Visual detection device and method for building engineering wall quality

Abstract

The invention provides a visual detection device and method for building engineering wall quality, and the method comprises the steps: firstly obtaining a multi-angle image set of a to-be-detected wall, which comprises surface images and internal structure scanning images under different illumination conditions, and then carrying out the visual feature extraction of the multi-angle image set, according to the method, a comprehensive visual detection feature set covering surface integrity, internal crack distribution and material uniformity features is obtained, defect area positioning processing is performed on the comprehensive visual detection feature set based on a preset quality evaluation rule, a defect distribution map is generated, and then a visual quality report is generated according to the defect distribution map. And finally, the visual quality report is transmitted to terminal equipment to trigger wall maintenance operation, so that comprehensive and accurate detection and visual presentation of the wall quality are realized, wall maintenance can be effectively guided, and the construction engineering wall quality guarantee level is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart construction sites, and in particular to a device and method for visually detecting the quality of wall surfaces in construction projects. Background Art

[0002] In the field of construction engineering, wall quality inspection is a key link to ensure the safety and durability of buildings. Traditional construction wall quality inspection methods have many limitations.

[0003] In the early days, inspectors primarily relied on visual inspection of wall surfaces to assess quality. This method could only detect obvious surface defects, such as large cracks and peeling, but was unable to detect hidden quality issues within the wall, such as fine cracks and uneven material. Furthermore, visual inspection is subject to significant subjective influence from the inspector, resulting in varying results from different inspectors, and low accuracy and reliability.

[0004] To address this, some inspection methods have begun using a single type of image for wall inspection, such as capturing only surface images under a single lighting condition. While this approach improves inspection efficiency to a certain extent, due to its limited focus on the surface and the single lighting condition, it fails to fully reflect the wall's actual quality. Such methods are even more incapable of addressing internal structural issues. Furthermore, some inspection technologies utilize specialized equipment to scan internal structures, but their analysis relies solely on the scan results, lacking comprehensive consideration of the wall's surface characteristics, resulting in an incomplete and inaccurate assessment of wall quality.

[0005] Furthermore, traditional inspection methods also have shortcomings in the presentation and processing of inspection results. Inspection results are typically presented in text reports, which are abstract and lack intuitiveness, making them difficult for non-professionals to understand. Furthermore, there is a lack of effective connection between inspection results and subsequent wall maintenance operations. Manual information transmission is often required, which is prone to information delays and errors, leading to untimely wall maintenance, thereby affecting the overall quality and service life of the building. Summary of the Invention

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for visually detecting the quality of a wall in a construction project, the method comprising: Acquire a multi-angle image set of the wall to be inspected, wherein the multi-angle image set includes surface images and internal structure scan images under different lighting conditions; Performing visual feature extraction on the multi-angle image set to obtain a comprehensive visual detection feature set of the wall to be inspected, wherein the comprehensive visual detection feature set includes surface integrity features, internal crack distribution features, and material uniformity features; Perform defect area positioning processing on the comprehensive visual inspection feature set based on preset quality assessment rules to generate a defect distribution map of the wall to be inspected; Generate a visual quality report based on the defect distribution map; The visual quality report is transmitted to a terminal device to trigger a wall maintenance operation.

[0007] On the other hand, an embodiment of the present invention also provides a visual detection device for the quality of wall in a construction project, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0008] Based on the above aspects, the embodiment of the present invention obtains surface images and internal structure scanning images of the wall to be inspected under different lighting conditions to form a multi-angle image set, and on this basis, performs visual feature extraction on the multi-angle image set to obtain a comprehensive visual detection feature set, which covers surface integrity features, internal crack distribution features and material uniformity features, and deeply mines the key feature information of the wall from multiple dimensions, thereby realizing all-round and accurate capture of wall quality features. Based on preset quality assessment rules, the comprehensive visual detection feature set is processed to locate the defect area and generate a defect distribution map, which can accurately mark the location and severity level of potential quality hazards, making wall quality problems intuitive and quantified, and generating a visual quality report based on the defect distribution map, presenting complex wall quality information in a clear and easy-to-understand visual form, and finally transmitting the visual quality report to the terminal device to trigger the wall maintenance operation, thereby realizing the automation and efficiency of the entire process from wall quality inspection to maintenance, greatly improving the accuracy, comprehensiveness and timeliness of wall quality inspection in construction projects, and effectively reducing the safety risks and economic losses caused by wall quality problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 The present invention provides a flowchart of a method for visually detecting the quality of a wall in a construction project.

[0010] Figure 2 Schematic diagram of exemplary hardware and software components of a visual detection device for construction wall quality provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 The present invention provides a flow chart of a method for visually detecting the quality of a wall in a construction project. The method for visually detecting the quality of a wall in a construction project is introduced in detail below.

[0012] Step S110: Acquire a multi-angle image set of the wall to be detected, wherein the multi-angle image set includes surface images and internal structure scanning images under different lighting conditions.

[0013] In the actual scenario of building wall quality inspection, in order to obtain accurate and comprehensive wall information, it is necessary to obtain a multi-angle image set. The specific steps for obtaining this multi-angle image set will be detailed below.

[0014] Step S111: collecting a surface thermal imaging image of the wall to be inspected by an infrared camera device carried by a drone, and collecting laser three-dimensional point cloud data of the wall to be inspected by a ground mobile scanning device.

