Deep learning-based picture frame information intelligent identification method and device

Abstract

The invention relates to the technical field of constructional engineering, and discloses an intelligent picture frame information identification method and device based on deep learning, and the method comprises the steps: obtaining data in an engineering drawing as original data; performing picture frame range extraction and filtering processing on the original data to obtain a target picture frame range; a deep learning model is adopted to carry out picture signature area and cell recognition on the target picture frame range, and picture signature areas and corresponding cell information are obtained; and performing picture label information association analysis on the picture label area and the corresponding cell information to obtain final picture label information. The method has the remarkable advantages in the field of construction engineering drawing label information identification: in a drawing frame extraction link, through layer classification, geometric algorithm filtering and multi-feature fusion verification, interference lines and non-standard drawing frame layout influence in a complex drawing can be effectively eliminated, a target drawing frame really containing effective information is screened out, and the identification efficiency of the target drawing frame is improved. And invalid calculation of a non-frame area is avoided.

Description

Technical Field

[0001] This invention relates to the field of architectural engineering, and more specifically to a method and apparatus for intelligent recognition of frame information based on deep learning. Background Technology

[0002] In the current construction engineering field, project complexity and the number of drawings continue to increase. Traditional drawing processing relies on manual visual inspection and manual input, which has certain problems: In existing technologies, simple image recognition algorithms are difficult to handle interference lines and non-standard frame layouts in complex drawings, and lack an understanding of the relationship between text and frame structure. They cannot meet the requirements for high-precision and automated extraction of image label information. Therefore, it is urgent to combine deep learning technology to build an end-to-end intelligent recognition solution. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a method and apparatus for intelligent recognition of frame information based on deep learning, which solves the aforementioned technical problems.

[0004] Firstly, a method for intelligent recognition of frame information based on deep learning is provided, including: Obtain data from the engineering drawings as raw data; The original data is subjected to frame range extraction and filtering to obtain the target frame range; A deep learning model is used to identify the label region and cell information of the target frame range, thereby obtaining the label region and the corresponding cell information. The image signature area and its corresponding cell information are parsed to obtain the final image signature information.

[0005] Furthermore, the original data undergoes frame range extraction and filtering to obtain the target frame range, including: The layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be map frames; The set of line primitives of the suspected frame is subjected to non-closed path elimination processing to obtain a set of candidate frames; The candidate path set of the image frame is filtered and verified to obtain the target image frame range.

[0006] Furthermore, the layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be title frames, including: Based on the layer name attribute and block name attribute of the original data, a classification mapping relationship is constructed; Based on the classification mapping relationship, the line primitives in the original data are divided into different sets; Based on the attribution characteristics of the outer boundary lines of the drawing frame in the engineering drawings, a set of line elements suspected to be drawing frames is selected from the different sets.

[0007] Furthermore, the set of line primitives of the suspected frame is subjected to non-closed path elimination processing to obtain a set of candidate paths for the frame, including: A specific geometric algorithm is used to reconstruct the paths of the line primitives of the suspected frame; Based on the path reconstruction, identify and eliminate non-closed paths; The internal loops in the closed path are cleared to obtain a set of candidate paths for the graph frame.

[0008] Furthermore, a deep learning model is used to identify the label region and cell within the target frame, obtaining the label region and corresponding cell information, including: The target frame area is cropped to obtain a cropped frame image; The first deep learning model is used to locate cells in the cropped image of the frame to obtain cell coordinates and structure type; A second deep learning model is used to predict the title region of the cropped image frame to obtain the title region coordinates. The cell coordinates and the title area coordinates are optimized by feature fusion to obtain the title area and the corresponding cell information.

[0009] Furthermore, a first deep learning model is used to perform cell localization on the cropped image of the image frame to obtain cell coordinates and structure type, including: The cropped image of the frame is normalized; The normalized image is input into a deep learning model built on depthwise separable convolution and feature pyramid network; The deep learning model extracts table line features from the image and fuses multi-scale features to obtain cell coordinates and structure type.

