifically includes the following steps: Obtain distance consistency, orientation consistency, connectivity consistency, and ROI constraint consistency, and combine them with the geometric consistency score calculation formula to obtain the geometric consistency score; The formula for calculating the geometric consistency score is as follows: ;in, For distance consistency, For consistency of direction, For connectivity consistency, Consistency of ROI constraints; , and The corresponding weighting coefficients; Obtain field schema consistency, contextual semantic consistency, and semantic confidence, and combine them with the semantic consistency score calculation formula to obtain the semantic consistency score; The formula for calculating the semantic consistency score is as follows: ;in, For field schema consistency, Contextual semantic consistency For semantic confidence, , and The corresponding weighting coefficients; The joint score Score(s,t) is calculated using the joint score calculation formula; the joint score calculation formula is as follows: ; The weights for geometric consistency scores, The weights for semantic consistency scores.

[0015] Preferably, semantic determination is performed on multiple candidate triples to output the comprehensive confidence score of each candidate triple, specifically including: For each candidate triple (s,t,b), the corresponding analysis data is encapsulated to form a multimodal evidence package. The analysis data includes visual evidence containing the tube support symbol recognition result s, the lead wire and the local image slice of the text block t, textual evidence containing the OCR recognition original text of the text block t and its character-level confidence, tabular evidence containing the field data of the candidate row item b and its coordinate position in the material table, and rule evidence containing preset field format templates, unit specifications and common character confusion rules. Semantic analysis of multimodal evidence packages is performed using a large multimodal model to perform the following operations: Correcting erroneous characters identified by OCR based on visual evidence context; Determine whether the specifications and material information in text block t match the candidate row item b based on textual and tabular evidence, and remove mismatched candidate row items b; complete missing units based on rule-based evidence; Output the binding confidence of each candidate triple, and select the binding confidence of the candidate triple with the highest binding confidence as the final binding confidence; The comprehensive confidence of the candidate triple (s,t,b) is calculated based on the detection confidence, OCR recognition confidence, and final binding confidence of the tube stent symbol recognition result s; The formula for calculating the overall confidence level is: ; Where λ1, λ2, and λ3 are preset as fusion weights, To determine the detection confidence level of the tube support symbol recognition result s, To determine the confidence level for OCR recognition, To bind confidence levels.

[0016] Preferably, when the overall confidence level is less than or equal to the first preset threshold, the initial candidate region screening K value is increased, a high-precision OCR model is called to recognize text blocks, or the field consistency retrieval conditions are adjusted to obtain multiple new candidate triples {s,t,b}. Semantic determination is performed on multiple new candidate triples {s,t,b}, and a comprehensive confidence score is calculated and judged.

[0017] The beneficial effects of the technical solutions provided in this application include: A comprehensive technical framework integrating layout analysis, candidate area selection, and anti-interference detection was constructed to address the low accuracy of pipe support identification in complex chemical drawings. Its core lies in changing the inefficient traditional blind inspection model, instead employing a strategy of prior guidance combined with localized precision inspection. First, layout analysis decouples unstructured drawings into material lists and drawing areas, isolating irrelevant noise at the data source. Second, pre-defined prior features are used to generate initial candidate areas and select high-confidence target areas, significantly narrowing the search space for subsequent detection and avoiding wasted computational resources in large blank or purely pipeline areas. Finally, background interference suppression and multi-scale detection are performed on the target areas to specifically address the pain point of overlapping isometric lines and symbols in chemical drawings. This layered and progressive processing flow not only significantly reduces the false activation of small symbols by long straight backgrounds but also improves the signal-to-noise ratio of the detector by focusing on key areas. Thus, while ensuring processing speed, it fundamentally improves the detection rate and positioning accuracy of pipe support symbols in dense line environments, laying a solid foundation for subsequent information association. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the initial candidate region for the chemical single-line diagram pipe support information extraction method of this application; Figure 2 This is a schematic diagram of the target candidate region after background interference suppression in the chemical single-line diagram pipe support information extraction method of this application; Figure 3 This is a schematic diagram illustrating the general process of the chemical single-line diagram pipe support information extraction method of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present application.

[0020] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0021] In engineering construction fields such as chemical and petrochemical industries, isometric drawings are crucial design documents used to guide pipeline prefabrication, installation, and material statistics. Isometric drawings typically express the three-dimensional spatial routing of pipelines using isometric projection, and describe component information within the pipeline system through graphic symbols and textual annotations. The drawing content mainly includes: pipeline routing, fittings, valves, welds, instrument points, pipe support symbols and their numbers, specifications, elevations, and other engineering attributes.

[0022] In current engineering practice, pipe support information is mainly used to generate support ledgers, material procurement lists, and budget statistics tables. Its fields typically include support number, support type, specifications, elevation, material information, and quantity. Due to the high information density of single-line diagrams, the variety of symbols, and the differences in standards among different design institutes, manual identification and statistics are labor-intensive, time-consuming, and prone to problems such as missed detections, misreadings, and duplicate statistics.