[0015] In this step, the drone equipped with an infrared camera is a key tool for obtaining thermal images of the wall surface. The drone's flight path and parameter settings are crucial for acquiring high-quality thermal images. First, the drone's flight path needs to be planned based on factors such as the shape, size, and geographical location of the wall to be inspected. Assuming that the wall to be inspected is the facade of a rectangular building, the flight path can be designed as multiple straight line trajectories parallel to the wall surface, with a certain spacing d between adjacent trajectories. The flight height h needs to be determined based on the performance of the infrared camera equipment and the inspection requirements of the wall. Generally speaking, too high a height may result in reduced image resolution, while too low a height may not cover the entire wall. The flight speed v also needs to be reasonably controlled. Too fast a speed may blur the image, while too slow a speed will increase the inspection time.

[0016] During the acquisition process, the infrared camera captures images at fixed intervals Δt. Each captured image has a certain pixel resolution, assumed to be m×n. The pixel value in the image represents the thermal radiation intensity at the corresponding location on the wall surface, which is related to the wall surface temperature. Because different lighting conditions affect the temperature distribution on the wall surface, performing multiple acquisitions at different time periods can provide more comprehensive information. For example, acquisitions can be performed during direct sunlight during the day and in the evening when the light is weaker.

[0017] Mobile ground-based scanning equipment is used to collect 3D laser point cloud data from walls. This device is typically mounted on a movable platform and moves along a pre-set path at the base of the wall. The pre-set path can be designed based on the wall's shape. For example, for a rectangular wall, the device can move along the two long and two short sides of the wall. The principle of collecting 3D laser point cloud data is that the device emits a laser beam onto the wall surface and determines the 3D coordinates of each point on the wall surface by receiving the reflected laser beam. As the device moves, it emits laser beams at set angle intervals Δθ and distance intervals Δs. Each time a laser beam is emitted, the time t and angle θ of the reflected light are recorded. Based on the propagation speed c of the laser in air, the distance r = c × t / 2 from the device to the corresponding point on the wall surface can be calculated. Combining the device's current position and angle information, the 3D coordinates (x, y, z) of the point on the wall surface can be determined. By continuously emitting laser beams and recording the relevant information, the 3D coordinates of a large number of points on the wall surface can be obtained, forming 3D laser point cloud data.

[0018] Step S112: performing temperature calibration processing on the surface thermal imaging image to generate a set of calibrated thermal imaging images.

[0019] The temperature data collected in surface thermal images may contain errors due to various factors, including the inherent characteristics of the infrared imaging equipment and environmental factors. To improve the accuracy of temperature data, temperature calibration is required. During temperature calibration, a standard reference object with a known temperature is introduced. The temperature stability and accuracy of the standard reference object are crucial to the calibration results. Assume that the actual temperature of the standard reference object is T_ref, and the average temperature of the pixel area corresponding to the standard reference object in the thermal image is T_meas_ref. For each pixel in the thermal image, the collected temperature is T_meas. By calculating the calibration coefficient k = T_ref / T_meas_ref, and then multiplying the collected temperature T_meas for each pixel by the calibration coefficient k, the calibrated temperature T_cal = T_meas × k is obtained. This calibration process is repeated for all pixels in the thermal image to produce a set of calibrated thermal images. In practice, to improve calibration accuracy, the temperature of the standard reference object and the corresponding temperature in the thermal image can be measured multiple times, and the average values are calculated as T_ref and T_meas_ref.

[0020] Step S113: performing registration and redundancy removal processing on the laser three-dimensional point cloud data to generate target structure model data.

[0021] During the acquisition process of laser 3D point cloud data, due to factors such as equipment movement and measurement errors, there may be problems such as incomplete data, overlap, and noise. Therefore, it is necessary to perform registration and redundancy removal. This process includes the following sub-steps: Step S1131: performing outlier filtering processing on the laser 3D point cloud data, filtering out noise points whose spacing exceeds a preset threshold based on an adjacent point spacing threshold, and obtaining preliminary point cloud data.

[0022] In laser 3D point cloud data, outliers are often caused by measurement errors, external interference, and other factors. These points are often far from surrounding points, which can affect subsequent data processing and analysis. To filter out outliers, a threshold d_th for the distance between adjacent points is set. For each point P (x, y, z) in the point cloud data, the distance to its surrounding adjacent points is calculated. This can be done by searching for all points within a spherical neighborhood with radius r_search centered on point P and calculating the Euclidean distance from point P to these points. Assuming the points in the neighborhood are P_i (x_i, y_i, z_i), the distance d_i from point P to point P_i is √[(x-x_i)^2+(y-y_i)^2+(z-z_i)^2]. If any distance d_i exceeds the preset threshold d_th, point P is considered an outlier and is removed from the point cloud data. By performing this judgment and processing on all points, preliminary point cloud data can be obtained.

[0023] Step S1132: Calling the iterative closest point algorithm to perform multi-frame rigid registration processing on the preliminary point cloud data, aligning each frame of point cloud data to a unified spatial coordinate system, and generating registered point cloud data.

