Historical building brick wall surface damage intelligent identification and diagnosis method

Abstract

The invention relates to the technical field of historical building protection, and discloses an intelligent identification and diagnosis method for historical building brick wall surface damage, which breaks through the limitations of low efficiency, strong subjectivity and easy detail omission of traditional manual exploration by fusing unmanned aerial vehicle high-precision image acquisition, deep learning semantic segmentation and multi-modal feature analysis technologies. And automatic rapid positioning and accurate classification of large-range wall surface damage are realized. Environmental distortion is eliminated by using a self-adaptive image registration algorithm, fine texture features are extracted in combination with a damage probability prediction model, and complex damage forms such as salting-out crystallization, plant root erosion and weathering spalling are effectively distinguished; through three-dimensional space clustering and visual marking technologies, the damage distribution density and the evolution trend are visually presented, and a reliable basis is provided for scientifically formulating a grading repair strategy. Besides, through establishment of a damage database and iterative optimization of an intelligent diagnosis model, the building health condition can be tracked for a long time, and technical support is provided for preventive protection and sustainable management of cultural heritage.

Description

Technical Field

[0001] The present invention relates to the technical field of historical building protection, and in particular to a method for intelligently identifying and diagnosing damage to brick walls of historical buildings. Background Art

[0002] Brick walls, a common exterior facade of historical buildings, are inevitably subject to various types of damage due to long-term exposure to the elements, including wind, rain, sun, and biological action. Weathering, alkali erosion, and vegetation erosion are the three most typical forms of damage, each with its own unique visual characteristics:

[0003] Weathering damage: Under the action of long-term weathering, the internal structure of the brick surface material is destroyed, resulting in a loose surface. As time goes by, part of the surface material gradually falls off and disappears, making the brick wall appear bumpy and rough, seriously affecting the structural integrity and aesthetics of the wall.

[0004] Alkali damage: Soluble salts in building materials, under the action of water migration, precipitate white crystals on the brick wall surface as the water evaporates. These crystals are usually irregularly distributed and attached to the surface of the bricks, which not only changes the original color of the wall, but may also further erode the bricks and aggravate wall damage.

[0005] Vegetation erosion damage: Under suitable temperature and humidity conditions, gaps or loose areas in brick walls can easily become the habitat of plant seeds. As the plants grow, their roots gradually penetrate into the bricks, causing physical damage to the brick structure. At the same time, plants and green mosses cover the walls, which on the one hand obscures the original appearance of the walls, and on the other hand, their metabolic process may also cause chemical erosion to the wall material.

[0006] When it comes to restoring large-scale brick walls in historic buildings, traditional methods that rely on manual on-site inspections and empirical judgment are inefficient and highly subjective, making them difficult to meet the demands of large-scale, high-precision restoration. With the rapid development of computer vision technology and artificial intelligence, it is imperative to efficiently and accurately identify and diagnose brick wall damage through intelligent methods such as image recognition and neural networks. To this end, a method for intelligent identification and diagnosis of brick wall damage in historic buildings was proposed. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention aims to overcome the limitations of existing brick wall damage detection methods and provide an intelligent identification and diagnosis method for brick wall damage in historical buildings based on advanced imaging technology and intelligent algorithms, so as to achieve rapid and accurate determination of the type and scope of damage to large-area brick walls, provide a reliable basis for the subsequent scientific formulation of repair plans, and improve the efficiency and quality of historical building protection work.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligently identifying and diagnosing damage to brick walls of historical buildings, comprising the following steps:

[0009] Step 1: Image sample collection and training: Take a large number of images of brick walls damaged by weathering, alkali efflorescence, and vegetation erosion, and use image recognition training to distinguish the visual features of different damage types.

