A Deep Learning-Based Method for Defect Identification in Building Curtain Walls

Abstract

This invention belongs to the field of intelligent inspection technology for building curtain walls. It discloses a deep learning-based method for identifying defects in building curtain walls. The method involves acquiring raw inspection data of the building curtain wall and generating a spatial inspection sequence based on this data. The spatial inspection sequence is then subjected to image standardization processing to obtain standardized curtain wall image data. This standardized image data is input into a curtain wall component segmentation model to extract the spatial regions of the curtain wall components. A spatial region map of the curtain wall components is generated based on these regions. The method utilizes the building curtain wall spatial inspection sequence and image standardization processing principles to unify curtain wall images under different viewing angles, distances, and lighting conditions. Furthermore, it establishes a correspondence between defect identification results and specific curtain wall component regions using curtain wall component segmentation and component region constraint principles. Finally, it employs multi-view spatial matching, defect spatial evidence chains, and re-updating principles.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology for building curtain walls, and in particular relates to a method for identifying defects in building curtain walls based on deep learning. Background Technology

[0002] As an important component of the external envelope of high-rise buildings, building curtain walls are subject to long-term effects from wind loads, temperature changes, rainwater erosion, ultraviolet radiation, and structural deformation. This can easily lead to defects such as glass cracks, panel damage, sealant delamination, water seepage, contamination buildup, localized deformation, and loose connections. If these defects are not detected and located in a timely manner, they may affect the appearance quality, sealing performance, structural safety, and subsequent maintenance efficiency of the building curtain wall. Therefore, a deep learning-based method for identifying building curtain wall defects is needed. This method would standardize curtain wall inspection images, identify component regions, detect candidate defect regions, correlate evidence from multiple perspectives, and update defect results.

[0003] Existing methods for detecting defects in building curtain walls mainly rely on manual visual inspection, single-image recognition, or simple target detection models for judgment. In actual inspections, the curtain wall surface is easily affected by reflections, obstructions, changes in shooting distance, shooting angle deviations, and repetitive textures of the facade structure, which can cause defective areas to be confused with the edges of normal components, shadow areas, and reflective areas, resulting in insufficient accuracy and stability in defect identification.

[0004] Furthermore, existing defect identification schemes typically rely on detection boxes or segmented regions within a single image as the basis for judgment, lacking a mechanism to correlate shooting pose data, building facade area numbers, continuous inspection videos, and spatial areas of curtain wall components. This makes it difficult to form a multi-view evidence chain for the same defect. When the reliability of a defect is low, existing schemes also lack a processing method to automatically determine the re-sampling area and re-sampling angle, resulting in insufficient verifiability of the identification results and low efficiency in subsequent maintenance and positioning. Summary of the Invention

[0005] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a deep learning-based method for identifying building curtain wall defects. It utilizes building curtain wall spatial inspection sequences and image standardization processing principles to unify curtain wall images under different viewing angles, distances, and lighting conditions. By employing curtain wall component segmentation and component region constraint principles, it establishes a correspondence between defect identification results and specific curtain wall component regions. Furthermore, by utilizing multi-view spatial matching, defect spatial evidence chains, and re-updating principles, it performs multi-view evidence fusion and dynamic correction of low-confidence results for the same defect, achieving stable identification, accurate positioning, and verifiable output of building curtain wall defects. This solves the problems of existing detection methods being susceptible to reflections, occlusions, and viewing angle deviations, insufficient accuracy of single-image recognition, and the lack of a re-updating mechanism.

[0006] This invention provides a deep learning-based method for identifying defects in building curtain walls. This method addresses the technical problems of existing building curtain wall defect identification methods, which mainly rely on static detection of single images. These methods are difficult to adapt to situations such as curtain wall facade reflection, occlusion, long-distance shooting, viewing angle deviation, and complex component structures. As a result, the accuracy of defect identification, such as cracks, damage, delamination, water seepage marks, pollution deposits, deformation, and loose connections is insufficient, defect location is unstable, and the identification results lack multi-view evidence support.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying defects in building curtain walls based on deep learning includes: acquiring raw inspection data of building curtain walls; generating a spatial inspection sequence of building curtain walls based on the raw inspection data; performing image standardization processing on the spatial inspection sequence of building curtain walls to obtain standardized curtain wall image data; inputting the standardized curtain wall image data into a curtain wall component segmentation model to extract the spatial regions of curtain wall components; and generating a spatial region map of curtain wall components based on the spatial regions of curtain wall components.

