Intelligent building plan analysis system based on deep learning

Abstract

This invention discloses a deep learning-based intelligent architectural floor plan parsing system, relating to drawing recognition technology. This deep learning-based intelligent architectural floor plan parsing system includes: a data acquisition and preprocessing module for acquiring, converting, and enhancing architectural floor plans from multiple sources; a multi-scale feature extraction module for parallel multi-scale extraction of architectural segmented image sets; a multi-scale feature fusion module for performing scale anomaly compensation processing and feature fusion path selection; an object detection and classification module for detecting reliability through fused features, probability matrices, and cross-scale detail discrimination; a spatial relationship reasoning module for constructing candidate topologies based on vector elements and constructing topology consistency discrimination through relationship reasoning; and a result post-processing and optimization module for achieving cross-segment alignment and topology cleaning. This system solves the problems of unstable extraction and alignment of multi-scale features in architectural floor plans, leading to missed detection of small components and unreliable spatial relationships.

Description

Technical Field

[0001] This invention relates to the field of drawing recognition technology, specifically to a deep learning-based intelligent analysis system for architectural floor plans. Background Technology

[0002] Architectural floor plans are the core expressive medium in architectural design, construction, and operation and maintenance management, comprehensively describing room zoning, load-bearing and non-load-bearing walls, door and window openings, staircases, electrical and plumbing layouts, and annotation information. With the widespread use of architectural design tools and scanning and acquisition equipment, early architectural floor plan parsing solutions mostly adopted rule-based geometric recognition and template matching methods. These methods used line segment extraction, character recognition, symbol font matching, and topological rule libraries to identify wall lines, door and window outlines, room contours, and common components.

[0003] For example, invention patent CN113469086A discloses a method, apparatus, device, and medium for region division in architectural floor plans, including: extracting multi-channel basic image features at a set low-dimensional scale from the architectural floor plan to be identified; performing convolution processing on these features using convolution kernels with different numbers of channels to obtain multi-channel high-dimensional image features at multiple scales; fusing the multi-channel high-dimensional image features to obtain multi-channel first fused image features, and using a dynamic convolution algorithm to enhance boundary features; fusing the multi-channel first fused image features to obtain multi-channel second fused image features while ensuring that the first fused image features are not lost; and obtaining the region division result of the architectural floor plan based on the multi-channel second fused image features. In the above technical solution, the features extracted from the architectural floor plan are processed based on artificial intelligence algorithms to obtain the region division result, achieving accurate region division of the architectural floor plan and improving production efficiency.

[0004] For example, invention patent CN112926392A discloses a method for room recognition in architectural floor plans based on contour filtering. It uses image correlation algorithms to identify all contours on the plan, then filters these contours using filtering rules to eliminate interfering contours that do not meet the requirements, thus selecting the contours most likely to be rooms, which are represented by bounding rectangles. Finally, a non-maximum suppression algorithm is used to further filter the rectangles to obtain the rectangles most likely to be rooms. This method has a certain generalization ability for room recognition in architectural floor plans, meaning it can identify rooms on different types of architectural floor plans, providing a new approach to room recognition in architectural floor plans.

[0005] However, existing research on deep learning for architectural floor plan analysis still has several limitations. On the one hand, the scale of walls, room outlines, and elevator lobbies in architectural floor plans is much larger than that of electrical switches, small windows, and small-sized symbols. Convolutional downsampling and fixed receptive field design can easily cause small target textures to be submerged, and multi-scale structures are mixed in the feature space, resulting in insufficient positioning accuracy of fine-grained components and prominent category confusion. On the other hand, different projects use significantly different drawing scales, line widths, and symbol libraries, leading to large differences in drawing noise levels, scanning clarity, and contrast. Unified preprocessing and feature extraction cannot simultaneously take into account thick wall and thick line areas as well as thin line and symbol areas, easily resulting in recognition biases caused by scale inconsistencies.

[0006] Therefore, in order to address the above problems, there is an urgent need for an intelligent architectural floor plan parsing system based on deep learning. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent analysis system for building floor plans based on deep learning, which solves the problems of unstable extraction and alignment of multi-scale features in building floor plans, leading to missed detection of small components and unreliable spatial relationships. Technical solution

[0008] To achieve the above objectives, this invention provides the following technical solution: a deep learning-based intelligent architectural floor plan parsing system, comprising: a data acquisition and preprocessing module for acquiring architectural floor plans from multiple sources, converting and enhancing them, outputting a set of usable architectural slice images, storing them, and then constructing a drawing parsing link database; a multi-scale feature extraction module for performing multi-scale parallel extraction on the architectural slice image set and constructing a multi-scale feature extraction model, outputting a fine-grained feature matrix and a semantic feature matrix; and a multi-scale feature fusion module for constructing a cross-scale correlation consistency matrix and establishing a cross-scale alignment criterion based on correlation energy and difference residuals, and outputting the execution scale according to the alignment criterion. The system includes: anomaly compensation processing and feature fusion path selection; an object detection and classification module, which determines the reliability of detection by fusing features, category probability matrices, and cross-scale detail-preserving data, and performs small object enhancement sampling, soft / hard label switching, and vector element output control based on the reliability determination results; a spatial relationship reasoning module, which constructs candidate topologies based on vector elements and constructs topology consistency determination through relationship reasoning, and performs spatial connection relationship verification and correction and topology stability structure solidification based on the consistency determination results; and a result post-processing and optimization module, which completes cross-segment alignment, geometric repair, and topology cleanup based on vector elements and spatial relationship structures, and generates an archived result set.

[0009] Further, the specific steps for acquiring, converting, and enhancing multi-source architectural floor plans, outputting a usable set of architectural slice images, and storing them to construct a drawing parsing link database are as follows: First, raw architectural floor plan data is collected from scanned documents, files exported from drawing software, and screenshots. Then, multiple formats such as PDF, JPG, PNG, and TIFF are converted to a preset target resolution and grayscale normalization range through a format parsing and rendering process. First, linear stretching and histogram equalization are performed within the grayscale domain to complete grayscale processing, ensuring a uniform brightness distribution across different source drawings. Finally, noise suppression and... Contrast enhancement suppresses residual background textures and grid lines, highlighting wall lines, door and window outlines, and symbolic line segments. For ultra-large format drawings, segmentation and overlapping window generation are performed to form a set of architectural segment images that can be used for feature extraction. At the same time, the positional mapping relationship of each segment in the original image is recorded to support result write-back. The main structural regions are extracted based on connected component analysis, generating structural region indexes and candidate region masks to limit the processing range and reduce invalid background interference. After processing, the segment images, segment position mapping tables, structural region indexes, and candidate region masks are output, and the segment indexes, mapping relationships, and processing logs are stored in the drawing parsing link database.

[0010] Furthermore, the specific steps for performing multi-scale parallel extraction on the architectural patch image set and constructing a multi-scale feature extraction model to output fine-grained feature matrices and semantic feature matrices are as follows: Input the architectural patch image set; use an image segmentation algorithm to divide the original image into several windows and local regions; employ a pyramid hierarchy to simultaneously extract features at multiple scale layers, allowing large-scale structural information such as walls and room layouts to be simultaneously incorporated into feature representation along with small-scale component information such as electrical switches, legend symbols, and small windows and doors; introduce dilated convolutions during feature extraction to expand the receptive field, enabling the network to obtain more complete contextual information without reducing spatial details, avoiding the problem of texture smoothing and edge swallowing of small components due to multiple downsampling; further introduce multi-channel convolution kernels at each scale layer, allowing different channels to respond differently... Structural elements of the same size but different orientations are used to enhance the differential expression of long straight wall lines, diagonal wall lines, door and window opening curves, and symbolic line segment combinations. An independent multi-scale extraction process is adopted for window clipping images, separating feature extraction and encoding processes for different regions to reduce expression bias caused by small-scale components being diluted by large background areas and large structures being fragmented by local clipping. Random scaling, rotation, and flipping enhancements are introduced during the training phase to enable the network to learn stable feature patterns of the same structure at different scales, orientations, and layout densities, improving adaptability across different drawing scales. Based on the above processing, a multi-scale feature extraction model is constructed, outputting fine-grained feature maps and semantic feature maps. Linear transformations and scale alignment are performed on the channel and spatial dimensions using attention and sampling operators to map the semantic feature maps to the fine-grained feature maps. Figure 1The scale and number of channels are determined, and then the semantic feature matrix is ​​obtained by expanding it according to spatial location.

