Building defect identification method and system based on image identification

Abstract

The invention discloses a building defect recognition method and system based on image recognition, and particularly relates to the technical field of computer vision. Comprising a data acquisition and preprocessing module, a building defect knowledge graph construction module, a building defect identification module, a building defect quantitative analysis module, a building defect risk assessment module and a building defect risk early warning module. According to the method, the extracted building defect information and structured domain knowledge are deeply fused, the reason for forming the defect can be further reasoned, deep understanding and diagnosis of the defect are achieved, the interpretability of model output is remarkably improved, quantitative parameters, diagnosis causes, component importance and historical change trends of the defect are calculated, and the accuracy of the model output is improved. According to the method, the quantitative building defect risk index is obtained, early warning of different levels is automatically triggered according to the preset value, early recognition and predictive maintenance of the structure risk are achieved, and active early warning and risk prediction of the building defect are achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, more particularly, the present application relates to a building defect identification method and system based on image recognition. BACKGROUND

[0002] As the core carrier of infrastructure and livelihood security, the structural safety and performance of buildings are directly related to life and property safety and social and economic stability. In the whole life cycle of buildings, affected by multiple factors such as material performance degradation, construction process deviation, environmental factor erosion, etc., various defects such as cracks, hollowing, leakage, spalling, and steel corrosion are prone to occur. If not detected and treated in time, it may lead to building function failure, or even structural collapse and other major safety accidents. Therefore, efficient, accurate and timely detection of building defects is a key link to ensure building safety, prolong building service life and reduce operation and maintenance costs.

[0003] However, in actual use, it still has some disadvantages, such as the existing building defect identification method only identifies the defects existing in the image through pattern matching, and cannot understand the internal causal relationship between these defects, so that the output result of the prior art defect identification is of limited value and difficult to be directly used to guide maintenance decision; The severity of the defect depends not only on its own size, but also on its cause and its impact on the overall safety of the building. The prior art cannot quantify the physical size of the defect to assess the risk it brings to the overall building, and lacks the ability to understand the defect in depth. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a building defect identification method and system based on image recognition to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a building defect identification method and system based on image recognition, comprising: A data acquisition and preprocessing module is used to collect multi-source image data related to target building defects using a mounted device, and output standardized data sets through preprocessing; A building defect knowledge graph construction module includes a knowledge graph pattern layer, a knowledge graph data layer, and a graph storage and query layer, and is used to store and construct a structured knowledge base containing entities and their associated relationships; A building defect identification module receives image data preprocessed by the data acquisition and preprocessing module, and uses a pre-trained deep convolutional neural network model to identify defects in the input image; Building defect quantitative analysis module: receiving the structured defect feature data output by the building defect identification module, calculating the quantitative parameters of the defect based on the calibration relationship between the image pixels and the actual physical size of the defect body within the bounding box; Building defect risk assessment module: including a fusion intelligent diagnosis unit and a building defect risk index calculation unit, for deeply fusing the image recognition result with the knowledge graph to obtain a building defect risk index; Building defect risk warning module: obtaining the calculated building defect risk index, comparing it with a preset building defect risk index interval, and generating warning information.

[0006] Preferably, the data acquisition and preprocessing module specifically comprises: The device configured to acquire the target building surface image data includes, but is not limited to, a high-definition digital camera, a drone-mounted aerial camera, a smartphone, a wall-climbing robot, or a fixed monitoring camera. The pre-processing of the collected original multi-source image data includes denoising, image enhancement, defect area preliminary segmentation, and data standardization. The denoising removes environmental noise by using Gaussian filtering and eliminates salt and pepper noise by using median filtering. The image enhancement improves the contrast of low-light images by histogram equalization and corrects the uneven light problem by Retinex algorithm. The defect area preliminary segmentation extracts the suspected defect area by threshold segmentation method to reduce background interference. The data standardization unifies the image size to 640x640 and normalizes the pixel value to [0, 1].

