Building quality dynamic evaluation method and system based on computer vision

Abstract

The invention provides a building quality dynamic evaluation method and system based on computer vision, and relates to the technical field of image processing, and the method comprises the steps: collecting a to-be-monitored surface image of a target building, and carrying out the high-recall change candidate region analysis of the to-be-monitored surface image and a historical reference image, traversing a plurality of candidate regions and a plurality of region change features to perform multi-model game verification, determining a reliable candidate region set, a change feature set and a consensus confidence coefficient set, generating a current building surface quality state diagram in combination with a spatial position, calling a historical building surface quality state diagram sequence to perform analysis and fitting, and obtaining a current building surface quality state diagram; and integrating the independent building quality evaluation result and the combined building quality evaluation result to obtain a building quality dynamic evaluation result. The technical problem that an existing building quality evaluation method is poor in building quality recognition accuracy and reliability is solved. The technical effects of accurately identifying fine and progressive building quality changes and improving the accuracy and reliability of a building quality evaluation result are achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method and system for dynamic evaluation of building quality based on computer vision. Background Technology

[0002] As urban building scale continues to expand, the structural safety and surface quality of buildings throughout their service life have gradually become important concerns in urban operation and maintenance management. Existing building quality inspection methods mostly rely on manual inspections or local non-destructive testing, which have problems such as long inspection cycles, strong subjectivity, and difficulty in achieving long-term continuous monitoring.

[0003] In recent years, computer vision-based building surface quality assessment methods have been gradually applied to defect detection scenarios such as cracks, peeling, and corrosion. However, existing technologies often use fixed threshold differential, single-scale features, or single models for change detection. When lowering the threshold to improve sensitivity, a large number of false changes caused by lighting changes, shooting noise, or environmental occlusion are introduced. When raising the threshold to suppress noise, it is easy to miss real subtle quality changes. It is difficult to balance change recall and detection reliability, resulting in poor stability of building quality assessment results and problems such as false detection, missed detection, or fluctuation in assessment results.

[0004] In summary, existing building quality assessment methods suffer from technical problems such as insufficient ability to distinguish subtle, gradual changes in quality, susceptibility to environmental interference and model instability, resulting in poor accuracy and reliability of identification results. Summary of the Invention

[0005] The purpose of this application is to provide a computer vision-based dynamic assessment method and system for building quality, which addresses the technical problems of existing building quality assessment methods, such as insufficient discriminability of subtle quality variations and susceptibility to environmental interference and model instability, resulting in poor accuracy and reliability of the identification results.

[0006] In view of the above problems, this application provides a method and system for dynamic evaluation of building quality based on computer vision.

[0007] The first aspect of this application provides a computer vision-based dynamic assessment method for building quality, comprising: periodically acquiring images of the target building's surface to be monitored using image acquisition equipment deployed at fixed monitoring points to obtain the current surface image to be monitored; performing high-recall change candidate region analysis based on the current surface image to be monitored and the previous historical baseline image to obtain multiple candidate regions and multiple region change features; traversing the multiple candidate regions and multiple region change features to perform multi-model game verification to determine a reliable candidate region set, a reliable region change feature set, and a consensus confidence set; aggregating the reliable candidate region set, the reliable region change feature set, and the consensus confidence, and combining the spatial location of each candidate region to generate a current building surface quality status map; performing independent building quality assessments based on the current building surface quality status map to obtain independent building quality assessment results, and retrieving historical building surface quality status map sequences from multiple historical monitoring periods to perform time-series analysis and trend fitting on the current building surface quality status map to obtain a joint building quality assessment result; and integrating the independent building quality assessment results and the joint building quality assessment results to obtain the dynamic assessment result of building quality.

[0008] Optionally, a dual-channel difference perception network is constructed to perceive the feature differences between the current surface image to be monitored and the previous historical baseline image, thereby obtaining a first-channel feature map and a second-channel feature map. An attention weight network layer is used to perform weighted fusion of the first-channel feature map and the second-channel feature map to obtain a difference heatmap. Based on the difference heatmap, high-recall change candidate regions are analyzed to obtain multiple candidate regions and multiple candidate region change features.

[0009] Optionally, the first channel of the dual-channel difference-aware network uses a Gabor filter bank to extract multi-directional, multi-scale texture features and calculates the residuals between these features and the previous historical baseline image; the second channel uses a deformable convolution module to capture local deformations for structural contour change analysis.

[0010] Optionally, the differential heatmap is traversed to perform adaptive dual-threshold segmentation to obtain multiple initial candidate regions and multiple initial candidate region change features; based on the positions of the multiple initial candidate regions, the historical state of the same position within a preset historical period is retrieved, and instantaneous interference authentication is performed in combination with the multiple candidate region change features. If the instantaneous interference authentication is passed, multiple candidate regions and multiple candidate region change features are added.

[0011] Optionally, a first difference threshold and a second difference threshold are obtained; strong change seed points are identified on the difference heatmap based on the first difference threshold to obtain multiple strong change seed points, wherein the first difference threshold is greater than the second difference threshold; region growing is performed on the multiple strong change seed points based on the second difference threshold to obtain multiple initial candidate regions; according to the difference heatmap, feature recognition is performed on the multiple initial candidate regions according to preset change indicators to obtain multiple initial candidate region change features, wherein the preset change indicators include morphological features, texture and color features, and positional context features.

[0012] Optionally, three heterogeneous pre-trained visual analysis models are selected, wherein the three heterogeneous pre-trained visual analysis models are respectively constructed based on Transformer, CNN and vision-language joint architecture, and a role decision head and an opinion output head are added to each of the three heterogeneous pre-trained visual analysis models; multiple game training sample pairs are constructed according to multiple candidate regions and multiple region change features, each game training sample including a real sample and a counterfactual sample; a first game training sample pair is extracted from the multiple game training sample pairs, and multi-model game verification is performed on the three heterogeneous pre-trained visual analysis models with the first game training sample as input to obtain a first game verification result; the first game verification result is added to the multiple game verification results, and the multiple candidate regions and multiple region change features are mapped and filtered according to the multiple game verification results to obtain a reliable candidate region set and a reliable region change feature set.

[0013] Optionally, based on a historical real defect sample library, defect morphology and texture parameters are extracted and categorized to establish a defect parameterization template library; multiple regional change features are matched with multiple matching defect parameterization templates from the defect parameterization template library; counterfactual perturbations are applied to the multiple matching defect parameterization templates to obtain multiple counterfactual samples; the multiple regional change features are used as multiple real samples, and the multiple counterfactual samples are jointly paired to obtain multiple game training sample pairs.

[0014] Optionally, multiple counterfactual samples are obtained by traversing the multiple matching defect parameterized templates to extend the crack morphology along the principal stress direction with pixel-level steps, diffuse the rust texture in the selected corrosion center according to the exponential decay law, and randomly perturb the mean and variance of the color channel in the stain area.

