Building three-dimensional model local redrawing generation method based on spatial confidence map guidance

Abstract

The application provides a building three-dimensional model local redrawing generation method based on spatial confidence map guidance. Building three-dimensional data and associated observation data are acquired, multi-source confidence evidence is extracted, and a spatial confidence map is generated; multi-scale confidence representation, layered redrawing masks, and scale labels are generated based on the spatial confidence map; local surface direction representation is extracted according to the layered redrawing masks and direction modulation noise is generated; redrawing condition representation is generated based on the building three-dimensional data and the associated observation data, redrawing guidance representation is formed by limiting the layered redrawing masks and the scale labels, and local redrawing representation is generated in cooperation with the direction modulation noise; the local redrawing representation is rechecked based on the spatial confidence map, local fine-tuning results and reserved redrawing representation are generated, fine-tuning redrawing representation and the reserved redrawing representation are associated to generate local redrawing results.

Description

Technical Field

[0001] This application relates to the field of architectural 3D reconstruction and 3D scene generation and repair technology, specifically, to a method for generating local redrawing of architectural 3D models based on spatial confidence maps. Background Technology

[0002] With the application of laser scanning, image reconstruction, and other 3D data acquisition technologies in digital building scenarios, 3D building data can relatively completely represent the spatial location, surface morphology, and local structure of building components. However, in the actual acquisition and 3D reconstruction process, due to the influence of observation occlusion, surface reflection characteristics, differences in sampling conditions, and differences in the quality of data from different sources, local voids, surface anomalies, or missing local structures can easily form in 3D building data. Since the defect range and surface geometry vary in different parts of the building, local defect repair not only requires determining the spatial range that needs to be regenerated, but also requires ensuring that the regenerated part maintains a good connection with the original 3D building data in terms of local structure and surface morphology.

[0003] Existing methods for local restoration of architectural 3D data typically involve setting judgment boundaries for local quality indicators in the 3D data and identifying anomalous spatial regions based on these indicators. After identifying the anomalous spatial regions, they are expanded within a fixed spatial range. A uniform form of random perturbation is then applied to the expanded spatial region, and local 3D content is generated based on the original architectural data surrounding the anomalous spatial regions. Once the local 3D content is generated, the generated portion directly replaces the original anomalous spatial regions to form the restored architectural 3D data.

[0004] However, the aforementioned solutions are insufficiently adaptable to situations where architectural 3D data contains observation errors from different sources, defects of varying spatial scales, and surface orientation characteristics. On one hand, a single quality index cannot simultaneously reflect the differences in local reliability resulting from different data sources, and a fixed spatial range cannot simultaneously accommodate large-scale structural defects and local detail deficiencies, easily leading to a mismatch between the repair area and the actual defect area. On the other hand, uniform random perturbations do not utilize the directional differences of local architectural surfaces, easily generating additional perturbations to surface orientations that do not require significant changes during local generation. Furthermore, directly outputting results after a single local generation makes it difficult to adjust remaining local anomalies based on the actual situation after generation, while areas that have already met repair requirements are easily unnecessarily affected during subsequent overall processing. Therefore, existing solutions struggle to simultaneously address the reliability assessment of local defect areas in architectural 3D data, spatial scale adaptation, surface orientation adaptation, and the local review and targeted re-repair of the generated results. Summary of the Invention

[0005] This application provides a method for generating local redrawing of architectural 3D models based on spatial confidence maps, which at least alleviates the technical problems existing in the local repair process of architectural 3D data, such as one-sided judgment of local confidence level, insufficient adaptation between repair range and defect spatial scale, failure to fully utilize surface orientation differences in the local generation process, and difficulty in performing local repair on residual defects after one generation.

[0006] A method for generating local redrawing of a 3D building model based on spatial confidence maps includes: Acquire 3D building data and associated observation data, extract multi-source confidence evidence from the 3D building data and associated observation data, and generate a spatial confidence map based on the multi-source confidence evidence; Multi-scale confidence representations are generated based on spatial confidence maps, hierarchical redraw masks are generated based on multi-scale confidence representations, and scale labels are generated based on multi-scale confidence representations and hierarchical redraw masks. Local surface orientation representations are extracted from building 3D data based on hierarchical redraw mask, and orientation modulation noise is generated based on local surface orientation representations and scale labels. Redrawing condition representations are generated based on building 3D data and associated observation data. Redrawing condition representations are limited by layered redrawing masks and scale labels to generate redrawing guidance representations. Local redrawing representations are generated based on directional modulation noise and redrawing guidance representations. The process involves re-examining local redrawing representations based on spatial confidence maps to generate redrawing re-examination results; generating local refinement results based on the redrawing re-examination results; generating refined redrawing representations based on the local refinement results; extracting and retaining redrawing representations from the local redrawing representations based on the redrawing re-examination results; and then associating the retained redrawing representations with the refined redrawing representations to generate local redrawing results.

[0007] Optionally, the extraction of multi-source confidence evidence from building 3D data and associated observation data includes: Local geometric observation representations are extracted from architectural 3D data and associated observation data, and evidence areas are delineated based on these local geometric observation representations. Extract semantic recognition evidence from associated observation data according to evidence regions; Read the data source identifier corresponding to the evidence area from the associated observation data, extract the cross-source observation results corresponding to the evidence area based on the data source identifier, and then compare the cross-source observation results to generate cross-source consistent evidence; Read the collection time identifier corresponding to the evidence area from the associated observation data, generate the collection time interval based on the collection time identifier, and generate time-sensitive evidence based on the collection time interval; Multi-source confidence evidence is generated based on local geometric observation representations, semantic recognition evidence, cross-source consistent evidence, and time-sensitive evidence.

[0008] Optionally, the step of extracting local geometric observation representations from building 3D data and associated observation data, and dividing the evidence region based on the local geometric observation representations, includes: Read the 3D sampling positions from the 3D building data, and generate a spatial neighborhood based on the 3D sampling positions; Based on the neighborhood distribution of the three-dimensional sampling locations in the spatial neighborhood, sampling sufficiency is generated; A neighborhood normal representation is generated based on the 3D sampling position in the spatial neighborhood, and a local orientation consistency is generated based on the directional dispersion of the neighborhood normal representation. Depth observation data is read from associated observation data according to spatial neighborhood, and observation signal-to-noise representation is generated based on the depth observation data; Local geometric observation representations are generated based on sampling sufficiency, local orientation consistency, and observation signal-to-noise ratio. Evidence regions are then divided based on the correspondence between local geometric observation representations and three-dimensional sampling positions.

[0009] Optionally, the generation of a spatial confidence map based on multi-source confidence evidence includes: Acquire historical building quality labeling data, extract historical confidence evidence and corresponding quality labels from the historical building quality labeling data, and generate evidence contribution representation based on the correspondence between historical confidence evidence and corresponding quality labels; Multi-source confidence evidence is modulated based on the contribution characteristics of the evidence to generate contribution-modulated evidence; Based on contribution modulation evidence, a statistical representation of the evidence neighborhood is generated, and then an initial spatial confidence map is generated based on the contribution modulation evidence and the statistical representation of the evidence neighborhood. Confidence boundary gradients are generated based on the initial spatial confidence map, and the initial spatial confidence map is smoothed by maintaining the boundaries based on the confidence boundary gradients to generate a spatial confidence map.

[0010] Optionally, the generation of multi-scale confidence representations based on spatial confidence maps includes: Maintain the spatial sampling resolution of the spatial confidence map to generate a fine-scale confidence map; Reduce spatial sampling resolution based on fine-scale confidence maps to generate mesoscale confidence maps; Based on the mesoscale confidence map, the spatial sampling resolution is further reduced to generate a coarse-scale confidence map; Scale location correspondence is generated based on the spatial sampling resolution correspondence between fine-scale confidence maps, mesoscale confidence maps, and coarse-scale confidence maps. The fine-scale confidence maps, mesoscale confidence maps, and coarse-scale confidence maps are then associated based on the scale location correspondence to generate multi-scale confidence representations.

[0011] Optionally, the step of generating a hierarchical redraw mask based on multi-scale confidence representation, and then generating scale labels based on the multi-scale confidence representation and the hierarchical redraw mask, includes: Based on the coarse-scale confidence map, continuously distributed low-confidence regions are identified, and candidate regions for structural defects are generated. Based on the candidate regions of structural defects, the spatial regions corresponding to the candidate regions of structural defects are excluded from the fine-scale confidence map to generate a fine-scale residual confidence map. Then, based on the fine-scale residual confidence map, isolated low-confidence regions are identified to generate candidate regions of detailed defects. Based on the mesoscale confidence map, candidate regions for structural defects and candidate regions for detailed defects are detected respectively, and scale verification results are generated. Based on the scale verification results, candidate regions for structural defects and candidate regions for detailed defects are selected to generate hierarchical candidate regions; Based on the layered candidate regions, boundary expansion results are generated. Based on the boundary expansion results, a layered redrawing mask is generated. Then, based on the scale position correspondence, the layered candidate regions are associated with the corresponding regions in the layered redrawing mask to generate scale labels.

[0012] Optionally, the extraction of local surface orientation representations from architectural 3D data based on layered redraw masks includes: The mask coverage area is located by redrawing the mask layer by layer, and the local sampling position is read from the 3D building data based on the mask coverage area; Local surface adjacency relationships are generated based on local sampling locations, and local normal directions are generated based on these local surface adjacency relationships. Based on the local normal direction, the local tangent direction orthogonal to the local normal direction is determined, and then the local normal direction and the local tangent direction are associated to generate a local surface orientation representation.

[0013] Optionally, generating directional modulation noise based on local surface orientation characterization and scale labels includes: Local normal direction and local tangent direction are extracted from the local surface orientation representation. Normal direction noise component is generated based on local normal direction and tangent direction noise component is generated based on local tangent direction. Normal direction noise component and tangent direction noise component are associated according to the relationship that the noise intensity of normal direction noise component is less than that of tangent direction noise component to generate initial orientation noise. Based on the scale labels, the initial directional noise is adjusted according to the relationship that the spatial frequency corresponding to the candidate region of structural defects is less than that corresponding to the candidate region of detailed defects, thus generating scale-modulated noise. Based on scale labels, adjacent mask regions corresponding to different scale labels are identified in the layered redraw mask, and scale boundary positions are generated. Transition modulation representations are generated based on the relative positions of sampling points corresponding to scale boundary positions between adjacent mask regions corresponding to different scale labels. Scale modulation noise is adjusted based on transition modulation representations to generate directional modulation noise.

[0014] Optionally, the step of generating a redrawing condition representation based on building 3D data and associated observation data, and limiting the redrawing condition representation according to a layered redrawing mask and scale labels to generate a redrawing guidance representation; generating a local redrawing representation based on directional modulation noise and the redrawing guidance representation, includes: Based on the layered redrawing mask, the mask coverage area and the area outside the mask are divided, and the original three-dimensional representation is read from the building three-dimensional data according to the area outside the mask; According to the mask coverage area, the observation condition representation is extracted from the associated observation data, and the original 3D representation and the observation condition representation are associated to generate the redrawing condition representation; The effective area of ​​the redraw condition representation is limited by the layered redraw mask, the spatial attention range is generated based on the scale label, and then the spatial association range of the redraw condition representation is limited based on the spatial attention range to generate the redraw guidance representation. The initial redrawing representation is generated based on the directional modulation noise and the redrawing guidance representation. The three-dimensional representation of the mask-covered area is then recovered step by step based on the initial redrawing representation to generate the local redrawing representation.

[0015] Optionally, the step of generating an initial redrawing representation based on directional modulation noise and redrawing guidance representation, and progressively recovering the three-dimensional representation of the mask-covered area based on the initial redrawing representation to generate a local redrawing representation, includes: Based on the redrawing guidance characterization, directional modulation noise is applied to the mask coverage area to generate the initial redrawing characterization; The process of hierarchically reconstructing the 3D representation of the masked area is divided into multiple reconstruction levels to generate reconstruction levels; Extract the mask boundary positions from the layered redraw mask, and generate boundary transition representations based on the mask boundary positions; Based on the restoration level, the first restoration level's 3D representation restoration is performed based on the initial redrawing representation to generate the first-level redrawing 3D representation; the first-level redrawing 3D representation and the original 3D representation are fused based on the boundary transition representation to generate inter-level redrawing representations; The three-dimensional representation of the next recovery level is restored sequentially based on the inter-level redrawing representation corresponding to the previous recovery level to generate the redrawing three-dimensional representation corresponding to the next recovery level. Then, the redrawing three-dimensional representation corresponding to the next recovery level and the original three-dimensional representation are fused according to the boundary transition representation to update the inter-level redrawing representation. After the final recovery stage is completed, candidate redrawing 3D representations are generated based on the inter-stage redrawing representations. The candidate redrawing 3D representations and the original 3D representations are then fused based on the boundary transition representations to generate local redrawing representations.

[0016] Optionally, the step of generating an initial redrawing representation based on directional modulation noise and redrawing guidance representation, and progressively recovering the three-dimensional representation of the mask-covered area based on the initial redrawing representation to generate a local redrawing representation, further includes: Based on the scale verification results, the region to be verified is extracted, and the 3D representation to be redrawn is extracted from the candidate 3D representation to be verified based on the region to be verified. Based on the detection of the 3D representation to be redrawn for review using mesoscale confidence maps, a consistent redrawing result is generated. Based on the redrawing consistency results, key review markers are generated and associated with local redrawing representations, so that key review markers and local redrawing representations jointly participate in the generation of redrawing review results.

[0017] Optionally, the step of reviewing the local redrawing representation based on the spatial confidence map to generate the redrawing review result includes: The redrawing area is determined based on the local redrawing representation. The original three-dimensional representations corresponding to the adjacent positions of the redrawing area are extracted from the building three-dimensional data. Based on the local redrawing representation and the original three-dimensional representation, a coarse-scale structural continuity result is generated. Based on the spatial confidence map, the original confidence representation corresponding to the redrawn region is extracted. The local integrity of the redrawn region is detected based on the local redrawn representation to generate the local integrity result. Then, a fine-scale residual confidence representation is generated based on the original confidence representation and the local integrity result. Extract the mask boundary positions from the layered redraw mask, and detect the boundary continuity relationship between the local redraw representation and the original 3D representation based on the mask boundary positions to generate boundary continuity results; Redrawing and re-inspection results are generated based on coarse-scale structural continuity results, fine-scale residual confidence characterization, boundary continuity results, and key verification markers.

[0018] Optionally, generating local refinement results based on the redrawing and re-inspection results includes: Extract the non-compliant areas and reasons for non-compliance from the redrawing and re-inspection results; Based on the reasons for non-compliance, mask adjustment information is generated. Based on the mask adjustment information, the layered redraw mask corresponding to the non-compliance area is locally adjusted to generate a fine-tuned redraw mask. Based on multi-scale confidence representation, fine-refinement redraw mask and non-compliant areas, fine-refinement scale labels are generated. Based on the fine-refinement redraw mask, fine-refinement local surface orientation representation is extracted from the building 3D data. Then, fine-refinement orientation modulation noise is generated based on the fine-refinement local surface orientation representation and fine-refinement scale labels. Based on the refinement redraw mask, refinement original 3D representation is extracted from the building 3D data. Refinement observation condition representation is extracted from the associated observation data according to the refinement redraw mask. Refinement original 3D representation and refinement observation condition representation are associated to generate refinement redraw condition representation. Refinement redraw condition representation is limited according to refinement redraw mask and refinement scale label to generate refinement redraw guidance representation. Based on the precision direction modulation noise and precision redrawing guidance representation, an initial precision redrawing representation is generated. Based on the initial precision redrawing representation, the three-dimensional representation corresponding to the unqualified area is restored step by step to generate local precision results.

[0019] Optionally, the step of generating a refined redrawing representation based on the local refinement result; extracting the retained redrawing representation from the local redrawing representation based on the redrawing review result, and then associating the retained redrawing representation with the refined redrawing representation to generate a local redrawing result includes: Extract the qualified and unqualified areas from the redrawing and re-inspection results. Based on the qualified areas, extract the corresponding three-dimensional representations from the local redrawing representations to generate retained redrawing representations. The scope of local fine-tuning is determined based on the areas that do not meet the standards. Based on the scope of local fine-tuning, the corresponding 3D fine-tuning representation is extracted from the local fine-tuning results to generate a fine-tuning redrawing representation. Based on the qualified and unqualified areas, determine the spatial continuity between the retained redrawing representation and the refined redrawing representation, and associate the retained redrawing representation and the refined redrawing representation according to the spatial continuity to generate local redrawing results.