[0023] To improve the efficiency of drawing information processing, several well-known technical approaches for intelligent drawing recognition and information extraction have been developed in the existing technology, mainly including the following categories: OCR text recognition technology The readability of the text is improved by image preprocessing (denoising, binarization, tilt correction, etc.), and then the text content is recognized by an OCR engine. Finally, fields such as number, specification, and elevation in the drawing are extracted by keywords or regular expression rules.

[0024] Page Layout Analysis and Region Segmentation Techniques Engineering drawings often include structures such as title blocks, explanatory areas, main drawing areas, and table areas. Existing technologies typically use layout detection models or geometric rule-based methods to divide these areas, allowing for the application of different recognition strategies in different regions.

[0025] Symbol detection and classification techniques For graphic symbols such as valves, instruments, and brackets, existing technologies typically use target detection networks (such as YOLO, Faster R-CNN, etc.) to locate and classify the symbols in order to obtain the symbol's category and coordinate information.

[0026] Leader / Arrow Analysis and Spatial Association Techniques In engineering drawings, leader lines or arrows are often used to point text labels to corresponding components. Existing technologies can extract leader lines using methods such as line segment detection, skeletonization, and connected component analysis, and then match text with component symbols based on spatial neighborhood relationships.

[0027] Table recognition and bill of materials parsing technology Single-line diagrams often contain bills of materials or component lists. Existing technologies typically obtain table row fields, such as component name, specifications, and quantity, through table structure recognition and OCR parsing.

[0028] The above-mentioned technologies constitute the technical basis of this application. However, due to the characteristics of chemical single-line diagrams, such as extremely small support symbols, severe overlap with isometric lines, and dispersed information expression across modalities, the existing technologies still have significant shortcomings in extracting information from pipe supports.

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0030] Firstly, a chemical engineering single-line diagram pipe support information extraction system is provided, comprising: an intelligent OCR and layout analysis engine (perception layer), a multimodal large model engine (cognitive layer), and a chemical engineering professional algorithm and knowledge graph engine (expert layer). These three engines are coordinated through a collaborative control module and transmit data using a unified intermediate representation, achieving a closed-loop process of "drawing upload—layout analysis—multimodal recognition and association—professional verification—structured output—visualized report." This solves the following problems: 1) The problem of low accuracy in identifying pipe supports in complex chemical single-line diagrams; Due to the dense information in chemical single-line diagrams, pipelines, isometric lines, component lines and text annotations intersect and overlap, and the pipe support symbols are extremely small and easily overlap with isometric lines, resulting in problems such as missed detection, false detection and missing fields in existing OCR or single image detection methods, making it difficult to reliably identify pipe support symbols and their related information.

[0031] 2) The problem of the inability to form a complete support object due to the cross-modal dispersion of support information; the support information (symbols, leaders, text annotations, bill of materials items) is usually scattered in different locations on the drawings. Existing technologies can generally only identify some of the information and lack a reliable cross-modal association mechanism. It is impossible to automatically bind symbols, annotations and bill of materials items, thus making it difficult to generate a complete structured attribute table of the support.

[0032] 3) The lack of professional verification and error correction makes it difficult to deliver the results to the project. Most existing methods are limited to the identification level and lack rule verification and constraint reasoning error correction mechanisms based on common sense in chemical engineering. This results in errors such as numbering, specifications, elevation, and model in the identification results, which are difficult to detect and correct automatically. A lot of manual review is still required, and they cannot be directly used in the project delivery process such as budgeting, procurement and construction management.

[0033] Based on a chemical single-line diagram pipe support information extraction system, this application embodiment provides a method for extracting chemical single-line diagram pipe support information, and will explain the above-mentioned problems one by one: Step 100: Utilize the perception layer to perform format parsing and standardization conversion on the chemical engineering single-line diagram, generating a standard image with uniform resolution and orientation, while preserving the mapping relationship from the original coordinates to the standard coordinates for subsequent result writing and visualization. Preprocessing may include noise reduction, contrast enhancement, tilt correction, binarization / grayscale dual-channel generation, etc., followed by layout analysis to obtain row information in the material table area and image and annotation text information in the drawing area. The perception layer performs layout analysis on the drawing, dividing it into at least the following areas: The drawing area includes pipelines, support symbols, leader lines / arrows, and local annotations; The materials section contains the bill of materials (BOM) table; The title bar / notes area contains metadata such as drawing number and project number.

[0034] The bounding box coordinates (bbox) and region category label (r) of each output region are then fed into different recognition strategies for the material table area and the drawing area, respectively.

[0035] The identification strategy includes: The drawing area is recognized by OCR to obtain multiple text blocks t, and the multiple text blocks t form the labeled text information. in The text content, location, and confidence level are respectively identified; and preliminary field type annotations are performed on text that conforms to the pipe support field pattern (number / elevation / specification, etc.).

[0036] In the material table area, table structure recognition (row and column boundaries, cell division) is performed, and row item fields are extracted using OCR. This yields field data for multiple row items b, which together form the row item information for the material table area. .