[0010] Furthermore, the image signature area and its corresponding cell information are parsed to obtain the final image signature information, including: The text bounding box corresponding to the cell information is associated with the cell, and the confidence level is evaluated to filter out high-confidence cells. The text within the high-confidence cells is integrated to obtain the integrated text. If no high-confidence cells exist, the multimodal model is called to perform text prediction and obtain the predicted text. The matching degree between the integrated text and the predicted text is evaluated to obtain the final image tag information.

[0011] Furthermore, the association confidence of the text bounding boxes corresponding to the cell information and the cells is evaluated, and cells with high confidence are selected, including: Calculate the text clarity corresponding to the text bounding box, the cell integrity corresponding to the cell, and the positional overlap between the text bounding box and the cell; Based on the text clarity, cell integrity, and positional overlap, a comprehensive confidence level is calculated using preset weights. The overall confidence level is filtered to select cells with high confidence levels.

[0012] Secondly, a deep learning-based intelligent frame information recognition device is provided, based on any one of the deep learning-based intelligent frame information recognition methods described above, including: The acquisition module is configured to acquire data from the engineering drawings as raw data. The processing module is configured to extract and filter the frame range of the raw data to obtain the target frame range; The recognition module is configured to use a deep learning model to recognize the label area and cell information of the target frame range, and obtain the label area and corresponding cell information. The parsing module is configured to perform image information association parsing on the image area and the corresponding cell information to obtain the final image information.

[0013] Thirdly, a terminal is provided, including a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the deep learning-based intelligent recognition method for frame information as described in any of the preceding claims.

[0014] The invention employing the above technical solution has the following advantages: In the drawing frame extraction stage, this invention uses layer classification, geometric algorithm filtering, and multi-feature fusion verification to eliminate the influence of interference lines and non-standard drawing frame layouts in complex drawings, and select the target drawing frames that truly contain effective information. This avoids invalid calculations for non-drawing frame areas, reduces false detections and missed detections in the drawing frame extraction process, and ensures the relevance and effectiveness of subsequent processing objects.

[0015] In the image tag recognition stage, this invention integrates an image recognition algorithm with an optimized YOLOv8 prediction model. This achieves accurate structural positioning of the image tag title bar cells and predicts the entire image tag area, improving the reliability of the text bounding box in key areas. At the same time, by cropping the image tag area, it reduces interference from irrelevant areas during subsequent multimodal model inference, reduces model resource consumption, and improves the smoothness and stability of the recognition process.

[0016] In the information parsing stage, this invention filters high-reliability cells through multi-dimensional confidence assessment, integrates text by combining the principle of semantic coherence, and supplements prediction with the semantic understanding capability of the Qwen_VL multimodal model. This effectively solves the problems of text fragmentation and incomplete extraction of key information in traditional technologies, ensuring the accurate parsing of core label information such as project name and drawing number. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly described below. In all the drawings, the elements or parts are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart of the intelligent recognition method for frame information based on deep learning according to the present invention; Figure 2 This is a flowchart of the intelligent frame information recognition device based on deep learning according to the present invention; Figure 3 This is a schematic diagram of the terminal structure in the intelligent recognition method and device for frame information based on deep learning of the present invention. Detailed Implementation

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

[0020] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, 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 these processes, methods, products, or apparatuses.

[0021] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0022] like Figures 1-3 As shown, the intelligent recognition method for frame information based on deep learning of the present invention includes: Step S01: Obtain data from the engineering drawing as raw data; Step S02: Extract and filter the frame range of the original data to obtain the target frame range; Step S03: Use a deep learning model to identify the label area and cell of the target frame range to obtain the label area and corresponding cell information; Step S04: Perform image tag information association parsing on the image tag area and the corresponding cell information to obtain the final image tag information.

[0023] Specifically, the PyCAD library is used to read DWG format engineering drawings in the field of architectural engineering, and the core data in the drawings is extracted as raw data, including: Layer information: Layer names (such as "frame layer", "annotation layer", "structure layer", "pipeline layer") and the relationship between the graphic elements of each layer; Block information: Block name (such as "Standard drawing frame block", "Custom drawing frame block", "Annotation block") and the set of line elements contained in the block; Line graph metadata: coordinate parameters (vertex coordinates, line segment length), line type attributes (straight line, dashed line) and the association identifier of the layer / block to which each line graph element belongs.