[0007] Preferably, the building defect knowledge graph construction module specifically comprises: Knowledge graph mode layer: presetting the mode of core entities, attributes, and relationships. The core entities include building components, defect types, defect causes, and risk consequences. Knowledge graph data layer: based on the mode layer, extracting industry standards, expert experience, and historical case data to build a rich knowledge graph between entities. Graph storage and query layer: using a graph database to store the knowledge graph and providing graph query and relationship traversal interfaces. It also automatically crawls the latest building defect detection data, updates the knowledge graph content through incremental knowledge extraction and fusion.

[0008] Preferably, the building defect identification module specifically comprises: Converting the pre-processed standardized image data into a tensor format required by a deep neural network model. The converted tensor image is input into a backbone network of a deep convolutional neural network trained by a large number of labeled building defect images. The network gradually extracts low-level edge and texture features and high-level semantic features through a series of convolution and pooling operations, and finally generates a set of multi-scale feature maps.

[0009] Preferably, the building defect recognition module further comprises: predicting potential defect target positions on the feature map and integrating multi-scale context information to improve small target detection capability, and the specific implementation is: For a single-stage detector: using an anchor box mechanism on the feature map, predicting the coordinate offset, confidence and class probability of multiple bounding boxes for each candidate region, and fusing deep semantic features and shallow high-resolution features to ensure positioning accuracy and classification accuracy at the same time; For a two-stage detector: first, a region proposal network is used to generate a series of candidate regions containing potential defects on the feature map; then, a RoI alignment layer is used to extract a fixed-size feature vector for each candidate region from the feature map; The feature vector of each candidate region or the feature of the anchor box is input into the fully connected layer or the convolutional layer of the neural network model, and a class probability distribution vector is output, each element in the vector corresponding to a defect type; A confidence threshold is set to filter out predicted boxes with a maximum confidence lower than the confidence threshold. For the remaining predicted boxes, for each defect type, the confidence is sorted, and the intersection over union between all boxes is calculated to filter out redundant boxes with an intersection over union exceeding a preset overlap threshold with the highest confidence box, and finally an optimal detection result is retained for each defect instance; Output structured defect feature data, which includes the specific name of the defect and the position bounding box represented by pixel coordinates.

[0010] Preferably, the building defect quantification analysis module specifically comprises: Based on the output bounding box, the defect area is extracted from the image according to the bounding box coordinates, and the defect pixels are separated from the background pixels through image processing algorithms: For crack defects: using an edge detection algorithm on the defect area image to obtain a binary image of the crack and extract a single-pixel-wide crack skeleton line; for spalling or corrosion defects: using a segmentation algorithm based on color and texture features on the defect area image to extract a continuous defect area pixel set; For each defect instance, a parameter set is generated, which specifically comprises: All pixel points on the crack skeleton line are traversed, the Euclidean distance between adjacent pixels is calculated and accumulated to obtain the total skeleton pixel length, and the defect length is calculated by ​Along the crack skeleton line, the gray profile of the normal direction is calculated every fixed interval, the edges of the crack on both sides are determined by finding the gradient extreme points of the profile, and the defect width of each point is calculated by , and the maximum defect width and the average defect width are output. The total number of pixels in the defect area pixel set is counted, and the defect area is calculated by . As an illustration, the pixel proportion coefficient k calculated in advance for the corresponding image is loaded from the system storage, which is calculated by the calibration object in the image, and the calculation formula is k = actual physical length of calibration object / pixel length of calibration object.

[0011] Preferably, the building defect risk assessment module specifically comprises: The intelligent diagnosis unit: according to the type, position and shape parameters of the defect, all defect cause paths connected with the defect in the knowledge graph are traversed, and the cause danger degree is obtained in combination with the properties of the building component; the defect severity degree is obtained by comparing the quantitative parameter value of the defect with the threshold value defined in the knowledge graph; the component importance factor is pre-set according to the importance of the component in the overall structure of the building; when there is historical detection data of the defect, the defect expansion factor is obtained by comparing the growth rate of the defect; The building defect risk index calculation unit: used for calculating the building defect risk index: Among them, represents the building defect risk index, represents the cause danger degree, represents the defect severity degree, represents the component importance factor, represents the defect expansion factor, represents the aggregation function.

