A Deep Learning-Based Method for Grading the Severity of Hollow Defects in Building Facades

By combining deep learning and neural network models with visible light and infrared images, the system automatically identifies and classifies hollow defects on building facades, solving the problem of the inability to determine the severity of defects in existing technologies and improving detection efficiency and safety.

CN120125498BActive Publication Date: 2026-11-10SOUTH CHINA UNIV OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510017766.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-01-06
Publication Date
2026-11-10
Estimated Expiration
2045-01-06

AI Technical Summary

Technical Problem

Existing methods for detecting hollow defects on building facades cannot effectively determine the severity of the defects, resulting in low detection efficiency, high costs, and risks.

Method used

Using a deep learning-based approach, a radial basis function neural network model based on the U-Net semantic segmentation model and the Grey Wolf optimization algorithm is combined with visible light and infrared images to automatically identify and classify the severity of hollow defects in building facades.

Benefits of technology

It enables rapid and intelligent classification of hollow defects on building facades, improves detection efficiency, reduces costs and risks, quantifies defect levels, and assists in the formulation of maintenance plans.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120125498B_ABST
    Figure CN120125498B_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based method for grading the severity of hollow defects in building facades. The method includes: acquiring infrared and visible light images of the building facade using a drone; preprocessing the images to obtain a dataset; manually labeling the dataset and generating corresponding defect mask images; constructing and training a deep learning-based U-Net semantic segmentation model to detect defect areas and generate mask images; calculating the defect area, height, and temperature difference between the defect and normal areas based on the mask images; manually grading the severity of defects using grading indicators and standards; and constructing a GWO-RBF neural network model using the manually graded results as labels to achieve intelligent grading of the severity of hollow defects in building facades. This invention improves detection efficiency, reduces costs and risks, and can accurately assess the severity of defects, solving the problems of low efficiency, high cost, and difficulty in quantifying the severity of defects in traditional detection methods.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent building inspection, specifically involving a method for classifying the severity of hollow defects in building facades based on deep learning. Background Technology

[0002] The different characteristics of hollow defects in building facades, such as their area and height, result in varying degrees of harm. Taking different measures in advance to address different levels of harm can reduce the possibility of injury caused by hollow defects in building facades and also reduce repair costs. However, current image detection methods can only detect the presence or absence of hollow defects in building facades, but cannot accurately determine their degree of harm.

[0003] An existing non-contact building facade defect detection device (202323295285.X) mainly uses drones equipped with both visible light and infrared cameras to achieve non-contact detection of facade defects, based on traditional methods such as visual inspection and tapping. However, it has some shortcomings. First, the defect objects are too general and not clearly defined. Defects such as cracks and peeling can be detected by visual inspection. Second, it can only detect the presence or absence of defects and cannot classify the degree of harm caused by defects.

[0004] The Development Trend of Intelligent Building Facade Damage Detection (Wang Pujin, Xiao Jianzhuang, Duan Zhenhua, Li Chunlai, College of Civil Engineering, Tongji University, Shanghai 200092; 2. Shanghai Institute of Technical Physics, Chinese Academy of Sciences): While digital and intelligent methods are used for detection, the final decision still requires judgment from experienced professionals. Simultaneously, on the software side, artificial intelligence algorithms capable of rapid damage identification and judgment are being developed, reducing the time to issue inspection reports to a few hours or even instantly. On the hardware side, robots and detection modules suitable for specific targets are being developed. This invention proposes a method using UAVs equipped with high-definition visible light and infrared cameras, combined with artificial intelligence methods, to detect damage. This invention utilizes these methods and, based on this, classifies the severity of detected damage, expanding the original target detection to target detection + target classification.

[0005] Existing research on building exterior wall hollowing and detachment identification methods based on visible light and infrared image fusion fuses visible light and infrared images of building exterior walls, combining image information from both modalities, and using UNet and Res images of different depths. The UNet model is used for building exterior wall defect identification based on fused images (Research on the identification method of building exterior wall hollowness and detachment based on visible light and infrared image fusion, Wang Wei, Mi Qingren, Xiao Yun, Yang Xincong, Guangzhou Metro Group Co., Ltd., Harbin Institute of Technology (Shenzhen), Shenzhen, Guangdong, China Railway Construction South China Construction Co., Ltd.). The shortcomings are as follows: First, the case analysis of the exterior wall defect identification results given are mainly for detachment defects, and the identification results for hollowness defects are very few. Second, it is still limited to defect identification, belonging to the field of target detection, and does not classify the severity of defects.