[0024] When collecting 3D laser point cloud data, it may be captured in multiple frames, each frame potentially in a different spatial coordinate system. To integrate this data into a unified spatial coordinate system, registration is required. The iterative closest point (ICP) algorithm is a commonly used point cloud registration algorithm. The basic idea of this algorithm is to find, through continuous iteration, an optimal rotation matrix R and translation vector t that minimizes the sum of the distances between corresponding points in two frames of point cloud data. The specific steps are as follows: First, two corresponding point sets P and Q are selected from the two frames of point cloud data. Then, the centroids C_P and C_Q of the two point sets are calculated. Next, the centroids C_P and C_Q of the two point sets are subtracted from each other to obtain the centroid-free point sets P' and Q'. The covariance matrix H between the point sets P' and Q' is calculated and singular value decomposition (SVD) is performed on H to obtain the rotation matrix R. Based on the rotation matrix R and the translation vector t = C_Q - R × C_P, point set P is rotated and translated to align it with point set Q. Repeat the above steps until the sum of the distances between the corresponding points converges to a preset threshold or the maximum number of iterations is reached. By performing this registration process on multiple frames of preliminary point cloud data, all frames can be aligned to a unified spatial coordinate system, generating the registered point cloud data.

[0025] Step S1133: performing uniform downsampling processing based on a voxel grid on the registered point cloud data, retaining the geometric center point in each voxel by setting the voxel size parameter, and generating downsampled point cloud data.

[0026] The registered point cloud data may contain a large number of points, which increases the computational complexity of subsequent processing. To reduce the data size while preserving the key geometric features of the point cloud data, downsampling is necessary. A commonly used downsampling method is the voxel-based uniform downsampling method. This method first divides the three-dimensional space containing the registered point cloud data into a number of voxels of equal size, where the voxel size is determined by the voxel size parameter v_size. For each voxel, the number of points contained within it is counted. If the number of points within a voxel is greater than 0, the geometric center of these points is calculated. The coordinates of the geometric center can be obtained by taking the average of the coordinates of all points within the voxel. Assuming that the points within a voxel are P_j(x_j, y_j, z_j), and the number of points within the voxel is n, the coordinates of the geometric center (x_c, y_c, z_c) = (∑x_j / n, ∑y_j / n, ∑z_j / n). The geometric center of each voxel is used as the downsampled point, ultimately generating the downsampled point cloud data.

[0027] Step S1134: performing Poisson surface reconstruction processing on the downsampled point cloud data, generating a closed triangular mesh surface according to the point cloud normal vector and density distribution, and retaining the non-closed topological structure of the internal crack area by adjusting the surface reconstruction parameters.

[0028] Poisson surface reconstruction is a method for generating three-dimensional surfaces from point cloud data. Based on the point cloud's normal vectors and density distribution, this method solves the Poisson equation to generate a closed triangular mesh surface. First, the normal vector of each point in the downsampled point cloud data must be estimated. This normal vector can be obtained by computing the principal component analysis (PCA) of the point's local neighborhood. Assuming that the set of points in the local neighborhood of point P is N, the covariance matrix C of point set N is calculated. Eigenvalue decomposition is performed on C, and the eigenvector corresponding to the minimum eigenvalue is the normal vector n of point P. Then, an implicit function is constructed based on the point cloud's normal vectors and density distribution. Solving the Poisson equation yields the zero level set of the implicit function, which is the resulting closed triangular mesh surface. During the reconstruction process, adjustments to surface reconstruction parameters, such as the smoothness parameter and sampling rate, can be made to preserve the non-closed topology of internal crack regions. For example, appropriately reducing the smoothness parameter can better reflect the crack's detailed features.

[0029] Step S1135: converting the triangular mesh surface into a parameterized continuous surface model to generate target structure model data containing internal defect features.

[0030] The generated triangular mesh surface is a discrete representation. To facilitate subsequent analysis and processing, it needs to be converted into a parameterized continuous surface model. This conversion can be performed using a spline function-based method. First, the triangular mesh surface is sampled to obtain a series of discrete points. These points are then fitted using a spline function to obtain a parameterized continuous surface model. During the fitting process, the smoothness and accuracy of the surface must be considered. By adjusting the parameters of the spline function, the fitted surface can be made to better approximate the triangular mesh surface. The resulting parameterized continuous surface model incorporates the internal defect features of the wall, such as cracks and voids, forming the target structural model data.

[0031] Step S114: When collecting the surface thermal imaging image and the laser three-dimensional point cloud data, the positioning target coordinates of the UAV and the ground scanning equipment are synchronously recorded.

[0032] To accurately align the calibration thermal image set with the target structure model data in three-dimensional space, it is necessary to simultaneously record the coordinates of the positioning targets of the drone and ground-based scanning equipment during data acquisition. Multiple positioning targets with known three-dimensional coordinates are set around the wall. The drone and ground-based scanning equipment are each equipped with positioning sensors. During the acquisition of surface thermal images and laser three-dimensional point cloud data, the positioning sensors record the position and attitude of the equipment relative to the positioning targets in real time. Assuming the coordinates of the positioning targets are T_i (x_ti, y_ti, z_ti), the positioning coordinates of the drone when acquiring a frame of thermal imagery are U(x_u, y_u, z_u), and the positioning coordinates of the ground-based scanning equipment when acquiring a frame of point cloud data are G(x_g, y_g, z_g). By recording these coordinates, a correspondence can be established between the thermal imagery and point cloud data and the actual position of the wall.