[0010] Step 2: Acquisition and preprocessing of brick wall image data: Acquisition of brick wall image data by panoramic image acquisition or oblique photography, and generation of orthographic projection of the brick wall by image registration;

[0011] Step 3: Brick wall image segmentation and potential damage interface analysis:

[0012] The brick wall surface is divided into regions by image segmentation technology to form several regional images;

[0013] Using the results of image learning in step 1, perform pixel recognition on the regional image and analyze the potential damage interface;

[0014] Step 4: Determine and mark the damaged area: Use a clustering algorithm to mark the color blocks, and finally stitch and screen several regional images to give the damaged area in the orthographic projection image, and use different color blocks to distinguish different damages.

[0015] Preferably, the step 1 specifically includes:

[0016] Collect at least 500 high-definition digital photographs of brick walls from historical buildings of different eras, regions, and architectural styles, showing signs of weathering, alkali efflorescence, and vegetation erosion under various environmental conditions, ensuring that the sample covers all possible damage morphological changes;

[0017] Professional image annotation tools are used to accurately annotate the damaged areas in each photo. The annotation content includes but is not limited to the damage type (weathering, alkali efflorescence, vegetation erosion) and boundary contour information.

[0018] The labeled image sample set is divided into training set, validation set and test set according to a certain ratio.

[0019] A convolutional neural network (CNN) model is built using a deep learning framework. The model structure includes multiple convolutional layers, pooling layers, and fully connected layers. The input layer receives image data, and the output layer outputs the damage type prediction results.

[0020] The CNN model was trained using the training set, and the model parameters were continuously adjusted through the back-propagation algorithm so that the accuracy of the model on the validation set reached the predetermined threshold. The final model performance was evaluated using the test set to ensure that the model could accurately distinguish the visual features of the three types of injuries.

[0021] Preferably, the brick wall image data in step 2 includes the following contents:

[0022] Based on the actual height, area, and surrounding environment of the brick wall, a panoramic camera or a drone equipped with oblique photography equipment is selected for image acquisition;

[0023] For low walls with open perimeters, a panoramic camera is used to shoot at multiple locations on the ground, ensuring a certain degree of overlap between adjacent shooting points.

[0024] For the brick walls of tall and complex buildings, drones are used for oblique photography, with appropriate flight altitudes and route planning set to capture wall images from multiple angles, while also ensuring that the overlapping areas between images meet subsequent processing requirements.

[0025] Preferably, the image preprocessing in step 2 specifically includes:

[0026] An image registration algorithm based on feature point matching is used to find feature points with the same name between adjacent images. By solving the transformation matrix, multiple images are unified into the same coordinate system, and then an orthographic projection of the brick wall is generated, eliminating the geometric deformation caused by the shooting angle and obtaining a planar image that conforms to the actual wall size ratio.

[0027] Preferably, the image segmentation technology used in step 3 specifically includes:

[0028] The orthographic projection image is processed by combining a variety of image segmentation methods such as threshold segmentation, region growing, and semantic segmentation. For areas with relatively clear boundaries between bricks and backgrounds, a simple threshold segmentation method is used to quickly separate the main part of the brick wall based on differences in features including but not limited to color and grayscale. For complex situations inside the brick wall, such as color gradients and stains, a region growing algorithm is used, starting from the seed point, and gradually merging adjacent pixels according to the similarity criterion to achieve fine regional division. At the same time, a semantic segmentation model based on deep learning is introduced to accurately classify different semantic categories such as bricks, damaged areas, and gaps, and divide the brick wall into several non-overlapping regional images with clear semantics, providing a basis for subsequent damage identification.

[0029] Preferably, the pixel identification in step 3 specifically includes:

[0030] The segmented regional image is fed into a pre-trained CNN model. The model predicts the damage type for each pixel and outputs a damage type probability map of the same size as the regional image. Each pixel in the probability map corresponds to a three-dimensional vector (representing the probability values of weathering, alkali efflorescence, and vegetation erosion).

[0031] According to the set probability threshold, potential damage pixels are determined, and adjacent potential damage pixels are connected through morphological processing to outline the contour of the potential damage interface, and the possible damage type and approximate range in each area are preliminarily judged.