[0011] Furthermore, the specific steps for constructing a cross-scale correlation consistency matrix and establishing a cross-scale alignment criterion based on correlation energy and difference residuals are as follows: Obtain the fine-grained feature matrix and semantic feature matrix; transpose the fine-grained feature matrix to obtain the transposed fine-grained matrix; perform matrix multiplication between the transposed fine-grained matrix and the semantic feature matrix to obtain the cross-scale correlation matrix; transpose the cross-scale correlation matrix to obtain the transposed cross-scale correlation matrix; perform matrix multiplication between the transposed cross-scale correlation matrix and the cross-scale correlation matrix to obtain the correlation cumulative matrix; construct the identity matrix; and perform matrix multiplication between the identity matrix and the correlation cumulative matrix... The matrix is ​​summed to obtain the cross-scale correlation consistency matrix; the determinant of the cross-scale correlation consistency matrix is ​​calculated and its absolute value is obtained to obtain the determinant term; the trace of the cross-scale correlation consistency matrix is ​​calculated to obtain the trace term, which is added to the numerical stability term to obtain the trace stability term; the difference operation is performed on the fine-grained feature matrix and the semantic feature matrix to obtain the cross-scale difference matrix; the Frobenius norm of the cross-scale difference matrix is ​​calculated to obtain the difference norm term, which is added to the numerical stability term to obtain the difference stability term; the trace stability term and the difference stability term are multiplied to obtain the alignment denominator term; the determinant term is divided by the alignment denominator term to obtain the cross-scale alignment value.

[0012] Furthermore, the specific steps for performing scale anomaly compensation processing and feature fusion path selection based on the alignment criterion output are as follows: By comparing the cross-scale alignment value and the alignment threshold in real time, when the cross-scale alignment value is less than the alignment threshold, a scale anomaly mask is generated in the fine-grained feature map. Local sub-blocks are extracted from the original fine-grained feature matrix according to the mask. After local upsampling, the area covered by the scale anomaly mask is cropped into fixed-size sub-blocks and fed into line enhancement convolution and symbol texture compensation convolution. The line enhancement convolution uses a 3x3 superimposed 5x5 learnable convolution kernel. The input is a feature block obtained by concatenating the fine-grained feature matrix and the edge guidance map by channel. The number of output channels is consistent with the fine-grained feature matrix. The symbol texture compensation convolution uses a 3x3 learnable convolution kernel and modulates the convolution response with a weight map generated by the inflection point guidance map. The parameters of the two types of convolutions are updated synchronously with the detection loss and the cross-scale alignment criterion during the joint training phase of the entire network, thereby reconstructing a set of fine-grained compensation features. The attention operator is used to improve the fine-grained compensation features in the fusion channels. The proportion of the semantic feature matrix is ​​reduced locally, and the scale anomaly mask and the enhanced features are written into the drawing parsing link database. Local secondary feature extraction is performed. In the secondary feature extraction stage, connectivity is determined based on the endpoint direction angle threshold and the endpoint spacing threshold, and endpoint connection is performed to form a line segment completion mask and write back the corresponding position of the fused feature. In the symbol region, missing short line segments and arc responses are filled in according to the consistency of line segment combination. After the write-back is completed, the cross-scale alignment value is recalculated. If the cross-scale alignment value is still less than the alignment threshold, it is marked as a cross-scale difficult area, further feature repair is stopped, the current fine-grained compensation feature is frozen as the reference upper limit, and the region index and local image slices are written into the difficult sample set. When the cross-scale alignment value is greater than or equal to the alignment threshold, only the standard pyramid fusion path is used, no additional enhancement branch is added, the corresponding position is marked as the scale alignment stable area, only one channel attention allocation is performed, and the fine-grained feature matrix and the semantic feature matrix are merged and the fused feature is directly output.

[0013] Furthermore, the specific steps for determining the reliability of the detection by fusing features, class probability matrices, and cross-scale detail-preserving data are as follows: Obtain the fusing features and cross-scale alignment values; input the fusing features into the candidate box generator to obtain a candidate box set; then perform region alignment sampling on each candidate box using the fusing features to obtain a fixed-size feature block; input the fixed-size feature block into the detection classification head to obtain a class confidence vector; stack the class confidence vectors of all candidate instances in instance order to obtain a class probability matrix; search for the maximum class probability of each candidate instance row by row in the class probability matrix, retaining only the maximum probability at the corresponding position and decrementing the values ​​of the other positions in the same row. Set the value to zero to obtain the principal class response matrix with the same shape as the class probability matrix; calculate the Frobenius norm of the class probability matrix to obtain the global energy term; subtract the principal class response matrix from the class probability matrix to obtain the non-principal class response matrix; calculate the Frobenius norm of the non-principal class response matrix to obtain the confusion energy term; add the confusion energy term to the numerical stability term to obtain the confusion stability term; divide the global energy term by the confusion stability term to obtain the class set ratio term; add the cross-scale alignment value to the numerical stability term to obtain the cross-scale stability term; take the reciprocal of the cross-scale stability term to obtain the cross-scale suppression term; multiply the class set ratio term by the cross-scale suppression term to obtain the confidence constraint value.

[0014] Further, the specific steps for performing small object enhancement sampling, switching between soft and hard labels, and vector feature output control based on the credible discrimination results are as follows: Real-time comparison of the detection credible constraint value and credible threshold; when the detection credible constraint value is less than the credible threshold, the candidate instance is added to the small object enhancement queue, retaining the dual outputs of the main classification branch and the symbol auxiliary branch as multi-label records; a consistency verification field is added to the candidate instances entering the small object enhancement queue; multiple sets of category confidence are obtained through repeated reasoning using zoomed and rotated views, and the confidence dispersion is recorded; candidate instances with excessive confidence dispersion are marked as targets to be verified, and an output delay is set; simultaneously, bounding box refinement regression is performed on the targets to be verified, and the coordinate difference before and after correction is written into the displacement field; if the recalculated detection credible constraint value is still less than the credible threshold, downgrade labeling is performed, and the record is written into the drawing parsing link database; when the detection credible constraint value is greater than or equal to the credible threshold, hard labels are used in the label generation stage, directly converting the detection instance into vector feature output. The vector features include vertex coordinates, orientation angles, category codes, and confidence fields, which are directly used by the spatial relationship reasoning module.

[0015] Furthermore, the specific steps for constructing candidate topologies based on vector features and constructing topology consistency judgment through relational reasoning are as follows: obtaining detection confidence constraint values ​​and vector features; constructing a node list based on vector features, performing candidate relation screening to generate a candidate edge set according to the rules of geometric proximity, direction consistency, overlap relationship and endpoint adsorption connectivity, and writing the candidate edge set into the adjacency matrix; constructing a relational reasoning network within the candidate edge range defined by the adjacency matrix, the relational reasoning network receiving node features and edge features and outputting the relational confidence and relational type of each candidate edge, thereby forming a structured relational record between elements and outputting a relational confidence matrix; The adjacency matrix and the relation confidence matrix are multiplied element-wise to obtain the relation coupling matrix; the Frobenius norm of the relation coupling matrix is ​​calculated to obtain the effective relation strength energy term; the Frobenius norm of the adjacency matrix is ​​calculated to obtain the adjacency structure energy term; the adjacency structure energy term is added to the numerical stability term to obtain the adjacency stability term; the effective relation strength energy term is divided by the adjacency stability term to obtain the relation strength ratio term; the detection confidence constraint value is added to the numerical stability term to obtain the confidence stability term; the detection confidence constraint value is divided by the confidence stability term to obtain the confidence constraint ratio term; the relation strength ratio term is multiplied by the confidence constraint ratio term to obtain the topology consistency verification value.