[0008] Furthermore, standardized curtain wall image data and spatial area maps of curtain wall components are input into a deep learning defect detection model to extract multi-view defect candidate regions. Based on the shooting pose data corresponding to the multi-view defect candidate regions, the building facade area number, the spatial area map of curtain wall components, and continuous inspection videos, a defect spatial evidence chain is constructed. Based on the defect spatial evidence chain, the spatial area map of curtain wall components, and standardized curtain wall image data, the building curtain wall defect identification result is generated.

[0009] Furthermore, based on the shooting pose data, building facade area numbering, and continuous inspection videos, the original data of building curtain wall inspection is sorted in time order, facade area is bound, and viewpoint order is associated to generate a building curtain wall space inspection sequence. Based on the viewpoint overlap rate and facade coverage rate between adjacent images in the building curtain wall space inspection sequence, the inspection coverage integrity is calculated, and the uncovered facade areas are marked according to the inspection coverage integrity.

[0010] Furthermore, standardized curtain wall image data is input into the curtain wall component segmentation model to extract curtain wall edge texture features, component outline features, and component spatial semantic features. Based on the curtain wall edge texture features, component outline features, and component spatial semantic features, a spatial region map of the curtain wall components is generated.

[0011] Furthermore, standardized curtain wall image data and spatial region maps of curtain wall components are input into a deep learning defect detection model to extract defect visual features; component region constraints are applied to the defect visual features based on the spatial region maps of curtain wall components to generate multi-view defect candidate regions; and the confidence level of the defect candidate regions is calculated based on the defect visual feature response value, component region matching degree, image quality score, and reflective interference intensity.

[0012] Furthermore, based on the spatial location, visual characteristics of the defects, spatial area of ​​the curtain wall component to which they belong, and differences in shooting angles, a cross-view defect matching score is calculated. Multi-view defect candidate areas with cross-view defect matching scores higher than the matching score threshold are associated to generate a multi-view evidence set. A defect spatial evidence chain is generated based on the multi-view evidence set, and an evidence chain integrity score is calculated based on the number of defect candidate areas, view coverage completeness, and component area consistency in the defect spatial evidence chain.

[0013] Furthermore, the defect type is identified based on the visual characteristics of the defect in the defect spatial evidence chain, the spatial area of ​​the curtain wall component to which it belongs, and the multi-view evidence set; the defect location is calculated based on the building facade area number, the spatial area map of the curtain wall component, and the shooting pose data; the defect area and defect length are calculated based on the defect candidate area boundary, perspective distortion correction results, and scale unification results; the defect severity score is calculated based on the defect type, defect location, defect area, defect length, evidence chain integrity score, and component importance; and the defect credibility score is calculated based on the defect candidate area confidence score, cross-view defect matching score, and evidence chain integrity score, and the building curtain wall defect identification result is generated.

[0014] Furthermore, when the defect credibility score is lower than the credibility threshold, the defect spatial evidence chain is marked as a low credibility evidence chain; the supplementary sampling area and supplementary sampling angle are determined based on the low credibility evidence chain, and a supplementary sampling inspection instruction is generated; supplementary sampling curtain wall image data is obtained based on the supplementary sampling inspection instruction, and the defect spatial evidence chain and building curtain wall defect identification results are updated based on the supplementary sampling curtain wall image data.