[0032] Preferably, the color block marking in step 4 specifically includes:

[0033] For the potential damaged pixels identified in each area image, a clustering algorithm is used to divide them into different clusters according to the color and spatial position characteristics of the pixels. Each cluster represents a relatively concentrated damaged area.

[0034] Preferably, the damage area and the damage area distinction in step 4 specifically include:

[0035] Assign unique color labels to different clusters, including but not limited to red for weathering damage, blue for alkali damage, and green for vegetation erosion damage, and visually display the distribution of damage areas on regional images through color labels;

[0036] All regional images are spliced according to their original positions in the orthographic projection, and duplicate marks in overlapping areas are removed to integrate the damaged area information on the complete orthographic projection image. Different types of damage are clearly distinguished by different color blocks, providing a damage distribution map for the repair work.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] By integrating high-precision drone image acquisition, deep learning semantic segmentation, and multimodal feature analysis technology, this invention overcomes the limitations of traditional manual surveys, which are low efficiency, strong subjectivity, and easy omission of details, and achieves automated rapid positioning and accurate classification of large-scale wall damage. An adaptive image registration algorithm is used to eliminate environmental distortion, and a damage probability prediction model is combined to extract subtle texture features, effectively distinguishing complex damage forms such as salt precipitation crystallization, plant root erosion, and weathering and peeling. Through three-dimensional spatial clustering and visual labeling technology, the damage distribution density and evolution trend are intuitively presented, providing a reliable basis for the scientific formulation of graded repair strategies. In addition, by establishing a damage database and iteratively optimizing the intelligent diagnostic model, the health status of buildings can be tracked over the long term, providing technical support for the preventive protection and sustainable management of cultural heritage.

[0039] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of the framework of the intelligent identification and diagnosis method for damage to brick walls of historical buildings of the present invention;

[0041] Figure 2 This is a specific flow chart of the intelligent identification and diagnosis method for damage to brick walls of historical buildings of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] See also Figure 1 and Figure 2 The present invention provides an intelligent identification and diagnosis method for damage to brick walls of historical buildings; the method includes the following contents:

[0044] Example 1:

[0045] This example takes a section of an ancient city wall from the Ming Dynasty as an example. The total length is 800 meters and the height is 12 meters. The wall is built with blue bricks, and the surface is damaged by weathering, alkali efflorescence, and vegetation erosion caused by long-term rain erosion, salt crystallization, and plant root invasion.

[0046] Step 1: Image sample collection and training

[0047] 1. Data collection and annotation

[0048] Data source:

[0049] Damage images of ancient city walls were collected from different climatic regions in China (the humid Jiangnan region, the arid Northwest region, and the freeze-thaw region in the Northeast), covering different damage types (weathering, alkali efflorescence, and vegetation erosion), damage degrees (mild, moderate, and severe), and lighting conditions (sunny, cloudy, and dawn and dusk); covering damage morphologies under different climatic environments to enhance the model's generalization ability.

[0050] Use a high-resolution digital camera (such as the Canon EOS 5D Mark IV) to capture images with a resolution of ≥4000 × 6000 pixels and a total sample size of 2000 images. High-resolution images ensure that textural details of subtle damage (such as alkali crystallization and tiny cracks) can be captured, providing a basis for subsequent pixel-level identification.

[0051] Marking method:

[0052] Use the LabelImg tool to perform polygon labeling. The labeling content includes:

[0053] Damage types: weathering (code 1), alkali erosion (code 2), and vegetation erosion (code 3).

[0054] Boundary contour: Accurately mark the scope of the damaged area with pixel coordinates.

[0055] Annotation example (XML format):

[0056]

[0057]

[0058] 2. Dataset Division and Enhancement

[0059] The labeled dataset was divided into a training set (1400 images), a validation set (400 images), and a test set (200 images) according to a 7:2:1 ratio. The training set was randomly rotated (±15°), horizontally flipped, brightness adjusted (±20%), and Gaussian noise (σ = 0.05) was added.