[0016] Furthermore, the specific steps for performing spatial connection relationship verification and correction and topological stability structure solidification based on the consistency judgment results are as follows: Real-time comparison of the topological consistency check value and the consistency threshold; when the topological consistency check value is less than the consistency threshold, the relationship verification and correction process is initiated; abnormal relationship edges related to door openings, window openings, wall segments, and room units are written into a conflict list and a conflict subgraph is constructed; candidate connection relationships are generated only within the conflict subgraph and relationship reasoning is re-executed; and geometric rules are introduced to verify and correct and eliminate abnormal combinations of spatial relationships. These geometric rules include that door openings must connect to both spatial boundaries, window openings must be attached to the outer wall boundary, and wall segments must be shaped... The room enclosure and closed chain, and the direction of the door opening must be consistent with the direction of the wall segment; the confidence of the relationship that does not meet the rules is multiplied by the attenuation coefficient to reduce the weight. After the weight reduction, if the topology consistency check value is still less than the consistency threshold, the relationship edge is deleted and the neighborhood is reconnected to complete it. The corrected spatial relationship structure, along with the conflict list and correction record, is written into the drawing parsing link database; when the topology consistency check value is greater than or equal to the consistency threshold, the corresponding relationship is included in the topology stable set, the spatial relationship structure is solidified and written into the drawing parsing link database. The spatial relationship structure includes feature identifier, feature geometry, relationship type, relationship confidence, connection endpoints and enclosure information.

[0017] Furthermore, the specific steps for completing cross-segment alignment, geometric repair, and topological cleanup based on vector features and spatial relationship structures, and generating an archived result set, are as follows: Input vector features and spatial relationship structures; first, perform deduplication, merging, and coordinate alignment in overlapping areas based on the segment location mapping table to form unique vector features and write them into the source traceability field; repair broken segments of wall lines and door / window opening boundaries based on line segment completion masks to ensure closed enclosure boundaries; replace feature versions in areas based on scale anomaly masks and write cross-scale difficult area markers; merge multi-label records of main classification branches and symbol auxiliary branches for targets to be reviewed and retain confidence dispersion, and write downgraded labels for downgraded annotation records; clean up abnormal relationship edges based on the conflict list and perform endpoint alignment and polyline smoothing on room enclosure chains, entrance / exit connection chains, and symbol attachment relationships to output a topologically stable set; finally, generate a result summary table and vector result files conforming to preset vector format specifications according to room connectivity, opening distribution, electrical and fire protection layout, and archive them in the drawing parsing link database.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) This invention performs format parsing, rendering conversion, grayscale processing, noise suppression, and contrast enhancement on scanned documents, files exported from drawing software, and screenshot resources. It also combines segmented cropping, overlapping window generation, and segmented position mapping table recording. Simultaneously, it writes the segmented index, mapping relationship, and processing log into the drawing parsing link database, thereby achieving the effect of consistent input drawings and traceable processing links. This effectively solves the problems of inconsistent scales of multiple source drawings, strong noise interference, and difficulty in stable processing of ultra-large format drawings in the prior art.

[0019] (2) In this invention, features are extracted synchronously at multiple scales through a pyramid hierarchy, and dilated convolution is introduced during the extraction process to expand the receptive field. At the same time, multi-channel convolution kernels respond to structural elements of different sizes and directions respectively, and an independent multi-scale extraction process is performed for the window clipping image. This achieves the effect of synchronously preserving large-scale structural information and small-scale component information, effectively solving the problems in the prior art where multiple downsampling leads to the smoothing of fine component textures, the swallowing of edges, and the offset of feature expression under the difference in drawing scale.

[0020] (3) In this invention, a candidate box set is generated by fusing features and a category probability matrix is ​​formed. Then, the main class response matrix, category confusion energy and cross-scale detail fidelity data are combined to determine the credibility of the detection. Based on the credible discrimination results, small object enhanced sampling, soft and hard label switching and vector element output control are performed. This achieves the effect of credible layering of detection results and controllable output granularity, effectively solving the problems of frequent missed detection of small targets, difficulty in constraining category confusion, and direct output of low-credibility results causing downstream inference errors in the prior art.

[0021] (4) This invention constructs candidate topologies based on vector elements and generates a relation confidence matrix by performing relation reasoning under adjacency constraints. Then, it combines the detection of confidence constraints to participate in topology consistency judgment, constructs conflict subgraphs, regenerates candidate connection relations, introduces geometric rules for verification, and applies attenuation and neighborhood reconnection to the relation confidence that does not meet the rules. This achieves the effect of verifiable spatial connection relations and solidifiable topology structure, effectively solving the problems of abnormal connections, broken enclosing chains, incorrect door and window attachments and unstable relation results caused by the lack of constraints in relation reasoning in the prior art.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] Figure 1 This is a structural diagram of the intelligent architectural floor plan parsing system based on deep learning, as described in this invention. Figure 2 This is a cross-scale correlation consistency spectrum distribution diagram of the architectural floor plan of this invention; Figure 3 This is a flowchart of the spatial relationship topology consistency verification and correction process of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figures 1-3This invention provides a technical solution: a deep learning-based intelligent architectural floor plan parsing system, comprising a data acquisition and preprocessing module for acquiring architectural floor plans from multiple sources, converting and enhancing them, outputting a set of usable architectural slice images, storing them, and then constructing a drawing parsing link database; a multi-scale feature extraction module for performing multi-scale parallel extraction on the architectural slice image set and constructing a multi-scale feature extraction model, outputting a fine-grained feature matrix and a semantic feature matrix; and a multi-scale feature fusion module for constructing a cross-scale correlation consistency matrix and establishing a cross-scale alignment criterion based on correlation energy and difference residuals, and outputting scale anomaly compensation according to the alignment criterion. The system includes: a processing and feature fusion path selection module; an object detection and classification module, which determines the reliability of detection by fusing features, category probability matrices, and cross-scale detail-preserving data, and performs small object enhancement sampling, soft / hard label switching, and vector feature output control operations based on the reliability determination results; a spatial relationship reasoning module, which constructs candidate topologies based on vector features and constructs topology consistency determination through relationship reasoning, and performs spatial connection relationship verification and correction and topology stability structure solidification operations based on the consistency determination results; and a result post-processing and optimization module, which completes cross-segment alignment, geometric repair, and topology cleanup based on vector features and spatial relationship structures, and generates an archived result set.

[0026] Specifically, the steps for acquiring, converting, and enhancing multi-source architectural floor plans, outputting a usable set of architectural slice images, and storing them to construct a drawing parsing link database are as follows: First, raw architectural floor plan data is collected, receiving raw architectural floor plan data from scanned documents, exported files from drawing software, and screenshots. During the receiving phase, page indexes, scale labels, and layer identifiers are simultaneously read to form source records. Then, multiple formats such as PDF, JPG, PNG, and TIFF are converted to preset target resolutions and grayscale normalization ranges through format parsing and rendering processes. Vector-based drawings are rendered and rasterized, and scanned documents undergo tilt and perspective correction. First, linear stretching and histogram equalization are performed within the grayscale domain to complete grayscale processing, ensuring that drawings from different sources have a uniform brightness distribution. Finally, noise suppression and contrast adjustments are performed on the normalized grayscale image. Enhance and suppress residual grid lines, stamp shadows, and crease textures to highlight wall lines, door and window outlines, and symbol lines. For ultra-large format drawings, perform segmented cropping and overlapping window generation. Segmented cropping uses a fixed step size and overlapping boundary preservation strategy to generate a set of architectural segmented images. At the same time, record the position mapping relationship of each segment in the original drawing to support result write-back. The position mapping relationship includes page number, top left corner offset of segment, segment size, and scaling ratio. Extract the main structural regions based on connected component analysis, and generate structural region indexes and candidate region masks according to wall line connected components, door and window opening connected components, and symbol connected components to limit the processing range and reduce invalid background interference. After processing, output segmented images, segmented position mapping tables, structural region indexes, and candidate region masks, and store the segmented indexes, mapping relationships, and processing logs to construct a drawing parsing link database.

[0027] In this implementation scheme, the original data of multi-source architectural floor plans are parsed and rendered, and the resolution is unified and grayscale normalization range is constrained. Linear stretching, histogram equalization, noise suppression and contrast enhancement are combined to strengthen the distinguishability of wall lines, door and window outlines and symbol line segments. In addition, the plotting and overlapping window generation of ultra-large format drawings are used to record the plot position mapping relationship. At the same time, connected component analysis is used to obtain structural region index and candidate region mask. Thus, the plotted images, plot position mapping table, structural region index and candidate region mask that can be directly used for subsequent feature extraction are output to form a drawing parsing link database. This makes subsequent processing have a consistent input scale, rewriteable positioning basis and effective background suppression foundation.