[0006] Therefore, there is an urgent need for an intelligent classification method for the severity of hollow defects in building facades, which can effectively improve the efficiency and accuracy of detecting hollow defects in building facades and realize intelligent classification of the severity of hollow defects in facades. Summary of the Invention

[0007] The main objective of this invention is to provide a deep learning-based method for classifying the severity of hollow defects in building facades. This method aims to address a series of problems in existing methods for detecting hollow defects in building facades, such as low efficiency, high cost, high risk, and inability to determine the degree of harm caused by defects. This intelligent classification method for building facades can be applied to the health inspection of facades of old buildings and the acceptance inspection of facades of completed buildings. It plays a role in improving the efficiency of facade inspection, reducing the cost of facade inspection, and designing facade maintenance plans. Furthermore, it can, to some extent, free up manpower and improve the safety of inspection.

[0008] The present invention is achieved by at least one of the following technical solutions.

[0009] A deep learning-based method for classifying the severity of hollow defects in building facades includes the following steps: S1. Input the building facade image into the trained U-Net semantic segmentation model and generate the corresponding defect region mask image; S2. Calculate the defect area, defect height, and temperature difference between the defect area and the normal area based on the defect area mask image; S3. Select the severity grading index for hollow defects on the building facade, determine the grading standard, manually grade the severity of hollow defects on the building facade, and obtain a database of facade defect grades. S4. Using the severity classification results of hollow defects in building facades as labels, a radial basis function neural network model based on the gray wolf optimization algorithm is used to intelligently classify the severity of hollow defects in building facades in the images to be classified.

[0010] Furthermore, training the deep learning-based U-Net semantic segmentation model includes the following steps: The drone, equipped with both a visible light camera and an infrared thermal imaging camera, acquires infrared and visible light images of the building facade. The building facade infrared and visible light images are preprocessed to obtain a building facade image dataset; The data from the building facade image dataset were manually annotated and corresponding mask images were generated. Further, the steps of preprocessing the acquired images to obtain the building facade image dataset include: The infrared thermal images obtained by taking separate shots of the building area are stitched together. Remove background areas unrelated to the building, retaining only the main body of the building. Select the temperature display range: 20°C ~ 40°C for images taken during the day and 5°C ~ 25°C for images taken at night. Use 2 to 3 colors to display the image and highlight the distribution of defects in the image. Remove temperature distribution unrelated to external wall defects caused by wall contamination and equipment heat sources; High-resolution visible light images are stitched together, and the stitched high-resolution visible light images are overlaid with infrared thermal images to obtain a dataset of building facade images.

[0011] Furthermore, the deep learning-based U-Net semantic segmentation model includes: The encoder is used for downsampling paths, extracting image features step by step through convolutional and pooling layers while reducing resolution; The decoder is used for upsampling paths, progressively restoring the resolution of the feature maps, and outputting a mask of the same size as the input image; Skip connections are used to connect corresponding layers in the encoder and decoder, passing high-resolution features from the encoder to the decoder.

[0012] Further, in step S1, the building facade image is input into the trained U-Net semantic segmentation model to generate the corresponding defect region mask image, including: Input the building facade image to be segmented into the trained U-Net semantic segmentation model, and output the binary prediction of each pixel. If the prediction result of the U-Net semantic segmentation model is 1, it means that the pixel belongs to the hollow defect area of ​​the building facade. If the classification result of the U-Net semantic segmentation model is 0, it means that the pixel belongs to the non-defect area, and the corresponding mask image of the defect area is obtained.

[0013] Furthermore, in step S2, the defect area for: ; in This represents the actual area of ​​the defective region. To count the number of pixels with a value of 1 in a masked image, The scale of an image in a real-world scene is determined by defining a known actual length within the image. And the pixel distance of that distance in the image. Sure; Defect height: ; ; in The actual height of the defect. For the camera's focal length, The distance from the building facade to the defect image captured by the drone using a laser rangefinder. The height of the defect in pixels; The coordinates of the top pixel of the defect in the image. The bottom pixel coordinates of the defect are obtained by using a mask image to find the top and bottom pixel coordinates of the defect in the image.