[0015] Optionally, counterfactual samples from the first game training samples are randomly assigned to any one of the three heterogeneous pre-trained visual analysis models, and reverse-labeled to obtain a reverse-labeled heterogeneous pre-trained visual analysis model. Real samples from the first game training samples are assigned to the other two heterogeneous pre-trained visual analysis models, and conventionally labeled to obtain two conventionally labeled heterogeneous pre-trained visual analysis models. The reverse-labeled heterogeneous pre-trained visual analysis model and the two conventionally labeled heterogeneous pre-trained visual analysis models, based on the assigned samples, evaluate the realism of change features through their respective role decision heads, and output the evaluation results through their respective opinion output heads, obtaining three evaluation results. These three evaluation results are synchronized to the reverse-labeled heterogeneous pre-trained visual analysis model and the two conventionally labeled heterogeneous pre-trained visual analysis models for cross-questioning, obtaining cross-questioning voting results. If the cross-questioning evaluation result is for the reverse-labeled heterogeneous pre-trained visual analysis model, then the two evaluation results corresponding to the two conventionally labeled heterogeneous pre-trained visual analysis models are true. The two evaluation results are averaged to obtain the first game verification result.

[0016] A second aspect of this application provides a computer vision-based dynamic assessment system for building quality, comprising: an image acquisition module for periodically acquiring images of the target building's surface to be monitored using image acquisition devices deployed at fixed monitoring points, to obtain the current surface image to be monitored; an image analysis module for performing high-recall change candidate region analysis based on the current surface image to be monitored and the previous historical baseline image, to obtain multiple candidate regions and multiple region change features; and a feature verification module for traversing the multiple candidate regions and multiple region change features to perform multi-model game verification, to determine a reliable candidate region set, a reliable region change feature set, and a consensus setting. The system comprises: a reliability set; a quality status map generation module, used to aggregate a reliable candidate region set, a reliable region change feature set, and consensus confidence, and combine the spatial location of each candidate region to generate a current building surface quality status map; a quality assessment module, used to perform independent building quality assessments based on the current building surface quality status map, obtain independent building quality assessment results, and retrieve historical building surface quality status map sequences from multiple historical monitoring periods to perform time-series analysis and trend fitting on the current building surface quality status map to obtain joint building quality assessment results; and a dynamic assessment result acquisition module, used to integrate the independent building quality assessment results and the joint building quality assessment results to obtain the dynamic building quality assessment results.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application uses image acquisition equipment deployed at fixed monitoring points to periodically acquire images of the target building's surface to be monitored, obtaining the current surface image to be monitored. Based on the current surface image to be monitored and the previous historical baseline image, high-recall change candidate region analysis is performed to obtain multiple candidate regions and multiple region change features. Multi-model game verification is performed on the multiple candidate regions and multiple region change features to determine a reliable candidate region set, a reliable region change feature set, and a consensus confidence set. The reliable candidate region set, reliable region change feature set, and consensus confidence are aggregated, and combined with the spatial location of each candidate region, to generate a current building surface quality status map. Independent building quality assessments are performed based on the current building surface quality status map to obtain independent building quality assessment results. Historical building surface quality status map sequences from multiple historical monitoring periods are retrieved, and time-series analysis and trend fitting are performed on the current building surface quality status map to obtain a joint building quality assessment result. The independent building quality assessment results and the joint building quality assessment results are integrated to obtain the dynamic building quality assessment result. This achieves the technical effect of accurately identifying subtle and gradual changes in building quality, improving the accuracy and reliability of building quality assessment results.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the computer vision-based dynamic assessment method for building quality provided in this application.

[0021] Figure 2 A schematic diagram of the structure of the computer vision-based dynamic assessment system for building quality provided in this application.

[0022] Figure labeling: Image acquisition module 11, Image parsing module 12, Feature verification module 13, Quality status map generation module 14, Quality assessment module 15, Dynamic assessment result acquisition module 16. Detailed Implementation

[0023] This application provides a computer vision-based dynamic assessment method and system for building quality, addressing the technical problems of existing building quality assessment methods, such as insufficient discriminability of subtle, gradual changes in quality, and susceptibility to environmental interference and model instability, leading to poor accuracy and reliability of identification results. It achieves the technical effect of accurately identifying subtle, gradual changes in building quality, thereby improving the accuracy and reliability of building quality assessment results.

[0024] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0025] Example 1, as Figure 1 As shown, this application provides a computer vision-based dynamic assessment method for building quality, which includes: By deploying image acquisition equipment at fixed monitoring points, images of the target building's surface to be monitored are periodically acquired to obtain the current surface image to be monitored.

[0026] Specifically, fixed monitoring points are rationally planned and deployed around the target building. These fixed monitoring points should comprehensively cover all key surface areas of the target building to be monitored, avoiding blind spots. Simultaneously, the locations of the monitoring points should be relatively stable and not easily affected by external factors such as construction or vandalism, ensuring the accuracy and consistency of subsequently acquired images. For example, fixed monitoring points can be set up in the middle of the four facades and at the corners of the target high-rise residential building.

[0027] Image acquisition equipment is installed at fixed monitoring points. High-definition cameras with a resolution of at least 1080P are suitable for this purpose, enabling clear capture of detailed information on building surfaces, such as minute cracks and peeling. Based on a pre-set acquisition cycle, the equipment automatically acquires images of the target building's surface to be monitored, obtaining the image of the surface to be monitored within the current monitoring cycle. The pre-set acquisition cycle can be determined empirically based on factors such as building type, age, and environment. For example, for buildings that age rapidly and experience frequent quality changes, a shorter acquisition cycle, such as once a day, can be set. For buildings with relatively stable quality and slow changes, the acquisition cycle can be appropriately extended, such as once a week.

[0028] By establishing fixed monitoring points and a periodic acquisition mechanism, a foundation of building surface image data with spatial consistency and temporal continuity is constructed, effectively reducing the interference of perspective drift and imaging randomness on change analysis and improving the effectiveness and reliability of the surface images to be monitored.

[0029] Based on the current expected monitoring surface image and the previous historical baseline image, high recall change candidate region analysis is performed to obtain multiple candidate regions and multiple region change features.

[0030] Furthermore, based on the current expected monitoring surface image and the previous historical baseline image, high-recall change candidate region analysis is performed to obtain multiple candidate regions and multiple region change features, including: constructing a dual-channel difference perception network, using the dual-channel difference perception network to perceive feature differences between the current expected monitoring surface image and the previous historical baseline image, obtaining a first channel feature map and a second channel feature map; using an attention weight network layer to perform weighted fusion of the first channel feature map and the second channel feature map to obtain a difference heatmap; and based on the difference heatmap, high-recall change candidate region analysis is performed to obtain the multiple candidate regions and multiple candidate region change features.

[0031] Furthermore, the first channel of the dual-channel difference-aware network uses a Gabor filter bank to extract multi-directional, multi-scale texture features and calculates their residuals with the previous historical baseline image; the second channel uses a deformable convolution module to capture local deformations for structural contour change analysis.

[0032] Specifically, high-recall change candidate region analysis refers to the process of analyzing acquired building surface images. Through feature difference perception and heatmap generation, it aims to identify all regions with potential structural or textural changes as completely as possible, without immediately distinguishing between true and false changes, thus ensuring that early, minor, or gradual defects are not missed. Feature difference perception is achieved by constructing a dual-channel difference perception network, which includes two feature extraction channels with different functions, processed in parallel. The first channel uses a Gabor filter bank to extract multi-directional, multi-scale texture features. Based on the complexity of the building surface texture and the required monitoring accuracy, the direction and scale parameters of the Gabor filter are determined. For example, five different directions are set: 0°, 45°, 90°, 135°, and 180°, to comprehensively capture textures in all directions. Three different scales are selected: small, medium, and large scales, to analyze textures of varying fineness. Small scales detect minute crack textures, while large scales reflect macroscopic features such as surface roughness. Based on the mathematical expression of the Gabor function, which includes variables such as direction and scale parameters, the corresponding values ​​are substituted into the Gabor function for each given combination of direction and scale. For example, when the direction is set to 0° and the scale is small, these two parameter values ​​are substituted into the function to obtain the kernel function expression of the Gabor filter for that specific direction and scale. In this way, all combinations of 5 directions and 3 scales are substituted and calculated, ultimately generating 15 (5×3) Gabor filter banks with different directions and scales for comprehensively extracting the texture features of building surfaces.