[0028] Specifically, the steps for multi-scale parallel extraction of architectural segmented image sets and construction of a multi-scale feature extraction model to output fine-grained feature matrices and semantic feature matrices are as follows: Input the architectural segmented image set; based on the segmented location mapping table and structural region index in the drawing parsing link database, use an image segmentation algorithm to divide the original image into several windows and local regions; prioritize the allocation of subsequent computational resources to windows within the candidate region mask coverage area; employ a pyramid-level approach to simultaneously extract features at multiple scale layers, allowing large-scale structural information such as walls and room layouts to be simultaneously incorporated into feature representation along with small-scale component information such as electrical switches, legend symbols, and small windows and doors; retain fine lines and corner responses in the shallow feature map, and precipitate the overall shape of the room and the distribution of functional blocks in the deep feature map; introduce dilated convolution during feature extraction to expand the receptive field, enabling the network to obtain more complete contextual information while preserving spatial details, avoiding the problem of texture smoothing and edge swallowing of small components due to multiple downsampling; and control the perception of the continuity of long-distance wall lines through dilation ratio combinations; further refine the feature extraction process at each scale layer. This paper introduces multi-channel convolutional kernels, enabling different channels to respond to structural elements of different sizes and orientations. This strengthens the differential representation of long straight wall lines, diagonal wall lines, door and window opening arcs, and symbolic line segment combinations. Normalization is performed in the channel dimension to suppress strong background responses. For window clipping images, an independent multi-scale extraction process is adopted, separating feature extraction and encoding processes for different regions. This reduces expression bias caused by small-scale components being diluted by large-area backgrounds and large structures being fragmented by local clipping. The discontinuity caused by clipping breaks is mitigated by retaining overlapping buffers at window boundaries. During training, random scaling, rotation, and flipping enhancements are introduced, allowing the network to learn stable feature patterns of the same structure at different scales, orientations, and layout densities. Parameters are continuously updated using historical fragment samples from the drawing parsing database, enhancing adaptability across drawing scale differences. Based on the above processing, a multi-scale feature extraction model is constructed, outputting fine-grained feature maps and semantic feature maps. Attention and sampling operators are used to perform linear transformations and scale alignment in the channel and spatial dimensions, mapping the semantic feature map to the fine-grained feature map. Figure 1 By determining the scale and number of channels, the fine-grained feature map is flattened along the spatial dimension and its shape aligned to obtain the fine-grained feature matrix. Then, it is expanded according to the spatial position to obtain the semantic feature matrix, providing a unified feature input basis for subsequent cross-scale correlation analysis and reliable detection discrimination.

[0029] In this implementation scheme, fine-grained feature maps and semantic feature maps of architectural patch images are extracted and aligned in a multi-scale space. Components of different scales, such as walls, room outlines, electrical switches, legend symbols, small windows and doors, are uniformly encoded into fine-grained feature matrices and semantic feature matrices. While preserving line details and overall layout shape, this reduces expression deviations caused by clipping boundaries and background interference. This provides a structurally complete, scale-uniform, and semantically clear feature input foundation for subsequent cross-scale correlation analysis, scale anomaly compensation, and reliable detection discrimination.

[0030] Specifically, the steps for constructing a cross-scale correlation consistency matrix and establishing a cross-scale alignment criterion based on correlation energy and difference residuals are as follows: obtaining a fine-grained feature matrix. With semantic feature matrix Then, for the fine-grained feature matrix Perform a transpose operation to obtain the transpose fine-grained matrix. Transpose the fine-grained matrix With semantic feature matrix Perform matrix multiplication to obtain the cross-scale correlation matrix. For cross-scale correlation matrices Transpose to get Then and Perform matrix multiplication to obtain the relevant cumulative matrix. Constructing and related cumulative matrices. identity matrices of the same dimensions ,Will and Adding them together yields the cross-scale correlation consistency matrix. For cross-scale correlation consistency matrices Calculate the determinant and take its absolute value to get the terms of the determinant. ;right Calculate the trace to obtain the trace term. The trace term is then added to the numerical stability term ε to form the trace-stable term. Difference operations are performed on the fine-grained feature matrix and the semantic feature matrix to generate a cross-scale difference matrix. The Frobenius norm is calculated on the cross-scale difference matrix to obtain the difference norm term, which is then added to the numerical stability term ε to obtain the difference-stable term. The trace-stable term is multiplied by the difference-stable term to obtain the alignment denominator term. Finally, the determinant term is divided by the alignment denominator term to obtain the final cross-scale alignment value. In the construction method of the correlation matrix, if it is necessary to ensure the correlation cumulative matrix... For a square matrix, it must satisfy the following conditions: If there is a discrepancy, it can be corrected. Add a linear transformation to make its channel number equal to Matching ensures the rationality and stability of subsequent determinant and trace calculations. The numerical stability term ε is a preset positive number, typically used to avoid zero denominators and improve the stability of numerical operations. Theoretically, the range of cross-scale alignment values ​​is a non-negative real number, and the alignment threshold is set empirically near the mean to distinguish the degree of feature alignment.

[0031] The cross-scale correlation consistency matrix is ​​calculated as follows: ; The specific calculation method for cross-scale alignment values ​​is as follows: ; In the formula, This represents the cross-scale correlation consistency matrix, used to characterize the coupling strength of cross-scale features in the correlation space; Represents the identity matrix, used to ensure matrix invertibility and suppress degenerate related structures; This represents a fine-grained feature matrix, indicating a continuous feature representation of the outline and local texture information of small components; The semantic feature matrix represents the projection representation of the semantics of structures such as rooms and walls at a fine scale. The value represents the cross-scale alignment, indicating the degree of alignment between the fine-grained representation and the semantic projection representation under the constraints of relevant structure and differential energy. Determines operations are used for measurement. The overall structural resolvability and coupling stability; The trace operation represents the measurement. The overall energy scale and suppressing local dimensional aberrations; express and The Frobenius norm difference term is used to measure the alignment residuals of features across scales in the spatial unfolding domain; This indicates a numerically stable term, set to less than Use positive numbers to avoid the denominator being zero and to ensure computational stability.

[0032] Table 1 shows the evaluation data for cross-scale feature alignment in this embodiment. The determinant of the cross-scale correlation consistency matrix for the first piece is 1.6437, the trace is 2.50, and the norm of cross-scale feature difference is 4.00, resulting in a calculated cross-scale alignment value of 0.164365. The determinant of the cross-scale correlation consistency matrix for the second piece is 0.6473, the trace is 3.00, and the norm of cross-scale feature difference is 4.00, resulting in a calculated cross-scale alignment value of 0.053944. The determinant of the cross-scale correlation consistency matrix for the third piece is 0.7152, and the trace is 2... The cross-scale feature difference norm is 5.00, and the calculated cross-scale alignment value is 0.057217. The determinant of the cross-scale correlation consistency matrix of the 4th segment is 1.5098, the trace of the cross-scale correlation consistency matrix is ​​2.20, the cross-scale feature difference norm is 0.45, and the calculated cross-scale alignment value is 1.525003. The determinant of the cross-scale correlation consistency matrix of the 5th segment is 0.9745, the trace of the cross-scale correlation consistency matrix is ​​2.80, the cross-scale feature difference norm is 3.00, and the calculated cross-scale alignment value is 0.116011.

[0033] Table 1. Evaluation Data of Piecewise Cross-Scale Feature Alignment

[0034] like Figure 2 The image shows a cross-scale correlation consistency spectrum distribution of architectural floor plans provided in an embodiment of this application. Table 1 and... Figure 2 Data shows that the cross-scale alignment values ​​obtained in this embodiment on different slices range from 0.05 to 1.53, and the curve trend is consistent with the combination relationship of determinant, trace, and feature difference norm in the table. The cross-scale alignment value of slice 4 is about 1.53, which is significantly larger, indicating that the correlation energy between fine-grained contours and semantic structures is more concentrated in this area, the difference residuals converge fully, the number of times the scale anomaly mask is triggered is significantly limited, and it is closer to the stable state of scale alignment. The alignment values ​​of slices 1 and 5 are around 0.11 to 0.16, corresponding to the slowly fluctuating position of the curve in the figure, reflecting that these areas have completed cross-scale alignment to a certain extent, while still relying on local upsampling, line enhancement convolution, and symbol texture compensation convolution to maintain stability. The alignment values ​​of segment 2 and segment 3 are in the range of 0.05 to 0.06, and the curve is close to the alignment threshold. This shows that these areas are more likely to be captured by the scale anomaly mask and cross-scale difficult area markers, and will be given priority in the review process in the subsequent stages of detection confidence constraint value calculation and spatial relationship reasoning, thereby improving the reliability and consistency of cross-scale feature analysis of building floor plans as a whole.