[0024] In this embodiment, the original data is subjected to frame range extraction and filtering to obtain the target frame range, including: The layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be map frames; The set of line elements suspected of being bounding frames is processed by removing non-closed paths to obtain a set of candidate paths for the bounding frames; The candidate path set for the target frame is filtered and verified to obtain the target frame range.

[0025] In this embodiment, the layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be graphic frames, including: Based on the layer name and block name attributes of the original data, construct a classification mapping relationship; Based on the classification mapping relationship, the line graph primitives in the original data are divided into different sets; Based on the attribution characteristics of the outer boundary lines of the drawing frame in the engineering drawings, a set of line elements suspected to be drawing frames is selected from different sets.

[0026] Specifically, the layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be map frames: Constructing a classification mapping relationship: Based on the layer name attribute and block name attribute of the original data, construct a "hash classification mapping table" and establish the correspondence between the key and the line primitive set with the "XOR result of layer name hash value and block name hash value" as the unique key; Line primitive partitioning: Traverse all line primitives in the original data, calculate the corresponding "hash value XOR key" based on the layer name and block name to which each line primitive belongs, and partition the line primitive into the set corresponding to the key; Screening suspected drawing frame line elements: Utilizing the inherent characteristic of engineering drawings that "the outer boundary lines of the drawing frame usually belong to the same layer or block", line elements belonging to the "drawing frame layer" or "standard drawing frame block" are screened from the divided set of line elements. At the same time, interfering line elements such as temporary annotation lines belonging to the "annotation layer" and pipeline line elements belonging to the "pipeline layer" are excluded, and finally a set of line elements suspected of being drawing frames is obtained.

[0027] In this embodiment, the set of line primitives suspected of being frame frames is subjected to non-closed path elimination processing to obtain a set of candidate frames, including: A specific geometric algorithm is used to reconstruct the paths of line primitives in suspected bounding boxes; Identify and eliminate non-closed paths based on path reconstruction; The internal loops in the closed path are cleared to obtain a set of candidate paths for the graph frame.

[0028] Specifically, the set of line primitives suspected of being part of a bounding box is processed to remove non-closed paths, resulting in a set of candidate paths for the bounding box: Path reconstruction: The CGALArrangement geometric algorithm is used to reconstruct the paths of line primitives that are suspected to be bounding boxes. This algorithm traverses the vertices and edges of each line primitive through the built-in "half-edge data structure" to establish the connection relationship between line primitives. Non-closed path elimination: Based on the path reconstruction results, automatically identify "bridge-shaped non-closed paths" (line segments are not connected to other line elements at both ends) and "antenna-shaped discontinuous paths" (line segments are connected to other line elements at only one end and have no extended endpoint at the other end), and eliminate the two types of non-closed paths, retaining only bounded and closed paths; Inner loop clearing: Call the remove_inner_loops() function of the CGAL Arrangement algorithm to clear the inner loops such as nested small rectangular frames in the closed path (such loops are mostly internal decoration frames or invalid separation frames in the title block), and obtain the set of candidate paths for the drawing frame.

[0029] Screen and verify the set of candidate paths for the drawing frame to obtain the target drawing frame range: Dynamic threshold screening: First, extract the resolution (R, in pixels) of the original engineering drawing and the common size features (average aspect ratio S, that is, the ratio of the width to the height of the drawing frame) of the engineering drawing frame; Calculate the dynamic screening threshold according to the formula Tdyn = α×R + β×S (where α is the resolution adaptation coefficient, which is adaptively adjusted according to the drawing clarity, and the value of α is 0.3 - 0.6; β is the size feature coefficient, and the value of β is 0.4 - 0.7, which is dynamically assigned according to the drawing type, and β takes 0.5 - 0.7 for structural drawings); Calculate the area (A) of the candidate path for the drawing frame and the average area (Aavg) of all candidate paths, and retain the paths that satisfy |A / Aavg - 1| < Tdyn, and preliminarily exclude non-drawing-frame paths such as too small annotation frames (the area is much smaller than Aavg) or too large drawing area boundaries (the area is much larger than Aavg).