[0015] The beneficial effects of this invention are as follows: (1) Using the principle of building curtain wall space inspection sequence and image standardization processing, a building curtain wall space inspection sequence is generated based on shooting pose data, building facade area number and continuous inspection video. The building curtain wall space inspection sequence is then subjected to sharpness detection, exposure detection, motion blur detection, perspective distortion correction, scale unification and curtain wall boundary clipping to obtain standardized curtain wall image data. This achieves unified processing of curtain wall images under different viewing angles, distances and lighting conditions, and solves the problem of large differences in image quality and significant deviations in shooting angle in existing curtain wall defect detection, which leads to unstable recognition results.

[0016] (2) Using the principle of curtain wall component segmentation and component area constraint, the standardized curtain wall image data is input into the curtain wall component segmentation model to extract the spatial area of ​​the curtain wall component and generate the spatial area map of the curtain wall component; then the standardized curtain wall image data and the spatial area map of the curtain wall component are input into the deep learning defect detection model to constrain the component area of ​​the defect visual features and generate multi-view defect candidate areas. This realizes the positional correspondence between the defect identification results and the glass panel area, metal frame area, sealant joint area, connection node area and operable sash area, and solves the problem that reflection, shadow, component edge and background texture are easily misjudged as defects in the existing defect detection.

[0017] (3) Using the principles of multi-view spatial matching, defect spatial evidence chain and re-update, based on the shooting pose data corresponding to the multi-view defect candidate area, the building facade area number, the curtain wall component spatial area map and continuous inspection video, the cross-view defect matching score is calculated and the defect spatial evidence chain is constructed; then, the building curtain wall defect identification result is generated based on the defect spatial evidence chain. When the defect confidence score is lower than the confidence threshold, a re-update inspection instruction is generated and the defect spatial evidence chain and building curtain wall defect identification result are updated. This realizes the fusion of multi-view evidence for the same defect and the dynamic correction of low confidence results, and solves the problems of lack of evidence support, unstable defect localization and insufficient verifiability of identification results in existing single image detection. Attached Figure Description

[0018] The accompanying drawings are provided to further understand the present invention and form part of the specification. They are used together with the embodiments of the present invention to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is an overall flowchart of the deep learning-based building curtain wall defect identification method proposed in this invention; Figure 2 This is a flowchart illustrating the defect identification result update process proposed in this invention. Detailed Implementation

[0020] Example 1, see Figures 1-2 The present invention provides a method for identifying defects in building curtain walls based on deep learning, comprising the following steps: Step S1: Acquire building curtain wall inspection image data; Collect first-view images, continuous inspection videos, shooting pose data, shooting time data and building facade area numbers of the building curtain wall through drone inspection equipment, ground mobile inspection equipment and fixed camera equipment to generate building curtain wall inspection raw data; Step S2: Construct a building curtain wall space inspection sequence; Based on the shooting pose data, building facade area numbering and continuous inspection videos, organize the original building curtain wall inspection data in time order, bind facade areas and associate viewpoint order to generate a building curtain wall space inspection sequence. Step S3: Generate standardized curtain wall image data; perform sharpness detection, exposure detection, motion blur detection, perspective distortion correction, scale unification, and curtain wall boundary clipping on the building curtain wall space inspection sequence to obtain standardized curtain wall image data; Step S4: Extract the spatial region of the curtain wall components; input the standardized curtain wall image data into the curtain wall component segmentation model, extract the glass panel region, metal frame region, sealant joint region, connection node region and operable sash region, and generate a spatial region map of the curtain wall components based on the glass panel region, metal frame region, sealant joint region, connection node region and operable sash region. Step S5: Generate multi-view defect candidate regions; Input standardized curtain wall image data and spatial region map of curtain wall components into deep learning defect detection model, extract candidate regions for cracks, damage, delamination, water seepage, contamination deposition, deformation, and loose connection, and generate multi-view defect candidate regions. Step S6: Construct a defect spatial evidence chain; Based on the shooting pose data corresponding to the multi-view defect candidate areas, the building facade area number, the curtain wall component spatial area map and continuous inspection video, calculate the spatial correspondence of the same defect under different views, and spatially associate the multi-view defect candidate areas under different views to generate a defect spatial evidence chain. Step S7: Generate building curtain wall defect identification results; Based on the defect spatial evidence chain, curtain wall component spatial area map and standardized curtain wall image data, calculate the defect type, defect location, defect area, defect length, defect severity score and defect confidence score, and generate building curtain wall defect identification results. Step S8: Perform defect re-sampling and result update; when the defect credibility score is lower than the credibility threshold, determine the re-sampling perspective and re-sampling area based on the defect spatial evidence chain, and generate a re-sampling inspection instruction; re-acquire the curtain wall image data based on the re-sampling inspection instruction, and update the defect spatial evidence chain and building curtain wall defect identification results based on the re-sampling curtain wall image data.