[0012] Preferably, the building defect risk warning module specifically comprises: The building defect risk index is obtained, compared with a preset building defect risk index interval, if the building defect risk index is less than or equal to the minimum value of the preset building defect risk index interval, it is low risk, indicating that the defect has no influence on the safety of the building structure at this time, if the building defect risk index is greater than the minimum value of the preset building defect risk index interval and less than or equal to the maximum value of the preset building defect risk index interval, it is medium risk, indicating that the defect has a significant influence on the safety of the building structure at this time, if the building defect risk index is greater than the maximum value of the preset building defect risk index interval, it is high risk, indicating that the defect is a serious danger to the safety of the building structure, when the risk level is determined to be medium risk or high risk, a warning process is automatically triggered, and defect structured warning information is generated in parallel and sent to the management terminal of the relevant responsible person. The defect structured warning information can be defect positioning and description, defect knowledge graph diagnosis result, building defect risk index evaluation result, and preliminary disposal suggestion for the cause and risk level associated and output from the knowledge graph.

[0013] Preferably, a building defect identification method based on image recognition comprises the following steps: Step S01: data acquisition and preprocessing: used for acquiring multi-source image data related to target building defects using a mounted device, and outputting a standardized data set through preprocessing; Step S02: building defect knowledge graph construction: including a knowledge graph mode layer, a knowledge graph data layer and a graph storage and query layer, used for storing and constructing a structured knowledge base containing entities and their associated relationships; Step S03: building defect identification: receiving image data preprocessed in the data acquisition and preprocessing step, using a pre-trained deep convolutional neural network model to identify defects in the input image; Step S04: building defect quantitative analysis: receiving structured defect feature data output by the building defect identification step, calculating the quantitative parameters of the defect based on the calibration relationship between the image pixels and the actual physical size of the defect body in the bounding box; Step S05: building defect risk assessment: including a fusion intelligent diagnosis sub-step and a building defect risk index calculation sub-step, used for deeply fusing the image recognition result with the knowledge graph to obtain a building defect risk index; Step S06: building defect risk warning module: obtaining the calculated building defect risk index, comparing it with a preset building defect risk index interval, and generating warning information.

[0014] The technical effects and advantages of the present application are: 1. The application provides a building defect identification method and system based on image recognition. By introducing a building defect knowledge graph, the apparent defect information extracted by the image recognition module is deeply fused with the structured domain knowledge, which can further infer the causes of the defects, realize the deep understanding and diagnosis of the defects, significantly improve the explainability of the model output, calculate the quantitative parameters, diagnosis causes, component importance and historical change trend of the defects, obtain the quantitative building defect risk index, and automatically trigger different levels of early warning according to the preset value, realize early identification and predictive maintenance of structural risk, realize the leap from passive identification to active early warning and risk prediction. 2. The application provides a building defect identification method and system based on image recognition. The target building defect related multi-source image data is collected by the mounted equipment, a pre-trained deep convolutional neural network model is used for defect identification of the input image, the quantitative parameters of the defect are calculated based on the calibration relationship between the image pixels and the actual physical size of the defect body in the boundary box, the building defect risk index is calculated, and early warning information is generated, thereby constructing a full-process automatic assembly line from image collection, defect identification, parameter quantization and risk assessment, and thus a complete digital archive is constructed for each defect, a large amount of manual inspection is replaced by automatic detection, and the long-term labor cost is significantly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 FIG. 1 is a structural schematic diagram of a building defect identification system based on image recognition according to the application.