[0033] Step S115: performing three-dimensional spatial projection alignment on the calibration thermal imaging image set and the target structure model data according to the positioning target coordinates, so as to match the surface defect position with the internal structure coordinates, and generate the multi-angle image set.

[0034] Using the previously recorded positioning target coordinates, the calibration thermal image set and the target structure model data are projected and aligned in three-dimensional space. First, based on the positioning target coordinates, the relative position and orientation of the calibration thermal image set and the target structure model data in three-dimensional space are determined. This process can be achieved by calculating a coordinate transformation matrix. Assume that the coordinate system corresponding to the calibration thermal image set is I, and the coordinate system corresponding to the target structure model data is M. The coordinates of the positioning target in coordinate system I are T_i_I (x_ti_I, y_ti_I, z_ti_I), and the coordinates in coordinate system M are T_i_M (x_ti_M, y_ti_M, z_ti_M). The coordinate transformation matrix can be obtained by solving the coordinate transformation equation T_i_M = R × T_i_I + t, where R is the rotation matrix and t is the translation vector. This coordinate transformation matrix is then used to project each pixel in the calibration thermal image set into the three-dimensional space of the target structure model data, ensuring that the surface defect locations match the internal structure coordinates. Finally, the calibrated thermal imaging image set and the target structure model data are integrated together to generate a multi-angle image set.

[0035] Step S120: extracting visual features from the multi-angle image set to obtain a comprehensive visual detection feature set of the wall to be detected, wherein the comprehensive visual detection feature set includes surface integrity features, internal crack distribution features, and material uniformity features.

[0036] After acquiring a multi-angle image set, it is necessary to extract visual features to obtain a comprehensive visual inspection feature set that can reflect the quality of the wall to be inspected. The specific process is as follows: Step S121: performing illumination compensation processing on the surface image to obtain a standardized surface image set, and performing noise filtering processing on the internal structure scan image to obtain a denoised scan image set, wherein the standardized surface image set and the denoised scan image set are uniformly scaled to the same resolution and aligned through spatial coordinates so that pixel positions match.

[0037] Surface images are collected under different lighting conditions, and there may be uneven lighting problems, which will affect subsequent feature extraction. Therefore, it is necessary to perform lighting compensation on the surface image. A method based on histogram equalization can be used for lighting compensation. First, the surface image is converted into a grayscale image. Then, the histogram of the grayscale image is calculated, and the grayscale distribution of the image is adjusted by equalizing the histogram to make the brightness of the image more uniform. For internal structure scanning images, there may be noise interference, and noise filtering is required. The median filtering method can be used for noise filtering. Median filtering is a nonlinear filtering method that replaces the grayscale value of each pixel with the median grayscale value of the pixels in its neighborhood. By performing lighting compensation and noise filtering on the surface image and the internal structure scanning image respectively, a standardized surface image set and a denoised scanning image set are obtained.

[0038] To facilitate subsequent feature extraction and analysis, the standardized surface image set and the denoised scanned image set need to be uniformly scaled to the same resolution. Bilinear interpolation can be used for this scaling. Bilinear interpolation is a commonly used image scaling method that calculates the grayscale values of scaled pixels by linearly interpolating the grayscale values of adjacent pixels in the image. Furthermore, to ensure a one-to-one correspondence between the pixel positions in the standardized surface image set and the denoised scanned image set, spatial coordinate alignment is required. This can be achieved by calculating the affine transformation matrix between the two images and then performing an affine transformation on one image to align its spatial coordinates with the other.

[0039] Step S122: calling a pre-trained deep convolutional network to perform feature extraction processing on the standardized surface image set and the denoised scan image set respectively, to generate a primary surface feature map and a primary structure feature map with the same spatial dimension and channel dimension.

[0040] Pretrained deep convolutional networks are highly capable of extracting image features. You can choose a convolutional neural network pretrained on a large-scale image dataset, such as ResNet or VGG. Input a set of normalized surface images and a set of denoised scanned images into the pretrained deep convolutional network. The network extracts features from the input images through a series of convolutional layers, pooling layers, and activation functions. Convolutional layers slide convolution kernels across the image to extract local features. Pooling layers reduce the spatial dimensionality of the feature map, reducing computational effort. Activation functions introduce nonlinearity to enhance the network's expressive power.

[0041] The final few layers of the network output primary surface feature maps and primary structural feature maps. To ensure that the primary surface feature maps and primary structural feature maps have the same spatial and channel dimensions, appropriate adjustments can be made to the network structure. For example, convolutional or pooling layers can be added to the network's output layer to adjust the feature map dimensions.

[0042] Step S123: performing feature interaction processing on the primary surface feature map and the primary structure feature map to generate a fusion channel weight matrix, and performing weighted superposition on the primary surface feature map and the primary structure feature map based on the fusion channel weight matrix to obtain a fusion feature map.

[0043] The purpose of feature interaction processing is to make full use of the information in the primary surface feature map and the primary structural feature map to generate a more representative fusion feature map. The specific process is as follows: Step S1231: performing channel dimension splicing on the primary surface feature map and the primary structure feature map to obtain a spliced feature map.