[0037] Step 200: The perception layer generates an initial candidate region in the drawing area based on preset prior features, and then the cognitive layer filters out the target candidate region through confidence evaluation; background interference suppression and multi-scale target detection are performed on the target candidate region to output the tube support symbol recognition result; the tube support symbol recognition result includes symbol category, bounding box coordinates and detection confidence.

[0038] The above constructs an overall technical framework of layout analysis, candidate area selection, and anti-interference detection, primarily addressing the low accuracy of pipe support identification in complex chemical drawings. Its core lies in changing the inefficient traditional full-map blind inspection model, instead adopting a strategy of prior guidance + local precision inspection. First, layout analysis decouples unstructured drawings into material lists and drawing areas, isolating irrelevant noise from the data source. Second, it uses preset prior features to generate initial candidate areas and selects high-confidence target areas, significantly narrowing the search space for subsequent detection and avoiding wasting computational resources in large blank or pure pipeline areas. Finally, it performs background interference suppression and multi-scale detection on target areas, specifically addressing the pain point of overlapping isometric lines and symbols in chemical drawings. This layered and progressive processing flow not only significantly reduces the false activation of small symbols by long straight backgrounds but also improves the signal-to-noise ratio of the detector by focusing on key areas. Thus, while ensuring processing speed, it fundamentally improves the detection rate and positioning accuracy of pipe support symbols in dense line environments, laying a solid foundation for subsequent information association.

[0039] In some preferred embodiments, the following steps are also included: Using the tube stent symbol recognition result as the anchor point, and based on the lead line direction and spatial topology of the corresponding target candidate region image, the corresponding candidate text set is identified from the labeled text information. At the same time, the candidate line item set corresponding to the candidate text set is retrieved from the line item information. The tube stent symbol recognition result, the candidate text set, and the candidate line item set are combined to generate multiple candidate triplets of the target candidate region. Semantic determination is performed on multiple candidate triples to output the comprehensive confidence score of each candidate triple; candidate triples with a comprehensive confidence score greater than a first preset threshold are used as target scaffold objects for the target candidate region; Multiple target scaffold objects are validated using a rule base and knowledge graph in the chemical engineering field to output the final target scaffold object.

[0040] This solution addresses two major challenges: the fragmented nature of support structure information across modalities, leading to the inability to form complete objects, and the lack of professional verification hindering project delivery. It overcomes the limitations of existing technologies that can only identify isolated elements. By using symbols as anchors and leveraging spatial topology and semantic consistency, it logically connects leader lines, annotation text, and material list items scattered across different locations on drawings, constructing structured objects with complete attributes (such as number, specification, and material). The introduction of a multimodal large-scale model for semantic judgment utilizes its powerful contextual understanding capabilities to automatically correct OCR recognition errors and eliminate logically mismatched candidates, resolving ambiguities in cross-modal association. Furthermore, final verification is performed using a chemical engineering rule base and knowledge graph, reviewing the recognition results with common-sense engineering principles (such as model compatibility and elevation rationality), automatically identifying and correcting errors that violate engineering logic. This ensures that the output data is not only formatted completely but also conforms to engineering realities, allowing direct use in budgeting and construction management, significantly reducing manual review costs.

[0041] Specifically, the rule base and knowledge graph in the chemical industry include: The expert layer receives the target scaffold object output by the cognitive layer, which is used for subsequent consistency verification of the target scaffold object based on chemical engineering standards; it automatically corrects and modifies fields for identifiable errors; it triggers backtracking and re-identification for uncertain items to form an engineering-level closed loop; and it outputs data that meets the requirements of project delivery (which can be used for budgeting, procurement, and construction management).

[0042] The expert layer incorporates a built-in chemical engineering common sense rule base and knowledge graph construction, which includes at least: The system includes: a valid set of stent models / types and alias mapping; a model-pipe diameter / pressure rating / pipeline attribute matching relationship; a reasonable range of elevations and unit consistency rules; numbering format rules and uniqueness constraints; integrity and consistency constraints of material list row items and fields; and a knowledge graph that can represent entity relationships such as "stent type-applicable conditions-material specifications-quantity rules" and be used for constraint reasoning and conflict resolution.

[0043] The expert layer performs rule validation on each scaffold object and outputs validation status codes and reason codes, realizing constraint reasoning validation and automatic error correction; typical validations include: Model Existence Verification: Identifies whether the model belongs to a valid set; Compatibility verification: Whether the bracket model matches the pipe diameter / grade / installation conditions; Numbering compliance check: Whether the numbering format conforms to the rules and whether it is duplicated; Elevation rationality verification: consistency of numerical range and units; Material list consistency check: Whether the object specifications / quantity are consistent with the material list row items.

[0044] When a typical misidentification is detected (e.g., "S1" is misidentified as "51"), the expert layer uses the rule base to determine that "51" is not a valid model or is not suitable for the current pipe diameter, thereby automatically correcting it to the most likely valid candidate and updating the confidence level and evidence chain.

[0045] When verification fails and cannot be directly corrected, the expert layer triggers a backtracking strategy, calling the cognitive layer or perception layer to perform local re-identification.