[0016] Figure 2 FIG. 2 is a flowchart of a building defect identification method based on image recognition according to the application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the application will be described clearly and completely in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0018] Please refer to Figure 1 The application provides a building defect identification system based on image recognition, which comprises a data acquisition and preprocessing module, a building defect knowledge graph construction module, a building defect identification module, a building defect quantization analysis module, a building defect risk assessment module and a building defect risk warning module.

[0019] The data acquisition and preprocessing module is connected with the building defect identification module, the building defect knowledge graph construction module is connected with the building defect identification module, the building defect identification module is connected with the building defect quantitative analysis module, the building defect quantitative analysis module is connected with the building defect risk assessment module, and the building defect risk assessment module is connected with the building defect risk early warning module.

[0020] The data acquisition and preprocessing module is used for collecting target building defect related multi-source image data using a mounted device, and outputting a standardized data set through preprocessing. In a possible design, the data acquisition and preprocessing module specifically comprises: The device configured to collect target building surface image data includes but is not limited to a high-definition digital camera, a drone-mounted aerial camera, a smart phone, a wall-climbing robot or a fixed monitoring camera. The collected original multi-source image data is preprocessed, including denoising, image enhancement, defect area preliminary segmentation and data standardization. The denoising removes environmental noise by using Gaussian filtering and eliminates salt and pepper noise by using median filtering. The image enhancement improves the contrast of low-light images by histogram equalization and corrects the problem of uneven light by Retinex algorithm. The defect area preliminary segmentation extracts the suspected defect area by threshold segmentation method to reduce background interference. The data standardization unifies the image size to 640*640 and normalizes the pixel value to [0, 1].

[0021] The building defect knowledge graph construction module comprises a knowledge graph mode layer, a knowledge graph data layer and a graph storage and query layer, and is used for storing and constructing a structured knowledge base containing entities and their associated relationships.

[0022] In a possible design, the building defect knowledge graph construction module specifically comprises: The knowledge graph mode layer presets the mode of core entities, attributes and relationships. The core entities include: Building components, such as beams, plates, columns, walls and foundations, have attributes such as material type, design strength, service life and spatial position; Defect types, such as cracks, peeling, leakage, corrosion and deformation, have attributes such as typical morphology and quantitative threshold (such as safe width and dangerous width); Defect causes, such as load overrunning, concrete carbonization, steel bar corrosion, temperature stress and construction quality, have attributes such as triggering conditions and action mechanisms; Risk consequences, such as bearing capacity reduction, durability reduction and aesthetic damage, have attributes such as risk level and influence range; Knowledge graph data layer: based on the pattern layer, through the extraction of industry standards, expert experience and historical case data, a rich knowledge graph between entities is constructed; Graph storage and query layer: use a graph database to store the knowledge graph, and provide graph query and relationship traversal interface. Automatically crawl the latest building defect detection data, update the knowledge graph content through incremental knowledge extraction and fusion.

[0023] The building defect recognition module: receives the image data preprocessed by the data acquisition and preprocessing module, and uses a pre-trained deep convolutional neural network model to recognize defects in the input image.

[0024] In one possible design, the building defect recognition module specifically comprises: Convert the preprocessed standardized image data into the tensor format required by the deep neural network model; Input the converted tensor image into the backbone network of the deep convolutional neural network trained on a large number of labeled building defect images. The network gradually extracts low-level edge and texture features to high-level semantic features through a series of convolution and pooling operations, and finally generates a set of multi-scale feature maps; Predict the potential defect target position on the feature map and integrate multi-scale context information to improve small target detection capability. The specific implementation is as follows: For single-stage detectors: use anchor box mechanism on the feature map to predict the coordinate offset, confidence and class probability of multiple bounding boxes in each candidate region. Fuse deep semantic features and shallow high-resolution features to ensure positioning accuracy and classification accuracy at the same time; For two-stage detectors: first, generate a series of candidate regions containing potential defects on the feature map through the region proposal network; then, extract a fixed-size feature vector for each candidate region from the feature map through the RoI alignment layer; Input the feature vector of each candidate region or the feature of the anchor box into the fully connected layer or convolutional layer of the neural network model, and output a class probability distribution vector. Each element in the vector corresponds to a defect type; Set a confidence threshold to filter out predicted boxes with a maximum confidence lower than the confidence threshold. For the remaining predicted boxes, sort them according to the confidence for each defect type, and calculate the intersection over union between all boxes. Filter out redundant boxes whose intersection over union with the highest confidence box exceeds the preset overlap threshold, and finally retain one optimal detection result for each defect instance; Output structured defect feature data, which includes the specific name of the defect and the position bounding box represented in pixel coordinates. As an illustration, the generation of the prediction frame is a mechanism based on each candidate region or anchor frame of the deep convolutional neural network, which directly predicts the target object boundary frame through regression calculation.