[0035] In this implementation scheme, matrix transposition, multiplication, differencing, and norm calculations are performed on the fine-grained feature matrix and the semantic feature matrix. Combined with the identity matrix and the numerical stability term ε, a cross-scale alignment value is ultimately obtained, achieving a measure of consistency in the spatial distribution and structural representation of features at different scales. This alignment value comprehensively reflects the degree of alignment of spatial information between the main-scale and auxiliary-scale features during the fusion process, effectively capturing global correlations and local differences between features. Through this metric, the system can dynamically adjust the fusion weights of multi-scale features, improving the balance of representation between small objects and large structures in the same feature space. This provides a highly consistent and low-biased feature foundation for subsequent architectural plan target detection and spatial relationship reasoning, thereby enhancing the stability and accuracy of the overall analysis.

[0036] Specifically, the steps for performing scale anomaly compensation processing and feature fusion path selection based on the alignment criterion output are as follows: By comparing the cross-scale alignment value and the alignment threshold in real time, when the cross-scale alignment value is less than the alignment threshold, a scale anomaly mask is generated in the fine-grained feature map. Local sub-blocks are extracted from the original fine-grained feature matrix according to the mask. After local upsampling, the area covered by the scale anomaly mask is cropped into fixed-size sub-blocks and then fed into line enhancement convolution and symbol texture compensation convolution. The line enhancement convolution uses a 3x3 stacked with a 5x5 learnable convolution kernel. The input consists of a feature block obtained by concatenating a fine-grained feature matrix and an edge guidance map by channel. The number of output channels is consistent with that of the fine-grained feature matrix. The symbolic texture compensation convolution uses a 3x3 learnable convolution kernel and modulates the convolution response with a weight map generated by the inflection point guidance map. During the joint training phase of the entire network, the parameters of the two types of convolutions are updated synchronously with the detection loss and cross-scale alignment criteria, thereby reconstructing a set of fine-grained compensation features. The attention operator is used to increase the proportion of fine-grained compensation features in the fused channels, while reducing the weight of the semantic feature matrix locally and masking scale anomalies. The membrane and enhanced features are written into the drawing parsing link database. Local secondary feature extraction is performed. In the secondary feature extraction stage, gradient responses are extracted from the fine-grained feature map to form an edge guidance map and locate endpoint candidates. Corner responses are extracted from the fine-grained feature map to form an inflection point guidance map and locate inflection point candidates. Connectivity is determined based on the endpoint direction angle threshold and the endpoint spacing threshold, and endpoint connection is performed to form a line segment completion mask and write back the corresponding position of the fused feature. Missing short line segments and arc responses are filled in the symbol region according to the consistency of line segment combination. After the write-back is completed, the cross-scale alignment value is recalculated. If the cross-scale alignment value is still less than the alignment threshold, it is marked as a cross-scale difficult area, further feature repair is stopped, the current fine-grained compensation feature is frozen as the reference upper limit, and the region index and local image slices are written into the difficult sample set. When the cross-scale alignment value is greater than or equal to the alignment threshold, only the standard pyramid fusion path is used, no additional enhancement branches are added, the corresponding position is marked as the scale alignment stable area, only one channel attention allocation is performed, and the fused feature is directly output after merging the fine-grained feature matrix and the semantic feature matrix.

[0037] In this implementation scheme, cross-scale mismatched regions are identified from the fusion link and compensated in a targeted manner by comparing cross-scale alignment values ​​with alignment thresholds in real time. When alignment is insufficient, a scale anomaly mask is used to lock the distorted local area, reconstruct fine-grained compensation features, and improve the detail contribution and suppress the interference of semantic projection in the local area during channel fusion. At the same time, edge guidance maps and inflection point guidance maps are used to complete line segment completion and local symbol repair, making the continuity of wall lines, door and window opening boundaries, and symbol line segments more stable. The abnormal regions and compensation records are written into the drawing parsing link database. When the standard is still not met after multiple compensations, the region is marked as a cross-scale difficult area and precipitated into a difficult sample set to provide a basis for subsequent training and quality backtracking. When the alignment standard is met, the standard pyramid fusion path is directly used to output the fusion features, so that downstream detection and spatial relationship reasoning obtain a more consistent input representation.

[0038] Specifically, the steps for determining the reliability of the detection by fusing features, category probability matrices, and cross-scale detail fidelity data are as follows: First, obtain the fusing features and cross-scale alignment values. Use the fusing features output from the previous stage as the detection input and the cross-scale alignment values ​​as a cross-scale stability reference to constrain the consistency of the current building segments at the semantic and fine-grained layers. Second, input the fusing features into the candidate box generator to obtain a candidate box set. Third, cover the areas where walls, doors, windows, symbols, and small components are located by multi-scale receptive fields and dense candidate box sampling. Finally, perform region alignment sampling on the fusing features for each candidate box to obtain a fixed-size feature block. Candidate regions of different scales and aspect ratios are aligned with features at a uniform size. Fixed-size feature blocks are input into the detection and classification head to obtain a class confidence vector. The class confidence vectors of all candidate instances are stacked in instance order to obtain a class probability matrix, providing a complete class response distribution for subsequent confidence constraints. The maximum class probability of each candidate instance is searched row-by-row in the class probability matrix. Only the maximum probability is retained at the corresponding position, and the values ​​at other positions in the same row are set to zero, resulting in a dominant class response matrix with the same shape as the class probability matrix, thus explicitly characterizing the dominant class response of each candidate instance. The class probability matrix is ​​then calculated... The Frobenius norm yields the overall energy term, used to measure the overall response strength of the current detection batch in the class space. Subtracting the principal class response matrix from the class probability matrix yields the non-principal class response matrix, used to focus on the confusion components of candidate instances in non-dominant classes. Calculating the Frobenius norm on the non-principal class response matrix yields the confusion energy term, reflecting the degree of class confusion by characterizing the overall strength of the non-principal class responses. Adding the confusion energy term to the numerical stability term yields the confusion stability term, providing numerical correction for extreme minimum values ​​and avoiding denominator instability. Dividing the overall energy term by the confusion stability term yields the class set ratio term, used to measure the main... The degree of concentration of class responses among all class responses; the cross-scale stability term is obtained by adding the cross-scale alignment value to the numerical stability term, and the cross-scale structural alignment is introduced into the detection credibility constraint; the cross-scale suppression term is obtained by taking the reciprocal of the cross-scale stability term, which is used to suppress overly aggressive class concentration evaluation when cross-scale alignment is insufficient; the credibility constraint value is obtained by multiplying the class concentration ratio term and the cross-scale suppression term, and the credibility constraint value is obtained by comprehensively reflecting the joint credibility level of the current candidate instance set in terms of class distribution concentration and cross-scale structural stability in numerical form, providing a criterion basis for subsequent small object enhanced sampling, label strategy adjustment and vector element output control.

[0039] The calculation method for the credibility constraint value is as follows: ; In the formula, This represents the credibility constraint value, indicating the credibility of the detection output under the constraints of category concentration and cross-scale bias. The category probability matrix represents the confidence distribution of each candidate instance across all categories. This represents the overall energy term of the classification probability matrix, used to characterize the response strength of the candidate set in the classification space; The main class response matrix represents the centralized representation of the main class response. This represents the energy of non-primary class responses, used to characterize the degree of class confusion and to penalize confusion. This represents the cross-scale detail fidelity value, used to apply cross-scale bias suppression to the detection confidence level; This represents a numerically stable term, obtained by setting it to less than... A positive number is used to ensure the stability of the denominator.

[0040] In this implementation scheme, a category probability matrix is ​​constructed by utilizing fusion features and extracting the main class response. This comprehensively measures the response strength and confusion level of candidate targets in the category space. Combined with cross-scale alignment values, a numerical credibility constraint index is formed to characterize the reliability level of target detection results in terms of semantic consistency and structural stability. Based on this index, the detection stage can distinguish between targets with clear structure, concentrated categories, and cross-scale consistency, and targets with complex structure, easy confusion, and significant cross-scale deviation. This provides a unified criterion for subsequent small object enhancement sampling, soft label and hard label switching, and vector element output screening, thereby improving the accuracy and stability of target detection results in building floor plans as a whole.