[0060] Simulate the lighting changes and shooting angle differences that may be encountered in actual applications to avoid model overfitting; the validation set is used for hyperparameter tuning, and the test set ensures the objectivity of model performance evaluation.

[0061] 3. Convolutional Neural Network (CNN) Model Construction

[0062] Based on the improvement of ResNet-50, an attention mechanism module (SE Block) is added with the following structure:

[0063] Input layer: 224×224×3 RGB image (obtained by center cropping and scaling the original image).

[0064] SE Block: After inserting each residual block, global average pooling (GAP) is used to compress spatial information and generate channel attention weights. The formula is:

[0065] F out =F in ·σ(W2·ReLU(W1·GAP(F in )))

[0066] Among them, GAP takes the mean of the spatial dimensions of the feature map, W1 and W2 are the weights of the fully connected layer, and σ is the Sigmoid function;

[0067] SE Block adopts a targeted design: for local texture roughening, SE Block enhances the model's sensitivity to edge gradients through channel attention; for uneven distribution of white crystals, SE Block suppresses background interference and focuses on high-brightness areas.

[0068] Classification head: global average pooling layer + fully connected layer (output 3D vector, corresponding to three types of damage probabilities);

[0069] Loss function: Cross entropy loss:

[0070]

[0071] Where N is the batch size (32), y i,c is the true label, p i,c Predict probabilities for the model.

[0072] Optimization strategy: Adam optimizer (initial learning rate 0.001, decaying by 0.1 times every 30 rounds).

[0073] Step 2: Brick wall image data acquisition and preprocessing

[0074] 1. Implementation of UAV tilt photography technology

[0075] Equipment configuration and flight parameters

[0076] Drone model: DJI M300 RTK, equipped with a Zenmuse P1 camera (full-frame 35mm, 45 megapixels), equipped with an RTK (Real-Time Kinematic) module, with a horizontal positioning accuracy of ±1 cm and an elevation accuracy of ±1.5 cm.

[0077] Flight Planning:

[0078] Flight altitude: 40 meters (guaranteed ground resolution 2cm / pixel);

[0079] Route setting: Five parallel routes were planned along the wall, with a route spacing of 15 meters, a single route overlap of 40%, and a tilt angle of 70° to ensure complete coverage of the wall texture details;

[0080] Data output: A total of 1,200 RAW images were collected, each with a size of 7,952 × 5,304 pixels, and stored in DNG format to preserve the original information.

[0081] The 2cm / pixel resolution can clearly capture subtle damage such as brick joints (width ≥ 0.5cm) and alkali crystals (diameter ≥ 1cm); the 70° tilt angle avoids wall texture compression, retains vertical details, and provides rich feature points for subsequent alignment.

[0082] 2. Image registration and orthographic projection generation

[0083] 2.1 Preprocessing operations:

[0084] Brightness balancing: Histogram matching is used to eliminate illumination differences. The cumulative distribution function (CDF) is calculated for each image, and a mapping function is used to adjust its histogram to be consistent with the reference image.

[0085] Denoising: Use the non-local means denoising (NL-Means) algorithm, filter strength h = 10, search window 21×21, and similarity window 7×7. Formula:

[0086]

[0087] Where Z(x) is the normalization factor; x is the coordinate of the pixel to be denoised; Ω is the search window; I(x) is the original grayscale value or RGB vector of pixel x; I(y) is the original grayscale value or RGB vector of pixel y; h is the filter strength parameter; NL-Means(x) is the denoised value of pixel x.

[0088] 2.2 Feature point matching and registration:

[0089] SIFT feature extraction:

[0090] Scale-space extrema detection: locating key points through Difference of Gaussian (DoG) pyramid;

[0091] Key point description: Generate a 128-dimensional directional histogram descriptor.