[0030] Multi-feature weighted fusion verification: Text feature extraction: Identify the text in the path after preliminary screening through the Tesseract lightweight OCR model, count the occurrence frequencies of key texts such as "Project Name", "Check and Approval", "Drawing Number", "Design Unit", etc., and calculate the text matching degree Ftext = number of matching keywords / total number of keywords; Graphic feature extraction: Use the Hough transform to detect the table lines (horizontal and vertical lines) in the path, count the number of detected valid table lines, and compare it with the standard number of table lines in the title block of the drawing title (set in advance according to industry standards), and calculate the table structure integrity Fgraph = number of detected table lines / standard number of table lines; Feature fusion score calculation: Calculate the fusion score according to the formula Sfusion = ω1×Ftext + ω2×Fgraph (where ω1 is the text feature weight and ω2 is the graphic feature weight); Determine the target drawing frame: Retain the paths with Sfusion > 0.7, and finally obtain the target drawing frame range (this range can accurately cover the area containing valid title block information, without missing or misdetecting).

[0031] In this embodiment, a deep learning model is used to identify the title block area and cells in the target drawing frame range, and obtain the title block area and corresponding cell information, including: The target frame area is cropped to obtain a cropped frame image; The first deep learning model is used to locate cells in the frame-cropped image to obtain cell coordinates and structure type; A second deep learning model is used to predict the title region of the cropped image frame to obtain the title region coordinates. Feature fusion optimization is performed on cell coordinates and icon area coordinates to obtain icon area and corresponding cell information.

[0032] In this embodiment, a first deep learning model is used to locate cells in the frame-cropped image to obtain cell coordinates and structure type, including: Normalize the cropped image; The normalized image is input into a deep learning model built on depthwise separable convolution and feature pyramid network; By extracting table line features from images using a deep learning model and fusing multi-scale features, cell coordinates and structure types are obtained.

[0033] Specifically, the target frame area is cropped to obtain a cropped frame image: Based on the coordinate boundaries (x1, y1, x2, y2) of the target drawing frame, the original engineering drawing image is cropped to remove irrelevant areas outside the target drawing frame (such as drawing areas and blank areas), resulting in a cropped image containing only the target drawing frame, thus reducing the computational load for subsequent model inference.

[0034] The first deep learning model is used to locate cells in the image cropped by the image frame, and the cell coordinates and structure type are obtained: Image preprocessing normalizes the cropped image of the drawing frame to a fixed size (e.g., 640×640 pixels) to eliminate the impact of differences in drawing sizes on model recognition. Model Construction and Training: The first deep learning model was built based on MobileNetV3. This model reduces computational complexity through depthwise separable convolutions and introduces a Feature Pyramid Network (FPN) to fuse multi-scale feature layers C3, C4, and C5 (C3 is a shallow feature layer, and C5 is a deep feature layer; after fusion, both cell details and global positional information can be taken into account). 500 bounding box images with labeled cells (labeled using LabelMe, with annotations including cell coordinates and structure type) were used as training data. The AdamW optimizer was selected, with a learning rate of 1e-4, and training was conducted for 50 epochs. Cell location inference: Input the normalized frame cropped image into the trained first deep learning model. The model outputs the coordinate frame (x1, y1, x2, y2) and structure type of each cell (divided into "title cell", such as the cell corresponding to "project name"; "content cell", such as the cell corresponding to the specific content of the project name).

[0035] The second deep learning model is used to predict the title region of the cropped image to obtain the title region coordinates: Training data preparation: Collect 300 images of engineering drawing frames covering different building types and different frame styles, label the “figure label column” category using the labelimg tool, and divide the training set and validation set in an 8:2 ratio; Model optimization and training: A second deep learning model was built based on YOLOv8n, and hyperparameters were optimized: the classification loss weight was reduced to 0.3, and the bounding box regression loss weight was increased to 0.7; the CIoU (Complete Intersectionover Union) loss function was used to optimize the bounding box regression accuracy. The CIoU loss formula is: LCIoU=1−IoU+c 2 / d 2 +αv (where IoU = area of ​​intersection between predicted and ground truth boxes / area of ​​union between predicted and ground truth boxes; d is the Euclidean distance between the center points of the predicted and ground truth boxes; c is the diagonal length of the smallest rectangle enclosing both boxes; α = v / (1−IoU+v); v = π 2 / 4×(arctan (actual box height / actual box width)−arctan (predicted box height / predicted box width)) 2 During training, offline data augmentation (image flipping, rotation, brightness adjustment) and online data augmentation (random cropping, color jittering) were introduced to avoid model overfitting. After 80 rounds of training, the model achieved an IoU of over 0.92 for localization of image tags on the validation set. Image tag region prediction inference: Input the cropped image of the image frame into the trained second deep learning model, and the model outputs the coordinates (x1, y1, x2, y2) of the image tag region, which can accurately cover the entire range of the image tag.