[0033] After extracting the texture features of the first channel, the texture feature matrix extracted from the current period image is subtracted element-wise from the corresponding texture feature matrix extracted from the previous period's historical reference image. The previous period's historical reference image refers to a building surface image collected and quality-confirmed in the previous monitoring period, used as a reference benchmark for change detection in the current period. The texture feature matrix is ​​a matrix formed by extracting multi-directional, multi-scale texture features from the surface image to be monitored using a Gabor filter bank, and then organizing the filter response results according to certain rules. If the texture feature matrix of the current period image is A, and the texture feature matrix of the previous period's historical reference image is B, the residual matrix R is calculated as R = A − B. Each element value in the residual matrix represents the degree of difference between the current period image and the previous period image at the corresponding texture feature position. If the residual value at a certain position is large, it indicates that the texture feature at that position has changed significantly compared to the previous period; if the residual value is close to zero, it indicates that the texture feature at that position has changed little. By calculation, a residual matrix reflecting the degree of difference in texture between the image to be monitored and the previous historical baseline image is obtained, and a first channel feature map is obtained. The first channel feature map reflects the change information of the image to be monitored from the perspective of texture.

[0034] The second channel employs a deformable convolution module to capture local deformations, used for analyzing structural contour changes in architectural images, such as crack propagation, edge shifting, and structural deformation caused by stress on building surfaces. The deformable convolution module is an improvement on a conventional convolutional network. Its network structure adds an offset generation branch before the standard convolutional layer. This offset generation branch consists of conventional convolutional operations and receives the current image to be monitored and the previous historical baseline image as input. Through convolution operations, an offset field with the same size as the input image is learned and generated. The value of each element in the offset field represents the offset distance and direction of the sampling point of the convolution kernel at the corresponding position. Based on the generated offset field, the standard convolutional layer breaks the limitation of the fixed sampling position of the traditional convolution kernel and adaptively adjusts the sampling point position of the convolution kernel. This allows the convolution kernel to flexibly sample the surrounding area according to the differences in image content, such as the local deformation and structural contour changes of the building surface in the current cycle compared to the previous cycle. This better captures these changing features, enhances the ability to analyze complex building surface structures, and generates a second channel feature map. The second channel feature map reflects the spatial feature information of the local structural deformation and contour changes of the building surface from the perspective of deformation.

[0035] The first and second channel feature maps are input into the attention weight network layer, which performs weighted fusion of the two feature maps. The attention weight network layer employs an encoder-decoder structure. In the encoder section, multiple convolutional blocks are stacked in an ordered manner. Each convolutional block contains a convolutional layer, a batch normalization layer, and an activation function layer. The convolutional layers use 3×3 kernels with a stride of 1 and the same padding method. Through convolution, preliminary feature extraction is performed on the input feature map, capturing feature patterns in local regions, such as the local direction of surface textures in buildings and the local shape of structural contours. The batch normalization layer normalizes the convolutional feature map, adjusting the feature value distribution of each channel to a range with a mean of 0 and a variance of 1. This accelerates the training convergence speed of the attention weight network and improves the model's generalization ability. The activation function layer uses the ReLU function, introducing a non-linear factor to accelerate the learning of more complex feature relationships by the attention weight network.

[0036] Attention mechanisms are introduced in the middle or at the end of the encoder, including channel attention and spatial attention modules. The channel attention module first averages the feature values ​​of each channel using global average pooling to obtain a global feature vector. This vector contains global information for each channel, reflecting its overall feature performance across the entire feature map. Then, a fully connected layer is used to learn the global feature vector. The number of neurons in the fully connected layer can be set according to actual needs, for example, 32. Through non-linear transformations in the fully connected layer, the weights of each channel are learned, highlighting the features of important channels. The spatial attention module uses convolution operations, employing 1×1 convolution kernels on the feature map. The number of convolution kernels can be set as needed, for example, one. After convolution, a spatial attention weight map is generated. Each value in the spatial attention weight map represents the importance of the corresponding spatial location in the feature map. In building surface images, areas with significant changes, such as new cracks or shifts in structural contours, will have higher values ​​in the spatial attention weight map, thus emphasizing these important spatial locations. The decoder consists of transposed convolutional layers or upsampling layers. The transposed convolutional layers use 3×3 kernels with a stride of 2 and the same padding method. The transposed convolution operation restores the high-dimensional feature map to the size of the original feature map. Simultaneously, the learned attention weights are used to weightedly fuse the features. The fused feature map is then normalized so that its values ​​range from 0 to 1, with darker colors indicating greater variation. This generates a difference heatmap, visually displaying the areas and degrees of difference between the surface image to be monitored and the previous historical baseline image, including texture differences and structural deformation information.

[0037] By using a dual-channel differential sensing network to analyze the current monitoring image and the previous historical baseline image, it is possible to simultaneously capture and accurately locate subtle changes in texture and structural contour deformation of the building surface, achieving highly sensitive detection of progressive degradation. This improves the accuracy and reliability of the overall dynamic assessment of building quality, timely detects potential quality problems on the building surface, and has broad adaptability.

[0038] Furthermore, based on the difference heatmap, high-recall change candidate region analysis is performed to obtain multiple candidate regions and multiple candidate region change features, including: traversing the difference heatmap to perform adaptive dual-threshold segmentation to obtain multiple initial candidate regions and multiple initial candidate region change features; based on the positions of the multiple initial candidate regions, historical states of the same positions within a preset historical period are retrieved; instantaneous interference authentication is performed in combination with the multiple candidate region change features; if the instantaneous interference authentication is passed, the multiple candidate regions and multiple candidate region change features are added.

[0039] Furthermore, adaptive dual-threshold segmentation is performed by traversing the difference heatmap to obtain multiple initial candidate regions and multiple initial candidate region change features, including: obtaining a first difference threshold and a second difference threshold; identifying strong change seed points in the difference heatmap based on the first difference threshold to obtain multiple strong change seed points, wherein the first difference threshold is greater than the second difference threshold; performing region growing on the multiple strong change seed points based on the second difference threshold to obtain multiple initial candidate regions; and performing feature recognition on the multiple initial candidate regions according to the difference heatmap and preset change indicators to obtain multiple initial candidate region change features, wherein the preset change indicators include morphological features, texture and color features, and positional context features.

[0040] Specifically, a first difference threshold and a second difference threshold are set, with the first difference threshold being greater than the second difference threshold. The difference threshold is a boundary value used to measure the degree of change of pixel values ​​in the difference heatmap, and is used to accurately divide areas with different degrees of change. It can be based on the statistical distribution of pixel values ​​in the difference heatmap, for example, by analyzing the average change value and standard deviation of non-defect areas in historical monitoring periods to determine the adaptive threshold, or it can be set based on historical data experience.