[0092] Feature matching: Use the nearest neighbor distance ratio (NNDR) to screen matching pairs, with a threshold of 0.6. Formula:

[0093]

[0094] Among them, d1 is the nearest neighbor descriptor; d2 is the next nearest neighbor descriptor.

[0095] Affine transformation solution:

[0096] For each pair of adjacent images, the affine transformation matrix is estimated by the RANSAC algorithm:

[0097]

[0098] Where af is the affine parameter, x,y are the original image coordinates, and x′,y′ are the transformed coordinates. The number of RANSAC iterations is set to 1000, and the inlier threshold is 2 pixels.

[0099] 3. Global optimization and orthographic projection generation

[0100] The bundle adjustment is used to optimize the reprojection error: minimize the reprojection error. The objective function is:

[0101]

[0102] Among them, R i is the rotation matrix of the i-th image (3×3); t i is the translation vector of the i-th image (3×1); P j is the coordinate of the j-th three-dimensional point (3×1); m ij is the observed coordinate of the jth point on image i.

[0103] Orthographic projection generation:

[0104] The optimized image is projected onto the wall plane coordinate system to generate an orthophoto with a resolution of 2 cm / pixel and a size of 40,000 × 6,000 pixels (corresponding to an 800 m × 12 m wall). Bilinear interpolation is used to fill missing pixels to ensure image continuity.

[0105] Orthographic projection corrects the perspective distortion of oblique photography and eliminates geometric distortion; and through Bundle Adjustment optimization, it reduces global coordinate consistency errors.

[0106] Step 3: Image segmentation and damage interface analysis

[0107] 1. Implementation of multi-strategy image segmentation technology

[0108] 1.1 Threshold segmentation:

[0109] Background culling: For the sky and vegetation background, set the RGB threshold:

[0110] Sky: R<50, G<50, B>150 (extract blue channel);

[0111] Vegetation: Convert to HSV space, set H∈[35°,85°], S>40, V>30;

[0112] Quickly separate the main body of the brick wall to reduce the interference area for subsequent processing.

[0113] 1.2 Region Growing Algorithm:

[0114] Seed point selection: Manually select low gray value points (gray value <30) at the brick joints as the starting point for growth;

[0115] Growth criterion: Merging adjacent pixels must satisfy the grayscale difference |I(x,y)-I sced |<10;

[0116] Iterative process: Use queue data structure to implement breadth-first search until there are no new pixels to merge;

[0117] Finely divide the brick area to avoid mis-segmentation of gradient textures (such as color transitions caused by weathering) by threshold segmentation.

[0118] 1.3U-Net semantic segmentation:

[0119] Model Architecture:

[0120] Encoder: 4 layers of downsampling, each layer contains two 3×3 convolutions (stride 1, padding 1) + ReLU + 2×2 maximum pooling;

[0121] Decoder: 4 layers of upsampling, each layer uses transposed convolution (kernel 2×2, stride 2) to restore the resolution and concatenate with the feature map of the corresponding layer of the encoder (skip connection);

[0122] Output layer: 1×1 convolution + Softmax activation, outputting a three-channel probability map (brick body, damage, gap).

[0123] Loss function: weighted cross entropy + Dice Loss, formula:

[0124]

[0125] Cross Entropy Loss:

[0126]

[0127] Dice coefficient:

[0128]

[0129] Where c is the category index (such as weathering, alkali efflorescence, vegetation erosion); y c is the one-hot encoding of the true label; p c is the category probability predicted by the model; w c is the category weight;

[0130] Using the Adam optimizer (learning rate 1e-4), batch size 8, training for 100 epochs, the Dice coefficient on the validation set reached 0.89;

[0131] Cross entropy loss directly optimizes the degree of match between the model output probability and the true label, and after weighting, alleviates the impact of sample imbalance on the model; Dice Loss optimizes region overlap and is particularly suitable for pixel-level segmentation tasks (such as damage area identification) and is highly robust to category imbalance problems.