[0014] Temperature difference between the defective area and the normal area: ; in The temperature difference between the defective area and the normal area. This represents the average temperature value of the defect area. This represents the average temperature value of the normal area.

[0015] Furthermore, in step S3, the specific grading indicators for the severity of hollow defects in the building facade include: The following parameters were selected as evaluation indicators: the proportion of defect area in the building facade area to be inspected, the proportion of defect height to ground height, and the temperature difference between the defect area and the normal area. The proportion of defect area in the building facade area to be inspected is the building facade hollow defect area ratio coefficient. m The ratio of the defect height to the ground height is the building facade hollow defect height ratio coefficient. l Temperature difference between defective and normal areas The temperature difference variable is the area of ​​hollow defects on the building facade. The threshold temperature difference between the defective area and the normal area is T.

[0016] Furthermore, in step S3, the severity grading standard for hollow defects in the building facade includes: Grade A: Proportion coefficient of hollow area in building facade m ≤10%, building facade hollow defect height ratio coefficient l ≤0.3, temperature difference variation in the area of ​​hollow defects on the building facade ≤T, Grade A: Exterior facade is intact, poses little public hazard, and requires no observation or repair; Grade B: Proportion of hollow areas in building facade Number m ≤25%, building facade hollow defect height ratio coefficient l ≤0.45, temperature difference variation in areas of hollow defects on building facade ≤T+2.5, Class B facade has minor defects, posing little threat to the public; observation and use or minor repairs are recommended. Grade C: The proportion of hollow areas in the building facade is less than 25%. m <50%, the height ratio coefficient for hollow defects on the building facade is 0.45 < l <0.6, Temperature difference variable T+2.5 in the hollow defect area of ​​the building facade <T+5, Grade C facade has many defects and poses a certain public hazard; repairs are recommended. Grade D: Percentage of hollow area defects on building facade m ≥50%, the proportion coefficient of hollow defects in building facade height l ≥0.6, temperature difference variation in areas of hollow defects on building facade The facade of grade D (≥T+5) has many defects and poses a significant threat to the public. It is recommended to obtain a detailed inspection report and carry out facade renovation or large-scale repairs.

[0017] Furthermore, in step S4, the radial basis function (GWO-RBF) neural network model based on the gray wolf optimization algorithm is a radial basis function neural network model based on the gray wolf optimization algorithm: The radial basis function neural network model includes: The input layer is used to receive input data and pass it to the hidden layer; Hidden layers are used to map input data to a high-dimensional space. The activation function is a Gaussian function, in the form of: = ; in, It is the input vector. It is the center vector. This is the width parameter, used to control the extent of expansion of the radial basis functions and determine the shape of the activation function. It is the input vector With the center vector The distance between them It is a center The Gaussian distribution function with symmetry center, when and The closer it is to 1, the closer the function value is to 1; when Leave At greater distances, the function value rapidly decays to near zero; The output layer is used to calculate the weighted sum and output the result. The Gray Wolf optimization algorithm treats the center vector, width parameter, and weights of the radial basis function neural network as variables to be optimized, and uses the mean squared error (MSE) to continuously optimize the parameters of the radial basis function neural network. ; in, It is the actual value. It is the model output. The number of samples used to evaluate model error is used to continuously optimize the center, width, and weights of the radial basis function neural network through the Grey Wolf Optimization Algorithm. When the Grey Wolf Optimization Algorithm reaches the maximum number of iterations or the error reaches the target threshold, the optimized radial basis function neural network is output.