[0020] Optionally, the step of acquiring historical building quality labeling data, extracting historical confidence evidence and corresponding quality labels from the historical building quality labeling data, and generating an evidence contribution representation based on the correspondence between historical confidence evidence and corresponding quality labels includes: Acquire historical building quality labeling data, and extract historical confidence evidence and corresponding quality labels from the historical building quality labeling data; Based on the correspondence between historical confidence evidence and corresponding quality labels, the historical building quality label data is divided into multiple data partitions to generate a data partitioning scheme; The data partitions are selected sequentially from the data partitioning scheme to generate validation partitions. The validation partitions are excluded based on the data partitioning scheme to generate training partitions. Then, the validation partitions and training partitions are associated to generate a partition rotation relationship. Candidate evidence contribution representations are generated based on the historical confidence evidence and corresponding quality labels contained in the training partitions in the partition rotation relationship. Candidate evidence contribution representations are detected based on the verification partitions in the same partition rotation relationship to generate verification errors. Based on the degree of dispersion of verification errors corresponding to different partition rotation relationships, a fusion reliability is generated. Candidate evidence contribution characteristics are then selected based on the fusion reliability to generate evidence contribution characteristics.

[0021] The technical advantages of the technical solution provided in this application are: This application provides a method for generating partial redrawing of architectural 3D models based on spatial confidence maps. Addressing the technical shortcomings of existing methods that rely on single quality indicators during partial restoration of architectural 3D data, which can lead to biased judgments of local credibility, this method simultaneously acquires architectural 3D data and associated observational data. It then extracts multi-source confidence evidence from both data and generates a spatial confidence map based on this evidence. Compared to methods that determine the restoration area solely based on a single local indicator in the architectural 3D data, this application ensures that the spatial confidence map originates from both the architectural 3D data and associated observational information. This multi-source confidence evidence establishes a shared evaluation basis for the same architectural spatial location before the spatial confidence map is formed. Therefore, when the reflection of a local area by a particular source of data is affected by occlusion, noise, or observational conditions, the spatial confidence map is not solely determined by that single source but is formed by multiple sources of confidence evidence. This allows the subsequently regenerated architectural spatial area to be based on a more complete representation of local credibility, thereby reducing the impact of deviations in judging the local restoration range caused by anomalies in a single observational factor.

[0022] To address the technical limitation that a fixed spatial range cannot simultaneously accommodate local defects of different spatial scales, this application does not directly form a single redrawing range based on the spatial confidence map. Instead, it generates a multi-scale confidence representation based on the spatial confidence map, then generates a layered redrawing mask based on the multi-scale confidence representation, and further generates scale labels based on the multi-scale confidence representation and the layered redrawing mask. The multi-scale confidence representation allows the same architectural spatial location to reflect its local confidence level from different spatial scales. The layered redrawing mask uses this to define the architectural spatial areas that need to be regenerated in layers, and the scale labels further preserve the correspondence between different redrawing areas and their corresponding spatial scales. Compared to delineating all local defects according to a fixed neighborhood or fixed spatial range, this process allows the subsequent local redrawing process to obtain the corresponding redrawing range based on the spatial scale of the defect. This reduces the situation where large-scale local defects leave missing areas due to insufficient redrawing range, and also reduces the situation where small-scale local defects affect the surrounding original 3D data of buildings due to excessively large redrawing ranges.

[0023] To address the shortcomings of existing methods that use uniform random perturbations in local generation processes, which fail to reflect differences in building surface orientation and easily lead to inconsistencies between the generated area and the original surface morphology, this application extracts local surface orientation representations from 3D building data based on layered redrawing masks. Then, it generates orientation modulation noise based on these local surface orientation representations and scale labels. The local surface orientation representations are directly derived from the 3D building data corresponding to the area to be regenerated, establishing a spatial orientation relationship between the orientation modulation noise and the existing surface morphology of that area. The scale labels also participate in the generation of the orientation modulation noise, ensuring that it is not only constrained by the local surface orientation but also related to the spatial scale of the corresponding redrawn area. Compared to applying the same perturbation in all spatial directions, this application gives the noise entering the subsequent redrawing process directional characteristics that are compatible with the local surface orientation and the spatial scale of defects. This reduces the probability of additional morphological undulations forming in surface directions that do not require significant changes during the redrawing process, resulting in better coordination between the local redrawing representation and the geometry of the original building surface.

[0024] To address the technical shortcomings of relying solely on the data of the area to be repaired during local regeneration, which can easily lead to insufficient integration between the regenerated portion and the surrounding original building data, this application further generates a redrawing condition representation based on building 3D data and associated observation data. This representation is further defined by a layered redrawing mask and scale labels to generate a redrawing guidance representation. Finally, directional modulation noise and the redrawing guidance representation jointly generate the local redrawing representation. The redrawing condition representation allows the local regeneration process to continue utilizing information related to the area to be redrawn from the building 3D data and associated observation data. The layered redrawing mask defines the spatial area where the redrawing condition representation applies, and the scale labels further define the redrawing conditions at the corresponding spatial scale. The resulting redrawing guidance representation ensures that the local regeneration process is constrained by the redrawing area, spatial scale, and the original building data conditions. Furthermore, directional modulation noise provides initial change conditions adapted to the local surface orientation, and the redrawing guidance representation defines the association between the regenerated content and the existing building 3D data. This prevents the local redrawing representation from being solely determined by random perturbations, thereby improving the continuity between the regenerated portion and the surrounding original 3D form, and enhancing the coordination of the local structure.

[0025] To address the issue that existing partial repair schemes, once generated, are directly output, leaving residual local defects that are difficult to address specifically, while regenerating the entire scheme can easily affect technical defects in areas that have already met repair requirements, this application uses a spatial confidence map to re-examine the partial redrawing representation, generating a redrawing re-examination result. Based on this result, a partial refinement result is generated, followed by a refined redrawing representation. Simultaneously, based on the redrawing re-examination result, retained redrawing representations are extracted from the partial redrawing representations. Finally, the retained and refined redrawing representations are correlated to generate the final partial redrawing result. Therefore, the spatial confidence map not only determines the scope of partial redrawing at the front end but also re-applies to the re-examination of the partial redrawing representations, allowing the regenerated architectural space area to be reassessed based on an evaluation that aligns with the initial confidence level assessment. The redrawing re-examination result further distinguishes between areas requiring further adjustment and those that can be retained, allowing the partial refinement result to focus on areas still needing adjustment, while maintaining the existing effective redrawing content by retaining the redrawing representations. Compared to reprocessing the entire redrawn area every time residual defects are discovered, this application can reduce the morphological fluctuations caused by repeatedly generating areas that have met the repair requirements, and allow computational resources to be more concentrated on architectural space areas that still need refinement.

[0026] Therefore, this application begins by forming a spatial confidence map from multi-source confidence evidence, then defines redrawing areas at different spatial scales through multi-scale confidence representation, and combines this with local surface orientation representation to form directional modulation noise. Furthermore, it utilizes architectural 3D data, associated observation data, layered redrawing masks, and scale labels to form a redrawing guidance representation. After generating local redrawing representations, the spatial confidence map is used to generate redrawing review results, and the refined redrawing representations requiring further processing are correlated with the retained redrawing representations that do not require further processing. Each of the above technical steps uses the processing results of the previous step as the basis for the processing of the next step, maintaining a continuous technical relationship between local confidence evaluation, redrawing area delineation, initial redrawing conditions formation, local 3D content regeneration, and result review. This approach is more suitable for local defect repair scenarios in architectural 3D data, where different observation conditions, different defect spatial scales, and different surface orientations coexist. Attached Figure Description

[0027] Figure 1 This application provides an embodiment of a method for generating a local redraw of a 3D building model based on a spatial confidence map. Figure 2 This application provides an embodiment of a device for generating a partial redraw of a 3D architectural model guided by a spatial confidence map. Figure 3 An electronic device is described in an embodiment of this application; Figure 4 This application provides an embodiment of a computer-readable storage medium. Figure 5 This application provides an embodiment of a system for generating local redrawing of a 3D architectural model based on a spatial confidence map. Detailed Implementation

[0028] like Figure 1 As shown in the figure, this application provides an embodiment of a method for generating a local redraw of a 3D building model guided by a spatial confidence map, comprising: Acquire 3D building data and associated observation data, extract multi-source confidence evidence from the 3D building data and associated observation data, and generate a spatial confidence map based on the multi-source confidence evidence; Multi-scale confidence representations are generated based on spatial confidence maps, hierarchical redraw masks are generated based on multi-scale confidence representations, and scale labels are generated based on multi-scale confidence representations and hierarchical redraw masks. Local surface orientation representations are extracted from building 3D data based on hierarchical redraw mask, and orientation modulation noise is generated based on local surface orientation representations and scale labels. Redrawing condition representations are generated based on building 3D data and associated observation data. Redrawing condition representations are limited by layered redrawing masks and scale labels to generate redrawing guidance representations. Local redrawing representations are generated based on directional modulation noise and redrawing guidance representations. The process involves re-examining local redrawing representations based on spatial confidence maps to generate redrawing re-examination results; generating local refinement results based on the redrawing re-examination results; generating refined redrawing representations based on the local refinement results; extracting and retaining redrawing representations from the local redrawing representations based on the redrawing re-examination results; and then associating the retained redrawing representations with the refined redrawing representations to generate local redrawing results.

[0029] Specifically, in this application, 3D architectural data is used to represent the spatial form of the building to be partially redrawn. The 3D architectural data can be point clouds, triangular meshes, or implicit representations that allow reading the spatial form based on 3D sampling positions. Spatial locations in the 3D architectural data and observation locations in the associated observation data are established according to a unified coordinate relationship, enabling subsequent multi-source confidence evidence extracted from the 3D architectural data and associated observation data to correspond to the same architectural spatial location. Associated observation data are observation data that have a spatial correspondence with the 3D architectural data. Associated observation data includes spatial observation information corresponding to the acquisition viewpoint, depth observation data, and data related to building component identification. The 3D architectural data and associated observation data enter the subsequent confidence evaluation process according to the spatial correspondence, allowing the geometric observations, semantic recognition, cross-source observations, and acquisition time of the same architectural spatial location to jointly participate in the generation of multi-source confidence evidence.

[0030] Specifically, in this application, multi-source confidence evidence represents a multi-dimensional basis for credibility surrounding the same architectural spatial location. After the multi-source confidence evidence is generated, the local geometric observation representation, semantic recognition evidence, cross-source consistency evidence, and timeliness evidence in the multi-source confidence evidence all maintain a spatial correspondence with the corresponding architectural spatial location. This allows for the formation of location-by-location confidence evaluation inputs in the three-dimensional space corresponding to the architectural three-dimensional data. When generating a spatial confidence map based on multi-source confidence evidence, a unified evaluation is performed on the multi-source confidence evidence associated with the same architectural spatial location. Then, the comprehensive confidence level formed by the unified evaluation is written into the corresponding three-dimensional space according to the architectural spatial location to generate a spatial confidence map covering the entire spatial space of the architectural three-dimensional data. Each architectural spatial location in the spatial confidence map corresponds to a corresponding comprehensive confidence level, enabling the spatial confidence map to continue participating in the generation of multi-scale confidence representations.

[0031] Specifically, in this application, the multi-scale confidence representation is confidence information formed by expressing the same spatial confidence map at different spatial sampling resolutions. Different spatial sampling resolutions cause the same local confidence anomaly to exhibit different degrees of spatial continuity at different spatial scales. Continuous confidence anomalies with larger coverage areas remain spatially connected even after reducing the spatial sampling resolution, while confidence anomalies with smaller local areas are more likely to be preserved in fine-scale confidence maps. Based on this scale difference, a hierarchical redraw mask is generated, which then defines the redrawn regions of different spatial scales according to the multi-scale confidence representation. Scale labels further establish a correlation between the corresponding regions in the hierarchical redraw mask and the spatial scale that forms those regions, allowing the scale labels to continue participating in the definition of directional modulation noise and subsequent redrawing condition representations.

[0032] Specifically, in this application, the local surface orientation representation is generated from the 3D building data corresponding to the spatial location defined by the layered redraw mask. The local surface orientation representation describes the local normal and local tangent directions of the building surface within the mask-covered area, thus enabling the subsequently generated orientation modulation noise to distinguish between the local normal and local tangent directions of the building surface. Scale labels are simultaneously incorporated into the orientation modulation noise generation process, allowing redrawn areas of different spatial scales to develop noise spatial characteristics adapted to their corresponding spatial scales. The orientation modulation noise thus maintains its correspondence with the layered redraw mask, the local surface orientation representation, and the scale labels, and further serves as input to the local redraw representation generation process.

[0033] Specifically, in this application, the redrawing condition representation is generated from architectural 3D data and associated observation data. It is used to preserve the original 3D representation still available around the area to be redrawn, as well as the observation condition representation corresponding to the mask-covered area. Layered redrawing masks define the area of ​​effect corresponding to the redrawing condition representation, and scale labels further define the spatial focus for different mask areas to generate a redrawing guidance representation. Directional modulation noise provides the initial change conditions for the mask-covered area, and the redrawing guidance representation provides the conditional constraints between the mask-covered area and the original 3D representation. Directional modulation noise and the redrawing guidance representation jointly participate in the generation of local redrawing representations, enabling the local redrawing representation to continue utilizing the original 3D representation corresponding to the area outside the mask and the observation condition representation corresponding to the mask-covered area when restoring the 3D representation of the mask-covered area.

[0034] Specifically, in this application, the spatial confidence map is also used to review the local redrawing representation. After the local redrawing representation is generated, the corresponding confidence information is read from the spatial confidence map based on the building spatial location corresponding to the local redrawing representation, and combined with the redrawing result of the local redrawing representation to form the redrawing review result. The redrawing review result is used to distinguish between areas that still need further processing and areas that have met the current redrawing requirements. Areas that still need further processing are used to generate local refinement results, and then refinement redrawing representations are generated from the local refinement results; areas that have met the current redrawing requirements are used to extract and retain redrawing representations from the local redrawing representations. By associating the retained redrawing representations and the refinement redrawing representations to form local redrawing results, areas that still need further processing receive further processing, while avoiding repeated changes to areas that have already formed valid redrawing content.

[0035] Optionally, the extraction of multi-source confidence evidence from building 3D data and associated observation data includes: Local geometric observation representations are extracted from architectural 3D data and associated observation data, and evidence areas are delineated based on these local geometric observation representations. Extract semantic recognition evidence from associated observation data according to evidence regions; Read the data source identifier corresponding to the evidence area from the associated observation data, extract the cross-source observation results corresponding to the evidence area based on the data source identifier, and then compare the cross-source observation results to generate cross-source consistent evidence; Read the collection time identifier corresponding to the evidence area from the associated observation data, generate the collection time interval based on the collection time identifier, and generate time-sensitive evidence based on the collection time interval; Multi-source confidence evidence is generated based on local geometric observation representations, semantic recognition evidence, cross-source consistent evidence, and time-sensitive evidence.

[0036] Specifically, in this application, local geometric observation representation is used to reflect the sufficiency of sampling, surface orientation stability, and depth observation quality of architectural 3D data and associated observation data within a local spatial range. First, a spatial neighborhood is established according to the 3D sampling positions in the architectural 3D data. Then, associated observation data corresponding to the 3D sampling positions is read according to the spatial neighborhood, ensuring that the spatial morphological information in the architectural 3D data and the depth observation data in the associated observation data can correspond to the same spatial neighborhood. After forming a local geometric observation representation based on the architectural 3D data and associated observation data within the same spatial neighborhood, the evidence region is divided according to the correspondence between the local geometric observation representation and the 3D sampling positions, ensuring a continuous spatial correspondence among the local geometric observation representations within the evidence region. The evidence region thus becomes a common processing area for spatial alignment of semantic recognition evidence, cross-source consistency evidence, and time-sensitive evidence.

[0037] Specifically, in this application, semantic recognition evidence originates from the building component categories and corresponding confidence levels in the associated observation data that correspond to the evidence region. After retrieving the building component categories and corresponding confidence levels from the associated observation data according to the evidence region, semantic recognition evidence is generated based on the building component categories and corresponding confidence levels. The semantic recognition evidence further establishes a spatial correspondence with the evidence region, enabling the comparison of the credibility of the same evidence region reflected in local geometric observation representation and semantic recognition evidence when generating multi-source confidence evidence subsequently.