[0036] Feature fusion optimization is performed on cell coordinates and title area coordinates to obtain title area and corresponding cell information: Feature extraction involves extracting cell structure features (with dimensions of C×H×W, such as 256×80×80, where C is the number of feature channels and H and W are the height and width of the feature map) from the output of the first deep learning model, and extracting tag region features (with dimensions consistent with the cell structure features to ensure fusion compatibility) from the output of the second deep learning model. Cross-modal feature fusion: A "cross-modal feature fusion module" is introduced. Cell structure features (Fcell) and tag region features (Ftag) are input into this module, and the fused features are calculated according to the formula Fmerge=σ(W1Fcell+W2Ftag+b) (where W1 and W2 are feature weight matrices, which are set to W1=0.6 and W2=0.4 according to experimental verification; b is the bias term, set to 0.1; σ is the Sigmoid activation function, which maps the fused feature values ​​to the 0~1 interval). Coordinate Correction and Association: Based on the fusion feature, the positional deviation between the cell coordinate frame and the title area is corrected to ensure that each cell falls precisely within the title area, and finally the title area and the corresponding cell information are obtained (including the title area coordinates, cell coordinates and the association relationship of cell structure type).

[0037] In this embodiment, the image signature area and its corresponding cell information are parsed to obtain the final image signature information, including: The confidence level of the association between the text bounding box corresponding to the cell information and the cell is evaluated, and cells with high confidence are selected. The text in high-confidence cells is integrated to obtain the integrated text. If no high-confidence cells exist, the multimodal model is called to perform text prediction and obtain the predicted text. The matching degree between the integrated text and the predicted text is evaluated to obtain the final image tag information.

[0038] In this embodiment, the association confidence of the text bounding box corresponding to the cell information and the cell is evaluated, and cells with high confidence are filtered out, including: Calculate the text clarity corresponding to the text bounding box, the cell integrity corresponding to the cell, and the positional overlap between the text bounding box and the cell; The overall confidence level is calculated based on text clarity, cell integrity, and positional overlap, combined with preset weights. Filter the overall confidence level to select cells with high confidence levels.

[0039] Specifically, the confidence level of the association between the text bounding box corresponding to the cell information and the cell is evaluated, and cells with high confidence are selected: Multi-dimensional index calculation, text clarity (Ctext): The standard deviation of gray level (σgray) within the text bounding box is calculated using image grayscale analysis tools, and is calculated according to the formula Ctext=σgray / 255 (σgray values ​​range from 0 to 255, the closer Ctext is to 1, the clearer the text). Cell integrity (Ccell): Identifies missing areas (such as damaged or obscured areas) in cells using image segmentation tools, calculates the ratio of the missing area (Smissing) to the total area of ​​the cell (Scell), and uses the formula Ccell=1−Smissing / Scell ​​(the closer Ccell is to 1, the more complete the cell). Position overlap (Coverlap): Calculates the ratio of the intersection area to the union area of ​​the text bounding box (Atext) and the cell area (Acell), using the formula Coverlap=Atext∩Acell / Atext∪Acell (the closer Coverlap is to 1, the higher the positional match between the text bounding box and the cell). Overall confidence score calculation: The overall confidence score is calculated according to the formula Conf=ω1Ctext+ω2Ccell+ω3Coverlap (where ω1=0.4, ω2=0.3, and ω3=0.3 are fixed weights to balance the influence of each dimension indicator). High-confidence cell filtering: Based on the confidence threshold, cells with Conf > Confthresh are retained as high-confidence cells, with priority given to cells corresponding to core information such as project name, drawing number, design date, and design unit.