[0041] The difference heatmap is traversed based on a first difference threshold to identify strong change seed points. Strong change seed points are those whose pixel values ​​in the difference heatmap are greater than or equal to the first difference threshold, indicating a significant change at that location. Multiple strong change seed points are obtained through this traversal. Based on a second difference threshold, a region growing operation is performed on these strong change seed points. Region growing is an image segmentation method that uses a strong change seed point as the center and merges adjacent pixels with similar characteristics into the region where the seed point is located, thus obtaining multiple initial candidate regions. For example, in a building surface difference heatmap, the first difference threshold is set to 0.8, and the second difference threshold is set to 0.5. When the pixel value of a certain pixel in the difference heatmap is 0.85, it exceeds the first difference threshold and is identified as a strong change seed point. The determination of whether adjacent pixels have similar features to the strong change seed point is based on the relationship between the pixel value in the difference heatmap and the second difference threshold. For a certain adjacent pixel around the strong change seed point, if its pixel value is greater than or equal to 0.5, it is considered that the adjacent pixel has similar features to the strong change seed point and is merged into the region where the seed point is located. By continuously repeating this process, all adjacent pixels that meet the conditions are merged to obtain multiple initial candidate regions.

[0042] Based on the difference heatmap, feature recognition is performed on multiple initial candidate regions according to preset change indicators. These indicators include morphological features, texture and color features, and location context features. Morphological feature recognition includes region area, perimeter, aspect ratio, density, and the direction of the minimum bounding rectangle. Texture and color feature recognition involves calculating the contrast and homogeneity of the gray-level co-occurrence matrix of the corresponding initial candidate regions in the difference heatmap, and calculating the mean shift relative to the reference image in the Lab color space. Location context feature recognition includes the proportional position of the center coordinates of the initial candidate regions relative to the entire image. The initial candidate regions are compared with pre-stored structural diagrams to analyze their closest distance to known key structures, such as beam-column joints and window edges. By performing feature recognition on multiple initial candidate regions, the change characteristics of these regions are obtained.

[0043] To eliminate false changes caused by lighting, shadows, or other environmental interference, based on the locations of multiple initial candidate regions, historical states of the same locations are retrieved within a preset historical period. For example, the change characteristics of the current initial candidate region are compared with the historical states of the same location within the past three periods to perform transient interference authentication. Transient interference refers to changes that appear in the current period but did not persist in the past three periods, exhibiting a fleeting appearance and disappearance. If a candidate region is determined to be transient interference, it is filtered out. Authentication-compliant candidate regions are added to the multiple candidate regions and their change characteristics, resulting in multiple candidate regions and their change characteristics.

[0044] By combining dual-threshold segmentation with region growing and historical state retrieval, high recall analysis of potential change areas in the difference heatmap is achieved. At the same time, transient interference is eliminated, effectively removing interfering factors and improving the accuracy and comprehensiveness of candidate region recall. This provides a reliable data foundation for subsequent analysis and judgment of the actual changes on building surfaces, and helps to more accurately monitor changes on building surfaces.

[0045] The multiple candidate regions and multiple region change features are traversed to perform multi-model game verification, and a reliable candidate region set, a reliable region change feature set, and a consensus confidence set are determined.

[0046] Furthermore, multi-model game validation is performed by traversing the multiple candidate regions and multiple region change features to determine a reliable candidate region set, a reliable region change feature set, and a consensus confidence set. This includes: selecting three heterogeneous pre-trained visual analysis models, wherein the three heterogeneous pre-trained visual analysis models are respectively constructed based on Transformer, CNN, and a vision-language joint architecture, and adding a role decision head and an opinion output head to each of the three heterogeneous pre-trained visual analysis models; constructing multiple game training sample pairs according to the multiple candidate regions and multiple region change features, each game training sample including a real sample and a counterfactual sample; extracting a first game training sample pair from the multiple game training sample pairs, using the first game training sample as input, performing multi-model game validation on the three heterogeneous pre-trained visual analysis models to obtain a first game validation result; adding the first game validation result to the multiple game validation results, and mapping and filtering the multiple candidate regions and multiple region change features according to the multiple game validation results to obtain a reliable candidate region set and a reliable region change feature set.

[0047] Specifically, three heterogeneous pre-trained visual analysis models were selected. Heterogeneous means that the three models can analyze the input data from different perspectives. The three heterogeneous pre-trained visual analysis models are built based on Transformer, CNN, and a vision-language joint architecture, respectively. Among them, the heterogeneous pre-trained visual analysis model based on Transformer is used to capture long-distance dependencies and global features in the data, and can better understand the relationship between different parts of the image when processing images. The heterogeneous pre-trained visual analysis model based on CNN (Convolutional Neural Network) can effectively identify local information such as texture and edges in the image through the sliding operation of convolution kernels on the image. The heterogeneous pre-trained visual analysis model based on vision-language joint architecture integrates information from both visual and linguistic modalities, and can use language descriptions to provide additional semantic guidance for visual analysis, enhancing the analysis model's understanding of image content and judgment of the rationality of changes. Three heterogeneous pre-trained visual analysis models are arranged in parallel. Each heterogeneous pre-trained visual analysis model independently processes the input data and generates preliminary results. Each heterogeneous pre-trained visual analysis model is attached with a role decision head and an opinion output head. The role decision head is used to evaluate the authenticity or significance of changes in the input sample, and the opinion output head is used to output evaluation opinions on candidate regions or change features.

[0048] The input data consists of high-recall candidate regions of change and their change features. Each candidate region image, along with its texture, morphology, color, and location context information, forms a training sample. To improve the robustness of the three heterogeneous pre-trained visual analysis models in judging the authenticity of changes, each training sample is constructed as a game-theoretic training sample pair, containing one real sample and one counterfactual sample. The real sample refers to a real defect or unchanged region, while the counterfactual sample refers to pseudo-changes generated through a defect parameterization template, such as crack propagation, local corrosion, or color perturbation. In the Transformer model training, the candidate region image first extracts global features through a multi-layer self-attention encoder to capture long-distance dependencies between regions. The encoded features are then input into the role decision head for authenticity scoring, and into the opinion output head to output the evaluation opinion for the candidate region. The CNN model extracts local texture and morphological features through multi-layer convolutional and pooling layers, and similarly outputs authenticity scores and opinions through the role decision head and opinion output head. The visual-language joint model inputs the image into the visual encoder and the semantic description of the candidate region into the text encoder. A joint feature representation is generated through a visual-language alignment mechanism, and then the evaluation result is output through the role decision head and opinion output head.

[0049] During training, each heterogeneous pre-trained visual analytics model's role decision head uses a cross-entropy loss function to supervise the training of truth labels, while the opinion output head maintains consistency with other model outputs through a consistency constraint loss. In the game-theoretic training mechanism, counterfactual samples are randomly assigned to one heterogeneous pre-trained visual analytics model as reverse labels, while the other two models receive true samples. The heterogeneous pre-trained visual analytics models cross-question through their role decision heads and opinion output heads, achieving multi-model game-theoretic training. Parameters are iteratively updated until the three heterogeneous pre-trained visual analytics models achieve convergence in their ability to distinguish between true and counterfactual samples. After training, the three heterogeneous pre-trained visual analytics models process candidate regions in parallel during the inference phase. Each model independently generates truth judgments and opinion outputs. Through a multi-model game-theoretic verification mechanism, the truthfulness and consistency of changes in candidate regions are evaluated, yielding a first game-theoretic verification result. This first game-theoretic verification result includes each heterogeneous pre-trained visual analytics model's support or opposition attitude towards candidate regions and region change features, as well as corresponding confidence information.