[0021] Through the above steps, this method transforms the identification of building curtain wall defects from static detection of a single image to joint identification of multi-view spatial evidence. By continuously inspecting images, segmenting component regions, detecting defect candidate regions, constructing spatial evidence chains, and updating and supplementing data, it improves the stability and accuracy of curtain wall defect identification in complex building facade scenarios.

[0022] Example 2: This example is based on all the above examples. The construction of the building curtain wall space inspection sequence specifically includes: Building facade area division: Based on the floor number, axis number, facade direction and component arrangement of the building curtain wall, the building curtain wall is divided into multiple building facade areas, and a building facade area number is assigned to each building facade area. Inspection image binding: Bind the first-view image, continuous inspection video, shooting pose data and shooting time data according to the building facade area number to obtain the inspection image data after area binding; Viewpoint sequence association; based on shooting time data and shooting pose data, the inspection image data after area binding is arranged in viewpoint sequence to generate a building curtain wall space inspection sequence with continuous viewpoints; Viewpoint coverage detection: Based on the viewpoint overlap rate and facade coverage rate between adjacent images in the building curtain wall space inspection sequence, the inspection coverage integrity is calculated; when the inspection coverage integrity is lower than the coverage integrity threshold, the uncovered facade area is marked and sent to the subsequent supplementary acquisition step.

[0023] Regarding parameter adjustments: Step 1: When the reflectivity of the building curtain wall surface increases, increase the exposure detection threshold and reduce the weight of the influence of the bright areas in the image on the integrity of the inspection coverage. Step 2: When the fluctuation range of the shooting pose data increases, the requirement for the overlap rate of the viewpoint between adjacent images is increased to keep the building curtain wall space inspection sequence continuous. Step 3: When the uncovered facade areas are concentrated in the corners of high-rise buildings, increase the priority of the supplementary sampling corresponding to that location, so that the supplementary sampling inspection instructions will prioritize covering the uncovered facade areas.

[0024] Example 3: This example is based on all the above examples. The generation of standardized curtain wall image data specifically includes: Image quality screening: The image frames below the image quality threshold are removed from the building curtain wall space inspection sequence to obtain valid curtain wall image data; sharpness detection, exposure detection and motion blur detection are performed on the building curtain wall space inspection sequence to obtain effective curtain wall image data. Perspective distortion correction: Based on the shooting pose data and the curtain wall boundary line, perspective distortion correction is performed on the effective curtain wall image data to obtain the perspective-corrected curtain wall image data. Scale unification: Based on the boundaries of the glass panels, the metal frame, and the sealant joints, the curtain wall image data after perspective correction is scaled to obtain scale-unified curtain wall image data. Curtain wall boundary clipping: Based on the building facade area number and curtain wall boundary line, the curtain wall image data after scale unification is clipped to obtain standardized curtain wall image data; Illumination consistency processing: Based on the brightness differences and shadow distribution between adjacent images, the standardized curtain wall image data is normalized for brightness and shadow suppressed to obtain standardized curtain wall image data with consistent illumination.