[0025] The building defect quantitative analysis module receives the structured defect feature data output by the building defect identification module, calculates the quantitative parameters of the defects based on the calibration relationship between the image pixels and the actual physical size of the defect main body in the boundary frame.

[0026] In a possible design, the building defect quantitative analysis module specifically comprises: Based on the output boundary frame, the defect area in the image is intercepted according to the boundary frame coordinates, and the defect pixels are separated from the background pixels through an image processing algorithm: For crack defects: the binary image of the crack is obtained by using an edge detection algorithm on the image of the defect area, and a single-pixel-wide crack skeleton line is extracted; for spalling or corrosion defects: a continuous defect area pixel set is extracted based on a color and texture feature segmentation algorithm on the image of the defect area; A parameter set is generated for each defect instance, and the parameter set specifically comprises: All pixel points on the crack skeleton line are traversed, the Euclidean distance between adjacent pixels is calculated and accumulated to obtain the total pixel length of the skeleton, and the defect length is calculated through ; Along the crack skeleton line, the gray profile in the normal direction is calculated at a fixed interval, the edges on both sides of the crack are determined by finding the gradient extreme points of the profile, and the defect width of each point is calculated through ; The total number of pixels in the defect area pixel set is counted, and the defect area is calculated through ; As an illustration, the pixel proportion coefficient k calculated in advance for the corresponding image is loaded from the system storage, and the pixel proportion coefficient is calculated through a calibration object in the image, and the calculation formula is k = actual physical length of calibration object / pixel length of calibration object.

[0027] The building defect risk assessment module comprises a fusion intelligent diagnosis unit and a building defect risk index calculation unit, and is configured to deeply fuse the image recognition result and the knowledge graph to obtain a building defect risk index.

[0028] In a possible design, the building defect risk assessment module specifically comprises: The fusion intelligent diagnosis unit: according to the type, position and shape parameters of the defect, traversing all defect cause paths connected with the defect in the knowledge graph, and combining the properties of the building component, a cause danger degree is obtained; according to the quantitative parameter value of the defect and the threshold value defined in the knowledge graph, a defect severity degree is obtained; according to the importance of the component in the overall structure of the building, a component importance factor is pre-set; when there is historical detection data of the defect, by comparing the growth rate of the defect, a defect expansion factor is obtained; The building defect risk index calculation unit: used for calculating the building defect risk index: Wherein, represents the building defect risk index, represents the cause danger degree, represents the defect severity degree, represents the component importance factor, represents the defect expansion factor, represents the aggregation function; The building defect risk warning module: obtaining the calculated building defect risk index, comparing with the preset building defect risk index interval, and generating warning information.

[0029] In a possible design, the building defect risk warning module specifically comprises: obtaining the building defect risk index, comparing with the preset building defect risk index interval, if the building defect risk index is less than or equal to the minimum value of the preset building defect risk index interval, it is low risk, indicating that the defect has no effect on the safety of the building structure at this time, if the building defect risk index is greater than the minimum value of the preset building defect risk index interval and less than or equal to the maximum value of the preset building defect risk index interval, it is medium risk, indicating that the defect has a significant impact on the safety of the building structure at this time, if the building defect risk index is greater than the maximum value of the preset building defect risk index interval, it is high risk, indicating that the defect is a serious danger to the safety of the building structure, when the risk level is determined as medium risk or high risk, the warning process is automatically triggered, and the defect structured warning information is generated in parallel and sent to the management terminal of the relevant responsible person.