[0038] Specifically, in this application, data source identifiers are used to distinguish different observation sources corresponding to the same evidence area. The same evidence area can simultaneously correspond to 3D building data generated by laser scanning and 3D building data generated by image reconstruction. The data source identifiers corresponding to each observation source are recorded in the associated observation data. After extracting the cross-source observation results corresponding to the evidence area based on the data source identifiers, a correspondence is established between the cross-source observation results from different data sources according to the spatial location of the buildings in the evidence area. Then, the geometric differences between the cross-source observation results in the corresponding spatial locations of the buildings are compared to generate cross-source consistent evidence. The closer the spatial morphology between the cross-source observation results, the higher the degree of consistency between the sources reflected by the cross-source consistent evidence; when there are continuous spatial deviations between the cross-source observation results, the cross-source consistent evidence reflects differences between the sources in the corresponding evidence area. After the cross-source consistent evidence is generated, a spatial correspondence is further established with the corresponding evidence area to participate in the generation of multi-source confidence evidence.

[0039] Specifically, in this application, the collection time identifier is used to describe the collection time of the associated observation data corresponding to the evidence region. In cases where multiple collection results exist for the same evidence region, the collection time identifier corresponding to each collection is read from the associated observation data, and a collection time interval is generated based on the time difference between the collection time identifiers. Time-sensitive evidence is generated based on the collection time interval, ensuring that associated observation data with larger collection time intervals contribute to the subsequent generation of the spatial confidence map to a degree corresponding to the collection time interval. Time-sensitive evidence does not solely determine whether the evidence region needs to be redrawn; rather, it participates in the generation of multi-source confidence evidence along with local geometric observation representations, semantic recognition evidence, and cross-source consistency evidence.

[0040] Furthermore, local geometric observation representations, semantic recognition evidence, cross-source consistent evidence, and time-sensitive evidence all maintain spatial correspondence according to the evidence region. After forming local geometric observation representations, semantic recognition evidence, cross-source consistent evidence, and time-sensitive evidence for the same evidence region, multi-source confidence evidence corresponding to that evidence region is generated based on these representations. Different evidence regions generate their own multi-source confidence evidence. Thus, multi-source confidence evidence retains both the differences between evidence from different sources and the spatial correspondence with the building's spatial location in the 3D building data, thereby continuing to participate in the generation of the spatial confidence map.

[0041] Optionally, the step of extracting local geometric observation representations from building 3D data and associated observation data, and dividing the evidence region based on the local geometric observation representations, includes: Read the 3D sampling positions from the 3D building data, and generate a spatial neighborhood based on the 3D sampling positions; Based on the neighborhood distribution of the three-dimensional sampling locations in the spatial neighborhood, sampling sufficiency is generated; A neighborhood normal representation is generated based on the 3D sampling position in the spatial neighborhood, and a local orientation consistency is generated based on the directional dispersion of the neighborhood normal representation. Depth observation data is read from associated observation data according to spatial neighborhood, and observation signal-to-noise representation is generated based on the depth observation data; Local geometric observation representations are generated based on sampling sufficiency, local orientation consistency, and observation signal-to-noise ratio. Evidence regions are then divided based on the correspondence between local geometric observation representations and three-dimensional sampling positions.

[0042] Specifically, in this application, the three-dimensional sampling position represents a discrete position in the architectural three-dimensional data where spatial coordinates can be determined. When the architectural three-dimensional data is represented using a point cloud, the three-dimensional sampling position corresponds to a spatial point in the point cloud; when the architectural three-dimensional data is represented using a triangular mesh, the three-dimensional sampling position can correspond to a sampling position at a mesh vertex or mesh surface; when the architectural three-dimensional data uses an implicit representation, the three-dimensional sampling position corresponds to a discrete query position where the corresponding spatial form can be read. By uniformly using the three-dimensional sampling position to describe the spatial position in different forms of architectural three-dimensional data, the spatial neighborhood, sampling sufficiency, neighborhood normal representation, and local orientation consistency can all form a continuous technical relationship based on the three-dimensional sampling position.

[0043] Specifically, in this application, a spatial neighborhood is generated by reading other spatially adjacent 3D sampling locations centered on a 3D sampling location. The spatial extent of the neighborhood is configured according to the sampling interval of the building's 3D data, such that the neighborhood includes multiple 3D sampling locations capable of describing the local surface morphology around the central 3D sampling location. For example, if the building's 3D data uses indoor laser point clouds and the average spacing between adjacent 3D sampling locations is on the order of centimeters, a spatial extent covering several times the average spacing can be configured as the spatial neighborhood. This example is only for illustrating the formation of the spatial neighborhood; the actual spatial extent is configured based on the sampling density of the building's 3D data. After the spatial neighborhood is generated, the 3D sampling locations within it simultaneously participate in the generation of sampling sufficiency and neighborhood normal representation.

[0044] Specifically, in this application, sampling adequacy is used to describe the degree to which 3D building data covers local building surfaces within a spatial neighborhood. When statistically analyzing the neighborhood distribution of 3D sampling locations, the number of 3D sampling locations within the spatial neighborhood is counted, and the sampling adequacy is determined by combining this count with the continuity of the distribution of 3D sampling locations within the spatial neighborhood. Sampling adequacy is reduced when the number of 3D sampling locations is small or when there are continuous gaps in the 3D sampling locations within the spatial neighborhood; conversely, sampling adequacy is increased when the 3D sampling locations are continuously distributed within the coverage area of ​​the spatial neighborhood. After being generated, sampling adequacy continues to serve as an input for local geometric observation characterization to reflect insufficient local data caused by occlusion or observation blind spots.

[0045] Specifically, in this application, the neighborhood normal representation is used to describe the local surface orientation distribution within a spatial neighborhood. A local surface is fitted based on the 3D sampling positions within the spatial neighborhood, and then the neighborhood normal representation is generated from the normal directions of the local surface. For locally continuous building surfaces such as walls and floors, the normal directions in the neighborhood normal representation are typically concentrated in similar directions; for local areas affected by reflective noise or anomalous sampling, the distribution of normal directions in the neighborhood normal representation is more discrete. Local orientation consistency is generated based on the degree of directional dispersion of the neighborhood normal representation, allowing the local orientation consistency to continue participating in the generation of the local geometric observation representation.

[0046] Specifically, in this application, the observation signal-to-noise (SNR) representation is generated from depth observation data in the associated observation data. Depth observation data corresponding to the 3D sampling position is read according to the spatial neighborhood. Then, the deviation of adjacent depth observation data on the same local surface from the local surface change trend is compared to distinguish between depth observation components that continuously change with the local surface and those that deviate from the local surface change trend. The observation SNR representation is generated based on the relative relationships between these depth observation components. The observation SNR representation reflects the degree to which the depth observation data corresponding to the spatial neighborhood is affected by observation noise and continues to participate in the generation of the local geometric observation representation.

[0047] Furthermore, based on sampling sufficiency, local orientation consistency, and observation signal-to-noise ratio, local geometric observation representations corresponding to the same spatial neighborhood are generated, enabling these representations to simultaneously reflect local sampling coverage, surface orientation stability, and depth observation quality. According to the correspondence between local geometric observation representations and 3D sampling positions, spatially adjacent 3D sampling positions with continuously changing local geometric observation representations are grouped into the same evidence region to form an evidence region. This evidence region continues to serve as the spatial basis for semantic recognition evidence, cross-source consistency evidence, and timeliness evidence.

[0048] In one embodiment, taking the 3D architectural data of an existing office building lobby as an example, the number of 3D sampling locations in the spatial neighborhood corresponding to the area obscured by suspended components is lower than that in the unobstructed wall area, resulting in reduced sampling sufficiency. The directional dispersion of the neighborhood normal representation corresponding to the reflective area of ​​the glass curtain wall is higher than that of ordinary walls, resulting in reduced local directional consistency. When the reflection of the glass curtain wall also causes fluctuations in depth observation data, the observation signal-to-noise representation further reflects the changes in the depth observation quality of the corresponding spatial neighborhood. Sampling sufficiency, local directional consistency, and observation signal-to-noise representation jointly participate in the generation of local geometric observation representation, enabling local anomalies caused by different reasons to enter the same multi-source confidence evidence generation process.

[0049] Optionally, the generation of a spatial confidence map based on multi-source confidence evidence includes: Acquire historical building quality labeling data, extract historical confidence evidence and corresponding quality labels from the historical building quality labeling data, and generate evidence contribution representation based on the correspondence between historical confidence evidence and corresponding quality labels; Multi-source confidence evidence is modulated based on the contribution characteristics of the evidence to generate contribution-modulated evidence; Based on contribution modulation evidence, a statistical representation of the evidence neighborhood is generated, and then an initial spatial confidence map is generated based on the contribution modulation evidence and the statistical representation of the evidence neighborhood. Confidence boundary gradients are generated based on the initial spatial confidence map, and the initial spatial confidence map is smoothed by maintaining the boundaries based on the confidence boundary gradients to generate a spatial confidence map.

[0050] Specifically, in this application, the historical building quality labeling data is formed from the 3D building data with existing local quality assessment results and its associated observation data. The historical confidence evidence in the historical building quality labeling data and the multi-source confidence evidence corresponding to the building to be redrawn currently use the same evidentiary meaning, enabling the historical confidence evidence to reflect the consistency of local geometric observations, semantic recognition, and cross-source observations; the corresponding quality label is used to indicate the data quality of the building's spatial location corresponding to the historical confidence evidence. An evidence contribution characterization is generated based on the correspondence between historical confidence evidence and corresponding quality labels, allowing the evidence contribution characterization to describe the contribution relationship of different types of historical confidence evidence to the judgment of building spatial quality. After the evidence contribution characterization is generated, it continues to apply to the multi-source confidence evidence corresponding to the current 3D building data, thereby avoiding the direct combination of evidence from different sources according to a fixed contribution ratio.

[0051] Specifically, in this application, before adjusting multi-source confidence evidence based on evidence contribution characteristics, the numerical expression direction and range of different pieces of evidence in the multi-source confidence evidence are first unified. For evidence where an increase in value indicates improved credibility, its numerical expression direction is maintained; for evidence where an increase in value indicates improved anomaly, it is converted to an expression direction corresponding to improved credibility. Then, scale mapping is performed according to the effective numerical range of the corresponding type of historical confidence evidence in the historical building quality annotation data, so that the scale-mapped multi-source confidence evidence adopts a consistent credibility expression direction. Subsequently, the scale-mapped multi-source confidence evidence is adjusted based on evidence contribution characteristics to generate contribution-modulated evidence. Time-sensitive evidence acts on evidence with different collection time intervals in the above contribution adjustment process, causing the contribution of evidence with larger collection time intervals to decrease accordingly in the contribution-modulated evidence.

[0052] Specifically, in this application, the evidence neighborhood statistical representation reflects the distribution of contributing modulated evidence in adjacent evidence regions surrounding an evidence region. Centered on the current evidence region, contributing modulated evidence from adjacent evidence regions is read according to the spatial proximity relationships in the architectural 3D data. Then, the evidence neighborhood statistical representation is generated based on the local variation and distribution trend of various types of contributing modulated evidence in the adjacent evidence regions. The evidence neighborhood statistical representation and the contributing modulated evidence of the current evidence region jointly participate in the generation of the initial spatial confidence map, ensuring that the overall confidence level corresponding to the current evidence region is simultaneously influenced by the spatial variation relationships between the contributing modulated evidence of the current evidence region itself and the contributing modulated evidence of surrounding evidence regions.

[0053] In a preferred embodiment, an initial spatial confidence map is generated using evidence fusion mapping. The evidence fusion mapping is formed based on the correspondence between historical confidence evidence and corresponding quality labels in historical building quality labeling data, and converts the contribution modulation evidence and evidence neighborhood statistical representation corresponding to the current building spatial location into a comprehensive confidence level. For any three-dimensional sampling location in the building's three-dimensional space, the comprehensive confidence level satisfies the following formula:

[0054] Wherein, x represents the 3D sampling position in the 3D space of the building; c(x) represents the overall confidence level corresponding to the 3D sampling position x, which is expressed as a dimensionless numerical value; e_g(x) represents the contribution modulation geometric evidence corresponding to the 3D sampling position x, which is formed by adjusting the local geometric observation representation through evidence contribution representation and time-sensitive evidence, and is expressed as a dimensionless confidence level; e_s(x) represents the contribution modulation semantic recognition evidence corresponding to the 3D sampling position x, which is formed by adjusting the semantic recognition evidence through evidence contribution representation and time-sensitive evidence, and is expressed as a dimensionless confidence level; e_c(x) represents the contribution modulation cross-source consistency evidence corresponding to the 3D sampling position x, which is formed by adjusting the cross-source consistency evidence through evidence contribution representation and time-sensitive evidence, and is expressed as a dimensionless confidence level; s(x) represents the evidence neighborhood statistical representation corresponding to the evidence region where the 3D sampling position x is located, which is formed by the contribution modulation evidence of adjacent evidence regions; F represents the evidence fusion mapping formed based on the correspondence between historical confidence evidence and corresponding quality labels. Since e_g(x), e_s(x), and e_c(x) have already unified their confidence expression directions and numerical ranges before entering the evidence fusion map F, and s(x) is also formed by the unified contribution modulation evidence, there is no problem of inconsistent physical dimensions in the input processed by the evidence fusion map F. The comprehensive confidence level c(x) corresponding to each three-dimensional sampling position x is written into the three-dimensional architectural space according to the three-dimensional sampling position x to generate an initial spatial confidence map.

[0055] Specifically, in this application, the initial spatial confidence map still retains confidence fluctuations caused by local observation noise between adjacent evidence regions. Therefore, a confidence boundary gradient is further generated based on the initial spatial confidence map. The confidence boundary gradient is used to describe the spatial variation of the overall confidence level between adjacent spatial locations. Spatial locations with gentle changes in overall confidence level correspond to smaller confidence boundary gradients, while locations with continuous changes in overall confidence level correspond to larger confidence boundary gradients. When smoothing the initial spatial confidence map based on the confidence boundary gradient, the smoothing effect between adjacent overall confidence levels is enhanced in spatial ranges with smaller confidence boundary gradients, while the smoothing effect across locations with larger confidence boundary gradients is reduced. Thus, while reducing local confidence fluctuations caused by single-point observation noise, the confidence boundary corresponding to the actual quality change location is still retained, ultimately generating a spatial confidence map.

[0056] Furthermore, each spatial location in the spatial confidence map maintains its spatial correspondence with the 3D sampling location in the building's 3D data. The spatial confidence map is not a separate evaluation data point detached from the building's 3D space, but rather a location-by-location confidence representation covering the entire space of the building's 3D data. This spatial correspondence allows the spatial confidence map to directly enter the multi-scale confidence representation generation process, and enables fine-scale, mesoscale, and coarse-scale confidence maps at different spatial sampling resolutions to be traced back to the same architectural spatial location.

[0057] Optionally, the generation of multi-scale confidence representations based on spatial confidence maps includes: Maintain the spatial sampling resolution of the spatial confidence map to generate a fine-scale confidence map; Reduce spatial sampling resolution based on fine-scale confidence maps to generate mesoscale confidence maps; Based on the mesoscale confidence map, the spatial sampling resolution is further reduced to generate a coarse-scale confidence map; Scale location correspondence is generated based on the spatial sampling resolution correspondence between fine-scale confidence maps, mesoscale confidence maps, and coarse-scale confidence maps. The fine-scale confidence maps, mesoscale confidence maps, and coarse-scale confidence maps are then associated based on the scale location correspondence to generate multi-scale confidence representations.

[0058] Specifically, in this application, the fine-scale confidence map maintains the original spatial sampling resolution of the spatial confidence map, allowing it to retain local confidence variations and architectural detail boundaries. After the fine-scale confidence map is generated, the comprehensive confidence level corresponding to the mesoscale spatial location is generated based on the comprehensive confidence level corresponding to adjacent fine-scale spatial locations, thereby reducing the spatial sampling resolution and generating the mesoscale confidence map. The mesoscale confidence map further reduces its spatial sampling resolution according to the corresponding spatial aggregation relationship to generate the coarse-scale confidence map. By adopting a stepwise reduction in spatial sampling resolution, a continuous spatial source relationship is maintained between the fine-scale, mesoscale, and coarse-scale confidence maps, rather than regenerating independent confidence results from the architectural 3D data separately.

[0059] Specifically, in this application, the fine-scale confidence map mainly retains confidence anomalies with small local areas and confidence changes near the building surface boundaries; the mesoscale confidence map is used to describe confidence changes between local details and large-scale structures; and the coarse-scale confidence map, by reducing the spatial sampling resolution, weakens the influence of isolated local fluctuations, ensuring that confidence anomalies continuously covering a large spatial area still maintain spatial connectivity. The fine-scale, mesoscale, and coarse-scale confidence maps are formed hierarchically from the same spatial confidence map; therefore, they can describe the same building spatial location from different spatial scales.