[0018] Compared with the prior art, the present invention has the following advantages: This deep learning-based automatic severity classification method for hollow defects in building facades can quickly and intelligently identify, locate, and classify the severity of these defects. Furthermore, as the database expands and the learning depth increases, the classification accuracy and efficiency continuously improve. This method is in a continuous iterative optimization process, ultimately achieving a high level of accuracy. This method improves the efficiency of building facade defect detection while reducing the risk factor and cost of traditional facade inspection. It intelligently classifies the severity of hollow defects, making it significant for traditional building facade health monitoring, building facade completion acceptance, and building facade defect repair. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0020] Figure 1 This is a flowchart of the deep learning-based method for classifying the severity of hollow defects in building facades according to the present invention. Figure 2 A schematic diagram of the database for the level of hollow defects on the exterior facade; Figure 3 This is a schematic diagram showing the output results of a method for classifying the severity of hollow defects on building facades. Detailed Implementation

[0021] This invention discloses a method for classifying the severity of hollow defects in building facades based on deep learning. This intelligent classification method for building facades plays a certain role in improving the efficiency of building facade inspection, reducing the cost of building facade inspection, and assisting in facade maintenance. It also frees up manpower to a certain extent, reduces the risk factor of facade inspection and maintenance, and quantifies the level of hollow defects in facades.

[0022] The technical solutions of the invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a portion of the embodiments of the invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without creative effort are within the scope of protection of the invention.

[0023] like Figure 1 As shown, this invention discloses a method for classifying the severity of hollow defects in building facades based on deep learning, including the following steps: S1. Building facade image acquisition and preprocessing, including: S11. Acquire infrared and visible light images of the building facade using a drone equipped with both a visible light camera and an infrared thermal imaging camera, including: Data on the testing area and target buildings was collected, and on-site investigations were conducted in advance, including: 1) Verify the legality of the flight area; 2) Building structure, scale, finish, and usage period; 3) Architectural design drawings; 4) The building's location, orientation, sunlight exposure, and whether it is obstructed or reflected by the surrounding environment; 5) Location and operating status of the building's cold and heat sources; 6) The condition of the building's exterior walls, including leakage, cracking, peeling, and repairs; 7) Construction methods and maintenance records for building finishing materials; Based on the results of the on-site investigation and data collection, a testing technical solution should be developed to guide on-site implementation, and the technical solution should include the following: (a) Testing time and location; (ii) The location of the wall surface being inspected and the inspection period; (iii) The flight path of the drone shall be marked on the map of the flight plan; (iv) The drone's flight altitude and distance from the wall, shooting angle and number of shots (it should be ensured that there are no obstacles obstructing the surrounding environment of the building being measured, and that the obtained images are easy to identify). (v) The drone's flight altitude, distance from the wall, shooting angle, and number of shots should ensure that there are no obstructions in the surrounding environment of the building being measured, and that the resulting images are easy to identify; (vi) Handheld infrared thermal imaging shooting solution for walls that cannot be detected by drones; (vii) Verify the test results; Establish a ground control station that simultaneously serves as a hub for UAV flight path planning and an image control hub: a. Ground control station hardware setup, including: Mobile workstation: 13th generation i9-13900HX, RTX4090 graphics card, 16-inch 2.5K 240Hz high color gamut display, 32GB RAM, 1TB SSD, supports flight control software operation and image processing; Multi-rotor UAV: ​​Maximum takeoff weight 1100g, folded dimensions 214×91×84mm, maximum wind speed resistance level 5, equipped with both infrared and visible light cameras for precise positioning and flight path setting, meeting the infrared thermal imager technical requirements and multi-rotor UAV system performance requirements for detection missions. Infrared camera: Uncooled vanadium oxide microbolometer sensor with a resolution of 640×512 @30Hz, supports infrared imaging in the wavelength range of 8-14μm, has high-precision temperature measurement function, with a temperature measurement accuracy of ±2℃ or ±2% (whichever is greater), and provides two temperature measurement modes: high gain mode with a temperature measurement range of -40℃ to 150℃, and low gain mode with a temperature measurement range of -40℃ to 550℃. The lens focal length is about 9mm, and the equivalent focal length is 38mm. Display: Multi-screen display, making it easy for operators to view flight path and image information; Power system: Equipped with UPS (Uninterruptible Power Supply) to ensure stable system operation.

[0024] b. Ground control station software configuration, including: Route control software: WayPoint Master professional-grade customized route planning software; DJI Terra, an aerial data processing and analysis software, is used for processing and analyzing drone aerial data. It can generate high-precision 2D and 3D models by combining images captured by drones.