[0046] In some preferred embodiments, initial candidate regions are generated in the drawing area based on preset prior features, and then target candidate regions are selected through confidence evaluation, specifically including: The text hotspot, leader endpoint dense area, and primitive density response area of ​​the drawing area are used as candidate windows; (Refer to...) Figure 1 As shown, ROI1, ROI2, ROI3, and ROI4 are the initial candidate regions. The candidate windows are deduplicated and merged to obtain the initial candidate region; the deduplication and merging step includes: when the intersection-union ratio of any two adjacent candidate windows exceeds a second preset threshold θ iou When the value is 0.3 to 0.5, the two candidate windows are merged. Calculate the overall score for each initial candidate region. ; The initial candidate regions are sorted from highest to lowest based on their overall scores, and only the top K initial candidate regions are selected as target candidate regions.

[0047] The text hotspot, leader endpoint dense area, and primitive density response area of ​​the drawing area are used as candidate windows, and the specific steps include: The text block t that conforms to the format of bracket number-elevation-specification in the labeled text information is taken as the text hotspot, and then neighborhood expansion is performed to generate the first type of candidate window; the first type of candidate window is represented as ;in Indicates the first A text block, Δx and Δy can be adaptively set based on font height estimation or text confidence level to cover the area where potential symbols and lead lines connect. Obtain the densely populated areas of lead wire endpoints detected in the drawing area, and generate a second type of candidate window centered on each endpoint; the second type of candidate window is represented as... ; For the set of lead endpoints, w and h can be adaptively adjusted based on the density of line segments near the endpoints and the local scale. Obtain the detected primitive density response area in the drawing area, calculate the density response of the suspected symbol using a sliding window, and take the area corresponding to the maximum density response as the third type of candidate window; the density response is calculated based on corner density, number of small connected components, or number of thin line intersections; The first type of candidate window, the second type of candidate window, and the third type of candidate window are used as candidate windows; Overall Score The calculation formula is: Where S_text represents the weighted average of text pattern hit and OCR recognition confidence within the initial candidate region; S_lead represents the consistency between lead endpoint density and direction; S_dens represents the density response; S_shape represents the rapid morphological coarse screening response; and α, β, γ, and η are weights.

[0048] This system addresses the issue of missed and false detections of minute symbols in complex backgrounds. Its principle lies in fully utilizing domain knowledge from chemical engineering single-line diagrams: pipe supports are inevitably accompanied by annotation text, lead wire connections, or specific geometric shapes (such as corners and intersections). By fusing these three types of heterogeneous prior information, the system can comprehensively cover potential symbol locations without relying on a single feature, effectively avoiding missed detections caused by symbol ambiguity, occlusion, or broken leads. An IoU deduplication and merging mechanism is introduced to eliminate redundant detection boxes caused by overlapping multiple priors, preventing redundant calculations. More importantly, by constructing a comprehensive scoring formula that includes text hits, lead wire density, and morphological features, the system can intelligently evaluate the symbol probability of each candidate region and perform costly fine-tuning only on the top K regions with the highest scores. This coarse-screening + fine-tuning strategy filters out most meaningless background noise (such as pure pipe segments) and concentrates computational resources on the most likely targets, significantly improving the system's operating efficiency and detection robustness under high-resolution, large-scale images.

[0049] In some preferred embodiments, this application does not specifically limit the type of multimodal large model, as long as it can achieve the following functions. The multimodal large model here is merely a tool, therefore specific training details are not provided; see references. Figure 2 As shown, a multimodal large model of the cognitive layer is used to suppress background interference and detect targets at multiple scales in the target candidate region, outputting the tube support symbol recognition result, specifically including: Within the current target candidate region, detect the set of long straight lines {l n A line mask M_line is generated, and pixel-level attenuation is performed on the image of the current target candidate region to obtain the suppressed image; the attenuation formula is: Where I is the image of the current target candidate region, I′ is the suppressed image, λ is the attenuation coefficient, and ⊙ represents element-wise multiplication; The suppressed image I′ is scaled at different scales (e.g., 1.0× / 1.5× / 2.0×) to construct an image pyramid sequence, which is then input into an object detection model with a pyramid network (using a detection network with FPN) for inference, outputting a set of original predicted bounding boxes at each scale. During inference, feature extraction is performed using the resolution feature layer of the object detection model; one original predicted bounding box is generated. ,in These are the symbol category, bounding box coordinates, and detection confidence level, respectively. The original predicted bounding box set is deduplicated by category using non-maximum suppression or soft non-maximum suppression to remove redundant boxes whose spatial overlap exceeds a set threshold, thus obtaining the retained target boxes; that is, category-based NMS / Soft-NMS deduplication is performed on the multi-scale output box set; when the retained target boxes are overlapping boxes, multiple detection confidence scores are fused. The formula for fusing multiple detection confidence scores is as follows: ;in Let m be the confidence level at scale m. Scale weights; Based on the symbol category, bounding box coordinates, and detection confidence of the retained target box, the tube support symbol recognition result is generated.