[0050] The first game validation result is added to multiple game validation results. These multiple game validation results contain information related to both reliable and unreliable candidate regions generated during all previous game processes. Multiple candidate regions and their change features are mapped and filtered according to these multiple game validation results. Candidate regions and their change features that are consistently recognized by multiple models and have high confidence are selected, forming a set of reliable candidate regions and a set of reliable region change features. Simultaneously, a consensus confidence set is obtained based on the consistency of the three heterogeneous pre-trained visual analysis models during the game process. This consensus confidence set is used to measure the credibility of each reliable candidate region and its change features.

[0051] By introducing multiple heterogeneous pre-trained visual analysis models for game-theoretic validation, the advantages of different models are fully utilized. Candidate regions and their changing characteristics are analyzed and judged from multiple perspectives, effectively avoiding the biases and limitations that may exist with a single model, and significantly improving the accuracy and reliability of the detection results. Simultaneously, constructing game-theoretic training sample pairs and performing mapping filtering effectively eliminates spurious changes caused by lighting, noise, or model bias, thereby more accurately identifying the truly changed regions and features, and improving the accuracy and stability of dynamic building quality assessment.

[0052] Furthermore, multiple game training sample pairs are constructed based on multiple candidate regions and multiple region change features. Each game training sample includes one real sample and one counterfactual sample, including: extracting and classifying defect morphology and texture parameters based on a historical real defect sample library, and establishing a defect parameterization template library; matching multiple region change features with multiple matching defect parameterization templates from the defect parameterization template library; performing counterfactual perturbation on the multiple matching defect parameterization templates to obtain multiple counterfactual samples; and using the multiple region change features as multiple real samples, jointly pairing the multiple counterfactual samples to obtain multiple game training sample pairs.

[0053] Furthermore, the method also includes: traversing the multiple matching defect parameterized templates to extend the crack morphology along the principal stress direction with pixel-level steps, diffusing the rust texture at the selected corrosion center according to an exponential decay law, and randomly perturbing the mean and variance of the color channels in the stain area to obtain multiple counterfactual samples.

[0054] Specifically, the historical real-world defect sample library collects a large number of images and related data of various defects in actual building scenarios. Based on the historical real-world defect sample library accumulated through monitoring, defects are classified according to morphological and texture parameters. Defect morphology refers to the external manifestation of a defect in an image, such as the tortuous direction of a crack or the irregular shape of rust. Texture parameters are used to describe the texture features of the defect surface, such as roughness and particle distribution. By classifying these parameters, a defect parameterization template library is established, where each template records key parameters such as crack morphology, rust distribution, and surface stain characteristics.

[0055] The system quantifies and characterizes multiple regional variation features and templates from a defect parameterization template library. For defect morphological features, geometric parameters of their contours, such as aspect ratio and curvature, can be extracted. For texture features, gray-level co-occurrence matrix is ​​used to calculate statistical quantities such as contrast and correlation. For color features, the mean and variance of each Lab channel are calculated. In the Lab color space, the L channel represents brightness, the a channel represents the range from red to green, and the b channel represents the range from yellow to blue. For example, the morphological aspect ratio of a certain region is 2.5, the curvature is 0.3, the texture contrast is 0.6, the correlation is 0.4, and the Lab color mean is (L:60, a:15, b:-10) and the variance is (L:20, a:10, b:8). In the defect parameterization template library, a similarity metric is used to calculate the similarity between the feature vector of the candidate region and the feature vector of each template in the library. Commonly used similarity metrics include Euclidean distance, cosine similarity, or weighted distance. For example, weighted distance is used to calculate the absolute values ​​of differences in aspect ratio, curvature, and texture contrast, and then these differences are summed with weights to obtain a comprehensive difference value. The weights for morphological features are set to 0.4, texture features to 0.4, and color features to 0.2. By calculating the comprehensive difference value, multiple templates with comprehensive difference values ​​less than a threshold (e.g., 0.3) are selected as the most similar defect parameterization templates to the region's variation features. Alternatively, the templates can be sorted from high to low similarity, and the top k templates are selected to form the set of matching defect parameterization templates for the region, where k is an integer greater than or equal to 1.

[0056] Counterfactual perturbations are applied to multiple matched defect parameterized templates. These perturbations are used to generate samples that correspond to real samples but have different characteristics yet are logically related. The counterfactual perturbation operation involves traversing multiple matched defect parameterized templates and extending the crack morphology along the principal stress direction with pixel-level steps. The principal stress direction is the direction in materials mechanics that causes the main deformation and failure of an object. Extending the crack along this direction can simulate the propagation process under actual stress. If the initial length of the crack template is 50 pixels, and it extends along the principal stress direction with a step size of 2 pixels, extending 10 times, then the length of the extended crack = 50 + 10 × 2 = 70 pixels. In this way, counterfactual samples of cracks with different lengths and shapes can be generated to simulate the propagation of cracks under different stress stages or different material properties.

[0057] The rust texture diffuses according to an exponential decay law from the selected corrosion center. This exponential decay law accurately reflects the process of rust spreading from the center outwards at a gradually decreasing rate. Let the rust degree I0 of the pixel at the corrosion center be 1, and the rust degree I at a distance r from the corrosion center be... （r） =I0×e -λrλ is the attenuation coefficient, which controls the speed and range of rust diffusion. If λ is 0.1, and r is 10 pixels, the degree of rust is approximately 0.37. By adjusting the value of the attenuation coefficient, counterfactual samples of rust textures with different diffusion degrees can be generated to simulate rust conditions under different environmental conditions or material corrosion resistance.

[0058] The mean and variance of the color channels are randomly perturbed within the stained area. The mean of the color Lab is (L:60, a:15, b:-10) and the variance is (L:20, a:10, b:8). After adding the random perturbation, the color distribution of the stained area is regenerated based on the new mean and variance. A new set of mean and variance is randomly generated to generate counterfactual samples with different color characteristics, simulating different types and degrees of stains.

[0059] By performing counterfactual perturbations on crack morphology, rust texture, and stain color, multiple counterfactual samples under various conditions can be generated based on multiple matched defect parameterization templates. Multiple regional variation features are used as multiple real samples, and these counterfactual samples are jointly paired to obtain multiple game-theoretic training sample pairs. This pairing ensures that each real sample has a corresponding counterfactual sample for comparison, which is then used to train a multi-model game validation network.

[0060] By constructing game training sample pairs, we provide comparative samples that include both real changes and simulated pseudo-changes for multi-model game verification. This enables the model to learn to distinguish between real defects and potential interference during training, thereby improving the accuracy and robustness of candidate region authenticity judgment and providing a reliable input basis for dynamic assessment of building quality.