[0044] The primary surface feature map and the primary structural feature map are concatenated along the channel dimension, i.e., their number of channels is added together. Assuming the number of channels of the primary surface feature map is C_s and the number of channels of the primary structural feature map is C_s, the number of channels of the concatenated feature map is C = C_s + C_s. This channel-wise concatenation allows the surface and structural features to be fused together, providing richer information for subsequent feature interactions.

[0045] Step S1232: performing global average pooling processing on the spliced feature map to generate a channel statistical vector.

[0046] Global average pooling is a commonly used feature dimensionality reduction method that averages the feature maps across each channel to obtain a scalar value. Global average pooling is performed on the concatenated feature map. The average value of each channel's feature map is summed and divided by the total number of pixels to obtain the channel average. The average values of all channels are combined to form a channel statistics vector. The length of the channel statistics vector is equal to the number of channels, C, in the concatenated feature map.

[0047] Step S1233: Call a multilayer perceptron to perform nonlinear transformation processing on the channel statistical vector to generate a channel attention weight vector.

[0048] A multilayer perceptron (MLP) is a feedforward neural network consisting of an input layer, a hidden layer, and an output layer. The channel statistics vector is input to the MLP, which then undergoes nonlinear transformations in the hidden layer to output a channel attention weight vector. Activation functions, such as the ReLU function, are used in the hidden layer to introduce nonlinearity and enhance the network's expressiveness. The length of the channel attention weight vector is equal to the length C of the channel statistics vector, indicating the importance of each channel in the fusion process.

[0049] Step S1234: Normalize the channel attention weight vector to obtain the fused channel weight matrix, wherein each element of the fused channel weight matrix corresponds to the fusion ratio coefficient of the primary surface feature map and the primary structure feature map on the corresponding channel.

[0050] In order to make the element values in the channel attention weight vector within the range of [0, 1] and the sum of all elements equal to 1, it needs to be normalized. The Softmax function can be used for normalization. The Softmax function converts each element in the channel attention weight vector into a probability value, which represents the relative importance of each channel in the fusion process. The normalized channel attention weight vector becomes a normalized vector element, which is regarded as an element of the fusion channel weight matrix. Since the channels of the primary surface feature map and the primary structure feature map are arranged sequentially after splicing, the first part of the elements in the fusion channel weight matrix corresponds to the channels of the primary surface feature map, and the second part of the elements corresponds to the channels of the primary structure feature map. These elements respectively represent the fusion ratio coefficients of the primary surface feature map and the primary structure feature map on the corresponding channels.

[0051] After obtaining the fusion channel weight matrix, the primary surface feature map and the primary structure feature map are weightedly superimposed based on this. The feature map of each channel of the primary surface feature map is multiplied by the element corresponding to this channel in the fusion channel weight matrix, and the same operation is performed on the feature map of each channel of the primary structure feature map. Because the spatial dimensions of the primary surface feature map and the primary structure feature map have been unified before, the spatial dimensions of the weighted corresponding channel feature maps still match. The feature maps of the corresponding channels of the weighted primary surface feature map and the primary structure feature map are added pixel by pixel to obtain the fusion feature map. In this process, since the operation is performed on feature maps of the same spatial dimension and channel dimension, and the elements of the fusion channel weight matrix are dimensionless proportional coefficients, there will be no dimension mismatch problem, which ensures the consistency of the dimensions of the calculation results on both sides of the equal sign.

[0052] Step S124: performing multi-scale pooling processing on the fused feature map, extracting the surface integrity features, the internal crack distribution features and the material uniformity features, and associating and storing the surface integrity features, the internal crack distribution features and the material uniformity features as the comprehensive visual detection feature set.

[0053] Multi-scale pooling can extract feature information of different scales from the fused feature map, thereby more comprehensively reflecting the quality status of the wall to be inspected. The specific operation is as follows: Step S1241: performing average pooling processing of different sizes on the fused feature map to generate a set of multi-scale pooling feature maps corresponding to the pooling window size.

[0054] Average pooling is a commonly used pooling method that averages all pixel values within the pooling window to obtain the pooled pixel value. Average pooling is performed on the fused feature map using pooling windows of different sizes. For example, a series of pooling window sizes of different sizes are set, such as small-sized pooling windows, medium-sized pooling windows, and large-sized pooling windows. For each pooling window size, the pooling window is slid on the fused feature map, and the average value of the pixel values within each pooling window is calculated to obtain the corresponding pooling feature map. In this way, a set of multi-scale pooling feature maps corresponding to different pooling window sizes is generated. Pooling windows of different sizes can capture feature information of different scales. Small-sized pooling windows can retain more detailed features, while large-sized pooling windows can extract more macroscopic features.

[0055] Step S1242: splicing the pooling feature maps of each scale in the multi-scale pooling feature map set along the channel dimension to generate a multi-channel splicing feature map.

[0056] The pooled feature maps of each scale in the multi-scale pooled feature map set are concatenated along the channel dimension. Since each scale pooled feature map has a certain number of channels, the number of channels in the concatenated multi-channel feature map is equal to the sum of the number of channels in each scale pooled feature map. By concatenating along the channel dimension, feature information at different scales can be fused together to form a richer feature representation.

[0057] Step S1243: performing a channel dimensionality reduction operation on the multi-channel splicing feature map to obtain a reduced dimensionality feature map.