[0050] This paper addresses the challenge of false positives and false negatives of tiny symbols caused by background interference from long, straight lines in isometric projections. In chemical engineering single-line diagrams, dense isometric pipelines (long, straight lines) easily generate strong activations in convolutional neural networks, masking tiny support symbols. By detecting sets of long, straight lines to generate a mask Mline and performing suppression operations, the grayscale response of background lines is physically weakened at the pixel level. This is equivalent to putting a filter on the detector, allowing it to focus on the preserved symbol edge structures, thus significantly reducing the false positive rate. Simultaneously, considering the extremely small and highly variable size of support symbols, an image pyramid is constructed, and the P2 / P3 high-resolution feature layers of the FPN are used for inference. This preserves rich detail and texture information, overcoming the defect of deep networks losing small target features due to multiple downsampling. Multi-scale inference combined with NMS deduplication and confidence fusion ensures stable capture regardless of the symbol's scaling ratio in the image. This combined approach significantly improves the algorithm's signal-to-noise ratio and small target recall rate under strong background interference. Furthermore, if the recognition result of the stent symbol and the labeled text information in the current target candidate area meet any of the following triggering conditions, the rollback mechanism will be triggered. The triggering conditions include: First condition, no stent symbol recognition result is detected, but stent number text is detected; Second condition, the detection confidence of the stent symbol recognition result is lower than the preset detection confidence threshold. The rollback mechanism includes: Increase the initial candidate region screening K value, expand the candidate window size, adjust the attenuation coefficient λ, or increase the resolution of the resolution feature layer until the tube support symbol recognition result is detected or the maximum number of iterations is reached.

[0051] This solution primarily addresses the insufficient robustness of fixed-parameter detection strategies under conditions of occlusion, low resolution, or abnormal layouts. Its principle lies in introducing a dynamic closed-loop control logic of "detection-evaluation-adjustment-retry." Traditional static detection, if parameters are improperly set (e.g., too small a ROI, excessive background suppression), can lead to permanent missed detections. This solution intelligently determines the possible reasons for detection failure by monitoring two typical anomalous signals: "text without symbols" or "low confidence." It then adjusts the strategy accordingly: if insufficient search range is suspected, the K value is increased or the window is enlarged; if excessive background suppression is suspected of erasing symbols, the attenuation coefficient λ is adjusted; if insufficient resolution is suspected, super-resolution inference is enabled. This adaptive adjustment capability, similar to human experts "getting closer when it's hard to see, darkening when it's too bright," gives the system a self-correcting and fault-tolerant mechanism. It significantly reduces the missed detection rate caused by fluctuations in drawing quality or local special circumstances, ensuring that the system maintains a high level of recall performance even in complex and ever-changing engineering drawing environments.

[0052] In some preferred embodiments, the tube stent symbol recognition result is used as the anchor point, and the corresponding candidate text set is identified from the labeled text information based on the lead-line orientation and spatial topological relationship of the image of the corresponding target candidate region. At the same time, the candidate row item set corresponding to the candidate text set is retrieved from the row item information. Based on the tube stent symbol recognition result, the candidate text set, and the candidate row item set, multiple candidate triples are generated, specifically including: For each tube support symbol recognition result s, calculate its geometric consistency score and semantic consistency score with the neighboring text block t, and then calculate the joint score; retain the top N text blocks according to the joint score and use them as the candidate text set {t}. The joint score Score(s,t) is calculated by the following steps: The center of the tube support symbol is set as The center of text block t is ; Distance Consistency ; Directional consistency (based on lead) Let the direction vector of the lead wire be... The direction from symbol to text is ,but: ; Connectivity Consistency ; Constraint Consistency ; Geometric Consistency Total Score ; Calculate semantic consistency score Field schema consistency ; Contextual semantic consistency ; Large model semantic confidence ; Semantic consistency total score ; Joint rating .

[0053] in, , and The corresponding weight coefficients, , and The corresponding weighting coefficients; The weights for geometric consistency scores, The weights for semantic consistency scores.

[0054] For each text block in the candidate text set, retrieve the matching row item in the material list row item information, and calculate the field consistency score MatchScore(t,b). Based on the field consistency score, retain the top M candidate row items to form the candidate row item set {b}. The identification result s of each tube support symbol is combined with the elements in the corresponding candidate text set {t} and candidate line item set {b} to generate multiple candidate triples {s,t,b}.

[0055] That is, s→t is determined based on spatial relationships and the direction of the lead wire. For each s, select N text blocks t, and generate candidate symbol-to-text mappings based on the joint score Score(s,t). The specific steps are as follows: Select the N highest-scoring text blocks from the candidate set:

[0056] Symbol-to-text mapping relationship:

[0057] When there are multiple candidates for t, N is retained to proceed to the next step, and the geometric / semantic scores of the candidates are used as input features for subsequent large model decisions.

[0058] Determine t based on field consistency b For each candidate text t, retrieve candidate row item b from the row item information B in the material table area. The retrieval criteria include: model / specification field matching, name pattern matching, and material code / abbreviation dictionary matching.