[0041] Specifically, the steps for performing small object enhancement sampling, soft / hard label switching, and vector feature output control based on the confidence discrimination results are as follows: By comparing the confidence constraint value and confidence threshold in real time, when the confidence constraint value is less than the confidence threshold, the candidate instance is included in the small object enhancement queue. The dual outputs of the main classification branch and the symbol auxiliary branch are retained as multi-label records, and a consistency verification field is added to these instances. For candidate instances entering the small object enhancement queue, the system uses repeated inference in zoomed and rotated views to obtain multiple sets of category confidence distributions under N transformations, and calculates the confidence dispersion in the form of standard deviation. If the confidence dispersion exceeds the threshold, the instance is marked as a target to be verified, and an output delay is set. The delay mechanism can be to automatically trigger secondary inference and take the mean and confidence interval of the N inference results, where N is the number of repetitions of multi-view inference, with a value of 3 to 5, indicating that each candidate instance undergoes N independent inferences in both zoomed and rotated views. For targets to be verified, a bounding box refinement regression operation is further performed, and the coordinate differences before and after correction are written into the displacement field. If, after review and refinement, the detected confidence constraint value is still below the confidence threshold, a downgrade labeling is performed, marking the instance as uncertain or low-confidence. All relevant records are then synchronously written to the drawing parsing link database for subsequent error tracing and model optimization. When the detected confidence constraint value is greater than or equal to the confidence threshold, the label generation process uses hard labels, directly converting the detected instance into vector feature output, including vertex coordinates, orientation angles, category codes, and confidence fields. This allows the spatial relationship reasoning module to directly call upon these features, achieving high-confidence, standardized output of the target structure in the architectural floor plan.

[0042] In this implementation scheme, dynamic triage and fine-tuning of candidate instances are achieved by comparing the detection confidence constraint value with the confidence threshold in real time. When the detection confidence constraint value is lower than the confidence threshold, the system adds the instance to the small object enhancement queue, retains the dual-path multi-label output of the main classification branch and the symbol auxiliary branch, and enhances the expression of complex small targets. A consistency verification field is added to these instances, and N repeated inferences are performed by scaling and rotating the view to collect multiple sets of category confidence scores. The standard deviation is calculated as the confidence dispersion, and an output delay mechanism is triggered. The stability of the results is improved by using the quadratic inference mean and confidence interval. At the same time, bounding box refinement regression is performed, and the coordinate differences before and after correction are recorded. If the detection confidence constraint value still does not reach the confidence threshold after verification, the system downgrades the instance. All verification and downgrade records are synchronously written to the drawing parsing link database to provide data support for subsequent source tracing analysis and model optimization. When the detected confidence constraint value is greater than or equal to the confidence threshold, a hard labeling strategy is directly adopted to convert the instance into a vector feature output, including vertex coordinates, orientation angle, category code and confidence field. This ensures that the high-confidence structure can be efficiently utilized by the spatial relationship reasoning module, thereby improving the overall recognition accuracy of multi-category targets in the building floor plan and the traceability of the data link.

[0043] Specifically, the steps for constructing candidate topologies based on vector features and constructing topology consistency judgment through relational reasoning are as follows: Obtain detection confidence constraint values ​​and vector features. The vector features are derived from vertex coordinates, orientation angles, category codes, and confidence fields output by the detection and classification module, and are organized into a node attribute table according to feature identifiers; construct a node list based on vector features, registering room units, wall segments, door openings, window openings, and symbolic features as nodes and recording their geometric bounding boxes, orientation vectors, and endpoint sequences; perform candidate relation filtering based on geometric proximity, orientation consistency, overlap relationships, and endpoint adsorption connectivity rules to generate a candidate edge set, and write the candidate edge set into an adjacency matrix to limit the scope of relational reasoning; construct a relational reasoning network within the candidate edge range limited by the adjacency matrix. The relational reasoning network receives the relative distance, angle difference, and endpoint connection status between the node attribute table and the candidate edges as input and... Output the relation confidence and relation type for each candidate edge to form a structured relation record between elements and aggregate it to obtain the relation confidence matrix; perform element-wise multiplication of the adjacency matrix and the relation confidence matrix to obtain the relation coupling matrix; calculate the Frobenius norm of the relation coupling matrix to obtain the effective relation strength energy term; calculate the Frobenius norm of the adjacency matrix to obtain the adjacency structure energy term; add the adjacency structure energy term and the numerical stability term to obtain the adjacency stability term; divide the effective relation strength energy term by the adjacency stability term to obtain the relation strength ratio term; add the detection confidence constraint value and the numerical stability term to obtain the confidence stability term; divide the detection confidence constraint value by the confidence stability term to obtain the confidence constraint ratio term; multiply the relation strength ratio term and the confidence constraint ratio term to obtain the topology consistency verification value, and write the topology consistency verification value and relation type together into the parsed element traceability database.

[0044] The topology consistency check value is calculated as follows: ; In the formula, This represents the topology consistency check value, indicating the degree of topology consistency of the relation inference output under the combined effect of adjacency constraints and reliable detection constraints. The adjacency matrix is ​​used to record the candidate connection relationships between room units and wall segments, door openings, and window openings. It is obtained by constructing the matrix based on the coordinates, directions, and connectivity determination rules of the detection output. The relation confidence matrix represents the output of the relation network on the degree to which a node's relation holds true. It is obtained by inputting node features and edge features into the relation network. The relation coupling matrix represents the effective relation strength under adjacency constraints, and is obtained by... and Obtained by element-wise multiplication; The term representing the strength of effective relations is used to characterize the overall stability of inference relations. This represents the energy term of the adjacency structure, used to constrain the structural scale of the relationship strength. This represents the detection credibility constraint value, and the detection credibility basis for relational reasoning dependencies; This indicates a numerically stable term, set to less than A positive number is used to ensure the stability of the denominator.

[0045] In this implementation scheme, element-level topology construction and relation reasoning are completed based on vector elements. The geometric and semantic information of room units, wall segments, door openings, window openings, and symbol elements are organized into nodes and candidate connections. Within the candidate connection range, the relation type and relation confidence are output to form a searchable spatial relation structure record. At the same time, the effective relation strength under adjacency constraints and the detection credibility basis are jointly quantified to obtain the topology consistency verification result, thereby locating the connection segments that need to be verified and providing a basis for the subsequent solidification of the topologically stable set. Specifically, the steps for performing spatial connectivity verification and correction and topological stability solidification operations based on the consistency judgment results are as follows: Real-time comparison of topological consistency check values ​​and consistency thresholds, such as... Figure 3 This is a flowchart of the spatial relationship topology consistency verification and correction process in this embodiment. When the topology consistency verification value is less than the consistency threshold, the relationship verification and correction process is initiated. Abnormal relationship edges related to door openings, window openings, wall segments, and room units are written into a conflict list and a conflict subgraph is constructed. Candidate connection relationships are generated only within the conflict subgraph and relationship reasoning is re-executed. Geometric rules are introduced to verify and correct and eliminate abnormal combinations of spatial relationships. Geometric rules include that door openings must connect the two spatial boundaries, window openings must be attached to the outer wall boundary, wall segments must form a closed chain enclosing the room, and the opening direction of door openings must be consistent with the direction of wall segments. Relationships that do not meet the rules are multiplied by a decay coefficient to reduce their weight. Relationship edges whose topology consistency verification value is still less than the consistency threshold after weight reduction are deleted. In addition to reconnecting and completing the neighboring domains, the corrected spatial relationship structure, along with the conflict list and correction records, is written into the drawing parsing link database. When the topology consistency check value is greater than or equal to the consistency threshold, the corresponding relationship is assigned to the topology stable set, the spatial relationship structure is solidified, and written into the drawing parsing link database. The spatial relationship structure includes feature identifiers, feature geometry, relationship type, relationship confidence, connection endpoints, and enclosure information. The attenuation coefficient ranges from 0.3 to 0.8. A suitable range is selected by statistical analysis and grid search of the relationship confidence distribution in historical building samples. In implementation, 0.5 is preferred as the default attenuation intensity to achieve a balance between the risk of erroneous deletion of relationship edges and the ability to suppress abnormal relationships.