[0058] The multi-channel spliced feature map may have a large number of channels. To reduce data size and computational complexity, channel dimensionality reduction is necessary. This can be achieved using a convolutional layer. Design a convolutional layer with a smaller number of convolution kernels than the number of channels in the multi-channel spliced feature map and input the multi-channel spliced feature map into this convolutional layer. The convolutional layer extracts and fuses features from the multi-channel spliced feature map through convolution operations, outputting a reduced-dimensional feature map with a smaller number of channels. During the convolution process, the convolution kernels learn the correlations between different channels, thereby fusing information from multiple channels into a smaller number of channels.

[0059] Step S1244: performing global normalization processing on the dimension-reduced feature map to generate a standardized multi-scale feature vector.

[0060] Global normalization ensures that the eigenvalues in the reduced-dimensionality feature map have a uniform scale, facilitating subsequent analysis and comparison. Batch normalization can be used for global normalization. Batch normalization normalizes each channel in the reduced-dimensionality feature map, calculating the mean and variance of each channel. The eigenvalues are then subtracted from the mean and divided by the square root of the variance, and adjusted using learnable scaling and offset factors. After global normalization, the reduced-dimensionality feature map is converted into a standardized multi-scale feature vector. Each element in the standardized multi-scale feature vector represents feature information at different scales and channels, and has a uniform scale.

[0061] Step S1245: Segment the standardized multi-scale feature vector into a surface integrity feature vector, an internal crack distribution feature vector, and a material uniformity feature vector, wherein the surface integrity feature vector is composed of an edge sharpness parameter and a texture continuity parameter, the internal crack distribution feature vector is composed of a connected domain size parameter and a directional consistency parameter, and the material uniformity feature vector is composed of a grayscale variance parameter and a gradient change frequency parameter.

[0062] Based on the feature information represented by different elements in the normalized multi-scale feature vector, it is segmented into a surface integrity feature vector, an internal crack distribution feature vector, and a material uniformity feature vector. This can be achieved using pre-defined segmentation rules, such as segmentation based on the position or index of the elements in the feature vector.

[0063] The surface integrity eigenvector describes the integrity of the wall surface. The edge sharpness parameter reflects the clarity of the wall's edge. The sharper the edge, the clearer the boundary and the higher the likelihood of surface integrity. The texture continuity parameter reflects the continuity of the wall's surface texture. Texture continuity indicates a relatively even distribution of material on the wall surface, without noticeable breaks or defects.

[0064] The internal crack distribution eigenvector characterizes the distribution of cracks within the wall. The connected domain size parameter indicates the size of the connected region where the cracks form. A larger connected domain size indicates a more severe crack. The directional consistency parameter reflects the consistency of the crack extension direction. High directional consistency indicates that the cracks may have a certain regularity, which is important for determining the cause and development trend of the cracks.

[0065] The material uniformity eigenvector is used to assess the uniformity of wall materials. The grayscale variance parameter measures the dispersion of grayscale values in the wall image. A smaller grayscale variance indicates a more uniform density of the wall material. The gradient change frequency parameter indicates the frequency of grayscale gradient changes in the wall image. A lower gradient change frequency indicates a more uniform texture of the wall material, with no significant density variations.

[0066] The surface integrity feature vector, internal crack distribution feature vector, and material uniformity feature vector are stored together to form a comprehensive visual inspection feature set. This associative storage can be implemented using a data structure, such as storing the three feature vectors in a list or dictionary for easy subsequent use and management.

[0067] Step S130: performing defect area positioning processing on the comprehensive visual inspection feature set based on preset quality assessment rules to generate a defect distribution map of the wall to be inspected.

[0068] After obtaining the comprehensive visual inspection feature set, it needs to be processed according to the preset quality assessment rules to locate the defect area and generate a defect distribution map. The specific process is as follows: Step S131: performing edge detection processing on the surface integrity features, extracting surface crack profile data, and calculating crack length, width and extension direction parameters based on the surface crack profile data.

[0069] The surface integrity feature contains the edge and texture information of the wall surface. Through edge detection processing, the contour data of the surface cracks can be extracted. The specific steps are as follows: For example, step S1311: performing multi-directional gradient amplitude calculation on the surface integrity feature to generate an edge intensity distribution map.

[0070] To accurately detect the edges of surface cracks, it is necessary to calculate the gradient amplitude of the surface integrity feature in multiple directions. This can be done using gradient operators such as the Sobel operator or the Prewitt operator. These operators calculate the gradient of the image in the horizontal and vertical directions separately, and then synthesize them to obtain the gradient amplitude of each pixel. Multi-directional gradient amplitude calculations are performed on each pixel of the surface integrity feature to produce an edge intensity distribution map. Each pixel value in the edge intensity distribution map represents the edge intensity at that pixel. The greater the edge intensity, the more likely the pixel is an edge point.

[0071] Step S1312: Using a non-maximum suppression algorithm to refine the edge intensity distribution map, to obtain a set of candidate edge points with a single pixel width.

[0072] Edges in an edge intensity distribution map may have a certain width. To obtain more accurate edge information, a non-maximum suppression algorithm is used to refine them. This algorithm examines each pixel in the edge intensity distribution map and suppresses the pixel's gradient amplitude to zero if it is not the maximum value within its neighborhood. This refines the edge to a single-pixel width, resulting in a set of candidate edge points. Pixels in the candidate edge point set are all possible edge points, but further verification is required.