[0061] Further, a first game training sample pair is extracted from multiple game training sample pairs. Using the first game training sample as input, multi-model game verification is performed on the three heterogeneous pre-trained visual analysis models to obtain the first game verification result. This includes: randomly assigning counterfactual samples from the first game training sample to any one of the three heterogeneous pre-trained visual analysis models and performing reverse labeling on them to obtain a reverse-labeled heterogeneous pre-trained visual analysis model; assigning real samples from the first game training sample to the other two heterogeneous pre-trained visual analysis models and performing regular labeling on them respectively to obtain two regular-labeled heterogeneous pre-trained visual analysis models; reverse labeling... The heterogeneous pre-trained visual analysis model and two conventional heterogeneous pre-trained visual analysis models are evaluated for the authenticity of change features based on the assigned samples through the corresponding role decision heads, and the evaluation results are output through the corresponding opinion output heads, resulting in three evaluation results. The three evaluation results are then synchronized to the reverse heterogeneous pre-trained visual analysis model and the two conventional heterogeneous pre-trained visual analysis models for cross-questioning, resulting in cross-questioning voting results. If the cross-questioning evaluation result is for the reverse heterogeneous pre-trained visual analysis model, then the two evaluation results corresponding to the two conventional heterogeneous pre-trained visual analysis models are true. The two evaluation results are then averaged to obtain the first game verification result.

[0062] Specifically, the reverse-label heterogeneous pre-trained visual analysis model and two conventional-label heterogeneous pre-trained visual analysis models evaluate the authenticity of change features based on assigned samples through their respective role decision heads. For example, assuming the three models are Model A, Model B, and Model C, if a counterfactual sample is randomly assigned to Model A, then Model A acts as the reverse-label heterogeneous pre-trained visual analysis model, while Models B and C act as conventional-label heterogeneous pre-trained visual analysis models. They analyze and evaluate the samples through role decision heads, and then output their judgments on the authenticity of changes in candidate regions through their corresponding opinion output heads, thus obtaining three independent evaluation results.

[0063] The three evaluation results are synchronized to the reverse-label heterogeneous pre-trained visual analysis model and two conventional-label heterogeneous pre-trained visual analysis models, triggering a cross-questioning mechanism. Cross-questioning refers to allowing the three heterogeneous pre-trained visual analysis models to examine and judge each other's evaluation results, simulating a game-theoretic verification process. For example, models A, B, and C respectively give evaluation results R. A R B R CModels B and C will challenge the results of Model A, and Model A will also challenge the results of Models B and C, ultimately resulting in a cross-judgment voting result. If the cross-judgment evaluation result points to the reverse-label heterogeneous pre-trained visual analysis model, it means that the evaluations of the other two conventional-label heterogeneous pre-trained visual analysis models are more credible, i.e., the two evaluation results corresponding to the two conventional-label heterogeneous pre-trained visual analysis models are true. In this case, the average of these two evaluation results is calculated. If R... B =0.9, R C =0.8, and the average value obtained from the first game verification result is R = 0.85.

[0064] If the cross-questioning voting result is a heterogeneous pre-trained visual analysis model with any conventional identifier, then the game is considered to have failed to reach a clear consensus. In this case, the opinions of all models across all rounds are statistically analyzed, and a weighted fusion algorithm is used to generate a conclusion. The model weights are normalized based on the number of times each model was suspected of being a wrong judge during the game; their weights are inversely proportional to the number of times they were suspected. The normalization process ensures that the sum of all model weights is 1, meaning that the less suspected a model is, the higher its opinion weight. The weighted fusion algorithm then synthesizes the evaluation opinions output by each model across all rounds, validating the first game result.

[0065] By employing a multi-model game-based verification approach, the advantages of different heterogeneous pre-trained visual analysis models are fully utilized. This allows for mutual supervision and verification between models, effectively improving the accuracy and reliability of evaluating the authenticity of sample change features. It avoids the biases and limitations that may exist in a single model, provides credible quantitative results for candidate region change features, and enhances the reliability of identifying subtle architectural changes and the stability of evaluation results.

[0066] By aggregating a set of reliable candidate regions, a set of reliable region change characteristics, and consensus confidence, and combining the spatial location of each candidate region, a current building surface quality status map is generated.

[0067] Independent building quality assessments are conducted based on the current building surface quality status map to obtain independent building quality assessment results. Furthermore, historical building surface quality status map sequences from multiple historical monitoring periods are retrieved, and time-series analysis and trend fitting are performed on the current building surface quality status map to obtain joint building quality assessment results.

[0068] By integrating the results of independent building quality assessments and the joint building quality assessments, the dynamic building quality assessment results are obtained.

[0069] Specifically, the process involves aggregating the set of reliable candidate regions, the corresponding set of reliable region change features, and the consensus confidence score for each region. The reliable region change features corresponding to each reliable candidate region are then normalized to quantify their significance. Finally, the consensus confidence score is used as a weight to weight and fuse these features into the feature values, resulting in a comprehensive quality index for each region. After this aggregation, the aggregated feature values ​​are mapped to the corresponding regions on a pre-constructed building surface image template according to predefined rules, based on the spatial location information of the candidate regions on the building surface and their coordinates, generating a current building surface quality status map. For example, different feature value ranges correspond to different colors or markers. Regions with higher aggregated feature values, indicating poorer quality, are marked in a striking red, while regions with lower feature values, indicating better quality, are marked in green. Simultaneously, during the marking process, the transparency of the markers is adjusted based on the consensus confidence score; regions with low confidence scores are marked with higher transparency to visually reflect the credibility of the evaluation results.

[0070] Reliable candidate regions in the quality status map are accurately identified and counted to determine the number of candidate regions existing on the building surface within the current period. The distribution of these candidate regions on the building surface is analyzed, and spatial analysis algorithms are used to determine characteristics such as the concentration and uniformity of their distribution. For example, the variance of the candidate region distribution is calculated to measure the dispersion. Simultaneously, for each candidate region, reliable region variation characteristics are extracted, such as color variation amplitude and texture roughness variation. The average variation degree of all candidate regions is then calculated. Based on the number of candidate regions, distribution characteristics, and average variation degree, and compared with pre-set building quality assessment standards (e.g., specifying different quality levels for candidate regions exceeding a preset threshold, excessively concentrated distribution, or excessive variation), an independent assessment result is obtained. This independent building assessment result can intuitively reflect the overall quality status of the building surface in the current period.

[0071] Multiple historical monitoring periods' building surface quality status map sequences were retrieved, and these maps were aligned with the current building surface quality status maps and mapped onto a unified coordinate system. The change trajectory of each candidate region and its neighborhood was analyzed over time. The analysis included the rate of regional expansion, cumulative change, and direction of change trends. The rate of regional expansion refers to the growth rate of the changed area over time; the cumulative change refers to the cumulative intensity of long-term change; and the direction of change trends included crack propagation and corrosion spread. A time-series weighted average was used to fit the change trends of each region, and a comprehensive joint building quality assessment result reflecting long-term evolution characteristics was generated, comprehensively and dynamically reflecting the dynamic changes and development trends of building surface quality over time.

[0072] By integrating independent and joint building quality assessment results, a dynamic building quality assessment result is obtained. This dynamic assessment result includes direct quality information of the building surface in the current period and quality change trend information in historical periods. It can comprehensively and dynamically reflect the building quality status, accurately and timely reflect local defects and subtle changes, and reveal the evolution trend of building surface quality over time. This provides a comprehensive and quantitative basis for building maintenance, risk warning, and preventive management, thereby improving the pertinence and effectiveness of maintenance and repair strategies and ensuring the safety and durability of buildings.