[0060] In one embodiment, the spatial sampling resolution of the coarse-scale confidence map can be configured to be a fraction (e.g., one-quarter to one-eighth) of the spatial sampling resolution of the spatial confidence map, while the spatial sampling resolution of the meso-scale confidence map is configured between that of the fine-scale and coarse-scale confidence maps. These specific proportions illustrate one implementation of multi-scale confidence representation; the actual spatial sampling resolution is configured based on the original sampling density of the building's 3D data and the spatial extent of the local defects to be identified.

[0061] Specifically, in this application, each time the spatial sampling resolution is reduced, the source correspondence between the lower-level spatial location and the higher-level spatial location is recorded simultaneously. A scale location correspondence is generated from the spatial sampling resolution correspondence between the fine-scale confidence map, the mesoscale confidence map, and the coarse-scale confidence map. The scale location correspondence records at least one mesoscale spatial location corresponding to a coarse-scale spatial location, as well as the corresponding fine-scale spatial location, enabling the spatial region identified by the coarse-scale confidence map to be mapped back to the actual architectural space range in the fine-scale confidence map according to the scale location correspondence. The fine-scale confidence map, the mesoscale confidence map, and the coarse-scale confidence map are associated based on the scale location correspondence to generate a multi-scale confidence representation. This allows subsequent candidate regions for structural defects and candidate regions for detailed defects identified at different scales to generate layered redraw masks under the same architectural spatial coordinate relationship.

[0062] Optionally, the step of generating a hierarchical redraw mask based on multi-scale confidence representation, and then generating scale labels based on the multi-scale confidence representation and the hierarchical redraw mask, includes: Based on the coarse-scale confidence map, continuously distributed low-confidence regions are identified, and candidate regions for structural defects are generated. Based on the candidate regions of structural defects, the spatial regions corresponding to the candidate regions of structural defects are excluded from the fine-scale confidence map to generate a fine-scale residual confidence map. Then, based on the fine-scale residual confidence map, isolated low-confidence regions are identified to generate candidate regions of detailed defects. Based on the mesoscale confidence map, candidate regions for structural defects and candidate regions for detailed defects are detected respectively, and scale verification results are generated. Based on the scale verification results, candidate regions for structural defects and candidate regions for detailed defects are selected to generate hierarchical candidate regions; Based on the layered candidate regions, boundary expansion results are generated. Based on the boundary expansion results, a layered redrawing mask is generated. Then, based on the scale position correspondence, the layered candidate regions are associated with the corresponding regions in the layered redrawing mask to generate scale labels.

[0063] Specifically, in this application, before generating candidate regions for structural defects and candidate regions for detailed defects, coarse-scale confidence boundaries and fine-scale confidence boundaries adapted to the corresponding spatial sampling resolution are configured for the coarse-scale confidence map and the fine-scale confidence map, respectively. The coarse-scale and fine-scale confidence boundaries are determined based on the comprehensive confidence level distribution corresponding to local defect areas and normal areas in historical building quality annotation data, ensuring that the coarse-scale confidence boundaries are adapted to the identification of continuous spatial anomalies, and the fine-scale confidence boundaries are adapted to the identification of locally isolated anomalies. In one embodiment, the coarse-scale and fine-scale confidence boundaries can be configured in conjunction with the quality of the building's 3D data (e.g., the coarse-scale confidence boundary is set to 0.35, and the fine-scale confidence boundary is set to 0.25).

[0064] Specifically, in this application, when identifying continuously distributed low-confidence regions based on a coarse-scale confidence map, low-confidence spatial locations in the coarse-scale confidence map are first screened based on the coarse-scale confidence judgment boundary. Then, adjacent low-confidence spatial locations are connected based on the spatial adjacency relationship between coarse-scale spatial locations. Finally, the connected low-confidence spatial locations are mapped to the corresponding spatial range of the building's three-dimensional data through the scale-location correspondence to generate candidate regions for structural defects. Since the candidate regions for structural defects are formed by continuous low-confidence spatial locations in the coarse-scale confidence map, they are used to describe local defects that cover a large spatial range and have spatial connectivity.

[0065] Furthermore, based on the correspondence between structural defect candidate regions and their scale locations, the corresponding spatial regions in the fine-scale confidence map are determined. The spatial regions corresponding to structural defect candidate regions are then excluded from the fine-scale confidence map to generate a fine-scale residual confidence map. This process ensures that spatial locations already identified as structural defect candidate regions in the coarse-scale confidence map are no longer repeatedly identified as detailed defect candidate regions. In the fine-scale residual confidence map, low-confidence spatial locations are filtered based on the fine-scale confidence decision boundary. Then, low-confidence regions with small spatial extents that do not form large-scale continuous connections with other low-confidence spatial locations are identified according to spatial adjacency relationships to generate detailed defect candidate regions. Structural defect candidate regions and detailed defect candidate regions thus correspond to confidence anomalies of different spatial scales.

[0066] In a preferred embodiment, to further illustrate the formation process of candidate regions for structural defects, the intermediate mask generated from the coarse-scale confidence map can be denoted as the coarse-scale mask, which satisfies the following equation:

[0067] Where x represents the coarse-scale spatial location in the coarse-scale confidence map; M_coarse represents the coarse-scale mask, which is a binary spatial mask formed by coarse-scale spatial locations that satisfy the coarse-scale low-confidence condition and are spatially continuous, and has no physical dimensions; c_coarse(x) represents the coarse-scale comprehensive confidence level corresponding to the coarse-scale spatial location x, which is expressed using a dimensionless numerical value; τ_coarse represents the coarse-scale confidence decision boundary, which is expressed using the same dimensionless value as the coarse-scale comprehensive confidence level; 1[·] represents the indicator operation, which outputs the corresponding mask mark when the low-confidence condition in the parentheses is met, and does not output the corresponding mask mark when the low-confidence condition is not met; CC represents the connected component extraction operation, which is used to connect adjacent mask marks into a continuous region based on the spatial adjacency relationship between coarse-scale spatial locations. After being mapped to the three-dimensional building data space through the scale location correspondence, M_coarse is used to generate candidate regions for structural defects. Since both c_coarse(x) and τ_coarse are dimensionless confidence quantities, there is no dimension inconsistency in the comparison between the two.

[0068] Accordingly, to further illustrate the formation process of candidate regions for detailed defects, the intermediate mask formed by the fine-scale residual confidence map can be denoted as the fine-scale mask, which satisfies the following equation:

[0069] Where x represents the fine-scale spatial location in the fine-scale confidence map; M_fine represents the fine-scale mask, which is a binary space mask formed by low-confidence spatial locations identified from the fine-scale confidence map and located outside the spatial region corresponding to the coarse-scale mask, and has no physical dimension; c_fine(x) represents the fine-scale comprehensive confidence level corresponding to the fine-scale spatial location x, which is expressed using a dimensionless numerical value; τ_fine represents the fine-scale confidence decision boundary, which is expressed using the same dimensionless value as the fine-scale comprehensive confidence level; 1[·] represents the indicator operation; M_coarse represents the aforementioned coarse-scale mask; the horizontal line above M_coarse represents the complement region of the spatial region corresponding to the coarse-scale mask; ∩ represents the overlap relationship of the spatial regions. Thus, M_fine only retains the fine-scale spatial locations that meet the fine-scale low-confidence condition and do not fall into the spatial region corresponding to M_coarse, and the isolated low-confidence regions in M_fine are used to generate candidate regions for detailed defects. Since both c_fine(x) and τ_fine are dimensionless confidence quantities, there is no dimension inconsistency in the comparison between the two.

[0070] Specifically, in this application, the mesoscale confidence map is used to detect the confidence performance of structural defect candidate regions and detailed defect candidate regions at intermediate spatial scales to generate scale verification results. The corresponding positions of structural defect candidate regions and detailed defect candidate regions in the mesoscale confidence map are determined through scale location correspondence, and then the comprehensive confidence level of the corresponding position and its neighboring positions is read. When there is an inconsistency between the confidence judgment formed by the mesoscale confidence map for a certain structural defect candidate region and the coarse-scale and fine-scale confidence maps, the corresponding spatial position and the source of inconsistency are recorded in the scale verification results. When the coarse-scale and fine-scale confidence maps form a consistent low-confidence judgment at the same spatial position, while the mesoscale confidence map forms a different judgment, the scale verification results retain the consistency relationship between the coarse-scale and fine-scale confidence maps as a basis for subsequent screening. The scale verification results thus continue to participate in the generation of hierarchical candidate regions.

[0071] After screening candidate regions for structural defects and detailed defects based on the scale verification results, hierarchical candidate regions are generated. These hierarchical candidate regions retain the spatial scale attributes from which they originated. Subsequently, boundary expansion results are generated based on the spatial scale corresponding to each hierarchical candidate region. For hierarchical candidate regions originating from structural defect candidate regions, the boundary expansion results cover the surrounding space continuously related to the structural defect; for hierarchical candidate regions originating from detailed defect candidate regions, the boundary expansion results are limited to the vicinity of local anomalies to reduce the impact on the 3D data of adjacent normal buildings.

[0072] Furthermore, the layered redraw mask can be represented by the following formula:

[0073] Where M represents the layered redraw mask, which is a binary spatial mask formed by merging redraw regions of different spatial scales and has no physical dimension; M_coarse represents the aforementioned coarse-scale mask; M_fine represents the aforementioned fine-scale mask; D_{r_c}(·) represents the spatial boundary expansion operation performed according to the boundary expansion distance r_c corresponding to the structural defect candidate region; r_c represents the boundary expansion distance corresponding to the structural defect candidate region, whose physical dimension is length, and uses the same length unit as the spatial coordinates in the building 3D data; D_{r_f}(·) represents the spatial boundary expansion operation performed according to the boundary expansion distance r_f corresponding to the detailed defect candidate region; r_f represents the boundary expansion distance corresponding to the detailed defect candidate region, whose physical dimension is also length, and uses the same length unit as the spatial coordinates in the building 3D data; ∪ represents the merging relationship of two spatial mask regions. D_{r_c}(M_coarse) generates the boundary expansion results of candidate regions for structural defects, and D_{r_f}(M_fine) generates the boundary expansion results of candidate regions for detailed defects. The two are spatially merged to generate a layered redraw mask M. This formula only performs boundary expansion and region merging at the spatial mask level and does not involve direct numerical calculations between different physical dimensions.

[0074] In one embodiment, for continuous voids in the lobby of an office building caused by suspended components, a larger boundary extension range can be used to form the boundary extension result corresponding to the candidate area of ​​structural defects (e.g., a boundary extension distance of 0.3 meters); for local detail anomalies caused by reflective interference at the edge of a glass curtain wall, a smaller boundary extension range can be used to form the boundary extension result corresponding to the candidate area of ​​detail defects (e.g., a boundary extension distance of 0.05 meters). The specific values ​​mentioned above are only used to illustrate the boundary extension method at different spatial scales; the actual boundary extension distance depends on the sampling interval of the building's three-dimensional data and the spatial size configuration of the corresponding layered candidate areas.

[0075] Specifically, in this application, after establishing a correlation between the layered candidate regions and the corresponding regions in the layered redraw mask based on the scale position correspondence, a scale label is generated for each region in the layered redraw mask. The scale label records whether the corresponding mask region originates from a structural defect candidate region or a detailed defect candidate region, and continues to maintain its correlation with its spatial position in the layered redraw mask. The scale label can thus directly participate in the generation of directional modulation noise after local surface orientation characterization, enabling subsequent processing to select appropriate noise spatial characteristics based on the defect spatial scale corresponding to the mask region.

[0076] Optionally, the extraction of local surface orientation representations from architectural 3D data based on layered redraw masks includes: The mask coverage area is located by redrawing the mask layer by layer, and the local sampling position is read from the 3D building data based on the mask coverage area; Local surface adjacency relationships are generated based on local sampling locations, and local normal directions are generated based on these local surface adjacency relationships. Based on the local normal direction, the local tangent direction orthogonal to the local normal direction is determined, and then the local normal direction and the local tangent direction are associated to generate a local surface orientation representation.

[0077] Specifically, in this application, the mask coverage area is the actual spatial extent of the layered redraw mask in the architectural 3D data. After locating the mask coverage area according to the spatial correspondence between the layered redraw mask and the architectural 3D data, 3D sampling positions located within the mask coverage area and adjacent to its boundary are read from the architectural 3D data to generate local sampling positions. The 3D sampling positions adjacent to the mask coverage area boundary are also read, ensuring that the local sampling positions include not only the residual 3D sampling positions within the defect area but also the 3D sampling positions around the mask coverage area that can still be used to describe the orientation of the architectural surface. After generation, the local sampling positions continue to participate in the generation of local surface adjacency relationships.

[0078] Specifically, in this application, for each local sampling location, adjacent local sampling locations are connected according to the spatial distance between them and the continuity of the building surface to generate local surface adjacency relationships. The local surface adjacency relationship records which surrounding local sampling locations jointly describe the same local building surface. For local sampling locations with significant spatial breaks or obvious surface bends, local surface adjacency relationships are established without crossing these breaks or bends, thereby reducing the influence of adjacent different building surfaces on the local normal direction generation process. After the local surface adjacency relationship is generated, it continues to participate in the generation of the local normal direction.

[0079] When generating local normal directions based on local surface adjacency relationships, the current local sampling position and other local sampling positions that are adjacent to it are read. A local surface is fitted based on the distribution of these local sampling positions in 3D space, and then the local normal direction corresponding to the current local sampling position is generated based on the normal direction of the local surface. If local noise still exists at the local sampling position, the local surface is determined based on multiple adjacent local sampling positions, and then the local normal direction is determined based on the directional consistency between adjacent local surfaces, ensuring that isolated, anomalous 3D sampling positions do not independently determine the local normal direction. After generation, the local normal direction continues to participate in the determination of the local tangent direction.

[0080] Specifically, in this application, after obtaining the local normal direction, a local tangent direction is determined in a plane orthogonal to the local normal direction. The local tangent direction includes two mutually orthogonal surface extension directions, both of which are orthogonal to the local normal direction. For local areas with a clear building surface extension direction, the direction in which the local sampling location changes significantly within a plane orthogonal to the local normal direction can be determined as a local tangent direction. Then, another local tangent direction orthogonal to this local tangent direction is determined within the same plane. Thus, the same local sampling location has one local normal direction and two mutually orthogonal local tangent directions, which continue to jointly participate in the generation of the local surface orientation representation.

[0081] Furthermore, the local normal direction and local tangent direction corresponding to the same local sampling location are associated according to their spatial positions in the layered redraw mask to generate a local surface orientation representation. Each local sampling location in the local surface orientation representation has a corresponding local normal direction and local tangent direction, and continues to maintain the spatial correspondence with the scale label. Thus, subsequent orientation modulation noise can distinguish the local normal direction and local tangent direction based on the local surface orientation representation corresponding to the same local sampling location, and at the same time distinguish the defect scale of the corresponding local spatial region based on the scale label.

[0082] In a preferred embodiment, when the local sampling position is located in a candidate region of a structural defect with a large degree of missing information, resulting in insufficient local sampling positions within the mask coverage area to stably generate local normal directions, continuous local sampling positions outside the boundary of the mask coverage area and on the same building surface are read along the local surface adjacency relationship. Then, the local normal directions corresponding to the local sampling positions outside the boundary of the mask coverage area are continuously propagated into the mask coverage area to supplement the local normal directions corresponding to the missing positions. The corresponding local tangent direction is determined based on the supplemented local normal directions. The local normal directions used for propagation originate from the 3D building data adjacent to the same building surface, enabling large-scale missing areas to still form local surface direction representations that are continuous with the surrounding original building surface directions, and allowing the local surface direction representations to continue to participate in the generation of directional modulation noise.

[0083] Optionally, generating directional modulation noise based on local surface orientation characterization and scale labels includes: Local normal direction and local tangent direction are extracted from the local surface orientation representation. Normal direction noise component is generated based on local normal direction and tangent direction noise component is generated based on local tangent direction. Normal direction noise component and tangent direction noise component are associated according to the relationship that the noise intensity of normal direction noise component is less than that of tangent direction noise component to generate initial orientation noise. Based on the scale labels, the initial directional noise is adjusted according to the relationship that the spatial frequency corresponding to the candidate region of structural defects is less than that corresponding to the candidate region of detailed defects, thus generating scale-modulated noise. Based on scale labels, adjacent mask regions corresponding to different scale labels are identified in the layered redraw mask, and scale boundary positions are generated. Transition modulation representations are generated based on the relative positions of sampling points corresponding to scale boundary positions between adjacent mask regions corresponding to different scale labels. Scale modulation noise is adjusted based on transition modulation representations to generate directional modulation noise.