[0059] Field consistency score can be represented as: ; in, Indicates the similarity of specification fields. Indicates the semantic similarity of the name fields. select, , , This represents the weighting coefficient. The top M candidate rows, candidate row b, are retained, along with matching evidence (hit fields and cell coordinates).

[0060] The above describes how to construct candidate triples of symbol-text-line items using geometric consistency (leader lines, direction, distance) and semantic consistency (format, units), primarily addressing the problem of high association error rates caused by the dispersion of cross-modal information. The principle lies in simulating the cognitive process of an engineer reading an image: firstly, geometric constraints (such as leader line connectivity, vector angle, and spatial distance) lock the correspondence between symbols and text in physical space, eliminating interfering items with inconsistent spatial positions; then, semantic constraints (such as numbering regularization and unit validity) perform secondary filtering in the logical space, eliminating mismatches that are adjacent but semantically incompatible. This dual verification mechanism of "geometric + semantic" greatly compresses the candidate search space, ensuring that the candidate set entering subsequent large-scale model processing has extremely high purity. This step successfully aggregates scattered image elements, text strings, and tabular data into meaningful logical units, providing a reliable intermediate representation for generating complete structured scaffold objects, effectively solving the ambiguity and noise problems in cross-modal association.

[0061] In some preferred implementations, semantic determination is performed on multiple candidate triples to output the overall confidence score of each candidate triple, specifically including: For each candidate triple (s,t,b), the corresponding analysis data is encapsulated to form a multimodal evidence package. The analysis data includes visual evidence containing the tube support symbol recognition result s, the lead wire and the local image slice of the text block t, textual evidence containing the OCR recognition original text of the text block t and its character-level confidence, tabular evidence containing the field data of the candidate row item b and its coordinate position in the material table, and rule evidence containing the preset field format template, unit specifications and common character confusion rules. Semantic analysis of multimodal evidence packages is performed using a large multimodal model to perform the following operations: Correcting erroneous characters identified by OCR based on visual evidence context; Determine whether the specifications and material information in text block t match the candidate row item b based on textual and tabular evidence, and remove mismatched candidate row items b; complete missing units based on rule-based evidence; Output the binding confidence of each candidate triple, and select the binding confidence of the candidate triple with the highest binding confidence as the final binding confidence; The comprehensive confidence of the candidate triple (s,t,b) is calculated based on the detection confidence, OCR recognition confidence, and final binding confidence of the tube stent symbol recognition result s; The formula for calculating the overall confidence level is: ; Where λ1, λ2, and λ3 are preset as fusion weights, To determine the detection confidence level of the tube support symbol recognition result s, To determine the confidence level for OCR recognition, This represents the final binding confidence level.

[0062] This approach addresses the limitations of traditional rules in handling complex semantic errors and OCR misidentification. Its principle lies in fully leveraging the powerful multimodal understanding and reasoning capabilities of a large-scale model: by constructing an evidence package containing visual slices, OCR original text, tabular data, and engineering rules, the large model can "read images, decipher text, and look up tables" like a human expert. It uses visual context to correct similar-looking errors in OCR (such as '0' and 'O'), uses logical mutual exclusion to eliminate rows and items with mismatched specifications and materials, and uses domain knowledge to complete missing units. This context-based deep reasoning far surpasses the capabilities of traditional regular expression matching. Furthermore, the proposed comprehensive confidence fusion formula organically combines the underlying detection confidence and recognition confidence with the high-level semantic binding confidence, forming a global credibility index. This not only quantifies the reliability of the results but also provides a scientific basis for subsequent threshold screening and backtracking mechanisms, ensuring that the final output scaffold object is semantically self-consistent and data-complete, significantly improving the system's intelligence level and delivery quality.

[0063] Furthermore, when the overall confidence level is less than or equal to the first preset threshold, the initial candidate region screening K value is increased, a high-precision OCR model is called to recognize text blocks, or the field consistency retrieval conditions are adjusted to obtain multiple new candidate triples {s,t,b}. Semantic determination is performed on multiple new candidate triples {s,t,b}, and a comprehensive confidence score is calculated and judged.

[0064] To address the issues of single-pass recognition not guaranteeing 100% accuracy and the difficulty in eliminating long-tail errors, a "progressive" error correction mechanism is established. When the overall confidence level of the initial judgment is insufficient, the system does not directly abandon or report an error, but instead initiates different levels of remedial measures based on the possible sources of error. Level 1 backtracking addresses the "not found" problem by expanding the search range (increasing the K value); Level 2 backtracking addresses the "inaccurate recognition" problem by invoking high-precision OCR; and Level 3 backtracking addresses the "unconnected" problem by relaxing the search criteria. This hierarchical processing strategy avoids the resource waste caused by using high-cost strategies from the outset and ensures sufficient technical means to fall back when encountering difficult samples. By performing semantic judgment and confidence calculation again on newly generated candidate triples, the system forms a continuously iterative optimization closed loop until the results meet the delivery standards or are confirmed as difficult cases requiring manual intervention.