[0073] Example 2, based on the same inventive concept as the computer vision-based dynamic building quality assessment method in the foregoing examples, such as... Figure 2 As shown, this application provides a computer vision-based dynamic assessment system for building quality, wherein the computer vision-based dynamic assessment system for building quality includes: Image acquisition module 11 is used to periodically acquire images of the target building's surface to be monitored using image acquisition equipment deployed at fixed monitoring points, thereby obtaining the current surface image to be monitored. Image analysis module 12 is used to perform high-recall change candidate region analysis based on the current surface image to be monitored and the previous historical baseline image, thereby obtaining multiple candidate regions and multiple region change features. Feature verification module 13 is used to traverse the multiple candidate regions and multiple region change features to perform multi-model game verification, thereby determining a reliable candidate region set, a reliable region change feature set, and a consensus confidence set. Quality state map generation module 14 is used to aggregate the reliable candidate region set. The system combines a set of reliable regional change characteristics and consensus confidence levels with the spatial location of each candidate region to generate a current building surface quality status map. A quality assessment module 15 is used to perform independent building quality assessments based on the current building surface quality status map, obtaining independent building quality assessment results. It also retrieves historical building surface quality status map sequences from multiple historical monitoring periods, performs time-series analysis and trend fitting on the current building surface quality status map, and obtains joint building quality assessment results. A dynamic assessment result acquisition module 16 integrates the independent building quality assessment results and the joint building quality assessment results to obtain the dynamic building quality assessment results.

[0074] Furthermore, the image analysis module 12 is also used to: construct a dual-channel difference perception network, use the dual-channel difference perception network to perceive the feature differences between the current surface image to be monitored and the previous historical baseline image, and obtain a first channel feature map and a second channel feature map; use an attention weight network layer to perform weighted fusion of the first channel feature map and the second channel feature map to obtain a difference heatmap; and perform high-recall change candidate region analysis based on the difference heatmap to obtain the multiple candidate regions and the multiple candidate region change features.

[0075] Furthermore, the image parsing module 12 is also used for: the first channel of the dual-channel difference perception network uses a Gabor filter bank to extract multi-directional and multi-scale texture features and calculates its residual with the previous historical benchmark image; the second channel uses a deformable convolution module to capture local deformation for structural contour change analysis.

[0076] Furthermore, the image parsing module 12 is also used to: traverse the difference heatmap to perform adaptive dual-threshold segmentation, obtain multiple initial candidate regions and multiple initial candidate region change features; retrieve the historical state of the same position within a preset historical period based on the position of the multiple initial candidate regions, and perform instantaneous interference authentication in combination with the multiple candidate region change features; if the instantaneous interference authentication is passed, then multiple candidate regions and multiple candidate region change features are added.

[0077] Furthermore, the image parsing module 12 is also configured to: obtain a first difference threshold and a second difference threshold; identify strong change seed points in the difference heatmap based on the first difference threshold to obtain multiple strong change seed points, wherein the first difference threshold is greater than the second difference threshold; perform region growing on the multiple strong change seed points based on the second difference threshold to obtain multiple initial candidate regions; and perform feature recognition on the multiple initial candidate regions according to the difference heatmap and preset change indicators to obtain multiple initial candidate region change features, wherein the preset change indicators include morphological features, texture and color features, and positional context features.

[0078] Furthermore, the feature verification module 13 is also used to: select three heterogeneous pre-trained visual analysis models, wherein the three heterogeneous pre-trained visual analysis models are respectively constructed based on Transformer, CNN and visual-language joint architecture, and respectively add a role decision head and an opinion output head to the three heterogeneous pre-trained visual analysis models; construct multiple game training sample pairs according to multiple candidate regions and multiple region change features, each game training sample including a real sample and a counterfactual sample; extract a first game training sample pair from the multiple game training sample pairs, and use the first game training sample as input to perform multi-model game verification on the three heterogeneous pre-trained visual analysis models to obtain a first game verification result; add the first game verification result to the multiple game verification results, and map and filter the multiple candidate regions and multiple region change features according to the multiple game verification results to obtain a reliable candidate region set and a reliable region change feature set.

[0079] Furthermore, the feature verification module 13 is also used to: extract and classify defect morphology and texture parameters based on a historical real defect sample library, and establish a defect parameterization template library; match multiple regional change features with multiple matching defect parameterization templates from the defect parameterization template library; perform counterfactual perturbation on the multiple matching defect parameterization templates to obtain multiple counterfactual samples; and use the multiple regional change features as multiple real samples to jointly pair the multiple counterfactual samples to obtain multiple game training sample pairs.

[0080] Furthermore, the feature verification module 13 is also used to: traverse the multiple matching defect parameterized templates to extend the crack morphology along the principal stress direction with pixel-level steps, diffuse the rust texture in the selected corrosion center according to the exponential decay law, and randomly perturb the mean and variance of the color channel in the stain area to obtain multiple counterfactual samples.

[0081] Furthermore, the feature verification module 13 is also used to: randomly assign counterfactual samples from the first game training samples to any one of the three heterogeneous pre-trained visual analysis models, and reverse label them to obtain a reverse-labeled heterogeneous pre-trained visual analysis model; assign real samples from the first game training samples to the other two heterogeneous pre-trained visual analysis models, and perform regular labeling on them respectively to obtain two regular-labeled heterogeneous pre-trained visual analysis models; the reverse-labeled heterogeneous pre-trained visual analysis model and the two regular-labeled heterogeneous pre-trained visual analysis models evaluate the realism of change features based on the assigned samples through the corresponding role decision heads, and output the evaluation results through the corresponding opinion output heads to obtain three evaluation results; synchronize the three evaluation results to the reverse-labeled heterogeneous pre-trained visual analysis model and the two regular-labeled heterogeneous pre-trained visual analysis models for cross-questioning to obtain cross-questioning voting results; if the cross-questioning evaluation result is the reverse-labeled heterogeneous pre-trained visual analysis model, then the two evaluation results corresponding to the two regular-labeled heterogeneous pre-trained visual analysis models are true, and the two evaluation results are averaged to obtain the first game verification result.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The computer vision-based dynamic assessment method and specific examples in the aforementioned embodiment one are also applicable to the computer vision-based dynamic assessment system for building quality in this embodiment. Through the foregoing detailed description of the computer vision-based dynamic assessment method for building quality, those skilled in the art can clearly understand the computer vision-based dynamic assessment system for building quality in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0084] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A dynamic assessment method for building quality based on computer vision, characterized in that, The method includes: By periodically acquiring images of the target building's surface to be monitored using image acquisition equipment deployed at fixed monitoring points, the current surface image to be monitored is obtained. Based on the current expected monitoring surface image and the previous historical baseline image, high recall change candidate region analysis is performed to obtain multiple candidate regions and multiple region change features; The multiple candidate regions and multiple region change features are traversed to perform multi-model game verification to determine the set of reliable candidate regions, the set of reliable region change features, and the set of consensus confidence. By aggregating a set of reliable candidate regions, a set of reliable region change characteristics, and consensus confidence, and combining the spatial location of each candidate region, a current building surface quality status map is generated. Independent building quality assessments are conducted based on the current building surface quality status map to obtain independent building quality assessment results. Furthermore, historical building surface quality status map sequences from multiple historical monitoring periods are retrieved, and time-series analysis and trend fitting are performed on the current building surface quality status map to obtain joint building quality assessment results. By integrating the results of independent building quality assessments and the joint building quality assessments, the dynamic building quality assessment results are obtained.