[0084] Specifically, in this application, directional modulation noise is used to define the noise intensity and spatial frequency in different spatial directions within the mask coverage area according to local surface orientation characterization and scale labels. After locating the mask coverage area by redrawing the mask layer by layer, the local normal direction and local tangent direction corresponding to the same local sampling position are read, ensuring a one-to-one correspondence between the local normal direction and local tangent direction and the local sampling position. The local normal direction is used to describe the local vertical direction of the building surface outward or inward, and the local tangent direction is used to describe two mutually orthogonal directions extending along the building surface. The aforementioned local surface orientation characterization continues to participate in the generation of the normal direction noise component and the tangent direction noise component, enabling the directional modulation noise to adopt different noise intensities in different directions of the building surface.

[0085] In a preferred embodiment, for any local sampling location within the mask coverage area Directional modulation noise can be expressed as:

[0086] in, This indicates a local sampling location within the area covered by the layered redraw mask; Indicates the local sampling location The corresponding directional modulation noise has the same length dimension as the spatial coordinates when the building 3D data is expressed in spatial coordinates. Indicates the local sampling location The corresponding local normal direction, and These represent the local sampling positions. The corresponding two local tangent directions, , and All are unit direction vectors and have no physical dimensions; This represents a dimensionless noise scalar read along the local normal direction. and These represent the dimensionless noise scalars read along the two local tangent directions, respectively. Indicates the noise intensity coefficient in the normal direction. This represents the noise intensity coefficient in the tangential direction. When building 3D data is represented in actual spatial coordinates... and Having the same length dimension as the spatial coordinates of the building's 3D data, the sum of the noise components still forms directional modulation noise with a length dimension; when the building's 3D data uses normalized spatial coordinates... and A dimensionless representation consistent with normalized spatial coordinates is adopted. Configure the noise intensity coefficients in the normal direction and the tangential direction to make the disturbance amplitude of the directional modulation noise in the local normal direction smaller than the disturbance amplitude in the local tangential direction.

[0087] Specifically, in this application, the noise intensity coefficients in the normal direction and the tangential direction are configured based on the geometric variations of the building surface in the local normal and tangential directions. The geometric undulations of building surfaces such as walls, floors, and glass curtain walls along the local normal direction are generally smaller than the spatial extension variations along the local tangential direction. Therefore, configuring the noise intensity of the normal direction noise component to be lower than that of the tangential direction noise component can reduce the undulations in the local redrawn representation that are inconsistent with the adjacent original 3D representation along the local normal direction. In one embodiment, the noise intensity coefficients in the normal direction and the tangential direction can be configured in a ratio of approximately 1:3 to 1:5, for example, the noise intensity coefficient in the normal direction is 0.08 and the noise intensity coefficient in the tangential direction is 0.35. The above values ​​are only used to illustrate one noise intensity configuration method; the actual noise intensity coefficients are configured according to the spatial scale of the building's 3D data.

[0088] Specifically, in this application, the scale label continues to participate in the spatial frequency modulation of the initial directional noise. The scale label indicates whether the corresponding region in the layered redraw mask originates from a structural defect candidate region or a detailed defect candidate region. For the scale label corresponding to a structural defect candidate region, the initial directional noise is adjusted to form a scale modulation noise with a longer spatial variation period, making the noise change between adjacent local sampling positions more continuous; for the scale label corresponding to a detailed defect candidate region, the initial directional noise is adjusted to form a scale modulation noise with a shorter spatial variation period, enabling the local redrawing process to retain surface detail changes within a smaller spatial range. In one embodiment, the scale modulation noise corresponding to a structural defect candidate region can use a longer spatial wavelength (e.g., about 0.4 meters), and the scale modulation noise corresponding to a detailed defect candidate region can use a shorter spatial wavelength (e.g., about 0.05 meters). The specific spatial wavelength is configured based on the sampling interval of the building's 3D data and the spatial size of the layered candidate regions.

[0089] Furthermore, when the mask regions corresponding to structural defect candidate regions and detailed defect candidate regions are adjacent in the layered redraw mask, the adjacent positions of the two adjacent mask regions are identified based on the scale label to generate scale boundary positions. The center positions of the structural defect candidate regions and detailed defect candidate regions are generated based on the local sampling positions in the mask regions on both sides of the scale boundary position, respectively. Then, a transition modulation representation is generated based on the spatial distances from the local sampling positions corresponding to the scale boundary positions to the center positions of the structural defect candidate regions and detailed defect candidate regions, respectively. The transition modulation representation is used to continuously adjust the noise intensity and spatial frequency corresponding to the two scales at the scale boundary position, reducing the degree of abrupt changes in scale modulation noise on both sides of the scale boundary position, and continuing to participate in the generation of directional modulation noise.

[0090] In one embodiment, for the local sampling position corresponding to the scale boundary position Transition weights for candidate regions of structural defects Transition weights for candidate regions with detailed defects It can be represented as:

[0091] in, Indicates the local sampling location The spatial distance to the center of the candidate region for structural defects. Indicates the local sampling location The spatial distance to the center of the candidate region for detailed defects. and All use the same length unit as the spatial coordinates of the building's three-dimensional data; This represents the transition weight corresponding to the candidate region of structural defects. The transition weights represent candidate regions for detailed defects. Both transition weights are formed by the ratio of length to length and do not have physical dimensions, and satisfy the following conditions: The directional modulation noise corresponding to the scale boundary position can be further expressed as:

[0092] in, This represents the directional modulation noise corresponding to the scale boundary location; This represents scale-modulated noise generated according to the scale labels corresponding to the candidate regions of structural defects. This represents scale-modulated noise generated according to the scale labels corresponding to the candidate regions of detail defects. and Using the same spatial representation and physical dimensions, weighted correlation can be performed based on dimensionless transition weights. After adjusting the scale modulation noise based on the above transition modulation characterization, directional modulation noise is generated, and this directional modulation noise continues to participate in the generation of the initial characterization for redrawing.

[0093] Optionally, the step of generating a redrawing condition representation based on building 3D data and associated observation data, and limiting the redrawing condition representation according to a layered redrawing mask and scale labels to generate a redrawing guidance representation; generating a local redrawing representation based on directional modulation noise and the redrawing guidance representation, includes: Based on the layered redrawing mask, the mask coverage area and the area outside the mask are divided, and the original three-dimensional representation is read from the building three-dimensional data according to the area outside the mask; According to the mask coverage area, the observation condition representation is extracted from the associated observation data, and the original 3D representation and the observation condition representation are associated to generate the redrawing condition representation; The effective area of ​​the redraw condition representation is limited by the layered redraw mask, the spatial attention range is generated based on the scale label, and then the spatial association range of the redraw condition representation is limited based on the spatial attention range to generate the redraw guidance representation. The initial redrawing representation is generated based on the directional modulation noise and the redrawing guidance representation. The three-dimensional representation of the mask-covered area is then recovered step by step based on the initial redrawing representation to generate the local redrawing representation.

[0094] Specifically, in this application, the layered redraw mask is used to distinguish between the mask-covered area and the area outside the mask in the architectural 3D data. The mask-covered area corresponds to the architectural space range that needs to be locally redrawn, while the area outside the mask corresponds to the spatial range that retains the original architectural spatial form during this round of local redrawing. Based on the area outside the mask, the original 3D representation is read from the architectural 3D data, so that the original 3D representation continues to retain the spatial position, surface orientation, and local form adjacent to the mask boundary of the building surfaces around the mask-covered area. The original 3D representation subsequently participates in both the generation of the redrawing condition representation and the boundary fusion of each restoration level, so that the mask-covered area can continuously refer to the original form of the adjacent architectural space during the step-by-step restoration process.

[0095] Specifically, in this application, the observation condition characterization is generated from the observation information in the associated observation data that corresponds to the space of the mask-covered area.

[0096] Based on the masked area, spatial observation information, depth observation data, and data related to building component identification are read from the associated observation data corresponding to the acquisition viewpoint. Then, based on the building's spatial location, this observation information is associated to form an observation condition representation. This observation condition representation then establishes a spatial correspondence with the original 3D representation corresponding to the area outside the mask to generate a redrawing condition representation. The redrawing condition representation thus simultaneously includes the observation condition representation corresponding to the masked area and the original 3D representation corresponding to the area outside the mask, enabling the local redrawing process to utilize the original observation data of the defective area and the original spatial morphology of the surrounding unredrawn areas.

[0097] Specifically, in this application, the effective area of ​​the redrawing condition representation is limited based on the layered redrawing mask, preventing redrawing condition representations that are not spatially related to the current mask-covered area from entering the current local redrawing process. Subsequently, a spatial scope of interest is generated based on the scale label. For the scale label corresponding to a structural defect candidate region, the spatial scope of interest covers a large spatial neighborhood around the mask-covered area, allowing the stepwise recovery of larger missing areas to utilize a wider range of original 3D representations and observation condition representations. For the scale label corresponding to a detailed defect candidate region, the spatial scope of interest is concentrated in the vicinity of the mask-covered area, allowing local detail recovery to utilize more of the original 3D representations directly adjacent to the current defect. The spatial association range of the redrawing condition representation is limited based on the spatial scope of interest to generate a redrawing guidance representation, ensuring that the redrawing guidance representation maintains its correspondence with the layered redrawing mask and scale label.

[0098] Furthermore, the directional modulation noise is confined to the mask coverage area corresponding to the layered redraw mask, and the directional modulation noise and redraw condition representation are associated according to the spatial correlation range provided by the redrawing guidance representation to generate the initial redraw representation. The directional modulation noise in the initial redraw representation only acts on the mask coverage area and limits the initial change direction and amplitude of the mask coverage area; the area outside the mask continues to retain the original 3D representation. The redrawing guidance representation then limits the original 3D representation and observation condition representation that can be used during the step-by-step recovery process of the mask coverage area. Thus, the initial redraw representation continues to serve as the starting point for the step-by-step recovery of the 3D representation, ultimately generating the local redraw representation.

[0099] In a preferred embodiment, the step-by-step recovery process of the 3D representation of the masked area can be understood as, under the combined constraints of directional modulation noise, redrawing guided representation, and scale labels, gradually reducing noise components and restoring the architectural spatial form from the initial redrawing representation. This step-by-step recovery process can be represented by the following conditional relationship:

[0100] in, This represents the 3D representation of the area covered by the layered redraw mask after step-by-step reconstruction. This indicates the aforementioned layered redraw mask; This represents the initial redraw representation formed by the combination of directional modulation noise and redraw guidance representation. This indicates a redraw of the guiding representation; Indicates the scale label corresponding to the area covered by the mask; This represents the conditional probability distribution of the 3D representation of the masked area, given the initial redraw representation, the guided redraw representation, and the scale label; (symbol) This indicates that the three-dimensional representation after stepwise recovery is formed according to this conditional probability distribution. It represents the conditional distribution relationship given the initial redraw representation, the redraw guidance representation, and the scale label, and is not used to perform direct numerical addition or size comparison on different data channels; , and Each data channel maintains a numerical representation consistent with the building's 3D data or associated observation data. The stepwise restoration process continuously reduces the directional modulation noise component in the initial redrawing representation and continuously uses the redrawing-guided representation to limit the restoration direction, so that the local redrawing representation tends to be consistent with the architectural spatial form of the original 3D representation and associated observation data.

[0101] Optionally, the step of generating an initial redrawing representation based on directional modulation noise and redrawing guidance representation, and progressively recovering the three-dimensional representation of the mask-covered area based on the initial redrawing representation to generate a local redrawing representation, includes: Based on the redrawing guidance characterization, directional modulation noise is applied to the mask coverage area to generate the initial redrawing characterization; The process of hierarchically reconstructing the 3D representation of the masked area is divided into multiple reconstruction levels to generate reconstruction levels; Extract the mask boundary positions from the layered redraw mask, and generate boundary transition representations based on the mask boundary positions; Based on the restoration level, the first restoration level's 3D representation restoration is performed based on the initial redrawing representation to generate the first-level redrawing 3D representation; the first-level redrawing 3D representation and the original 3D representation are fused based on the boundary transition representation to generate inter-level redrawing representations; The three-dimensional representation of the next recovery level is restored sequentially based on the inter-level redrawing representation corresponding to the previous recovery level to generate the redrawing three-dimensional representation corresponding to the next recovery level. Then, the redrawing three-dimensional representation corresponding to the next recovery level and the original three-dimensional representation are fused according to the boundary transition representation to update the inter-level redrawing representation. After the final recovery stage is completed, candidate redrawing 3D representations are generated based on the inter-stage redrawing representations. The candidate redrawing 3D representations and the original 3D representations are then fused based on the boundary transition representations to generate local redrawing representations.

[0102] Specifically, in this application, the recovery level is used to represent the order in which the initial redrawn representation is gradually recovered to the candidate redrawn 3D representation. The number of recovery levels is configured according to the spatial scale and scale label of the mask coverage area, and the recovery levels are arranged in order of decreasing directional modulation noise to generate the recovery level. The recovery level corresponding to the candidate region of structural defects can adopt a longer step-by-step recovery process, allowing the 3D representation over a larger spatial range to gradually form a continuous structure; the recovery level corresponding to the candidate region of detailed defects can adopt a shorter step-by-step recovery process, allowing local details to be recovered within a smaller spatial scope of interest. In one embodiment, the number of recovery levels can be configured to be multiple, or fewer, such as 16 levels; this value is only used to illustrate one implementation of the recovery level.

[0103] Specifically, in this application, the locations adjacent to the mask-covered area and the area outside the mask are extracted from the layered redraw mask to generate mask boundary locations. Centered on the mask boundary locations, a boundary transition range with continuously varying weights is set within the mask-covered area, and a boundary transition representation is generated based on the distance from the local sampling locations within the boundary transition range to the mask boundary locations. The weights in the boundary transition representation increase the participation of the original 3D representation near the mask boundary locations, and increase the participation of the redrawn 3D representation corresponding to the current recovery level when moving away from the mask boundary locations and entering the mask-covered area, thereby ensuring that the 3D representation of each recovery level maintains spatial continuity with the original 3D representation at the mask boundary.

[0104] In a preferred embodiment, let the first The redrawn 3D representation generated at each restoration level is located at a local sampling position. place as The original three-dimensional representation of the region outside the mask, extending to the boundary transition range, is as follows: The fusion weights corresponding to the boundary transition representation are: Then the first The inter-level redrawing representation formed by each recovery level can be expressed as:

[0105] in, Indicates the recovery level number in the recovery level sequence; Indicates the local sampling location within the boundary transition range; Indicates the first Each recovery level corresponds to a redrawn 3D representation; Indicates the location of the local sampling The original three-dimensional representation corresponding to space; Indicates the location of local sampling Dimensionless fusion weights are generated relative to the distances at the mask boundary positions, and satisfy the following conditions: ; Indicates the first The inter-level redraw representation corresponding to each recovery level. and Using the same three-dimensional representation, the same data channels, and the same physical dimensions, weighted correlation can be achieved through dimensionless fusion weights. and and Maintain the same physical dimensions.

[0106] Specifically, in this application, the first recovery stage takes the initial redrawing representation as input, reduces some directional modulation noise components within the spatial interest range defined by the redrawing-guided representation, and recovers the preliminary spatial morphology of the mask-covered area based on the observation conditions corresponding to the associated observation data, thereby generating the first-level redrawing 3D representation. The first-level redrawing 3D representation and the original 3D representation are then fused based on the boundary transition representation to generate the corresponding inter-level redrawing representation. Subsequently, each subsequent recovery stage always takes the inter-level redrawing representation corresponding to the previous recovery stage as input, without returning to the initial redrawing representation, thus maintaining the unidirectional progressive relationship of the recovery stages.

[0107] As each recovery level is executed, the directional modulation noise component is further reduced based on the inter-level redrawing representation corresponding to the previous recovery level. The 3D representation within the current spatial scope of interest is then restored according to the redrawing guidance representation to generate the redrawn 3D representation corresponding to the next recovery level. Finally, the redrawn 3D representation corresponding to the next recovery level and the original 3D representation are fused based on the boundary transition representation to update the inter-level redrawing representation. This process is repeated according to the recovery levels until the last recovery level is completed, causing the inter-level redrawing representation to gradually change towards an architectural spatial form that is consistent with the redrawing condition representation.