[0025] Use drones equipped with both visible light cameras and infrared thermal imaging cameras to acquire images of building facades; S12. Preprocess the acquired images to obtain a dataset of building facade images, including: Accurately stitch together the infrared thermal images taken in sections of the building area; Remove the background, select an appropriate temperature display range, and use 2 to 3 colors to display the image to highlight the distribution of defects in the image; as an example, select 20°C to 40°C for images taken during the day and 5°C to 25°C for images taken at night.

[0026] Remove temperature distribution unrelated to external wall defects caused by wall contamination, equipment heat sources, etc. A dataset of building facade images is obtained by stitching together high-resolution visible light images and overlaying them with infrared thermal images.

[0027] It should be noted that the image acquisition process of the drone first requires the drone, equipped with an infrared thermal imager and a high-definition camera, to circle the target building and capture infrared thermal images and true-color photographs. The infrared thermal images and visible light images corroborate each other to locate facade defects. The image acquisition process module can be divided into three stages: pre-detection, detection execution, and detection. The pre-detection stage requires on-site investigation of the building's location, orientation, sunlight, and surrounding environmental obstructions or reflections, as well as data collection and technical solution development regarding the building's structural form, scale, finish, and usage time. The detection execution stage requires on-site work such as capturing and recording infrared thermal images and true-color images of the building under test at specific locations along a prescribed flight route, and recording information such as weather, temperature, sunlight, wind speed, surface temperature of the finish layer, and the height of visible defects. The post-detection stage requires image preprocessing such as stitching and noise reduction to output the data and information required for subsequent stages.

[0028] S2. Construct and train a deep learning-based U-Net semantic segmentation model, including: S21. Manually annotate the building facade image dataset and generate corresponding mask images; S22. Construct a deep learning-based U-Net semantic segmentation model, including: The encoder is used for downsampling paths, extracting image features step by step through convolutional and pooling layers while reducing resolution; The decoder is used for upsampling paths to progressively restore the resolution of the feature maps so that the output mask is the same size as the input image. Skip connections are used to connect corresponding layers in the encoder and decoder, passing high-resolution features from the encoder to the decoder. High-resolution feature maps are characterized by larger width and height, more closely resembling the original dimensions of the input image. Low-resolution feature maps, on the other hand, are smaller in width and height, representing the reduction of high-resolution features after downsampling. Features from the initial layer without multiple downsampling steps are called high-resolution features, while features obtained after multiple downsampling steps are called low-resolution features. High-resolution features primarily preserve the detailed features of the image, while low-resolution features primarily preserve abstract global features (such as classification information or overall contours). Therefore, by combining high- and low-resolution features through skip connections, both global semantic information and local detailed information can be preserved simultaneously, thereby improving segmentation accuracy.

[0029] S23. Train and evaluate the model using a dataset that includes the original image and the masked image; S24. Input the new building facade image into the trained U-Net semantic segmentation model and generate the corresponding defect region mask image, including: Input the image of the building facade to be segmented into the trained model, and the model automatically outputs a binary prediction (0 or 1) for each pixel, indicating whether the pixel belongs to the defect area, and obtains the mask image of the corresponding defect area.

[0030] S3. Defect region feature extraction and artificial defect level classification, including: S31. Calculate the defect area, defect height, and temperature difference between the defect area and the normal area based on the defect area mask image, including: The calculation of the defect area begins by counting the number of pixels with a value of 1 in the mask image. These pixels represent defect areas, and then the actual length of the defect is determined by a known value in the image. And the pixel distance of that distance in the image. Determine the scale of the image in the actual scene. Then we can obtain the actual area of ​​the defect region as follows: ; The calculation of the defect height first involves obtaining the pixel coordinates of the top and bottom of the defect in the image using a mask image. The top pixel coordinates of the defect in the image are then labeled as follows: The bottom pixel coordinates of the defect are: The difference between them is the height of the defect in pixels. Then, record the camera's focal length as... The distance from the laser rangefinder to the building facade when the drone captures defect images is recorded as . Then the actual height of the defect can be obtained. ; The temperature difference between the defective and normal regions is calculated by first extracting the temperature values ​​of the defective and normal regions from the infrared image based on the mask image, and then calculating their average temperatures. (Defect area) and (Normal area), the temperature difference between the defective area and the normal area is