[0132] 2. Pixel-level damage identification and morphological optimization

[0133] 2.1 CNN Model Inference

[0134] Input processing: The segmented region image (512×512 pixels) from step 1 is fed into the pre-trained ResNet-50 model.

[0135] Output analysis: Each pixel outputs a 3D probability vector [pweathering, palkali efflorescence, pvegetation][pweathering, palkali efflorescence, pvegetation];

[0136] Threshold screening: Set the probability threshold τ = 70%. If pc ≥ τ, it is marked as type c damage.

[0137] 2.2 Morphological post-processing

[0138] Closing operation:

[0139] Dilation: Use a 3×3 rectangular kernel to expand the damaged area and connect discrete pixels. Formula:

[0140]

[0141] Erosion: Use a 3×3 cross kernel to smooth the boundaries. Formula:

[0142]

[0143] Where A is the original binary image; B is the structural element; z is the pixel position coordinate in the image (z = x, y); (B) z The region after the structural element B is translated to position z.

[0144] Dilation followed by corrosion can connect the broken areas while maintaining the original shape and size, making it suitable for repairing discontinuous boundaries of damaged areas; eliminating noise points (such as single-pixel false detection) to form a continuous and closed damage contour.

[0145] Step 4: Damage area marking and output

[0146] 1. K-Means clustering and color block labeling

[0147] 1.1 Feature Engineering

[0148] Color features: Convert to Lab color space and extract L (lightness), a (green-red axis), and b (blue-yellow axis) components; consistent with human eye perception, better distinguishing between alkali (high L), vegetation (high b), and weathering (high a).

[0149] Spatial features: Normalize pixel coordinates x,y to the range [0,1]; constrain the spatial continuity of clustering results to avoid mistaken merging of distant similar color regions;

[0150] Eigenvector: x = [L, a, b, x, y] T , dimension 5.

[0151] 1.2 Clustering Algorithm Implementation

[0152] 1.21 Objective function: Minimize the intra-cluster squared error (SSE):

[0153]

[0154] Where C is; k=3 is the preset number of clusters (corresponding to three types of damage); μ i is the cluster center.

[0155] 1.22 Algorithm steps:

[0156] Initialization: K-Means++ algorithm selects the initial cluster center to avoid local optimality.

[0157] Iterative optimization: The maximum number of iterations is 300, and the convergence condition is that the SSE change is <1e-5.

[0158] 1.23 Color Mapping:

[0159] Weathering damage: red (RGB 255,0,0), corresponding to high aa value in Lab space (reddish);

[0160] Alkali damage: blue (RGB 0,0,255), corresponding to high LL values (bright areas);

[0161] Vegetation erosion: green (RGB 0,255,0), corresponding to a high BB value (greenish).

[0162] 2. Global stitching and geographic information integration

[0163] 2.1 Splicing strategy

[0164] Coordinate alignment: Based on the coordinate system of the orthographic projection, 500 regional images are stitched together according to their original positions;

[0165] Redundancy elimination: For the marking of overlapping areas, the damage category with the highest confidence (largest probability value) is retained.

[0166] 2.2 Output format and engineering application

[0167] GeoTIFF file: embeds geographic coordinate information (WGS84 coordinate system) and can be directly imported into ArcGIS or AutoCAD;

[0168] Damage statistics report: automatically generates the area, location coordinates and repair priority recommendations for each damage type.

[0169] Example 2: Damage identification of brick walls in small historic buildings

[0170] The exterior wall of a small Ming and Qing dynasty courtyard house, approximately 3 meters high and 10 meters long, was captured using a high-resolution digital SLR camera. Ten images were taken horizontally, 1-2 meters apart, at intervals of 1 meter along the wall. The overlap between the images was approximately 40%. The images were taken at different times of day, encompassing morning, afternoon, and overcast lighting conditions, to obtain a variety of original images.