[0108] After completing the final restoration level, candidate redrawn 3D representations are generated based on the inter-level redrawing representations corresponding to the final restoration level. These candidate redrawn 3D representations already include the 3D shape of the masked area after successive restorations and the boundary shape formed by the continuous fusion of the mask boundary positions across each restoration level. The candidate redrawn 3D representations and the original 3D representations are then fused again based on the boundary transition representations to generate local redrawing representations. In one embodiment, the boundary transition range can be configured as several spatial sampling unit widths (e.g., 2 voxel widths), with the specific width configured according to the spatial sampling resolution of the architectural 3D data.

[0109] Optionally, the step of generating an initial redrawing representation based on directional modulation noise and redrawing guidance representation, and progressively recovering the three-dimensional representation of the mask-covered area based on the initial redrawing representation to generate a local redrawing representation, further includes: Based on the scale verification results, the region to be verified is extracted, and the 3D representation to be redrawn is extracted from the candidate 3D representation to be verified based on the region to be verified. Based on the detection of the 3D representation to be redrawn for review using mesoscale confidence maps, a consistent redrawing result is generated. Based on the redrawing consistency results, key review markers are generated and associated with local redrawing representations, so that key review markers and local redrawing representations jointly participate in the generation of redrawing review results.

[0110] Specifically, in this application, the scale verification results record the spatial locations where there are discrepancies between the coarse-scale confidence map, the fine-scale confidence map, and the meso-scale confidence map during the generation of the layered redrawing mask from the aforementioned multi-scale confidence representation. These spatial locations are extracted based on the scale verification results to generate the region to be verified. The region to be verified continues to maintain its correlation with the scale location, enabling it to be mapped to the actual architectural spatial location in the candidate redrawing 3D representation. Based on the region to be verified, the corresponding 3D representation is extracted from the candidate redrawing 3D representation to generate the redrawing 3D representation to be verified.

[0111] Specifically, in this application, when detecting the three-dimensional representation to be redrawn based on the mesoscale confidence map, the spatial morphology corresponding to the three-dimensional representation to be redrawn is compared with the confidence change relationship given by the mesoscale confidence map at the same architectural location. If the local morphology after the three-dimensional representation to be redrawn is consistent with the continuity of the surrounding space, and the corresponding mesoscale confidence map no longer indicates that the location has a confidence change that contradicts the surrounding area, then a consistent redrawing result is generated; if the three-dimensional representation to be redrawn still has inconsistencies with the confidence change reflected by the mesoscale confidence map at the same spatial location, then the consistent redrawing result records that the area to be redrawn still needs further judgment in the subsequent redrawing review results.

[0112] Based on the consistency results of the redrawing, key verification markers are generated and associated with the local redrawing representation according to the architectural spatial location. These key verification markers do not alter the 3D form in the local redrawing representation; instead, they participate in the generation process of the redrawing review results along with the local redrawing representation. This ensures that spatial areas with scale discrepancies in the aforementioned multi-scale confidence representation retain independent review criteria after local redrawing. Key verification markers are removed when subsequent redrawing review results indicate that the corresponding area meets the redrawing requirements; if the redrawing review results still indicate that the corresponding area does not meet the current redrawing requirements, the key verification markers continue to participate in the local refinement process along with the non-compliant areas.

[0113] Optionally, the step of reviewing the local redrawing representation based on the spatial confidence map to generate the redrawing review result includes: The redrawing area is determined based on the local redrawing representation. The original three-dimensional representations corresponding to the adjacent positions of the redrawing area are extracted from the building three-dimensional data. Based on the local redrawing representation and the original three-dimensional representation, a coarse-scale structural continuity result is generated. Based on the spatial confidence map, the original confidence representation corresponding to the redrawn region is extracted. The local integrity of the redrawn region is detected based on the local redrawn representation to generate the local integrity result. Then, a fine-scale residual confidence representation is generated based on the original confidence representation and the local integrity result. Extract the mask boundary positions from the layered redraw mask, and detect the boundary continuity relationship between the local redraw representation and the original 3D representation based on the mask boundary positions to generate boundary continuity results; Redrawing and re-inspection results are generated based on coarse-scale structural continuity results, fine-scale residual confidence characterization, boundary continuity results, and key verification markers.

[0114] Specifically, in this application, the redrawing area is determined based on the spatial position of the local redrawing representation in the architectural 3D data. Then, the original 3D representation adjacent to the redrawing area is read along the boundary of the redrawing area to the area outside the mask. The coarse-scale structural continuity result is used to describe whether the large-scale surface orientation formed by the local redrawing representation remains continuous with the adjacent original 3D representation. For approximately planar architectural surfaces such as walls, the adjacent original surfaces are fitted from the adjacent original 3D representations, and then the spatial deviation of the local redrawing representation within the redrawing area relative to the adjacent original surfaces is calculated. For architectural surfaces with continuous curvature, the coarse-scale structural continuity result is generated based on the continuous change relationship of the surface orientation on both sides of the redrawing area.

[0115] In a preferred embodiment, for a wall redrawing area that is approximately planar, a reference plane can be fitted based on adjacent original 3D representations. And the three-dimensional sampling locations involved in coarse-scale detection in the redrawn region. to reference plane The average distance generates coarse-scale structural residuals. :

[0116] in, This represents the coarse-scale structural residual, whose physical dimension is length. It represents the number of three-dimensional sampling locations involved in the detection of coarse-scale structural continuity, and has no physical dimensions; Indicates the first element within the redrawn area. Three-dimensional sampling locations involved in the detection; This represents a reference plane formed by fitting the original 3D representations corresponding to adjacent positions in the redrawn region. Indicates the three-dimensional sampling position to reference plane The shortest spatial distance, with the physical dimension of length. Since the summation terms are all lengths and the divisor is... Dimensionless, therefore It remains a length measure. The coarse-scale structural continuity determination boundary is compared with a boundary configured using the same length unit to generate a coarse-scale structural continuity result. In one embodiment, the coarse-scale structural continuity determination boundary can be configured according to the sampling precision of the building's 3D data (e.g., 0.03 meters).

[0117] Specifically, in this application, spatial confidence information of the redrawn area before redrawing is extracted based on the spatial confidence map to generate an original confidence representation. When detecting the local integrity of the redrawn area based on the local redrawing representation, the spatial coverage continuity and local surface orientation continuity of the three-dimensional sampling positions within the redrawn area are re-examined to generate a local integrity result. Then, a correspondence is established between the local integrity result and the original confidence representation at the same architectural spatial location to generate a fine-scale residual confidence representation. For local locations where the original confidence representation before redrawing is low and the local integrity result still reflects a lack of spatial coverage, the fine-scale residual confidence representation continues to mark them as low-confidence locations that need to be re-examined; for local locations where the local integrity result already indicates continuous spatial coverage and stable surface orientation, the fine-scale residual confidence representation records that the location has been locally restored.

[0118] Specifically, in this application, the boundary continuity result is used to detect whether the local redrawn representation and the original 3D representation corresponding to the area outside the mask form a continuous spatial continuity at the mask boundary position. The mask boundary position is read from the layered redrawn mask, and then the local redrawn representation inside the mask boundary position and the original 3D representation outside the mask boundary position are read separately. The boundary continuity result is generated based on the spatial interval between the corresponding 3D sampling positions on both sides and the local surface orientation change. When the spatial interval on both sides of the mask boundary position is within a range suitable for the spatial sampling resolution of the building 3D data, and the local surface orientation change is continuous with the extension direction of the adjacent building surface, the boundary continuity result indicates that the mask boundary position forms a continuous continuity; otherwise, the boundary continuity result records the corresponding boundary position and continuity deviation type to participate in the generation of the redrawing review result.

[0119] Furthermore, redrawing and re-inspection results are generated based on coarse-scale structural continuity results, fine-scale residual confidence representations, boundary continuity results, and key verification markers. The redrawing and re-inspection results record the re-inspection conclusions for each redrawn area and the reasons for not meeting the current redrawing requirements according to the building's spatial location. Coarse-scale structural continuity results are used to determine whether a large area of ​​the building surface is continuous; fine-scale residual confidence representations are used to determine whether there are still local gaps within the redrawn area; boundary continuity results are used to determine whether there are boundary continuity anomalies between the local redrawn representation and the original 3D representation; and key verification markers are used to retain spatial areas where the aforementioned scale judgments differ. If any corresponding spatial location still has unmet re-inspection conditions, the redrawing and re-inspection results record that spatial location as a non-compliant area and record the reasons for non-compliance; when all re-inspection conditions are met, the corresponding spatial location is recorded as a compliant area.

[0120] In one embodiment, for a candidate area of ​​structural defects in the lobby of an office building formed by suspended components, the coarse-scale structural residual of the wall surface after local redrawing relative to the adjacent original wall surface can be 0.025 meters, which is lower than the coarse-scale structural continuity judgment boundary of 0.03 meters configured according to the sampling accuracy of the building's 3D data. Therefore, the coarse-scale structural continuity result is recorded as meeting the current redrawing requirements. For a candidate area of ​​detailed defects corresponding to the edge of the glass curtain wall, if the local integrity result still indicates the existence of a fine-scale residual area of ​​approximately 0.08 square meters, the fine-scale residual confidence characterization continues to record this fine-scale residual area, and the redrawing re-inspection result determines this fine-scale residual area as a non-compliant area. The above specific values ​​are only used to illustrate one implementation of the re-inspection process.

[0121] Optionally, generating local refinement results based on the redrawing and re-inspection results includes: Extract the non-compliant areas and reasons for non-compliance from the redrawing and re-inspection results; Based on the reasons for non-compliance, mask adjustment information is generated. Based on the mask adjustment information, the layered redraw mask corresponding to the non-compliance area is locally adjusted to generate a fine-tuned redraw mask. Based on multi-scale confidence representation, fine-refinement redraw mask and non-compliant areas, fine-refinement scale labels are generated. Based on the fine-refinement redraw mask, fine-refinement local surface orientation representation is extracted from the building 3D data. Then, fine-refinement orientation modulation noise is generated based on the fine-refinement local surface orientation representation and fine-refinement scale labels. Based on the refinement redraw mask, refinement original 3D representation is extracted from the building 3D data. Refinement observation condition representation is extracted from the associated observation data according to the refinement redraw mask. Refinement original 3D representation and refinement observation condition representation are associated to generate refinement redraw condition representation. Refinement redraw condition representation is limited according to refinement redraw mask and refinement scale label to generate refinement redraw guidance representation. The initial refinement redrawing representation is generated based on the refinement direction modulation noise and the refinement redrawing guidance representation. The three-dimensional representation corresponding to the unqualified area is restored step by step according to the initial refinement redrawing representation to generate the local refinement result.

[0122] Specifically, in this application, both the non-compliant areas and the reasons for non-compliance are directly derived from the redrawing and re-inspection results. Non-compliant areas record the locations of building spaces that still need to be redrawn, while the reasons for non-compliance record which re-inspection condition the building space location failed to meet: coarse-scale structural continuity results, fine-scale residual confidence characterization, boundary continuity results, or key review markers. When generating mask adjustment information based on the reasons for non-compliance, the reasons for non-compliance are converted into local adjustment methods executable by the layered redrawing mask, allowing the mask adjustment information to continue participating in the generation of the fine-tuning redrawing mask.

[0123] Specifically, in this application, when the cause of non-compliance corresponds to a low-confidence position recorded in the fine-scale residual confidence representation, the mask adjustment information increases the boundary extension distance of the adjacent position of the non-compliance area, and adjusts the fine-scale confidence judgment boundary according to the confidence distribution of the fine-scale confidence map near the non-compliance area, so that the fine-redrawing mask covers the local space that has not yet been fully restored; when the cause of non-compliance corresponds to the coarse-scale structural continuity result, the mask adjustment information extends the non-compliance area along the surface continuity direction recorded by the coarse-scale structural continuity result, so that the fine-redrawing mask can include the adjacent space that is continuously related to the non-compliance area; when the cause of non-compliance corresponds to the boundary continuity result, the mask adjustment information increases the boundary transition range along the corresponding mask boundary position, so that the subsequent fine-refinement process can re-establish the continuity with the original fine-refinement 3D representation at the corresponding mask boundary position. The above mask adjustment information only acts on the local range corresponding to the non-compliance area and does not change the layered redrawing mask corresponding to the qualified area.

[0124] Based on mask adjustment information, the layered redraw mask corresponding to the substandard area is locally adjusted to generate a refined redraw mask. Subsequently, refined scale labels are generated based on multi-scale confidence representation, the refined redraw mask, and the substandard area, allowing the refined scale labels to re-record the defect spatial scale of the spatial area corresponding to the refined redraw mask. When extracting refined local surface direction representations from the architectural 3D data based on the refined redraw mask, following the aforementioned generation process of local surface adjacency relationships, local normal directions, and local tangent directions, local sampling positions are reread within the spatial range corresponding to the refined redraw mask. Then, based on the reread local sampling positions, refined local surface direction representations are generated, ensuring that the refined local surface direction representation reflects the actual effective range of the substandard area after mask adjustment.

[0125] Furthermore, refinement direction modulation noise is generated based on the refinement local surface orientation representation and refinement scale labels. The refinement direction modulation noise adopts a noise intensity lower than the local tangent direction along the local normal direction in the refinement local surface orientation representation, and its spatial frequency is reconfigured based on the refinement scale labels. Therefore, the refinement direction modulation noise is not a direct reuse of the previous round of direction modulation noise, but rather a regeneration based on the refinement local surface orientation representation and refinement scale labels, allowing the refinement direction modulation noise to continue participating in the generation of the initial representation for refinement redrawing.

[0126] Specifically, in this application, based on the refinement redraw mask, three-dimensional representations that maintain their original shape around the refinement redraw mask are extracted from the architectural three-dimensional data to generate the original refinement three-dimensional representation. Data corresponding to the refinement redraw mask space is extracted from the associated observation data according to the refinement redraw mask to generate the refinement observation condition representation. The original refinement three-dimensional representation and the refinement observation condition representation are associated to generate the refinement redraw condition representation. Then, the effective area of ​​the refinement redraw condition representation is limited according to the refinement redraw mask, and the corresponding spatial focus area is limited according to the refinement scale label to generate the refinement redraw guidance representation. An initial refinement redraw representation is generated based on the refinement direction modulation noise and the refinement redraw guidance representation, ensuring that the initial refinement redraw representation only corresponds to the non-compliant area and its refinement redraw mask coverage area.

[0127] When progressively restoring the 3D representations of substandard areas based on the initial refined redrawing representation, a unidirectional progressive processing is performed along the aforementioned restoration levels. First, the boundary positions of the refined redrawing mask are extracted, and a refined boundary transition representation is generated based on these positions. In the first restoration level, the 3D representations of substandard areas are restored based on the initial refined redrawing representation to generate refined redrawing 3D representations. The refined redrawing 3D representations and the original refined 3D representations are then fused based on the refined boundary transition representations to generate inter-level refined redrawing representations. In subsequent restoration levels, the inter-level refined redrawing representations of the previous restoration level are used as input to continue restoring the 3D representations of substandard areas, updating the refined redrawing 3D representations. Then, the updated refined redrawing 3D representations and the original refined 3D representations are fused based on the refined boundary transition representations to update the inter-level refined redrawing representations. After completing the last restoration level, a local refined result is generated based on the inter-level refined redrawing representations of the last restoration level, allowing the local refined result to continue participating in the generation of refined redrawing representations.

[0128] In one embodiment, for areas with substandard details still existing at the edge of the glass curtain wall, the boundary expansion distance of the corresponding candidate areas for detailed defects can be adjusted from a smaller range to a larger range (e.g., from 0.05 meters to 0.08 meters) based on the reason for the substandard nature. The fine-scale confidence judgment boundary can also be adjusted based on the fine-scale confidence map (e.g., from 0.25 to 0.30). Then, the fine-tuning direction modulation noise and fine-tuning redrawing guidance characterization are regenerated based on the updated fine-tuning redrawing mask. The above values ​​are only used to illustrate one implementation method of locally adjusting the layered redrawing mask based on the redrawing review results.

[0129] Optionally, the step of generating a refined redrawing representation based on the local refinement result; extracting the retained redrawing representation from the local redrawing representation based on the redrawing review result, and then associating the retained redrawing representation with the refined redrawing representation to generate a local redrawing result includes: Extract the qualified and unqualified areas from the redrawing and re-inspection results. Based on the qualified areas, extract the corresponding three-dimensional representations from the local redrawing representations to generate retained redrawing representations. The scope of local fine-tuning is determined based on the areas that do not meet the standards. Based on the scope of local fine-tuning, the corresponding 3D fine-tuning representation is extracted from the local fine-tuning results to generate a fine-tuning redrawing representation. Based on the qualified and unqualified areas, determine the spatial continuity between the retained redrawing representation and the refined redrawing representation, and associate the retained redrawing representation and the refined redrawing representation according to the spatial continuity to generate local redrawing results.