[0031] S32. Select severity grading indicators for hollow surface defects on building facades, determine grading standards, manually grade the severity of hollow surface defects on building facades, and obtain a facade defect grade database, including: Severity rating index for hollow defects in building facades: According to relevant national and industry standards, the following evaluation indicators are selected: the proportion of the defect area in the exterior facade area of ​​the building to be inspected, the proportion of the defect height relative to the ground height, and the temperature difference between the defect area and the normal area. The proportion of the defect area in the exterior facade area of ​​the building to be inspected is the building facade hollow defect area ratio coefficient m, and the proportion of the defect height relative to the ground height is the building facade hollow defect height ratio coefficient. l The temperature difference between the defective area and the normal area is the variable temperature difference between the hollow defect areas of the building facade. The temperature difference threshold between the defective area and the normal area is T; As one example, the severity grading standard for hollow defects in building facades: Grade A: Area ratio of hollow defects on building facade (m) ≤ 10%, Height ratio of hollow defects on building facade l ≤0.3, temperature difference variation in the area of ​​hollow defects on the building facade ≤T, Grade A: Exterior facade is intact, poses little public hazard, and requires no observation or repair; Grade B: Area ratio of hollow defects on building facade (m) ≤ 25%, Height ratio of hollow defects on building facade l ≤0.45, temperature difference variation in areas of hollow defects on building facade ≤T+2.5, Class B facade has minor defects, posing little threat to the public; observation and use or minor repairs are recommended. Grade C: The proportion of hollow areas on the building facade is 25% < m < 50%, and the proportion of hollow heights on the building facade is 0.45 < l <0.6, Temperature difference variable T+2.5 in the hollow defect area of ​​the building facade <T+5, Grade C facade has many defects and poses a certain public hazard; repairs are recommended. Grade D: The proportion of hollow area defects on the building facade (m) ≥ 50%, and the proportion of hollow height defects on the building facade (m) ≥ 50%. l≥0.6, temperature difference variation in areas of hollow defects on building facade The facade of grade D (≥T+5) has many defects and poses a significant threat to the public. It is recommended to obtain a detailed inspection report and carry out facade renovation or large-scale repairs.

[0032] S4. Intelligent grading of the severity of hollow defects in building facades, including: Using the severity classification results of hollow defects in building facades as labels, a radial basis function (GWO-RBF) neural network model based on the Grey Wolf Optimization Algorithm is established to intelligently classify the severity of hollow defects in building facades to be classified. The GWO-RBF neural network model is a radial basis function (RBF) neural network model based on the Grey Wolf Optimization Algorithm (GWO), which includes an input layer, hidden layers, and an output layer. The input vector is X = [ , , , , , Representing the defect area, defect height, and temperature difference respectively, the output vector Y = [ , , , , , These represent the expected defect levels; The input layer is used to accept input data and pass it to the hidden layer; Hidden layers are used to map input data to a high-dimensional space. The activation function is usually a Gaussian function, in the form of: = ; in, It is the input vector. It is the center vector. It is the width parameter; The output layer is used to calculate the weighted sum and output the result; The GWO (Grey Wolf) optimization algorithm treats the center vector, width parameter, and weights of the RBF neural network as variables to be optimized, and uses the mean squared error (MSE) to continuously optimize the parameters of the RBF neural network. ; in, It is the actual value. It is the model output. The number of samples used to evaluate model error is used to continuously optimize the center, width, and weights of the RBF neural network through GWO iteration to update the parameters. When GWO reaches the maximum number of iterations or the error reaches the target threshold, the optimized RBF neural network is output.

[0033] See the diagram of the facade defect level database. Figure 2 As shown, in the early stages of database construction, a large number of hollow defects on the building facade were manually annotated. Defect areas were identified and marked using standard temperature difference and standard defect location temperature distribution as reference benchmarks. Indicators such as defect area, defect height, and temperature difference were also introduced, based on the hazard level standards for hollow defects on building facades. Figure 2 a, b, c, and d respectively indicate the four hazard levels of different types of facade defects, namely A, B, C, and D. The hazard level standard is given by combining current national industry standards, past accident and maintenance experience, and opinions of facade maintenance experts.