[0065] Thirdly, embodiments of this application provide a chemical single-line diagram pipe support information extraction device, which can be a personal computer (PC), laptop computer, server or other device with data processing capabilities.

[0066] In this embodiment, the chemical single-line diagram pipe support information extraction device may include a processor, a memory, a communication interface, and a communication bus.

[0067] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0068] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting internal components of the chemical single-line diagram pipe support information extraction device, as well as interfaces used for interconnecting the chemical single-line diagram pipe support information extraction device with other devices (such as other computing devices or user devices). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user devices can be displays, keyboards, etc.

[0069] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0070] The processor can be a general-purpose processor, which can call the chemical single-line diagram pipe support information extraction program stored in the memory and execute the chemical single-line diagram pipe support information extraction method provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the chemical single-line diagram pipe support information extraction program is called can refer to the various embodiments of the chemical single-line diagram pipe support information extraction method of this application, and will not be repeated here.

[0071] Fourthly, embodiments of this application also provide a computer-readable storage medium.

[0072] This application provides a computer-readable storage medium storing a chemical single-line diagram pipe support information extraction program, wherein when the chemical single-line diagram pipe support information extraction program is executed by a processor, it implements the steps of the chemical single-line diagram pipe support information extraction method described above.

[0073] The method implemented when the chemical single-line diagram pipe support information extraction program is executed can be referred to in various embodiments of the chemical single-line diagram pipe support information extraction method of this application, and will not be repeated here.

[0074] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0075] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0076] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0077] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0078] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0080] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for extracting information from a single-line diagram of a chemical engineering pipe support, characterized in that, It includes: Perform layout analysis on the chemical engineering single-line diagram to obtain row and item information in the material table area and image and annotation text information in the drawing area; Initial candidate regions are generated in the drawing area based on preset prior features, and then target candidate regions are selected through confidence evaluation. Background interference suppression and multi-scale target detection are performed on the target candidate region to output the tube support symbol recognition result; The results of tube support symbol recognition include symbol category, bounding box coordinates, and detection confidence.

2. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 1, characterized in that, It also includes the following steps: Using the tube support symbol recognition result as the anchor point, and based on the lead line direction and spatial topology of the image of the corresponding target candidate region, the corresponding candidate text set is identified from the labeled text information, and the candidate line set corresponding to the candidate text set is retrieved from the line item information. The results of tube support symbol recognition, candidate text set, and candidate line item set are combined to generate multiple candidate triples for the target candidate region; Semantic determination is performed on multiple candidate triples to output the overall confidence of each candidate triple; Candidate triples with a comprehensive confidence level greater than the first preset threshold are used as target scaffold objects for the target candidate region. Multiple target scaffold objects are validated using a rule base and knowledge graph in the chemical engineering field to output the final target scaffold object.

3. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 2, characterized in that, Initial candidate regions are generated in the drawing area based on preset prior features, and then target candidate regions are selected through confidence evaluation, specifically including: The text hotspot, leader endpoint dense area, and primitive density response area of ​​the drawing area are used as candidate windows; The candidate windows are deduplicated and merged to obtain an initial candidate region; the deduplication and merging step includes: when the intersection-union ratio of any two adjacent candidate windows exceeds a second preset threshold θ iou When the time comes, merge the two candidate windows; Calculate the overall score for each initial candidate region. ; The initial candidate regions are sorted from highest to lowest based on their overall scores, and only the top K initial candidate regions are selected as the target candidate regions.

4. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 3, characterized in that: Using the text hotspot, leader endpoint dense area, and primitive density response area of ​​the drawing area as candidate windows, the specific steps include: The text block t that conforms to the format of bracket number-elevation-specification in the labeled text information is taken as the text hot area, and then the neighborhood expansion is performed to generate the first type of candidate window; Obtain the dense area of ​​detected lead wire endpoints in the drawing area, and generate a second type of candidate window with each endpoint as the center; Obtain the detected primitive density response area in the drawing area, calculate the density response of the suspected symbol using a sliding window, and take the area corresponding to the maximum density response as the third type of candidate window; The first type of candidate window, the second type of candidate window, and the third type of candidate window are used as candidate windows; The overall score The calculation formula is: Where S_text represents the weighted average of text pattern hit and OCR recognition confidence within the initial candidate region; S_lead represents the consistency between lead endpoint density and direction; S_dens represents the density response; and S_shape represents the rapid morphological coarse screening response.

5. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 4, characterized in that, Background interference suppression and multi-scale target detection are performed on the target candidate region to output the tube support symbol recognition result, specifically including: Within the current target candidate region, detect the set of long straight lines {l n A line mask M_line is generated, and pixel-level attenuation is performed on the image of the current target candidate region to obtain the suppressed image; the attenuation formula is: Where I is the image of the current target candidate region, I′ is the suppressed image, λ is the attenuation coefficient, and ⊙ represents element-wise multiplication; The suppressed image I′ is scaled at different scales and an image pyramid sequence is constructed. This sequence is then input into an object detection model with a pyramid network for inference, outputting a set of original predicted bounding boxes at each scale. During inference, the resolution feature layer of the object detection model is used for feature extraction. An original predicted bounding box includes the symbol category, bounding box coordinates, and detection confidence. The original predicted bounding box set is deduplicated by category using non-maximum suppression or soft non-maximum suppression to remove redundant boxes whose spatial overlap exceeds a set threshold, thus obtaining the retained target boxes. Based on the symbol category, bounding box coordinates, and detection confidence of the retained target box, the tube support symbol recognition result is generated.

6. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 5, characterized in that: If the tube support symbol recognition result and the labeled text information in the current target candidate area meet any of the following triggering conditions, the rollback mechanism will be triggered; The triggering conditions include: a first condition, no stent symbol recognition result is detected, but stent number text is detected; and a second condition, the detection confidence of the stent symbol recognition result is lower than a preset threshold for detection confidence. The rollback mechanism includes: Increase the initial candidate region screening K value, expand the candidate window size, adjust the attenuation coefficient λ, or increase the resolution of the resolution feature layer until the tube support symbol recognition result is detected or the maximum number of iterations is reached.

7. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 2, characterized in that, Using the tube stent symbol recognition result as the anchor point, and based on the guide line direction and spatial topological relationship of the corresponding target candidate region image, the corresponding candidate text set is identified from the labeled text information. Simultaneously, the candidate row item set corresponding to the candidate text set is retrieved from the row item information. Based on the tube stent symbol recognition result, the candidate text set, and the candidate row item set, multiple candidate triples are generated, specifically including: For each tube support symbol recognition result s, calculate its geometric consistency score and semantic consistency score with the neighboring text block t, and then calculate the joint score; retain the top N text blocks according to the joint score and use them as the candidate text set {t}. For each text block in the candidate text set, retrieve the matching row item in the material list row item information, and calculate the field consistency score MatchScore(t,b). Based on the field consistency score, retain the top M candidate row items to form the candidate row item set {b}. The identification result s of each tube support symbol is combined with the elements in the corresponding candidate text set {t} and candidate line item set {b} to generate multiple candidate triples {s,t,b}.

8. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 7, characterized in that: Calculate the geometric consistency score and semantic consistency score between the text block t and its neighboring text block t, and then calculate the joint score Score(s,t), which includes the following steps: Obtain distance consistency, orientation consistency, connectivity consistency, and ROI constraint consistency, and combine them with the geometric consistency score calculation formula to obtain the geometric consistency score; The formula for calculating the geometric consistency score is as follows: ;in, For distance consistency, For consistency of direction, For connectivity consistency, Consistency of ROI constraints; , and The corresponding weighting coefficients; Obtain field schema consistency, contextual semantic consistency, and semantic confidence, and combine them with the semantic consistency score calculation formula to obtain the semantic consistency score; The formula for calculating the semantic consistency score is as follows: ;in, For field schema consistency, Contextual semantic consistency For semantic confidence, , and The corresponding weighting coefficients; The joint score Score(s,t) is calculated using the joint score calculation formula; the joint score calculation formula is as follows: ; The weights for geometric consistency scores, The weights for semantic consistency scores.

9. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 7, characterized in that, Semantic determination is performed on multiple candidate triples to output the overall confidence score of each candidate triple, specifically including: For each candidate triple (s,t,b), the corresponding analysis data is encapsulated to form a multimodal evidence package. The analysis data includes visual evidence containing the tube support symbol recognition result s, the lead wire and the local image slice of the text block t, textual evidence containing the OCR recognition original text of the text block t and its character-level confidence, tabular evidence containing the field data of the candidate row item b and its coordinate position in the material table, and rule evidence containing preset field format templates, unit specifications and common character confusion rules. Semantic analysis of multimodal evidence packages is performed using a large multimodal model to perform the following operations: Correcting erroneous characters identified by OCR based on visual evidence context; Determine whether the specifications and material information in text block t match the candidate row item b based on textual and tabular evidence, and remove mismatched candidate row items b; complete missing units based on rule-based evidence; Output the binding confidence of each candidate triple, and select the binding confidence of the candidate triple with the highest binding confidence as the final binding confidence; The comprehensive confidence of the candidate triple (s,t,b) is calculated based on the detection confidence, OCR recognition confidence, and final binding confidence of the tube stent symbol recognition result s; The formula for calculating the overall confidence level is: ; Where λ1, λ2, and λ3 are preset as fusion weights, To determine the detection confidence level of the tube support symbol recognition result s, To determine the confidence level for OCR recognition, To bind confidence levels.

10. The method for extracting information from a single-line diagram pipe support in a chemical industry as described in claim 9, characterized in that: When the overall confidence level is less than or equal to the first preset threshold, the initial candidate region screening K value is increased, a high-precision OCR model is called to recognize text blocks, or the field consistency retrieval conditions are adjusted to obtain multiple new candidate triples {s,t,b}. Semantic determination is performed on multiple new candidate triples {s,t,b}, and a comprehensive confidence score is calculated and judged.

Images