[0130] Specifically, in this application, the areas that meet the standards in the redrawing and re-inspection results correspond to the architectural space ranges that do not need to be changed again after this round of redrawing, while the areas that do not meet the standards correspond to the architectural space ranges that need to be replaced by the local refinement results of the previous round of local redrawing representation. Based on the areas that meet the standards, the corresponding three-dimensional representations are extracted from the local redrawing representations to generate retained redrawing representations. The retained redrawing representations maintain the three-dimensional forms that already meet the current redrawing requirements in the previous round of local redrawing representations, and do not enter the step-by-step restoration process corresponding to the areas that do not meet the standards, thus allowing subsequent calculations to focus on the areas that do not meet the standards.

[0131] When determining the scope of local refinement based on non-compliant areas, the spatial correspondence between the non-compliant areas and the coverage area of ​​the refinement redraw mask determines the spatial range that needs to be replaced by the local refinement results. Based on the scope of local refinement, the corresponding refinement 3D representation is extracted from the local refinement results to generate the refinement redraw representation. The refinement redraw representation and the retained redraw representation use the same architectural 3D data expression format, enabling them to directly establish a spatial correspondence based on the architectural spatial location.

[0132] Based on the compliant and non-compliant areas, the spatial continuity position between the retained redrawing representation and the refined redrawing representation is determined. A result continuity transition representation is generated based on this spatial continuity position. Then, the retained and refined redrawing representations are merged according to this transition representation to generate a local redrawing result. The result continuity transition representation continuously adjusts the participation level of the retained and refined redrawing representations along both sides of the spatial continuity position, ensuring that compliant areas in the local redrawing result originate from the retained redrawing representation, while non-compliant areas, after being limited by the local refined redrawing scope, originate from the refined redrawing representation. This ensures a continuous transition between the retained and refined redrawing representations at the spatial continuity position.

[0133] Furthermore, after the local redrawing results are generated, the previously non-compliant areas can be re-inspected using the same method as the redrawing review results. If a previously non-compliant area becomes a compliant area after the redrawing review, the corresponding refined redrawing representation is retained and the refinement of that area ends. If non-compliant areas still exist after the redrawing review, mask adjustment information and local refinement results are regenerated only for the new non-compliant areas. The number of repetitions is limited by a pre-configured maximum number of refinements, which is configured based on the scale of the 3D building data and the computational workload of a single local refinement. When the maximum number of refinements is reached, the latest generated local redrawing result is used as the output, and the remaining key review markers are retained in the redrawing review results. This process always updates non-compliant areas at the regional granularity and does not reprocess compliant areas.

[0134] In one embodiment, if the continuous occlusion area in the office building lobby has already met the standards in the first round of redrawing and review, while some areas at the edge of the glass curtain wall still remain unmet, the second round of local refinement only regenerates the refinement redraw mask, refinement direction modulation noise, and refinement redraw representation for the unmet areas at the edge of the glass curtain wall, without changing the retained redraw representation corresponding to the continuous occlusion area. After the second round of redrawing and review results show that the unmet areas at the edge of the glass curtain wall meet the current redrawing requirements, the retained redraw representation corresponding to the continuous occlusion area is associated with the refinement redraw representation corresponding to the edge of the glass curtain wall according to their spatial continuity to form a local redraw result. This implementation method maintains consistency with the processing relationship between the aforementioned redrawing and review results and the local refinement results.

[0135] Optionally, the step of acquiring historical building quality labeling data, extracting historical confidence evidence and corresponding quality labels from the historical building quality labeling data, and generating an evidence contribution representation based on the correspondence between historical confidence evidence and corresponding quality labels includes: Acquire historical building quality labeling data, and extract historical confidence evidence and corresponding quality labels from the historical building quality labeling data; Based on the correspondence between historical confidence evidence and corresponding quality labels, the historical building quality label data is divided into multiple data partitions to generate a data partitioning scheme; The data partitions are selected sequentially from the data partitioning scheme to generate validation partitions. The validation partitions are excluded based on the data partitioning scheme to generate training partitions. Then, the validation partitions and training partitions are associated to generate a partition rotation relationship. Candidate evidence contribution representations are generated based on the historical confidence evidence and corresponding quality labels contained in the training partitions in the partition rotation relationship. Candidate evidence contribution representations are detected based on the verification partitions in the same partition rotation relationship to generate verification errors. Based on the degree of dispersion of verification errors corresponding to different partition rotation relationships, a fusion reliability is generated. Candidate evidence contribution characteristics are then selected based on the fusion reliability to generate evidence contribution characteristics.

[0136] Specifically, in this application, historical building quality annotation data is used to generate the aforementioned evidence contribution characterization. The historical confidence evidence in the historical building quality annotation data and the multi-source confidence evidence corresponding to the current building's 3D data adopt the same evidentiary meaning. The corresponding quality annotations in the historical building quality annotation data record the data quality of the historical building's spatial location. To reduce the extent to which the evidence contribution characterization is only adapted to a single building type or a single observation source, the historical building quality annotation data is divided into multiple data partitions according to the correspondence between historical confidence evidence and corresponding quality annotations to generate a data partitioning scheme, ensuring that each data partition retains the correlation between historical confidence evidence and its corresponding quality annotations.

[0137] Specifically, in this application, one data partition is selected sequentially from the data partitioning scheme as the validation partition, and the remaining data partitions in the data partitioning scheme are used as training partitions. The validation partition and the training partitions are then associated to generate a partition rotation relationship. After completing one partition rotation relationship, the next data partition is selected from the data partitioning scheme as the new validation partition, and the remaining data partitions are re-determined as training partitions. The validation partitions are rotated sequentially in the above manner, ensuring that each data partition serves as a validation partition at least once, thus forming multiple partition rotation relationships.

[0138] When generating candidate evidence contribution representations based on historical confidence evidence and corresponding quality labels contained in training partitions within a partition rotation relationship, the relative contribution of different types of historical confidence evidence in forming the overall confidence level is determined according to the correspondence between changes in different types of historical confidence evidence and changes in corresponding quality labels in the training partitions, thus generating candidate evidence contribution representations. Subsequently, historical confidence evidence contained in validation partitions within the same partition rotation relationship is input into the evidence fusion mapping using this candidate evidence contribution representation, and the resulting overall confidence level is compared with the corresponding quality labels in the validation partitions to generate validation error. Validation error describes the degree of deviation of the candidate evidence contribution representation from the confidence evaluation on data partitions that did not participate in the current training partition generation process.

[0139] In a preferred embodiment, the data partitioning scheme includes The data partition, the first The verification error corresponding to each partition rotation relationship is: The average verification error corresponding to different partition rotation relationships is... and validation error variance It can be represented as:

[0140]

[0141] in, This represents the number of data partitions in the data partitioning scheme, and is a dimensionless integer. The sequence number indicating the partition rotation relationship; Indicates the first The verification error corresponding to each partition rotation relationship is formed by the overall confidence level and the deviation of the corresponding quality label within a uniform numerical range, and therefore is expressed as dimensionless. This represents the average verification error corresponding to all partition rotation relationships, and is also expressed in a dimensionless manner. This represents the dispersion of verification errors corresponding to different partition rotation relationships. It is formed by dimensionless verification errors and therefore does not involve physical dimension transformation. Verification error variance The smaller the value, the smaller the variation in validation error across different validation partitions in terms of the contribution of candidate evidence; validation error variance The larger the value, the more sensitive the contribution of candidate evidence is to changes in data partitioning.

[0142] When generating fusion reliability based on the dispersion of verification errors corresponding to different partition rotation relationships, the variance of verification errors will be... The fusion reliability is converted to an inverse relationship with the dispersion of the verification error, so that candidate evidence contribution representations with smaller verification error dispersion correspond to higher fusion reliability, and candidate evidence contribution representations with larger verification error dispersion correspond to lower fusion reliability. Candidate evidence contribution representations are then selected based on the fusion reliability to generate evidence contribution representations, which are then used to adjust the contribution of current building 3D data to multi-source confidence evidence. Thus, the evidence contribution representations not only originate from the correspondence between historical confidence evidence and corresponding quality labels, but also undergo rotational verification across different data partitions, ensuring a continuous technical relationship between the generation of the aforementioned spatial confidence map and historical building quality label data.

[0143] In an exemplary embodiment, the 3D building data is a laser point cloud of the lobby area of ​​an existing office building. The associated observation data also includes depth observation data of the lobby area, image reconstruction results, and data related to building component identification. In the multi-source confidence evidence, areas obscured by suspended components exhibit reduced sampling sufficiency, while reflective areas of the glass curtain wall show reduced local orientation consistency and observation signal-to-noise ratio. Furthermore, the laser scanning results and image reconstruction results form low cross-source consistency evidence in the glass curtain wall area. After adjusting the aforementioned multi-source confidence evidence based on the evidence contribution characterization and generating a spatial confidence map, multi-scale confidence characterization is used to form candidate regions for structural defects and candidate regions for detailed defects, respectively. Layered redrawing masks and scale labels are then generated.

[0144] Based on the layered redraw mask, local surface orientation representations are extracted from the 3D building data, and directional modulation noise is generated according to the relationship that the noise intensity in the local normal direction is less than the noise intensity in the local tangent direction. For candidate regions of structural defects formed by suspended components, scale labels make the directional modulation noise adopt a longer spatial variation period; for candidate regions of detailed defects formed by the edges of glass curtain walls, scale labels make the directional modulation noise adopt a shorter spatial variation period. After generating the initial redraw representation based on the directional modulation noise and the redraw guiding representation, the 3D representation of the mask-covered area is restored step by step along the restoration level. At each restoration level, the current redraw 3D representation and the original 3D representation are fused according to the boundary transition representation to generate a local redraw representation.

[0145] After the local redrawing representation is generated, a redrawing review result is generated based on the coarse-scale structural continuity result, fine-scale residual confidence representation, boundary continuity result, and key verification markers. If the coarse-scale structural continuity result corresponding to the area obscured by the suspended component indicates that the area meets the current redrawing requirements, the corresponding 3D representation is extracted from the local redrawing representation to form a retained redrawing representation. If there are still non-compliant areas recorded by the fine-scale residual confidence representation at the edge of the glass curtain wall, the layered redrawing mask corresponding to the area is adjusted according to the reason for non-compliance to generate a fine redrawing mask, and fine-tuning direction modulation noise, fine redrawing guidance representation, and local fine-tuning result are regenerated. After the redrawing review result is generated again and meets the current redrawing requirements, the fine-tuning 3D representation is extracted from the local fine-tuning result to form a fine redrawing representation. Then, the retained redrawing representation and the fine redrawing representation are associated according to the spatial continuity position to generate the local redrawing result. Throughout the process, each round of local fine-tuning only applies to the non-compliant areas recorded in the redrawing review result; the retained redrawing representation corresponding to the compliant areas does not enter the subsequent step-by-step recovery process.

[0146] like Figure 2 As shown in the figure, an embodiment of this application discloses a device for generating a partial redrawing of a 3D architectural model guided by a spatial confidence map, comprising: The multi-source confidence construction module is used to acquire building 3D data and associated observation data, extract multi-source confidence evidence from the building 3D data and associated observation data, and generate a spatial confidence map based on the multi-source confidence evidence; The scale layering module is used to generate multi-scale confidence representations based on spatial confidence maps, generate layered redraw masks based on multi-scale confidence representations, and then generate scale labels based on multi-scale confidence representations and layered redraw masks. The orientation modulation module is used to extract local surface orientation representations from building 3D data based on hierarchical redraw masks, and generate orientation modulation noise based on local surface orientation representations and scale labels. The local redraw module is used to generate redraw condition representations based on building 3D data and associated observation data. It limits the redraw condition representations according to the layered redraw mask and scale label to generate redraw guidance representations. It generates local redraw representations based on directional modulation noise and redraw guidance representations. The redrawing and re-inspection module is used to re-inspect local redrawing representations based on spatial confidence maps to generate redrawing and re-inspection results; generate local refinement results based on the redrawing and re-inspection results; generate refined redrawing representations based on the local refinement results; extract and retain redrawing representations from local redrawing representations based on the redrawing and re-inspection results, and then associate the retained redrawing representations and refined redrawing representations to generate local redrawing results.

[0147] In this embodiment, the spatial confidence map generated by the multi-source confidence construction module is passed to the scale layering module to participate in the generation of multi-scale confidence representation; the layered redraw mask and scale label generated by the scale layering module are passed to the directional modulation module and the local redraw module to participate in the generation of directional modulation noise and redraw guidance representation, respectively; the directional modulation noise generated by the directional modulation module is passed to the local redraw module to generate local redraw representation together with the redraw guidance representation; the local redraw representation is passed to the redraw review module to generate local redraw result, thereby enabling each module to be executed sequentially according to the unidirectional processing relationship corresponding to the aforementioned method.

[0148] like Figure 3 As shown, an electronic device according to an embodiment of this application includes a processor and a memory; The memory is used to store computer programs and the building 3D data, associated observation data, spatial confidence maps, layered redraw masks, scale labels and local redraw results generated during the operation of the computer programs; The processor is connected to the memory and is used to read and execute the computer program to implement the method for generating a local redraw of a three-dimensional building model based on a spatial confidence map as described in any one of this application. During the execution of the computer program, the processor reads the building 3D data and associated observation data from the memory, and writes the spatial confidence map, layered redrawing mask, scale label, and local redrawing results generated by executing any of the methods described in this application into the memory, so that the building 3D data and the associated observation data form corresponding local redrawing results through the building 3D model local redrawing generation method guided by the spatial confidence map.

[0149] In one embodiment, the electronic device can be a computing device for repairing 3D building reconstruction data. The 3D building data can be a building point cloud, a building triangular mesh, or a 3D implicit representation of the building. The associated observation data can originate from depth observation data that has a spatial correspondence with the 3D building data, spatial observation information corresponding to the acquisition viewpoint, and building component identification-related data. After reading the 3D building data and the associated observation data, the electronic device completes local redrawing according to any one of the methods described in this application and outputs the local redrawing result.

[0150] like Figure 4 As shown, this is a computer-readable storage medium according to an embodiment of the present application, wherein a computer program is stored in the computer-readable storage medium; When the computer program is read and executed by an electronic device, it implements the method for generating a local redrawing of a three-dimensional building model based on a spatial confidence map, as described in any one of this application. During the execution of the computer program, the three-dimensional building data and associated observation data are used as input. A spatial confidence map is generated according to any one of the methods described in this application. Then, based on the spatial confidence map, a layered redrawing mask and scale label are generated. Directional modulation noise is formed according to the local surface orientation representation. Then, a local redrawing representation is generated by combining the redrawing guidance representation to obtain the local redrawing result.

[0151] In one embodiment, the computer-readable storage medium may be a non-volatile storage medium capable of storing a computer program and being read by an electronic device. After the computer program is read, the building 3D data and associated observation data are invoked according to the processing order of any of the methods described in this application, so that the technical processing results generated by the previous technical processing process continue to participate in the subsequent technical processing process until a partial redrawing result is generated.

[0152] like Figure 5 As shown, this is an embodiment of a building 3D scene partial redrawing generation system, which includes a building 3D data acquisition device, an associated observation data acquisition device, a building 3D model partial redrawing device, and a 3D scene output device; the building 3D model partial redrawing device is communicatively connected to the building 3D data acquisition device, the associated observation data acquisition device, and the 3D scene output device, respectively. The building 3D data acquisition device is used to acquire 3D spatial data corresponding to the target building space to form building 3D data, and send the building 3D data to the building 3D model local redrawing device; The associated observation data acquisition device is used to acquire observation data that has a spatial correspondence with the building's three-dimensional data to form associated observation data, and to send the associated observation data to the building's three-dimensional model local redrawing device; The building 3D model local redrawing device is used to receive building 3D data and associated observation data, extract multi-source confidence evidence from the building 3D data and associated observation data, and generate a spatial confidence map based on the multi-source confidence evidence; The building 3D model local redrawing device is also used to generate multi-scale confidence representations based on spatial confidence maps, generate layered redrawing masks based on multi-scale confidence representations, and then generate scale labels based on multi-scale confidence representations and layered redrawing masks. The building 3D model local redrawing device is also used to extract local surface orientation representations from building 3D data based on layered redrawing masks, and generate orientation modulation noise based on local surface orientation representations and scale labels; The building 3D model local redrawing device is also used to generate redrawing condition representations based on building 3D data and associated observation data, limit the redrawing condition representations according to the layered redrawing mask and scale label, and generate redrawing guidance representations; and generate local redrawing representations based on directional modulation noise and redrawing guidance representations. The building 3D model local redrawing device is also used to re-examine the local redrawing representation based on the spatial confidence map to generate a redrawing re-examination result; generate a local refinement result based on the redrawing re-examination result, generate a refinement redrawing representation based on the local refinement result; extract the retained redrawing representation from the local redrawing representation based on the redrawing re-examination result, and then associate the retained redrawing representation with the refinement redrawing representation to generate a local redrawing result, and send the local redrawing result to the 3D scene output device; The 3D scene output device is used to receive the local redrawing results and update the 3D scene of the building corresponding to the target building space based on the local redrawing results.