2. The computer vision-based dynamic assessment method for building quality as described in claim 1, characterized in that, Based on the current surface image to be monitored and the previous historical baseline image, high-recall change candidate region analysis is performed to obtain multiple candidate regions and multiple region change features, including: A dual-channel difference sensing network is constructed, and the dual-channel difference sensing network is used to perform feature difference sensing between the current surface image to be monitored and the previous historical baseline image to obtain the first channel feature map and the second channel feature map. The first channel feature map and the second channel feature map are weighted and fused using an attention weight network layer to obtain a difference heatmap. Based on the difference heatmap, high-recall change candidate regions are analyzed to obtain multiple candidate regions and multiple candidate region change characteristics.

3. The computer vision-based dynamic assessment method for building quality as described in claim 2, characterized in that, The first channel of the dual-channel difference-aware network uses a Gabor filter bank to extract multi-directional, multi-scale texture features and calculates the residual between it and the previous historical benchmark image. The second channel uses a deformable convolution module to capture local deformations for structural contour change analysis.

4. The computer vision-based dynamic assessment method for building quality as described in claim 2, characterized in that, Based on the difference heatmap, high-recall change candidate regions are analyzed to obtain multiple candidate regions and multiple candidate region change features, including: The differential heatmap is traversed to perform adaptive dual-threshold segmentation, resulting in multiple initial candidate regions and multiple initial candidate region change features. Based on the positions of the multiple initial candidate regions, the historical state of the same position within a preset historical period is retrieved, and instantaneous interference authentication is performed in combination with the change characteristics of the multiple candidate regions. If the instantaneous interference authentication is passed, multiple candidate regions and multiple candidate region change characteristics are added.

5. The computer vision-based dynamic assessment method for building quality as described in claim 4, characterized in that, Adaptive dual-threshold segmentation is performed by traversing the difference heatmap to obtain multiple initial candidate regions and multiple initial candidate region change features, including: A first difference threshold and a second difference threshold are obtained. Based on the first difference threshold, strong change seed points are identified in the difference heatmap to obtain multiple strong change seed points, wherein the first difference threshold is greater than the second difference threshold. Based on the second difference threshold, region growing is performed on the multiple strongly changing seed points to obtain multiple initial candidate regions; Based on the difference heatmap, feature recognition is performed on the multiple initial candidate regions according to preset change indicators to obtain multiple initial candidate region change features, wherein the preset change indicators include morphological features, texture and color features, and positional context features.

6. The computer vision-based dynamic assessment method for building quality as described in claim 1, characterized in that, The multiple candidate regions and multiple region change features are traversed to perform multi-model game verification, determining a reliable candidate region set, a reliable region change feature set, and a consensus confidence set, including: Three heterogeneous pre-trained visual analysis models are selected. The three heterogeneous pre-trained visual analysis models are constructed based on Transformer, CNN and visual-language joint architecture respectively. A role decision head and an opinion output head are added to each of the three heterogeneous pre-trained visual analysis models. Multiple game training sample pairs are constructed based on multiple candidate regions and multiple region change characteristics. Each game training sample includes one real sample and one counterfactual sample. The first game training sample pair is extracted from multiple game training sample pairs. Using the first game training sample as input, multi-model game verification is performed on the three heterogeneous pre-trained visual analysis models to obtain the first game verification result. The first game verification result is added to multiple game verification results. The multiple candidate regions and multiple region change features are mapped and filtered according to the multiple game verification results to obtain a reliable candidate region set and a reliable region change feature set.

7. The computer vision-based dynamic assessment method for building quality as described in claim 6, characterized in that, Multiple game training sample pairs are constructed based on multiple candidate regions and the changing characteristics of multiple regions. Each game training sample includes one real sample and one counterfactual sample, including: Based on a historical real defect sample library, defect morphology and texture parameters are extracted and classified to establish a defect parameterization template library. Multiple regional variation features are matched with multiple matching defect parameterization templates from the defect parameterization template library; The multiple matching defect parameterized templates are subjected to counterfactual perturbations to obtain multiple counterfactual samples. The multiple regional change features are used as multiple real samples, and the multiple counterfactual samples are jointly paired to obtain multiple game training sample pairs.

8. The computer vision-based dynamic assessment method for building quality as described in claim 7, characterized in that, By traversing the multiple matching defect parameterized templates, extending the crack morphology along the principal stress direction with pixel-level steps, spreading the rust texture in the selected corrosion center according to the exponential decay law, and randomly perturbing the mean and variance of the color channel in the stain area, multiple counterfactual samples are obtained.

9. The computer vision-based dynamic assessment method for building quality as described in claim 6, characterized in that, From multiple game training sample pairs, a first game training sample pair is extracted. Using the first game training sample as input, multi-model game validation is performed on the three heterogeneous pre-trained visual analysis models to obtain the first game validation result, including: Randomly assign the counterfactual samples from the first game training samples to any one of the three heterogeneous pre-trained visual analysis models, and reverse label them to obtain the reverse-labeled heterogeneous pre-trained visual analysis model. The real samples in the training samples of the first game are assigned to two other heterogeneous pre-trained visual analysis models, and they are labeled in a conventional way to obtain two conventionally labeled heterogeneous pre-trained visual analysis models. The reverse-label heterogeneous pre-trained visual analysis model and two conventional-label heterogeneous pre-trained visual analysis models are based on the assigned samples. The change feature authenticity is evaluated through the corresponding role decision head, and the evaluation results are output through the corresponding opinion output head to obtain three evaluation results. The three evaluation results are synchronized to the reverse label heterogeneous pre-trained visual analysis model and two conventional label heterogeneous pre-trained visual analysis models for cross-questioning, and the cross-questioning voting results are obtained. If the cross-questioning evaluation result is a reverse-labeled heterogeneous pre-trained visual analysis model, then the two evaluation results corresponding to the two conventional labeled heterogeneous pre-trained visual analysis models are true. The two evaluation results are averaged to obtain the first game verification result.

10. A computer vision-based dynamic assessment system for building quality, characterized in that, The step of implementing the computer vision-based dynamic building quality assessment method according to any one of claims 1 to 9, wherein the computer vision-based dynamic building quality assessment system comprises: The image acquisition module is used to periodically acquire images of the target building's surface to be monitored through image acquisition devices deployed at fixed monitoring points, thereby obtaining the current surface image to be monitored. The image analysis module is used to analyze high-recall change candidate regions based on the current expected monitoring surface image and the previous historical baseline image, and obtain multiple candidate regions and multiple region change features; The feature verification module is used to traverse the multiple candidate regions and multiple region change features to perform multi-model game verification, and determine the reliable candidate region set, the reliable region change feature set, and the consensus confidence set. The quality status map generation module is used to aggregate a set of reliable candidate regions, a set of reliable region change characteristics, and consensus confidence, and combine the spatial location of each candidate region to generate a current building surface quality status map. The quality assessment module is used to conduct independent building quality assessments based on the current building surface quality status map, obtain independent building quality assessment results, and retrieve historical building surface quality status map sequences from multiple historical monitoring periods to perform time series analysis and trend fitting on the current building surface quality status map to obtain joint building quality assessment results. The dynamic assessment result acquisition module is used to integrate the independent building quality assessment results and the joint building quality assessment results to obtain the dynamic assessment results of the building quality.

Images