[0030] The defect structured warning information can be defect positioning and description, defect knowledge graph diagnosis result, building defect risk index evaluation result, preliminary disposal suggestion for the cause and risk level associated from the knowledge graph.

[0031] Please refer to Figure 2 The present application provides a building defect identification method based on image recognition, comprising the following steps: Step S01: data acquisition and preprocessing: used for collecting multi-source image data related to target building defects using a mounted device, and outputting a standardized data set through preprocessing; Step S02: building defect knowledge graph construction: including a knowledge graph mode layer, a knowledge graph data layer, and a graph storage and query layer, used for storing and constructing a structured knowledge base containing entities and their associated relationships; Step S03: building defect identification: receiving image data preprocessed in the data acquisition and preprocessing step, using a pre-trained deep convolutional neural network model to identify defects in the input image; Step S04: building defect quantitative analysis: receiving structured defect feature data output by the building defect identification step, calculating the quantitative parameters of the defect based on the calibration relationship between image pixels and actual physical dimensions of the defect body within the bounding box; Step S05: building defect risk assessment: including a fusion intelligent diagnosis sub-step and a building defect risk index calculation sub-step, used for deeply fusing the image recognition result with the knowledge graph to obtain a building defect risk index; Step S06: building defect risk warning module: obtaining the calculated building defect risk index, comparing it with a preset building defect risk index interval, and generating warning information.

[0032] In this embodiment, it is specifically pointed out that the present application introduces a building defect knowledge graph, deeply fuses the apparent defect information extracted by the image recognition module with the structured domain knowledge, can further infer the causes of the defects, realizes the deep understanding and diagnosis of the defects, significantly improves the explainability of the model output, calculates the quantitative parameters of the defects, the diagnosis causes, the importance of the components, and the historical change trend, obtains a quantitative building defect risk index, and automatically triggers different levels of warning according to the preset value, realizes early identification and predictive maintenance of the structural risk, realizes the leap from passive identification to active warning and risk prediction; The present application uses a mounted device to collect multi-source image data related to target building defects, uses a pre-trained deep convolutional neural network model to identify defects in the input image, calculates the quantitative parameters of the defects based on the calibration relationship between image pixels and actual physical dimensions of the defect body within the bounding box, calculates the building defect risk index, and generates warning information, and constructs a full-process automation pipeline from image acquisition, defect identification, parameter quantization, and risk assessment, thereby constructing a complete digital archive for each defect, replacing a large amount of manual inspection with automated detection, and significantly reducing long-term labor costs.

[0033] Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the scope of protection of the present application.

Claims

1. A building defect identification system based on image recognition, characterized in that, include: Data acquisition and preprocessing module: used to acquire multi-source image data related to defects in the target building using the onboard equipment, and output a standardized dataset through preprocessing; The knowledge graph construction module for building defects includes a knowledge graph schema layer, a knowledge graph data layer, and a graph storage and query layer, which is used to build and store a structured knowledge base containing entities and their relationships. Building Defect Identification Module: Receives image data preprocessed by the data acquisition and preprocessing module, and uses a pre-trained deep convolutional neural network model to identify defects in the input image; Building Defect Quantification Analysis Module: Receives structured defect feature data output by the building defect identification module, and calculates the quantitative parameters of the defect based on the calibration relationship between image pixels and actual physical dimensions for the defect subject within the bounding box; Building Defect Risk Assessment Module: Includes an integrated intelligent diagnostic unit and a building defect risk index calculation unit, which is used to deeply integrate image recognition results with knowledge graphs to obtain a building defect risk index; Building Defect Risk Early Warning Module: Obtains the calculated building defect risk index, compares it with the preset building defect risk index range, and generates early warning information.