[0153] In one embodiment, the building 3D data acquisition device can use a laser scanning device to acquire point cloud data of the interior space of an existing building, and form building 3D data from the point cloud data; the associated observation data acquisition device acquires depth observation data and image reconstruction data corresponding to the point cloud data to form associated observation data. Taking the lobby of an office building as an example, when the suspension components cause partial loss of building 3D data, or when the glass curtain wall reflection causes noise interference to the partial building 3D data, the building 3D model local redrawing device performs the aforementioned local redrawing process on the building 3D data and associated observation data to generate local redrawing results, and then the 3D scene output device updates the corresponding lobby building 3D scene based on the local redrawing results.

[0154] Figures 2-5 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for generating local redrawing of a 3D architectural model based on spatial confidence maps, characterized in that, include: Acquire 3D building data and associated observation data, extract multi-source confidence evidence from the 3D building data and associated observation data, and generate a spatial confidence map based on the multi-source confidence evidence; Multi-scale confidence representations are generated based on spatial confidence maps, hierarchical redraw masks are generated based on multi-scale confidence representations, and scale labels are generated based on multi-scale confidence representations and hierarchical redraw masks. Local surface orientation representations are extracted from building 3D data based on hierarchical redraw mask, and orientation modulation noise is generated based on local surface orientation representations and scale labels. Redrawing condition representations are generated based on building 3D data and associated observation data. Redrawing condition representations are limited according to layered redrawing masks and scale labels to generate redrawing guidance representations. Local redraw representations are generated based on directional modulation noise and redraw guidance representations; The local redrawing representation is re-examined based on the spatial confidence map to generate redrawing and re-examination results; Based on the redrawing and re-inspection results, a local fine-tuning result is generated, and a fine-tuning redrawing representation is generated based on the local fine-tuning result. Based on the redrawing and re-inspection results, the retained redrawing representation is extracted from the local redrawing representation, and then the retained redrawing representation and the fine-tuning redrawing representation are associated to generate a local redrawing result.

2. The method for generating local redrawing of a 3D building model based on a spatial confidence map as described in claim 1, characterized in that, The extraction of multi-source confidence evidence from building 3D data and associated observation data includes: Local geometric observation representations are extracted from architectural 3D data and associated observation data, and evidence areas are delineated based on these local geometric observation representations. Extract semantic recognition evidence from associated observation data according to evidence regions; Read the data source identifier corresponding to the evidence area from the associated observation data, extract the cross-source observation results corresponding to the evidence area based on the data source identifier, and then compare the cross-source observation results to generate cross-source consistent evidence; Read the collection time identifier corresponding to the evidence area from the associated observation data, generate the collection time interval based on the collection time identifier, and generate time-sensitive evidence based on the collection time interval; Multi-source confidence evidence is generated based on local geometric observation representations, semantic recognition evidence, cross-source consistent evidence, and time-sensitive evidence.

3. The method for generating local redrawing of a 3D building model based on a spatial confidence map as described in claim 2, characterized in that, The process of extracting local geometric observation representations from building 3D data and associated observation data, and delineating evidence regions based on these local geometric observation representations, includes: Read the 3D sampling positions from the 3D building data, and generate a spatial neighborhood based on the 3D sampling positions; Based on the neighborhood distribution of the three-dimensional sampling locations in the spatial neighborhood, sampling sufficiency is generated; A neighborhood normal representation is generated based on the 3D sampling position in the spatial neighborhood, and a local orientation consistency is generated based on the directional dispersion of the neighborhood normal representation. Depth observation data is read from associated observation data according to spatial neighborhood, and observation signal-to-noise representation is generated based on the depth observation data; Local geometric observation representations are generated based on sampling sufficiency, local orientation consistency, and observation signal-to-noise ratio. Evidence regions are then divided based on the correspondence between local geometric observation representations and three-dimensional sampling positions.

4. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 2, characterized in that, The generation of spatial confidence maps based on multi-source confidence evidence includes: Acquire historical building quality labeling data, extract historical confidence evidence and corresponding quality labels from the historical building quality labeling data, and generate evidence contribution representation based on the correspondence between historical confidence evidence and corresponding quality labels; Multi-source confidence evidence is modulated based on the contribution characteristics of the evidence to generate contribution-modulated evidence; Based on contribution modulation evidence, a statistical representation of the evidence neighborhood is generated, and then an initial spatial confidence map is generated based on the contribution modulation evidence and the statistical representation of the evidence neighborhood. Confidence boundary gradients are generated based on the initial spatial confidence map, and the initial spatial confidence map is smoothed by maintaining the boundaries based on the confidence boundary gradients to generate a spatial confidence map.

5. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 4, characterized in that, The generation of multi-scale confidence representations based on spatial confidence maps includes: Maintain the spatial sampling resolution of the spatial confidence map to generate a fine-scale confidence map; Reduce spatial sampling resolution based on fine-scale confidence maps to generate mesoscale confidence maps; Based on the mesoscale confidence map, the spatial sampling resolution is further reduced to generate a coarse-scale confidence map; Scale location correspondence is generated based on the spatial sampling resolution correspondence between fine-scale confidence maps, mesoscale confidence maps, and coarse-scale confidence maps. The fine-scale confidence maps, mesoscale confidence maps, and coarse-scale confidence maps are then associated based on the scale location correspondence to generate multi-scale confidence representations.

6. The method for generating local redrawing of a 3D building model based on a spatial confidence map according to claim 5, characterized in that, The process of generating a hierarchical redraw mask based on multi-scale confidence representation, and then generating scale labels based on the multi-scale confidence representation and the hierarchical redraw mask, includes: Based on the coarse-scale confidence map, continuously distributed low-confidence regions are identified, and candidate regions for structural defects are generated. Based on the candidate regions of structural defects, the spatial regions corresponding to the candidate regions of structural defects are excluded from the fine-scale confidence map to generate a fine-scale residual confidence map. Then, based on the fine-scale residual confidence map, isolated low-confidence regions are identified to generate candidate regions of detailed defects. Based on the mesoscale confidence map, candidate regions for structural defects and candidate regions for detailed defects are detected respectively, and scale verification results are generated. Based on the scale verification results, candidate regions for structural defects and candidate regions for detailed defects are selected to generate hierarchical candidate regions; Based on the layered candidate regions, boundary expansion results are generated. Based on the boundary expansion results, a layered redrawing mask is generated. Then, based on the scale position correspondence, the layered candidate regions are associated with the corresponding regions in the layered redrawing mask to generate scale labels.

7. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 6, characterized in that, The extraction of local surface orientation representations from 3D building data based on layered redraw masks includes: The mask coverage area is located by redrawing the mask layer by layer, and the local sampling position is read from the 3D building data based on the mask coverage area; Local surface adjacency relationships are generated based on local sampling locations, and local normal directions are generated based on these local surface adjacency relationships. Based on the local normal direction, the local tangent direction orthogonal to the local normal direction is determined, and then the local normal direction and the local tangent direction are associated to generate a local surface orientation representation.

8. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 7, characterized in that, The generation of directional modulation noise based on local surface orientation characterization and scale labels includes: Local normal direction and local tangent direction are extracted from the local surface orientation representation. Normal direction noise component is generated based on local normal direction and tangent direction noise component is generated based on local tangent direction. Normal direction noise component and tangent direction noise component are associated according to the relationship that the noise intensity of normal direction noise component is less than that of tangent direction noise component to generate initial orientation noise. Based on the scale labels, the initial directional noise is adjusted according to the relationship that the spatial frequency corresponding to the candidate region of structural defects is less than that corresponding to the candidate region of detailed defects, thus generating scale-modulated noise. Based on scale labels, adjacent mask regions corresponding to different scale labels are identified in the layered redraw mask, and scale boundary positions are generated. Transition modulation representations are generated based on the relative positions of sampling points corresponding to scale boundary positions between adjacent mask regions corresponding to different scale labels. Scale modulation noise is adjusted based on transition modulation representations to generate directional modulation noise.

9. The method for generating local redrawing of a 3D building model based on a spatial confidence map according to claim 8, characterized in that, The redrawing condition representation is generated based on the building's 3D data and associated observation data. The redrawing condition representation is limited according to the layered redrawing mask and scale label to generate a redrawing guidance representation. Local redraw representations are generated based on directional modulation noise and redraw guidance representations, including: Based on the layered redrawing mask, the mask coverage area and the area outside the mask are divided, and the original three-dimensional representation is read from the building three-dimensional data according to the area outside the mask; According to the mask coverage area, the observation condition representation is extracted from the associated observation data, and the original 3D representation and the observation condition representation are associated to generate the redrawing condition representation; The effective area of ​​the redraw condition representation is limited by the layered redraw mask, the spatial attention range is generated based on the scale label, and then the spatial association range of the redraw condition representation is limited based on the spatial attention range to generate the redraw guidance representation. The initial redrawing representation is generated based on the directional modulation noise and the redrawing guidance representation. The three-dimensional representation of the mask-covered area is then recovered step by step based on the initial redrawing representation to generate the local redrawing representation.

10. The method for generating local redrawing of a 3D building model based on a spatial confidence map according to claim 9, characterized in that, The process of generating an initial redrawing representation based on directional modulation noise and redrawing guidance representation, and then progressively recovering the 3D representation of the mask-covered area based on the initial redrawing representation to generate a local redrawing representation, includes: Based on the redrawing guidance characterization, directional modulation noise is applied to the mask coverage area to generate the initial redrawing characterization; The process of hierarchically reconstructing the 3D representation of the masked area is divided into multiple reconstruction levels to generate reconstruction levels; Extract the mask boundary positions from the layered redraw mask, and generate boundary transition representations based on the mask boundary positions; Based on the restoration level, the first restoration level's 3D representation restoration is performed based on the initial redrawing representation to generate the first-level redrawing 3D representation; the first-level redrawing 3D representation and the original 3D representation are fused based on the boundary transition representation to generate inter-level redrawing representations; The three-dimensional representation of the next recovery level is restored sequentially based on the inter-level redrawing representation corresponding to the previous recovery level to generate the redrawing three-dimensional representation corresponding to the next recovery level. Then, the redrawing three-dimensional representation corresponding to the next recovery level and the original three-dimensional representation are fused according to the boundary transition representation to update the inter-level redrawing representation. After the final recovery stage is completed, candidate redrawing 3D representations are generated based on the inter-stage redrawing representations. The candidate redrawing 3D representations and the original 3D representations are then fused based on the boundary transition representations to generate local redrawing representations.

11. The method for generating local redrawing of a 3D architectural model guided by a spatial confidence map according to claim 10, characterized in that, The process of generating an initial redrawing representation based on directional modulation noise and redrawing guidance representation, and progressively recovering the 3D representation of the mask-covered area based on the initial redrawing representation to generate a local redrawing representation, further includes: Based on the scale verification results, the region to be verified is extracted, and the 3D representation to be redrawn is extracted from the candidate 3D representation to be verified based on the region to be verified. Based on the mesoscale confidence map detection, the 3D representation to be redrawn is to be reviewed and redrawn to generate consistent redrawing results; Based on the redrawing consistency results, key verification marks are generated and associated with local redrawing representations, so that key verification marks and local redrawing representations jointly participate in the generation of redrawing review results.

12. The method for generating local redrawing of a 3D building model based on a spatial confidence map as described in claim 11, characterized in that, The process of re-examining local redrawing representations based on spatial confidence maps to generate redrawing and re-examining results includes: The redrawing area is determined based on the local redrawing representation. The original three-dimensional representations corresponding to the adjacent positions of the redrawing area are extracted from the building three-dimensional data. Based on the local redrawing representation and the original three-dimensional representation, a coarse-scale structural continuity result is generated. Based on the spatial confidence map, the original confidence representation corresponding to the redrawn region is extracted. The local integrity of the redrawn region is detected based on the local redrawn representation to generate the local integrity result. Then, a fine-scale residual confidence representation is generated based on the original confidence representation and the local integrity result. Extract the mask boundary positions from the layered redraw mask, and detect the boundary continuity relationship between the local redraw representation and the original 3D representation based on the mask boundary positions to generate boundary continuity results; Redrawing and re-inspection results are generated based on coarse-scale structural continuity results, fine-scale residual confidence characterization, boundary continuity results, and key verification markers.

13. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 12, characterized in that, The process of generating local refinement results based on the redrawing and re-inspection results includes: Extract the non-compliant areas and reasons for non-compliance from the redrawing and re-inspection results; Based on the reasons for non-compliance, mask adjustment information is generated. Based on the mask adjustment information, the layered redraw mask corresponding to the non-compliance area is locally adjusted to generate a fine-tuned redraw mask. Based on multi-scale confidence representation, fine-refinement redraw mask and non-compliant areas, fine-refinement scale labels are generated. Based on the fine-refinement redraw mask, fine-refinement local surface orientation representation is extracted from the building 3D data. Then, fine-refinement orientation modulation noise is generated based on the fine-refinement local surface orientation representation and fine-refinement scale labels. Based on the refinement redraw mask, refinement original 3D representation is extracted from the building 3D data. Refinement observation condition representation is extracted from the associated observation data according to the refinement redraw mask. Refinement original 3D representation and refinement observation condition representation are associated to generate refinement redraw condition representation. Refinement redraw condition representation is limited according to refinement redraw mask and refinement scale label to generate refinement redraw guidance representation. Based on the precision direction modulation noise and precision redrawing guidance representation, an initial precision redrawing representation is generated. Based on the initial precision redrawing representation, the three-dimensional representation corresponding to the unqualified area is restored step by step to generate local precision results.

14. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 13, characterized in that, The process involves generating a refined redrawing representation based on local refinement results; extracting and retaining redrawing representations from the local redrawing representations based on the redrawing review results; and then associating the retained redrawing representations with the refined redrawing representations to generate local redrawing results, including: Extract the qualified and unqualified areas from the redrawing and re-inspection results. Based on the qualified areas, extract the corresponding three-dimensional representations from the local redrawing representations to generate retained redrawing representations. The scope of local fine-tuning is determined based on the areas that do not meet the standards. Based on the scope of local fine-tuning, the corresponding 3D fine-tuning representation is extracted from the local fine-tuning results to generate a fine-tuning redrawing representation. Based on the qualified and unqualified areas, determine the spatial continuity between the retained redrawing representation and the refined redrawing representation, and associate the retained redrawing representation and the refined redrawing representation according to the spatial continuity to generate local redrawing results.

15. The method for generating local redrawing of a 3D building model guided by a spatial confidence map according to claim 4, characterized in that, The process of acquiring historical building quality labeling data, extracting historical confidence evidence and corresponding quality labels from the historical building quality labeling data, and generating evidence contribution representations based on the correspondence between historical confidence evidence and corresponding quality labels includes: Acquire historical building quality labeling data, and extract historical confidence evidence and corresponding quality labels from the historical building quality labeling data; Based on the correspondence between historical confidence evidence and corresponding quality labels, the historical building quality label data is divided into multiple data partitions to generate a data partitioning scheme; The data partitions are selected sequentially from the data partitioning scheme to generate validation partitions. The validation partitions are excluded based on the data partitioning scheme to generate training partitions. Then, the validation partitions and training partitions are associated to generate a partition rotation relationship. Candidate evidence contribution representations are generated based on the historical confidence evidence and corresponding quality labels contained in the training partitions in the partition rotation relationship. Candidate evidence contribution representations are detected based on the verification partitions in the same partition rotation relationship to generate verification errors. Based on the degree of dispersion of verification errors corresponding to different partition rotation relationships, a fusion reliability is generated. Candidate evidence contribution characteristics are then selected based on the fusion reliability to generate evidence contribution characteristics.

Images