[0025] Example 4: This example is based on all the above examples. The extraction of the spatial area of ​​the curtain wall components specifically includes: Input to the component segmentation model: Input standardized curtain wall image data into the curtain wall component segmentation model to extract semantic features of the curtain wall components; Multi-scale component feature extraction; the curtain wall component segmentation model extracts the edge texture features of the curtain wall through shallow convolutional units, extracts the component contour features through medium convolutional units, and extracts the spatial semantic features of the component through deep convolutional units. The curtain wall edge texture features, component contour features and component spatial semantic features are then fused to generate curtain wall component fusion features. Component area generation: Based on the fusion characteristics of curtain wall components, glass panel area, metal frame area, sealant joint area, connection node area, and operable sash area are generated. Component spatial area diagram generation: The glass panel area, metal frame area, sealant joint area, connection node area and operable sash area are spatially bound according to the building facade area number to generate the curtain wall component spatial area diagram.

[0026] Example 5: This example is based on all the above examples. The specific steps for generating multi-view defect candidate regions include: Input to the defect detection model: Input standardized curtain wall image data and spatial area maps of curtain wall components into the deep learning defect detection model; Visual feature extraction of defects: The deep learning defect detection model extracts crack texture features, broken edge features, delamination boundary features, water seepage color features, contamination deposition morphology features, deformation contour features, and loose connection local structural features. Component constraint detection: Based on the spatial area diagram of the curtain wall components, the visual features of defects are constrained by the component area, so that the crack texture features, damaged edge features, delamination boundary features, water seepage trace color features, pollution deposition morphology features, deformation contour features, and loose connection local structural features are respectively established with the corresponding spatial areas of the curtain wall components. Multi-view defect candidate region generation: Based on the visual characteristics of defects after component region constraints, candidate regions for cracks, damage, delamination, water seepage, contamination, deformation, and loose connections are generated. These candidate regions are then aggregated to generate multi-view defect candidate regions. Calculation of confidence score for candidate defect regions: Based on the visual feature response value of the defect, the component region matching degree, and the image quality score, the confidence score of the candidate defect regions is calculated using the following formula: in, For the first Confidence of each defect candidate region; This is the normalization function; For the first Visual feature response values ​​of defects in each defect candidate region; For the first The component region matching degree of each defect candidate region; For the first Image quality score corresponding to each defect candidate region; For the first The intensity of reflective interference in each defect candidate region; , , and These are the weight coefficients for the corresponding items.

[0027] Example 6: This example is based on all the above examples. The construction of the defect space evidence chain specifically includes: Multi-view defect candidate area reception; based on the shooting pose data corresponding to the multi-view defect candidate area, the building facade area number, the curtain wall component spatial area map and continuous inspection video, determine the spatial location corresponding to each defect candidate area; Cross-view defect matching: Based on the spatial location of the defect candidate area, the visual characteristics of the defect, the spatial area of ​​the curtain wall component to which it belongs, and the difference in shooting angle, a cross-view defect matching score is calculated. Cross-view defect matching score calculation; the formula for calculating the cross-view defect matching score is as follows: in, For the first The candidate defect region and the first Cross-view defect matching score between defect candidate regions; For the first The candidate defect region and the first Spatial consistency among candidate defect regions; For the first The candidate defect region and the first The similarity of visual features of defects among candidate defect regions; For the first The candidate defect region and the first Consistency of component regions among candidate defect regions; For the first The candidate defect region and the first Differences in shooting perspectives between candidate defect regions; , , and These are the matching weight coefficients for the corresponding items; Defect space evidence chain generation: Defect candidate regions with cross-view defect matching scores higher than the matching score threshold are associated to form a multi-view evidence set corresponding to the same defect, and a defect space evidence chain is generated based on the multi-view evidence set. Evidence chain integrity score calculation: The evidence chain integrity score is calculated based on the number of candidate defect regions, view coverage integrity, and component region consistency in the defect space evidence chain.

[0028] Regarding parameter adjustments: Step 1: When the difference in shooting angle increases, increase the matching weight coefficient corresponding to spatial position consistency and decrease the matching weight coefficient corresponding to defect visual feature similarity; Step 2: When the intensity of reflective interference increases, increase the matching weight coefficient corresponding to the consistency of the component area and increase the matching score threshold; Step 3: When the evidence chain integrity score is lower than the integrity threshold, reduce the defect credibility score corresponding to the defect space evidence chain and trigger the supplementary collection step.