[0034] The severity grading results for hollow defects in the building facade are shown below. Figure 3 As shown, the GWO-RBF neural network model intelligently outputs the defect results, marking the location and boundary of the defect identified by deep learning in the infrared thermal image, explaining the degree of defect hazard determined by the model and the repair suggestions on the right, and finally outputting the inspection results.

[0035] This deep learning-based intelligent grading method for the severity of hollow defects in building facades rapidly and intelligently identifies, locates, and grades defects. This method improves the efficiency of hollow defect detection while reducing the risk and cost of traditional facade inspections. Furthermore, it intelligently grades the severity of defects based on identification, making it significant in the fields of building facade defect detection, building facade completion acceptance, and building facade maintenance. The accuracy of this method's identification and grading can be continuously improved as the database expands and the learning depth increases, leading to increased efficiency. This method is in a continuous iterative optimization process, ultimately achieving a high level of accuracy.

[0036] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for classifying the severity of hollow defects in building facades, characterized in that, Includes the following steps: S1. Input the building facade image into the trained U-Net semantic segmentation model and generate the corresponding defect region mask image; S2. Calculate the defect area, defect height, and temperature difference between the defect area and the normal area based on the defect area mask image; defect area for: ; in This represents the actual area of ​​the defective region. To count the number of pixels with a value of 1 in a masked image, The scale of an image in a real-world scene is determined by defining a known actual length within the image. And the actual length in the image pixel distance Sure; Defect height: ; ; in The actual height of the defect. For the camera's focal length, The distance from the building facade to the defect image captured by the drone using a laser rangefinder. The height of the defect in pixels; The coordinates of the top pixel of the defect in the image. Given the bottom pixel coordinates of the defect, obtain the top and bottom pixel coordinates of the defect in the image using the mask image; Temperature difference between the defective area and the normal area: ; in The temperature difference between the defective area and the normal area. This represents the average temperature value of the defect area. This represents the average temperature value of the normal area. S3. Select the severity grading index for hollow defects on the building facade, determine the grading standard, manually grade the severity of hollow defects on the building facade, and obtain a database of facade defect grades. The specific grading indicators for the severity of hollow defects in building facades include: The following parameters were selected as evaluation indicators: the proportion of defect area in the building facade area to be inspected, the proportion of defect height to ground height, and the temperature difference between the defect area and the normal area. The proportion of defect area in the building facade area to be inspected is the building facade hollow defect area ratio coefficient. m The ratio of the defect height to the ground height is the building facade hollow defect height ratio coefficient. l Temperature difference between defective and normal areas The temperature difference variable is the area of ​​hollow defects in the building facade, and the threshold temperature difference between the defective area and the normal area is T. S4. Using the severity classification results of hollow defects in building facades as labels, a radial basis function neural network model based on the gray wolf optimization algorithm is used to intelligently classify the severity of hollow defects in building facades in the images to be classified. The radial basis function neural network model includes: The input layer is used to receive input data and pass it to the hidden layer; Hidden layers are used to map input data to a high-dimensional space. The activation function is a Gaussian function, in the form of: = ; in, It is the input vector. It is the center vector. This is the width parameter, used to control the extent of expansion of the radial basis functions and determine the shape of the activation function. It is the input vector With the center vector The distance between them; It is a center The Gaussian distribution function with symmetry center, when and The closer it is to 1, the closer the function value is to 1; when Leave At greater distances, the function value rapidly decays to near zero; The output layer is used to calculate the weighted sum and output the result. The Grey Wolf optimization algorithm treats the center vector, width parameter, and weights of the radial basis function neural network as variables to be optimized, and uses the mean squared error (MSE) to continuously optimize the parameters of the radial basis function neural network. ; in, It is the actual value. It is the model output. The number of samples used to evaluate model error is used to continuously optimize the center, width, and weights of the radial basis function neural network through the Grey Wolf Optimization Algorithm. When the Grey Wolf Optimization Algorithm reaches the maximum number of iterations or the error reaches the target threshold, the optimized radial basis function neural network is output.