[0073] Step S1313: performing connectivity verification on the candidate edge point set based on a dual-threshold connectivity criterion to generate a binary edge image.

[0074] The dual-threshold connection criterion is a commonly used edge connection method that determines edge connectivity by setting two thresholds, namely a high threshold and a low threshold. First, the pixels in the candidate edge point set whose gradient amplitude is greater than the high threshold are marked as strong edge points, and the pixels whose gradient amplitude is between the high threshold and the low threshold are marked as weak edge points. Then, starting from the strong edge point, the weak edge points in its neighborhood are checked. If the weak edge point is connected to the strong edge point, it is also marked as an edge point. In this way, the edge points are connected to generate a binary edge image. In the binary edge image, the pixel value of the edge point is 1, and the pixel value of the non-edge point is 0.

[0075] Step S1314: traverse the connected areas in the binary edge image, extract the contour point sequence of each crack area, and record the coordinates of the contour points.

[0076] Perform connected region analysis on the binary edge image to identify connected regions. Each connected region may represent a crack. Traverse each connected region and extract its contour point sequence. A boundary tracking algorithm can be used to extract the contour point sequence. Starting from a boundary point in the connected region, follow the specified rules to sequentially track adjacent boundary points until returning to the starting point. During the tracking process, record the coordinates of each contour point.

[0077] Step S1315: performing minimum circumscribed rectangle fitting on each crack contour point sequence, and taking the major axis direction of the rectangle as the crack extension direction parameter.

[0078] For each crack contour point sequence, the minimum bounding rectangle (MBR) fitting method is used to determine the crack extension direction. The MBR is the smallest rectangle that completely encompasses the crack contour point sequence. By calculating the major axis of the MBR, the crack extension direction parameter can be obtained. The major axis direction can be expressed as an angle, which reflects the direction of the crack extension on the wall surface.

[0079] Step S1316: Calculate the cumulative value of the adjacent point spacing according to the contour point coordinates, multiply it by a preset pixel-to-actual size ratio coefficient, and generate a crack length parameter.

[0080] Based on the recorded coordinates of the crack contour points, calculate the spacing between adjacent contour points. Add up the spacing between all adjacent points to obtain the pixel length of the crack. Then, multiply the pixel length by a preset pixel-to-actual-size ratio to convert the pixel length to actual length, thus obtaining the crack length parameter. The preset pixel-to-actual-size ratio can be obtained through calibration experiments and represents the actual length represented by each pixel in the image.

[0081] Step S1317: Count the maximum value of the spacing between contour points along the short axis direction of the minimum circumscribed rectangle, multiply it by the proportional coefficient, and generate a crack width parameter.

[0082] Count the distances between the crack outline points along the short axis of the minimum circumscribed rectangle. Find the maximum value, which is the pixel width of the crack along the short axis. Multiply the pixel width by the preset pixel-to-actual-size ratio to convert the pixel width to actual width, obtaining the crack width parameter.

[0083] Step S132: performing connected domain analysis on the internal crack distribution characteristics to determine the starting position, number of branches, and depth parameters of the internal cracks.

[0084] The internal crack distribution characteristics contain the distribution information of cracks in the wall. Through the connected domain analysis, the relevant parameters of the internal cracks can be determined. The specific steps are as follows: For example, step S1321: performing threshold segmentation processing on the internal crack distribution characteristics to generate an internal crack binary mask.

[0085] Threshold segmentation is a commonly used image segmentation method that uses a threshold to classify image pixels into two categories: pixels above the threshold and pixels below the threshold. Threshold segmentation is performed on the internal crack distribution characteristics, setting the pixel value of the crack area to 1 and the pixel value of the non-crack area to 0, thus generating a binary mask of the internal cracks. The threshold can be adjusted based on the specific internal crack distribution characteristics to ensure accurate segmentation of the crack area.

[0086] Step S1322: Mark the independent connected regions in the mask and assign a unique identifier to each connected region.

[0087] Perform connected component analysis on the binary mask of internal cracks to identify independent connected regions. Each independent connected region represents an internal crack. Each connected region is assigned a unique identifier for subsequent processing and management. This process can be implemented using a connected component labeling algorithm, which starts with an unlabeled pixel and recursively labels its neighboring pixels until all pixels in the connected region are labeled.

[0088] Step S1323: extracting the skeleton line of each connected area, tracing back pixel by pixel along the skeleton line to the endpoint that coincides with the projection coordinates of the wall surface, and using the endpoint coordinates as the starting position parameters.

[0089] The skeleton line is the central axis of the connected region and reflects the primary direction of the crack. A skeletonization algorithm is used to extract the skeleton line of each connected region. Then, starting from one endpoint of the skeleton line, the algorithm backtracks pixel by pixel until it finds an endpoint that coincides with the projected coordinates of the wall surface. The coordinates of this endpoint serve as the starting position parameters for the internal crack.

[0090] Step S1324: Count the number of bifurcation points in the skeleton line, and add 1 to the total number of bifurcation points as a branch number parameter.