[0029] Example 7: This example is based on all the above examples. The specific steps for generating the building curtain wall defect identification result include: Defect type identification: Based on the visual characteristics of defects in the defect spatial evidence chain, the spatial area of ​​the curtain wall component to which they belong, and the multi-view evidence set, the defect type is identified; Defect location calculation: The defect location is calculated based on the building facade area number corresponding to the defect spatial evidence chain, the spatial area diagram of the curtain wall components, and the shooting pose data. Defect size calculation: Based on the defect candidate region boundary, perspective distortion correction results, and scale unification results in the defect spatial evidence chain, the defect area and defect length are calculated. Defect severity score calculation: Based on defect type, defect location, defect area, defect length, evidence chain integrity score, and component importance, the defect severity score is calculated using the following formula: in, Score the severity of the defect; Risk value for defect type; The defect area; The length of the defect; Importance of components; Scoring based on the completeness of the chain of evidence; , , , and These are the severity weighting coefficients for the corresponding items; Defect credibility score calculation: The defect credibility score is calculated based on the confidence of the defect candidate region, the cross-perspective defect matching score, and the evidence chain integrity score. Building curtain wall defect identification results are generated by binding defect type, defect location, defect area, defect length, defect severity score and defect credibility score.

[0030] Example 8: This example is based on all the above examples. The defect re-collection and result update specifically include: Supplementary sampling triggers judgment; when the defect credibility score is lower than the credibility threshold, the defect space evidence chain corresponding to the defect is marked as a low credibility evidence chain. The supplementary sampling area was determined based on the defect location corresponding to the low-confidence evidence chain, the building facade area number, and the spatial area diagram of the curtain wall components. The re-sampling angle was determined based on the differences in shooting angles in the low-confidence evidence chain, the evidence chain integrity score, and the intensity of reflective interference. Generate supplementary mining inspection instructions; generate supplementary mining inspection instructions based on the supplementary mining area and supplementary mining perspective. Acquire supplementary curtain wall image data; acquire supplementary curtain wall image data according to the supplementary inspection instructions and add the supplementary curtain wall image data to the building curtain wall space inspection sequence; Update the defect space evidence chain; update the defect space evidence chain based on supplementary curtain wall image data; The building curtain wall defect identification results are updated; the defect severity score and defect credibility score are recalculated based on the updated defect spatial evidence chain, and the building curtain wall defect identification results are updated based on the recalculated defect severity score and defect credibility score.

[0031] Through the above processing, this method can obtain supplementary visual evidence by supplementary inspection commands when curtain wall defects are affected by reflection, obstruction, long-distance shooting and viewing angle deviation. It can also use the supplementary curtain wall image data to update the spatial evidence chain of defects and the building curtain wall defect identification results, thereby improving the accuracy and verifiability of building curtain wall defect identification.

Claims

1. A method for identifying defects in building curtain walls based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain the original data of the building curtain wall inspection and generate the building curtain wall space inspection sequence based on the original data of the building curtain wall inspection; Step S2: Perform image standardization processing on the building curtain wall space inspection sequence to obtain standardized curtain wall image data; Step S3: Input the standardized curtain wall image data into the curtain wall component segmentation model, extract the spatial region of the curtain wall component, and generate a spatial region map of the curtain wall component based on the spatial region of the curtain wall component; Step S4: Input the standardized curtain wall image data and the spatial area map of the curtain wall components into the deep learning defect detection model to extract multi-view defect candidate areas; Step S5: Based on the shooting pose data corresponding to the multi-view defect candidate areas, the building facade area number, the curtain wall component spatial area map, and the continuous inspection video, construct the defect spatial evidence chain; Step S6: Generate building curtain wall defect identification results based on the defect space evidence chain, curtain wall component spatial area map, and standardized curtain wall image data; Step S7: When the defect credibility score in the building curtain wall defect identification result is lower than the credibility threshold, generate a supplementary inspection instruction based on the defect spatial evidence chain, and update the defect spatial evidence chain and the building curtain wall defect identification result based on the supplementary inspection instruction.