2. The building defect identification system based on image recognition according to claim 1, characterized in that: The data acquisition and preprocessing module specifically comprises: Configure devices for collecting image data of the target building surface, including but not limited to high-definition digital cameras, drone-mounted aerial cameras, smartphones, wall-climbing robots, or fixed surveillance cameras; Preprocessing is performed on the acquired raw multi-source image data, including: denoising, image enhancement, preliminary segmentation of defect areas, and data standardization. The denoising process removes environmental noise by using Gaussian filtering and eliminates salt-and-pepper noise by using median filtering. The image enhancement improves the contrast of low-light images by using histogram equalization and corrects uneven lighting by using the Retinex algorithm. The preliminary segmentation of defect areas uses threshold segmentation to extract suspected defect areas and reduce background interference. The data standardization unifies the image size to 640×640 and normalizes the pixel values ​​to [0,1].

3. The building defect identification system based on image recognition according to claim 1, characterized in that: The building defect knowledge graph construction module specifically includes: Knowledge graph pattern layer: a pattern of preset core entities, attributes and relationships, wherein the core entities include: building components, defect types, defect causes and risk consequences; Knowledge graph data layer: Based on the schema layer, a rich knowledge graph between entities is constructed by extracting industry standards, expert experience, and historical case data; Knowledge graph storage and query layer: Uses a graph database to store the knowledge graph, provides graph query and relation traversal interfaces, and automatically crawls the latest building defect detection data. Through incremental knowledge extraction and fusion, it updates the knowledge graph content.

4. The building defect identification system based on image recognition according to claim 1, characterized in that: The building defect identification module is specifically as follows: Convert the preprocessed, standardized image data into the tensor format required by the deep neural network model; The transformed tensor image is input into the backbone of a deep convolutional neural network trained on a large number of labeled building defect images. The network extracts features from low-level edge and texture features to high-level semantic features through a series of convolution and pooling operations, and finally generates a set of multi-scale feature maps.

5. A building defect identification system based on image recognition according to claim 1, characterized in that: The building defect identification module further includes: predicting the location of potential defect targets on the feature map and integrating multi-scale contextual information to improve the small target detection capability. The specific implementation method is as follows: For single-stage detectors: Anchor box mechanism is used on feature maps to predict the coordinate offset, confidence and class probability of multiple bounding boxes for each candidate region, and the deep semantic features are fused with the shallow high-resolution features to ensure both localization accuracy and classification accuracy. For the two-stage detector: first, a series of candidate regions containing potential defects are generated on the feature map through a region proposal network; then, a fixed-size feature vector is extracted from the feature map for each candidate region through a RoI alignment layer. For each candidate region's feature vector or anchor box feature input, the neural network model's fully connected layer or convolutional layer outputs a class probability distribution vector, where each element of the vector corresponds to a defect type. Set a confidence threshold, filter out predicted boxes with a maximum confidence score lower than the confidence threshold, sort the remaining predicted boxes according to their confidence scores for each type of defect, calculate the intersection-union ratio (IU) between all boxes, filter out redundant boxes whose IU with the highest confidence box exceeds a preset overlap threshold, and finally retain an optimal detection result for each defect instance. The system outputs structured defect feature data, which includes the specific name of the defect and its location bounding box in pixel coordinates.