[0171] We collected 500 images of weathered, alkali-wetting, and vegetation-eroded brick walls from similar local historic buildings. We used annotation tools to precisely mark the damaged areas and divided the images into training, validation, and test sets in a 70:20:10 ratio. We then built a five-layer convolutional neural network model and trained the model using the training set. After 100 iterations, the model achieved an accuracy of 92% on the validation set, and its performance was verified on the test set.

[0172] Import the photos of the courtyard's exterior wall into the image registration software, use the SIFT algorithm to extract feature points, and achieve image registration through affine transformation to generate an orthographic projection image;

[0173] Threshold segmentation combined with region growing algorithm was used to divide the orthographic projection image into regions, obtaining 20 regional images. These images were then input into the trained CNN model. Based on the damage probability map output by the model, the probability threshold was set to 65%, and the potential damage interface was determined. The contours were then refined through morphological processing.

[0174] The K-Means clustering algorithm was used for potential damaged pixels, and the weathering, alkali efflux, and vegetation erosion damage areas were marked with red, blue, and green colors respectively. The 20 regional images were spliced back into the orthographic projection map, clearly showing a total of 3 weathering damage areas (total area of approximately 1.5 square meters), 2 alkali efflux damage areas (total area of approximately 0.8 square meters), and 4 vegetation erosion damage areas (total area of approximately 2 square meters) on the wall, providing precise guidance for subsequent repairs.

[0175] This embodiment proposes a method for identifying and diagnosing damage to brick walls of historical buildings based on the collaboration of multi-source data fusion and intelligent algorithms. It uses oblique photography technology to obtain high-precision wall images and generate geometrically corrected orthographic projections. It combines an adaptive semantic segmentation model to perform fine regional division of bricks, damage, and background. It uses a pre-trained deep convolutional network to achieve pixel-level damage probability prediction, and uses multi-scale morphological optimization to eliminate noise interference and reconstruct the continuous boundary of damage. In view of the complex damage distribution characteristics, a clustering algorithm integrating Lab color space and spatial coordinates is used to realize automatic classification of damage types and color block marking, and finally generates a visual damage heat map that integrates geographic coordinate information. This method breaks through the temporal and spatial limitations of traditional manual surveys, and realizes millimeter-level damage positioning and multi-category accurate identification of large-area walls. It not only supports quantitative decision-making on repair plans, but also can analyze damage evolution trends through historical data comparison. It provides a scalable technical framework for the digital protection and preventive maintenance of cultural heritage, and has significant social benefits and engineering application value.

Claims

1. A method for intelligent identification and diagnosis of damage to brick walls of historical buildings, characterized by: The following steps are involved: Step 1: Image sample collection and training: Take a large number of images of brick walls damaged by weathering, alkali efflorescence, and vegetation erosion, and use image recognition training to distinguish the visual features of different damage types. Step 2: Acquisition and preprocessing of brick wall image data: Acquisition of brick wall image data by panoramic image acquisition or oblique photography, and generation of orthographic projection of the brick wall by image registration; Step 3: Brick wall image segmentation and potential damage interface analysis: The brick wall surface is divided into regions by image segmentation technology to form several regional images; Using the results of image learning in step 1, perform pixel recognition on the regional image and analyze the potential damage interface; Step 4: Determine and mark the damaged area: Use a clustering algorithm to mark the color blocks, and finally stitch and screen several regional images to give the damaged area in the orthographic projection image, and use different color blocks to distinguish different damages.

2. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 1 is characterized in that: The step 1 specifically includes: Collect at least 500 high-definition digital photographs of brick walls from historical buildings of different eras, regions, and architectural styles, showing signs of weathering, alkali efflorescence, and vegetation erosion under various environmental conditions, ensuring that the sample covers all possible damage morphological changes; Professional image annotation tools are used to accurately annotate the damaged areas in each photo. The annotation content includes but is not limited to the damage type (weathering, alkali efflorescence, vegetation erosion) and boundary contour information. The labeled image sample set is divided into training set, validation set and test set according to a certain ratio. A convolutional neural network (CNN) model is built using a deep learning framework. The model structure includes multiple convolutional layers, pooling layers, and fully connected layers. The input layer receives image data, and the output layer outputs the damage type prediction results. The CNN model was trained using the training set, and the model parameters were continuously adjusted through the back-propagation algorithm so that the accuracy of the model on the validation set reached the predetermined threshold. The final model performance was evaluated using the test set to ensure that the model could accurately distinguish the visual features of the three types of injuries.

3. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 1 is characterized in that: The brick wall image data in step 2 includes the following contents: Based on the actual height, area, and surrounding environment of the brick wall, a panoramic camera or a drone equipped with oblique photography equipment is selected for image acquisition; For low walls with open perimeters, a panoramic camera is used to shoot at multiple locations on the ground, ensuring a certain degree of overlap between adjacent shooting points. For the brick walls of tall and complex buildings, drones are used for oblique photography, with appropriate flight altitudes and route planning set to capture wall images from multiple angles, while also ensuring that the overlapping areas between images meet subsequent processing requirements.

4. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 3 is characterized in that: The image preprocessing in step 2 specifically includes: An image registration algorithm based on feature point matching is used to find feature points with the same name between adjacent images. By solving the transformation matrix, multiple images are unified into the same coordinate system, and then an orthographic projection of the brick wall is generated, eliminating the geometric deformation caused by the shooting angle and obtaining a planar image that conforms to the actual wall size ratio.

5. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 1 is characterized in that: The image segmentation technology used in step 3 specifically includes: The orthographic projection image is processed by combining a variety of image segmentation methods such as threshold segmentation, region growing, and semantic segmentation. For areas with relatively clear boundaries between bricks and backgrounds, a simple threshold segmentation method is used to quickly separate the main part of the brick wall based on differences in features including but not limited to color and grayscale. For complex situations inside the brick wall, such as color gradients and stains, a region growing algorithm is used, starting from the seed point, and gradually merging adjacent pixels according to the similarity criterion to achieve fine regional division. At the same time, a semantic segmentation model based on deep learning is introduced to accurately classify different semantic categories such as bricks, damaged areas, and gaps, and divide the brick wall into several non-overlapping regional images with clear semantics, providing a basis for subsequent damage identification.

6. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 5 is characterized in that: The pixel identification in step 3 specifically includes: The segmented regional image is fed into a pre-trained CNN model. The model predicts the damage type for each pixel and outputs a damage type probability map of the same size as the regional image. Each pixel in the probability map corresponds to a three-dimensional vector (representing the probability values of weathering, alkali efflorescence, and vegetation erosion). According to the set probability threshold, potential damage pixels are determined, and adjacent potential damage pixels are connected through morphological processing to outline the contour of the potential damage interface, and the possible damage type and approximate range in each area are preliminarily judged.

7. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 1 is characterized in that: The color block marking in step 4 specifically includes: For the potential damaged pixels identified in each area image, a clustering algorithm is used to divide them into different clusters according to the color and spatial position characteristics of the pixels. Each cluster represents a relatively concentrated damaged area.

8. The intelligent identification and diagnosis method for damage to brick walls of historical buildings according to claim 7 is characterized in that: The damage area and the damage area distinction in step 4 specifically include: Assign unique color labels to different clusters, including but not limited to red for weathering damage, blue for alkali damage, and green for vegetation erosion damage, and visually display the distribution of damage areas on regional images through color labels; All regional images are spliced according to their original positions in the orthographic projection, and duplicate marks in overlapping areas are removed to integrate the damaged area information on the complete orthographic projection image. Different types of damage are clearly distinguished by different color blocks, providing a damage distribution map for the repair work.

Images