2. The method for classifying the severity of hollow defects in building facades based on deep learning according to claim 1, characterized in that, Training a deep learning-based U-Net semantic segmentation model includes the following steps: The drone, equipped with both a visible light camera and an infrared thermal imaging camera, acquires infrared and visible light images of the building facade. The building facade infrared and visible light images are preprocessed to obtain a building facade image dataset; The data from the building facade image dataset were manually annotated and corresponding mask images were generated. The deep learning-based U-Net semantic segmentation model was trained and evaluated using the original image and the masked image.

3. The method for classifying the severity of hollow defects in building facades based on deep learning according to claim 2, characterized in that, The steps for preprocessing the acquired images to obtain a dataset of building facade images include: The infrared thermal images obtained by taking separate shots of the building area are stitched together. Remove background areas unrelated to the building, retaining only the main body of the building. Select the temperature display range: 20°C ~ 40°C for images taken during the day and 5°C ~ 25°C for images taken at night. Use 2 to 3 colors to display the image and highlight the distribution of defects in the image. Remove temperature distribution unrelated to external wall defects caused by wall contamination and equipment heat sources; High-resolution visible light images are stitched together, and the stitched high-resolution visible light images are overlaid with infrared thermal images to obtain a dataset of building facade images.

4. The method for classifying the severity of hollow defects in building facades based on deep learning according to claim 1, characterized in that, The deep learning-based U-Net semantic segmentation model includes: The encoder is used for downsampling paths, extracting image features step by step through convolutional and pooling layers while reducing resolution; The decoder is used for upsampling paths, progressively restoring the resolution of the feature maps, and outputting a mask of the same size as the input image; Skip connections are used to connect corresponding layers in the encoder and decoder, passing high-resolution features from the encoder to the decoder.

5. The method for classifying the severity of hollow defects in building facades based on deep learning according to claim 1, characterized in that, In step S1, the building facade image is input into the trained U-Net semantic segmentation model to generate the corresponding defect region mask image, including: Input the building facade image to be segmented into the trained U-Net semantic segmentation model, and output the binary prediction of each pixel. If the prediction result of the U-Net semantic segmentation model is 1, it means that the pixel belongs to the hollow defect area of ​​the building facade. If the classification result of the U-Net semantic segmentation model is 0, it means that the pixel belongs to the non-defect area, and the corresponding mask image of the defect area is obtained.

6. The method for classifying the severity of hollow defects in building facades based on deep learning according to claim 1, characterized in that, In step S3, the severity grading standard for hollow defects in the building facade includes: Grade A: Proportion coefficient of hollow area in building facade m ≤10%, building facade hollow defect height ratio coefficient l ≤0.3, temperature difference variation in the area of ​​hollow defects on the building facade ≤T, Grade A: Exterior facade is intact, poses little public hazard, and requires no observation or repair; Grade B: Proportion of hollow areas in building facade Number m ≤25%, building facade hollow defect height ratio coefficient l ≤0.45, temperature difference variation in the area of ​​hollow defects on the building facade ≤T+2.5, Class B facade has minor defects, posing little threat to the public; observation and use or minor repairs are recommended. Grade C: The proportion of hollow areas in the building facade is less than 25%. m <50%, the height ratio coefficient for hollow defects on the building facade is 0.45 < l <0.6, Temperature difference variable T+2.5 in the hollow defect area of ​​the building facade <T+5, Grade C facade has many defects and poses a certain public hazard; repairs are recommended. Grade D: Percentage of hollow area defects on building facade m ≥50%, the proportion coefficient of hollow defects in building facade height l ≥0.6, temperature difference variation in areas of hollow defects on building facade The facade of grade D (≥T+5) has many defects and poses a significant threat to the public. It is recommended to obtain a detailed inspection report and carry out facade renovation or large-scale repairs.

7. A computer device, characterized in that, It includes: a memory and a processor, and a computer program stored in the memory, which, when executed on the processor, implements the deep learning-based method for classifying the severity of hollow defects in building facades as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Historic building wall wooden column defect grading method based on multi-source data fusion and machine learning

    CN118570125A

  • Non-contact building facade defect detection device

    CN221725881U