6. A building defect identification system based on image recognition according to claim 1, characterized in that: The building defect quantitative analysis module specifically includes: Based on the output bounding box, the defect region is extracted from the image according to the bounding box coordinates, and the defect pixels are separated from the background pixels using image processing algorithms: For crack defects: use an edge detection algorithm to obtain a binary image of the crack in the image of the defect area, and extract the crack skeleton line with a width of one pixel; for peeling or corrosion defects: use a segmentation algorithm based on color and texture features to extract a continuous set of defect area pixels in the image of the defect area. A parameter set is generated for each defect instance, and the parameter set is specifically as follows: Traverse all pixels along the crack skeleton line, calculate and sum the Euclidean distances between adjacent pixels to obtain the total pixel length of the skeleton. Calculate the defect length; Along the crack skeleton line, the grayscale profile along its normal direction is calculated at fixed intervals. The edges of the crack are determined by finding the gradient extrema of the profile. Calculate the defect width at each point and output the maximum defect width and average defect width; The total number of pixels in the pixel set of the defective region is counted by... Calculate the defect area; As an explanation, the pixel ratio coefficient k, which was pre-calculated for the corresponding image, is loaded from the system storage. The pixel ratio coefficient is calculated using a calibration object in the image, and the calculation formula is k = actual physical length of the calibration object / pixel length of the calibration object.

7. A building defect identification system based on image recognition according to claim 1, characterized in that: The building defect risk assessment module specifically includes: The integrated intelligent diagnostic unit traverses all defect cause paths connected to the defect in the knowledge graph based on the defect type, location, and morphological parameters, and combines this with the attributes of the building components to obtain the cause risk level; it compares the quantitative parameter value of the defect with the threshold defined in the knowledge graph to obtain the defect severity; and it pre-sets the component importance factor based on the component's importance in the overall building structure. When historical detection data for the defect is available, the defect scalability factor is obtained by comparing the defect growth rate. Building Defect Risk Index Calculation Unit: Used to calculate the building defect risk index. in, This is represented as the building defect risk index. This is expressed as the degree of risk of cause. This is expressed as the severity of the defect. Represented as component importance factor, Represented as the defect extensibility factor, It is represented as an aggregate function.

8. A building defect identification system based on image recognition according to claim 1, characterized in that: The building defect risk early warning module is specifically as follows: The system obtains the building defect risk index and compares it with the preset building defect risk index range. If the building defect risk index is less than or equal to the minimum value of the preset building defect risk index range, it is considered low risk, indicating that the defect has no impact on the structural safety of the building. If the building defect risk index is greater than the minimum value of the preset building defect risk index range but less than or equal to the maximum value of the preset building defect risk index range, it is considered medium risk, indicating that the defect has a significant impact on the structural safety of the building. If the building defect risk index is greater than the maximum value of the preset building defect risk index range, it is considered high risk, indicating that the defect seriously endangers the structural safety of the building. When the risk level is determined to be medium or high risk, the early warning process is automatically triggered, and a structured early warning information for defects is generated in parallel and sent to the management terminal of the relevant responsible personnel. The structured early warning information for defects can include defect location and description, defect knowledge graph diagnostic results, building defect risk index assessment results, and preliminary handling suggestions based on the causes and risk levels of the defects, which are then linked from the knowledge graph.

9. A building defect identification method based on image recognition, using a building defect identification system based on image recognition as described in any one of claims 1-8, characterized in that: Includes the following steps: Step S01: Data Acquisition and Preprocessing: This step involves using the onboard equipment to acquire multi-source image data related to defects in the target building, and then preprocessing the data to output a standardized dataset. Step S02: Construction of Building Defect Knowledge Graph: This includes a knowledge graph schema layer, a knowledge graph data layer, and a graph storage and query layer, used to construct and store a structured knowledge base containing entities and their relationships. Step S03: Building Defect Identification: Receive the image data after preprocessing in the data acquisition and preprocessing steps, and use a pre-trained deep convolutional neural network model to identify defects in the input image; Step S04: Quantitative Analysis of Building Defects: Receive the structured defect feature data output from the building defect identification step, and calculate the quantitative parameters of the defect within the bounding box based on the calibration relationship between image pixels and actual physical dimensions. Step S05: Building Defect Risk Assessment: This includes a fusion intelligent diagnosis sub-step and a building defect risk index calculation sub-step, which are used to deeply fuse image recognition results with knowledge graphs to obtain the building defect risk index; Step S06: Building Defect Risk Early Warning Module: Obtain the calculated building defect risk index, compare it with the preset building defect risk index range, and generate early warning information.

Images