[0046] In this implementation scheme, by introducing a comparison between the topological consistency check value and the consistency threshold, abnormal connections are concentrated under the inherent geometric rules of the conflict subgraph to complete weight reduction, deletion and neighborhood reconnection. The correction intensity is controlled by the attenuation coefficient, and finally the topologically stable set is screened and solidified, so that the spatial connection relationship forms a traceable structured output between contradiction elimination, break completion and reliable preservation.

[0047] Specifically, the steps for completing cross-segment alignment, geometric repair, and topological cleanup based on vector features and spatial relationship structures, and generating an archived result set, are as follows: Inputting vector features and spatial relationship structures, the system first performs deduplication and merging and coordinate alignment in overlapping areas according to the segment location mapping table, ensuring the generation of unique vector features within the same region, and writing them into the source traceability field to record the feature source link. Based on the line segment completion mask, the system automatically repairs broken segments of wall lines and door / window opening boundaries, ensuring continuous closure of enclosure boundaries and improving overall topological integrity. For areas where scale anomaly masks are detected, the system automatically replaces the corresponding feature version and writes a cross-scale problematic area marker for easy subsequent retrieval and verification. For verification targets, the system merges multi-label records of the main classification branch and symbol auxiliary branch, retains the confidence dispersion field, and synchronously writes a degradation identifier for downgraded annotation instances, achieving result traceability and anomaly identification. Based on the conflict list, the system cleans up abnormal relationship edges, performs endpoint alignment and polyline smoothing on structures such as room enclosure chains, entrance / exit connection chains, and symbol attachment relationships, and outputs a topologically stable set. Finally, based on structural elements such as room connectivity, opening distribution, electrical and fire protection layout, a result summary table is generated, and all structural data is output as vector result files conforming to preset vector format specifications. Commonly used formats include DXF, SVG, GeoJSON, etc. The complete link and structural data are archived to the drawing parsing link database to meet the needs of subsequent engineering applications, structural traceability and batch management.

[0048] This implementation plan achieves deduplication and merging of vector elements, coordinate alignment, and topology repair, outputting vector result files with closed structures, clear classification, and standardized format, and archiving them to the drawing parsing link database, providing authoritative data support for subsequent engineering design, structural review, and batch management.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based intelligent analysis system for architectural floor plans, characterized in that: include: The data acquisition and preprocessing module is used to acquire multi-source building floor plans, convert and enhance them, output a set of usable building slice images, store them, and then build a drawing parsing link database. The multi-scale feature extraction module is used to perform multi-scale parallel extraction on the architectural tile image set and build a multi-scale feature extraction model, outputting a fine-grained feature matrix and a semantic feature matrix; The multi-scale feature fusion module is used to construct a cross-scale correlation consistency matrix and establish a cross-scale alignment criterion based on correlation energy and difference residuals. Based on the alignment criterion output, it performs scale anomaly compensation processing and feature fusion path selection operations. The object detection and classification module is used to determine the reliability of detection by fusing features, category probability matrices and cross-scale detail-fidelity data, and to perform small object enhancement sampling, soft and hard label switching and vector feature output control operations based on the reliability determination results. The spatial relationship reasoning module is used to construct candidate topologies based on vector elements and construct topology consistency discrimination through relationship reasoning. Based on the consistency discrimination results, it performs spatial connection relationship verification and correction and topology stability structure solidification operations. The post-processing and optimization module is used to perform cross-part alignment, geometric repair, and topology cleanup based on vector features and spatial relationship structures, and generate archived result sets.

2. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for acquiring multi-source building floor plans, converting and enhancing them, outputting a set of usable building fragment images and storing them to construct a drawing parsing link database are as follows: The system collects raw architectural floor plan data, receiving raw architectural floor plan data from scanned documents, exported files from drawing software, and screenshots. Through format parsing and rendering, it converts multiple formats such as PDF, JPG, PNG, and TIFF into preset target resolutions and grayscale normalization ranges. First, it performs linear stretching and histogram equalization within the grayscale domain to complete grayscale processing, ensuring a uniform brightness distribution across different source drawings. Then, it performs noise suppression and contrast enhancement on the normalized grayscale image, suppressing background textures and grid line residues, and highlighting wall lines, door and window outlines, and symbolic line segments. For ultra-large format drawings, it performs segmented cropping and overlapping window generation to form a set of architectural segmented images suitable for feature extraction. Simultaneously, it records the positional mapping relationship of each segment in the original image to support result write-back. Based on connected component analysis, the main structural regions are extracted, and a structural region index and candidate region mask are generated to limit the processing range and reduce invalid background interference. After processing, the output includes segmented images, segmented location mapping tables, structural region indexes, and candidate region masks. The segmented indexes, mapping relationships, and processing logs are then stored in the constructed drawing parsing link database.

3. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for performing multi-scale parallel extraction on the architectural segmented image set and constructing a multi-scale feature extraction model to output a fine-grained feature matrix and a semantic feature matrix are as follows: The system takes a set of architectural images as input and uses an image segmentation algorithm to divide the original image into several windows and local regions. A pyramid-level approach is used to extract features simultaneously across multiple scales, allowing large-scale structural information such as walls and room layouts to be included in the feature representation along with small-scale component information such as electrical switches, legend symbols, and small windows and doors. During feature extraction, dilated convolutions are introduced to expand the receptive field, enabling the network to obtain more complete contextual information without sacrificing spatial detail, thus avoiding the problems of texture smoothing and edge swallowing caused by multiple downsampling of small components. Furthermore, multi-channel convolutional kernels are introduced at each scale level, allowing different channels to respond to structural elements of different sizes and orientations, strengthening the differential representation of long straight wall lines, diagonal wall lines, door and window opening curves, and symbolic line segment combinations. For window-cropped images, an independent multi-scale extraction process is adopted to separate the feature extraction and encoding processes of different regions, reducing the expression bias caused by small-scale components being diluted by large-area backgrounds and large structures being fragmented by local cropping. During the training phase, random scaling, rotation, and flipping enhancements are introduced to enable the network to learn stable feature patterns of the same structure under different scales, orientations, and layout densities, improving its adaptability across different paper scales. Based on the above processing, a multi-scale feature extraction model is constructed, outputting fine-grained feature maps and semantic feature maps. By performing linear transformations and scale alignment on the channel and spatial dimensions through attention and sampling operators, the semantic feature map is mapped to the same scale and number of channels as the fine-grained feature map, and then expanded according to spatial location to obtain the semantic feature matrix.

4. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for constructing a cross-scale correlation consistency matrix and establishing a cross-scale alignment criterion based on correlation energy and difference residuals are as follows: Obtain the fine-grained feature matrix and the semantic feature matrix; transpose the fine-grained feature matrix to obtain the transposed fine-grained matrix; perform matrix multiplication between the transposed fine-grained matrix and the semantic feature matrix to obtain the cross-scale correlation matrix; transpose the cross-scale correlation matrix to obtain the transposed cross-scale correlation matrix; perform matrix multiplication between the transposed cross-scale correlation matrix and the cross-scale correlation matrix to obtain the correlation cumulative matrix; Construct the identity matrix; Adding the identity matrix to the correlation cumulative matrix yields the cross-scale correlation consistency matrix; The determinant of the cross-scale correlation consistency matrix is ​​calculated and its absolute value is taken to obtain the determinant term; the trace of the cross-scale correlation consistency matrix is ​​calculated to obtain the trace term, which is added to the numerical stability term to obtain the trace stability term; the difference operation is performed on the fine-grained feature matrix and the semantic feature matrix to obtain the cross-scale difference matrix; the Frobenius norm of the cross-scale difference matrix is ​​calculated to obtain the difference norm term, which is added to the numerical stability term to obtain the difference stability term; the trace stability term and the difference stability term are multiplied to obtain the alignment denominator term; the determinant term is divided by the alignment denominator term to obtain the cross-scale alignment value.

5. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for performing scale anomaly compensation processing and feature fusion path selection based on the alignment criterion are as follows: By comparing cross-scale alignment values ​​and alignment thresholds in real time, when the cross-scale alignment value is less than the alignment threshold, a scale anomaly mask is generated in the fine-grained feature map. Local sub-blocks are extracted from the original fine-grained feature matrix according to the mask. After local upsampling, the area covered by the scale anomaly mask is cropped into fixed-size sub-blocks and fed into line enhancement convolution and symbol texture compensation convolution. The line enhancement convolution uses a 3x3 stacked with a 5x5 learnable convolution kernel. The input is a feature block obtained by concatenating the fine-grained feature matrix and the edge guidance map by channel. The number of output channels is consistent with the fine-grained feature matrix. The symbol texture compensation convolution uses a 3x3 learnable convolution kernel and modulates the convolution response with a weight map generated by the inflection point guidance map. The parameters of the two types of convolutions are updated synchronously with the detection loss and cross-scale alignment criteria during the joint training phase of the whole network, thereby reconstructing a set of fine-grained features. The fine-grained compensation feature is improved by using an attention operator to increase the proportion of fine-grained compensation features in the fusion channel, while reducing the weight of the semantic feature matrix in the local area. The scale anomaly mask and the enhanced features are written into the drawing parsing link database. Local secondary feature extraction is performed. In the secondary feature extraction stage, connectivity is determined based on the endpoint direction angle threshold and the endpoint spacing threshold, and endpoint connection is performed to form a line segment completion mask and write back the corresponding position of the fusion feature. In the symbol area, missing short line segments and arc responses are filled in according to the consistency of line segment combination. After the write-back is completed, the cross-scale alignment value is recalculated. If the cross-scale alignment value is still less than the alignment threshold, it is marked as a cross-scale difficult area, further feature repair is stopped, the current fine-grained compensation feature is frozen as the reference upper limit, and the region index and local image slices are written into the difficult sample set. When the cross-scale alignment value is greater than or equal to the alignment threshold, only the standard pyramid fusion path is used, no additional enhancement branches are added, the corresponding position is marked as the scale alignment stable region, only one channel attention allocation is performed, and the fused features are directly output after merging the fine-grained feature matrix and the semantic feature matrix.

6. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for determining the credibility of the detection by fusing features, category probability matrices, and cross-scale detail-fidelity data are as follows: Obtain fused features and cross-scale alignment values; input the fused features into the candidate box generator to obtain a candidate box set, then perform region alignment sampling on the fused features for each candidate box to obtain a fixed-size feature block, input the fixed-size feature block into the detection and classification head to obtain a class confidence vector, stack the class confidence vectors of all candidate instances in the instance order to obtain a class probability matrix; search for the maximum class probability of each candidate instance in the class probability matrix row by row, retain only the maximum probability at the corresponding position and set the values ​​of other positions in the same row to zero, to obtain a main class response matrix with the same shape as the class probability matrix; The overall energy term is obtained by calculating the Frobenius norm on the category probability matrix; Subtracting the principal class response matrix from the class probability matrix yields the non-principal class response matrix; calculating the Frobenius norm of the non-principal class response matrix yields the confusion energy term; adding the confusion energy term to the numerical stability term yields the confusion stability term; dividing the overall energy term by the confusion stability term yields the class set ratio term; adding the cross-scale alignment value to the numerical stability term yields the cross-scale stability term; taking the reciprocal of the cross-scale stability term yields the cross-scale suppression term; multiplying the class set ratio term by the cross-scale suppression term yields the confidence constraint value.

7. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for performing small object enhanced sampling, soft / hard label switching, and vector feature output control operations based on the reliable discrimination results are as follows: By comparing the detection confidence constraint value and confidence threshold in real time, when the detection confidence constraint value is less than the confidence threshold, the candidate instance is added to the small object reinforcement queue. The dual output of the main classification branch and the symbol auxiliary branch is retained as a multi-label record. A consistency verification field is added to the candidate instances that enter the small object reinforcement queue. Multiple sets of category confidence are obtained by repeated inference using zoomed and rotated views and the confidence dispersion is recorded. Candidate instances with confidence dispersion exceeding the limit are marked as targets to be verified and the output delay is set. At the same time, bounding box refinement regression is performed on the targets to be verified and the difference between the coordinates before and after the correction is written into the displacement field. If the recalculated detection confidence constraint value is still less than the confidence threshold, perform downgrade annotation and write the record to the drawing parsing link database. When the detected confidence constraint value is greater than or equal to the confidence threshold, the label generation process uses hard labels to directly convert the detected instance into vector feature output. The vector features include vertex coordinates, orientation angles, category codes, and confidence fields, which can be directly used by the spatial relationship reasoning module.

8. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for constructing candidate topologies based on vector features and constructing topology consistency judgment through relational reasoning are as follows: Obtain the detection confidence constraint value and vector features; construct a node list based on the vector features; perform candidate relationship filtering to generate a candidate edge set according to the rules of geometric proximity, direction consistency, overlap relationship and endpoint adsorption connectivity; and write the candidate edge set into the adjacency matrix. A relation reasoning network is constructed within the candidate edge range defined by the adjacency matrix. The relation reasoning network receives node features and edge features and outputs the relation confidence and relation type of each candidate edge, thereby forming a structured relation record between elements and outputting a relation confidence matrix. Element-wise multiplication of the adjacency matrix and the relation confidence matrix yields the relation coupling matrix; The Frobenius norm is calculated on the relation coupling matrix to obtain the effective relation strength energy term; the Frobenius norm is calculated on the adjacency matrix to obtain the adjacency structure energy term. The adjacency structure energy term is added to the numerical stability term to obtain the adjacency stability term; the effective relation strength energy term is divided by the adjacency stability term to obtain the relation strength ratio term. The reliable stability term is obtained by adding the detected reliable constraint value to the numerical stability term; The trusted constraint ratio is obtained by dividing the trusted stability term by the trusted constraint value; the topology consistency verification value is obtained by multiplying the relation strength ratio term by the trusted constraint ratio term.

9. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for performing spatial connectivity verification and correction and topological stability solidification operations based on the consistency discrimination results are as follows: The topology consistency check value and consistency threshold are compared in real time. When the topology consistency check value is less than the consistency threshold, the relationship review and correction process is initiated. Abnormal relationship edges related to door openings, window openings, wall segments, and room units are written into the conflict list and a conflict subgraph is constructed. Candidate connection relationships are generated only within the conflict subgraph and relationship reasoning is re-executed. Geometric rules are introduced to verify and correct and eliminate abnormal combinations of spatial relationships. Geometric rules include that door openings must connect the two spatial boundaries, window openings must be attached to the outer wall boundary, wall segments must form a closed chain enclosing the room, and the opening direction of door openings must be consistent with the direction of wall segments. Relationships that do not meet the rules are multiplied by a decay coefficient to reduce their weight. Relationship edges whose topology consistency check value is still less than the consistency threshold after weight reduction are deleted and their neighborhood reconnection is completed. The corrected spatial relationship structure, along with the conflict list and correction record, is written into the drawing parsing link database. When the topology consistency check value is greater than or equal to the consistency threshold, the corresponding relationship is included in the topology stable set, the spatial relationship structure is solidified and written into the drawing parsing link database. The spatial relationship structure includes feature identifier, feature geometry, relationship type, relationship confidence, connection endpoints and enclosure information.

10. The intelligent architectural floor plan parsing system based on deep learning according to claim 1, characterized in that: The specific steps for completing cross-fragment alignment, geometric repair, and topology cleanup based on vector features and spatial relationship structures, and generating an archived result set, are as follows: The process begins by inputting vector elements and spatial relationship structures. First, in overlapping areas, deduplication, merging, and coordinate alignment are performed based on the patch location mapping table to create unique vector elements, which are then written into the source traceability field. Next, broken sections of wall lines and door / window opening boundaries are repaired using line segment completion masks to ensure closed enclosure boundaries. Then, feature versions in areas with scale anomalies are replaced using masks, and cross-scale problematic area markers are written. For targets to be reviewed, multi-label records of the main classification branch and symbol auxiliary branch are merged while preserving confidence dispersion, and downgraded annotation records are marked with downgrade identifiers. Based on the conflict list, abnormal relationship edges are cleaned up, and endpoint alignment and polyline smoothing are performed on room enclosure chains, entrance / exit connection chains, and symbol attachment relationships, outputting a topologically stable set. Finally, a result summary table and vector output files conforming to the preset vector format specifications are generated according to room connectivity, opening distribution, and electrical and fire protection layout, and archived in the drawing analysis link database.

Images