[0040] The text within the high-confidence cells is then integrated to obtain the integrated text: Single-cell text extraction: Using an OCR tool, extract the text content from a single high-confidence cell and remove invalid characters such as spaces and garbled characters from the text. Semantic coherence judgment of adjacent text: The semantic similarity of adjacent cell text is calculated by using a pre-trained BERT model. Specifically, the adjacent text (T1, T2) is input into the BERT model to obtain the corresponding word embedding vectors (emb(T1), emb(T2)). The cosine similarity is calculated according to the formula Sim=cos(emb(T1), emb(T2)) (Sim ranges from -1 to 1, the closer to 1, the more coherent the meaning). Text integration: If the semantic similarity Sim>0.5, the text of adjacent cells will be merged (e.g., “XX Residential Community” and “Structural Construction Drawing” will be merged into “XX Residential Community Structural Construction Drawing”); if Sim≤0.5, the text will be treated as independent text (e.g., “Design Unit” and “2024.05.01” are semantically unrelated and will be retained separately), and the final integrated text will be obtained.

[0041] If there are cells without high confidence, the multimodal model is called to perform text prediction, and the predicted text is obtained: Model invocation and prompt construction: If a certain label area lacks high-confidence cells that satisfy Conf>Confthresh (e.g., damaged cells, blurred text), the Qwen_VL multimodal model is invoked for text prediction; a "context-enhanced prompt" input model is constructed, with the prompt content being "based on the structural features of the label area (cell position: [x1,y1,x2,y2], adjacent text: [T-neighbor]), identify key information (project name, drawing number, design unit, design date) within the area, and output structured results" (where [x1,y1,x2,y2] are the coordinates of the cells within the label area, and [T-neighbor] are the identifiable blurred text around the cell or the text in the adjacent area). Prediction result verification: The predicted text output by Qwen_VL is verified for reasonableness by combining it with the standard structure of the title block (such as "Project Name" usually located in the upper left cell of the title block, "Drawing Number" located below the project name, and "Design Date" located in the lower right corner of the title block). Prediction results that obviously do not conform to the structural logic are excluded (such as predicting "Design Date" as "XX Building"), and the predicted text is obtained.

[0042] The matching degree between the integrated text and the predicted text is evaluated to obtain the final tag information: Matching degree calculation, construct the "text-tag value matching degree" evaluation model, calculate the length of the longest common substring (LCS (T,V)) between the integrated text (T) and the predicted text (V), and calculate the matching degree according to the formula M=LCS(T,V) / max(len(T),len(V)) (where len (⋅) is the string length, and M takes the value range of 0~1. The closer it is to 1, the higher the consistency between the two). Final tag value determination: If the matching degree M>0.7, the predicted text V is used as the final tag value (in which case the predicted text is more complete); if M≤0.7, the integrated text T is used as the final tag value (in which case the integrated text is more accurate). Structured output: The final drawing title values ​​are organized according to standard fields such as "Project Name", "Drawing Number", "Design Unit" and "Design Date" to form structured final drawing title information (such as "Project Name: XX Residential Community Structural Construction Drawing (Basement 1); Drawing Number: JS-001; Design Unit: XX Architectural Design Institute; Design Date: 2024.05.01"), completing the intelligent recognition of drawing frame information.

[0043] In other embodiments, a deep learning-based intelligent frame information recognition device is provided, and a deep learning-based intelligent frame information recognition method based on any of the preceding embodiments is provided, including: The acquisition module is configured to acquire data from the engineering drawings as raw data. The processing module is configured to extract and filter the frame range of the raw data to obtain the target frame range. The recognition module is configured to use a deep learning model to identify the label area and cell of the target frame, and obtain the label area and corresponding cell information. The parsing module is configured to perform image tag information association parsing on the image tag area and the corresponding cell information to obtain the final image tag information.

[0044] In other embodiments, a terminal is provided, including a processor, an input device, an output device, and a memory, which are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute a deep learning-based intelligent recognition method for frame information as described above.

[0045] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for intelligent recognition of frame information based on deep learning, characterized in that, include: Obtain data from the engineering drawings as raw data; The original data is subjected to frame range extraction and filtering to obtain the target frame range; A deep learning model is used to identify the label region and cell information of the target frame range, thereby obtaining the label region and the corresponding cell information. The image signature area and its corresponding cell information are parsed to obtain the final image signature information.