2. The method for identifying building curtain wall defects based on deep learning according to claim 1, characterized in that: Based on the shooting posture data, building facade area numbering and continuous inspection videos, the original data of building curtain wall inspection is sorted in time sequence, facade area is bound and viewpoint sequence is associated to generate a building curtain wall space inspection sequence. Based on the visual overlap rate and facade coverage rate between adjacent images in the building curtain wall space inspection sequence, the inspection coverage integrity is calculated, and the uncovered facade areas are marked according to the inspection coverage integrity.

3. The method for identifying building curtain wall defects based on deep learning according to claim 2, characterized in that: The building curtain wall spatial inspection sequence is subjected to sharpness detection, exposure detection, motion blur detection, perspective distortion correction, scale unification, and curtain wall boundary clipping to obtain standardized curtain wall image data. The standardized curtain wall image data is then input into the curtain wall component segmentation model to extract curtain wall edge texture features, component outline features, and component spatial semantic features. Based on the curtain wall edge texture features, component outline features, and component spatial semantic features, a spatial region map of the curtain wall components is generated.

4. The method for identifying building curtain wall defects based on deep learning according to claim 3, characterized in that: Standardized curtain wall image data and spatial area maps of curtain wall components are input into a deep learning defect detection model to extract visual features of defects. Based on the spatial region map of the curtain wall components, the visual characteristics of defects are constrained by the component regions to generate multi-view defect candidate regions. Based on the visual feature response value of the defect, the component region matching degree, the image quality score, and the intensity of reflective interference, the confidence level of the defect candidate region is calculated, and the confidence level of the defect candidate region is bound to the multi-view defect candidate region.

5. The method for identifying building curtain wall defects based on deep learning according to claim 4, characterized in that: Based on the spatial location, visual characteristics of the defects, spatial area of ​​the curtain wall component to which they belong, and differences in shooting angles, the cross-view defect matching score is calculated. The candidate regions of defects from multiple perspectives whose cross-perspective defect matching scores are higher than the matching score threshold are associated to generate a multi-perspective evidence set; a defect space evidence chain is generated based on the multi-perspective evidence set, and the integrity score of the evidence chain is calculated based on the number of candidate regions of defects, the completeness of perspective coverage, and the consistency of component regions in the defect space evidence chain.

6. The method for identifying building curtain wall defects based on deep learning according to claim 5, characterized in that: The defect type is identified based on the visual characteristics of the defect in the defect spatial evidence chain, the spatial region of the curtain wall component to which it belongs, and the multi-view evidence set. The location of the defect is calculated based on the building facade area number corresponding to the defect space evidence chain, the curtain wall component space area diagram, and the shooting posture data. The defect area and length are calculated based on the defect candidate region boundary, perspective distortion correction results, and scale unification results in the defect space evidence chain; the defect severity score is calculated based on the defect type, defect location, defect area, defect length, evidence chain integrity score, and component importance.

7. The method for identifying building curtain wall defects based on deep learning according to claim 6, characterized in that: Based on the confidence score of the candidate defect region, the cross-view defect matching score, and the evidence chain integrity score, the defect credibility score is calculated; the defect type, defect location, defect area, defect length, defect severity score, and defect credibility score are bound together to generate the building curtain wall defect identification result.

8. The method for identifying building curtain wall defects based on deep learning according to claim 7, characterized in that: When the defect confidence score is lower than the confidence threshold, the defect spatial evidence chain is marked as a low-confidence evidence chain; the supplementary sampling area is determined based on the defect location, building facade area number, and curtain wall component spatial area map corresponding to the low-confidence evidence chain; the supplementary sampling angle is determined based on the shooting angle difference, evidence chain integrity score, and reflection interference intensity in the low-confidence evidence chain; a supplementary sampling inspection instruction is generated based on the supplementary sampling area and supplementary sampling angle; supplementary curtain wall image data is obtained based on the supplementary sampling inspection instruction, and the defect spatial evidence chain and building curtain wall defect identification results are updated based on the supplementary curtain wall image data.

Images