2. The method for intelligent recognition of frame information based on deep learning according to claim 1, characterized in that, The original data is subjected to frame range extraction and filtering to obtain the target frame range, including: The layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be map frames; The set of line primitives of the suspected frame is subjected to non-closed path elimination processing to obtain a set of candidate frames; The candidate path set of the image frame is filtered and verified to obtain the target image frame range.

3. The method for intelligent recognition of frame information based on deep learning according to claim 2, characterized in that, The layers and elements in the original data are classified and filtered to obtain a set of line elements that are suspected to be map frames, including: Based on the layer name attribute and block name attribute of the original data, a classification mapping relationship is constructed; Based on the classification mapping relationship, the line primitives in the original data are divided into different sets; Based on the attribution characteristics of the outer boundary lines of the drawing frame in the engineering drawings, a set of line elements suspected to be drawing frames is selected from the different sets.

4. The method for intelligent recognition of frame information based on deep learning according to claim 2, characterized in that, The set of line primitives of the suspected frame is subjected to non-closed path elimination processing to obtain a set of candidate frames, including: A specific geometric algorithm is used to reconstruct the paths of the line primitives of the suspected frame; Based on the path reconstruction, identify and eliminate non-closed paths; The internal loops in the closed path are cleared to obtain a set of candidate paths for the graph frame.

5. The method for intelligent recognition of frame information based on deep learning according to claim 1, characterized in that, A deep learning model is used to identify the label region and cell information of the target frame, resulting in the label region and its corresponding cell information, including: The target frame area is cropped to obtain a cropped frame image; The first deep learning model is used to locate cells in the cropped image of the frame to obtain cell coordinates and structure type; A second deep learning model is used to predict the title region of the cropped image frame to obtain the title region coordinates. The cell coordinates and the title area coordinates are optimized by feature fusion to obtain the title area and the corresponding cell information.

6. The method for intelligent recognition of frame information based on deep learning according to claim 5, characterized in that, The first deep learning model is used to locate cells in the cropped image of the image frame, obtaining cell coordinates and structure type, including: The cropped image of the frame is normalized; The normalized image is input into a deep learning model built on depthwise separable convolution and feature pyramid network; The deep learning model extracts table line features from the image and fuses multi-scale features to obtain cell coordinates and structure type.

7. The method for intelligent recognition of frame information based on deep learning according to claim 1, characterized in that, The image signature area and its corresponding cell information are parsed to obtain the final image signature information, including: The text bounding box corresponding to the cell information is associated with the cell, and the confidence level is evaluated to filter out high-confidence cells. The text within the high-confidence cells is integrated to obtain the integrated text. If no high-confidence cells exist, the multimodal model is called to perform text prediction and obtain the predicted text. The matching degree between the integrated text and the predicted text is evaluated to obtain the final image tag information.

8. The method for intelligent recognition of frame information based on deep learning according to claim 7, characterized in that, The text bounding boxes corresponding to the cell information are associated with the cells to perform a confidence assessment, and high-confidence cells are selected, including: Calculate the text clarity corresponding to the text bounding box, the cell integrity corresponding to the cell, and the positional overlap between the text bounding box and the cell; Based on the text clarity, cell integrity, and positional overlap, a comprehensive confidence level is calculated using preset weights. The overall confidence level is filtered to select cells with high confidence levels.

9. A deep learning-based intelligent recognition device for frame information, characterized in that, The method for intelligent recognition of frame information based on deep learning according to any one of claims 1 to 8 includes: The acquisition module is configured to acquire data from the engineering drawings as raw data. The processing module is configured to extract and filter the frame range of the raw data to obtain the target frame range; The recognition module is configured to use a deep learning model to recognize the label area and cell information of the target frame range, and obtain the label area and corresponding cell information. The parsing module is configured to perform image information association parsing on the image area and the corresponding cell information to obtain the final image information.

10. A terminal, characterized in that, The device includes a processor, an input device, an output device, and a memory, which are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the deep learning-based intelligent recognition method for frame information as described in any one of claims 1 to 8.

Images