[0091] In the skeleton line, bifurcation points indicate the locations where cracks branch. Count the number of bifurcation points in the skeleton line and add 1 to the total number of bifurcation points to obtain the internal crack branch number parameter. The branch number parameter reflects the complexity of the crack; a greater number of branches indicates a more complex crack.

[0092] Step S1325: Based on the depth channel data of the internal structure scanning image, the depth sampling values of the corresponding positions of each connected area are extracted, and according to the preset pixel and actual depth ratio coefficient, the average value of the depth sampling values along the crack skeleton line path is calculated as the depth parameter.

[0093] The depth channel data of the internal structure scan image contains depth information for each point within the wall. For each connected region, a depth sample value is extracted at the corresponding location. Depth samples are collected along the crack skeleton path and averaged to obtain the average depth of the connected region. The average depth is then multiplied by a preset pixel-to-actual depth ratio to convert the pixel depth to actual depth, thus determining the depth parameter for the internal crack. This preset pixel-to-actual depth ratio can be obtained through calibration experiments and represents the actual depth represented by each pixel in the image.

[0094] Step S133: performing regional segmentation processing on the material uniformity characteristics to divide the material density abnormality region, and calculating the area ratio and distribution dispersion parameter of the material density abnormality region.

[0095] The material uniformity feature can reflect the density distribution of the wall material. Through regional segmentation processing, we can find the area with abnormal material density and calculate the relevant parameters. The specific steps are as follows: For example, step S1331: performing clustering processing based on a Gaussian mixture model on the material uniformity feature, and dividing the pixels into normal density area and abnormal density area categories.

[0096] The Gaussian Mixture Model (GMM) is a commonly used clustering algorithm that can classify data points into distinct categories. Material uniformity features are clustered using this model to classify pixels into regions with normal and abnormal density. The GMM assigns pixels to different Gaussian distributions based on their characteristic values, such as grayscale and gradient. By setting appropriate clustering parameters, pixels can be accurately classified into regions with normal and abnormal density.

[0097] Step S1332: performing a morphological closing operation on the abnormal density region category to fill holes and generate a continuous abnormal region mask.

[0098] Morphological closing is a commonly used morphological operation that can fill holes and small connected regions in an image. Morphological closing is performed on the abnormal density region class to fill the holes and generate a continuous abnormal region mask. Morphological closing involves two steps: dilation and erosion. Dilation is first performed on the abnormal density region, followed by erosion. This combination of steps fills holes and maintains the continuity of the abnormal region.

[0099] Step S1333: Count the number of pixels in each connected area in the continuous abnormal area mask, and multiply it by a preset pixel area conversion coefficient to obtain the absolute area value of each abnormal area.

[0100] Perform connected region analysis on the continuous anomaly region mask to identify the connected regions. Count the number of pixels in each connected region and multiply this number by a preset pixel area conversion factor to obtain the absolute area of each anomaly region. The preset pixel area conversion factor can be obtained through calibration experiments and represents the actual area represented by each pixel in the image.

[0101] Step S1334: taking the ratio of the sum of the absolute area values to the total projected area of the wall as an area ratio parameter.

[0102] Add the absolute area values of all abnormal regions to obtain the total area of the abnormal regions. Then, divide the total area of the abnormal regions by the total projected area of the wall to obtain the area ratio parameter. The area ratio parameter reflects the proportion of the abnormal material density area in the wall. The larger the ratio, the more serious the uneven material density.

[0103] Step S1335: extract the centroid coordinates of all abnormal areas, calculate the standard deviation of the Euclidean distance between the centroids, and generate the distribution dispersion parameter in combination with the coefficient of variation of the absolute area value.

[0104] For each abnormal area, calculate its centroid coordinates. The centroid coordinates can be obtained by calculating the average coordinates of all pixel points in the abnormal area. Extract the centroid coordinates of all abnormal areas and calculate the Euclidean distance between the centroids. The Euclidean distance represents the straight-line distance between two centroids. Calculate the standard deviation of the Euclidean distances between all centroids. The standard deviation reflects the degree of dispersion of the centroid distribution. At the same time, calculate the coefficient of variation of the absolute area value. The coefficient of variation is the ratio of the standard deviation to the average value, which can reflect the degree of dispersion of the absolute area value. Combine the standard deviation of the Euclidean distance between centroids and the coefficient of variation of the absolute area value to generate a distribution dispersion parameter. The distribution dispersion parameter can comprehensively reflect the distribution dispersion of the material density abnormal area. The greater the dispersion, the more dispersed the distribution of the abnormal area.

[0105] Step S134: According to the crack length, width and extension direction parameters, the starting position, number of branches and depth parameters, and the area ratio and distribution dispersion parameters, a preset defect level mapping table is matched to determine the defect level label of each detection area.

[0106] In order to accurately assess the defect level of each inspection area, it is necessary to match the preset defect level mapping table based on the various parameters calculated previously. The specific process is as follows: Step S1341: Obtain defect detection data and subsequent maintenance records of similar walls in historical projects, and construct a training sample set containing the correlation between defect parameter range and maintenance urgency.

[0107] Defect detection data for similar walls from historical projects was collected, including parameters such as crack length, width, extension direction, starting location, number of branches, depth, area percentage, and distribution dispersion. Subsequent maintenance records for these walls were also collected to understand the maintenance urgency of each inspected area. Maintenance urgency can be quantified based on factors such as maintenance time and difficulty. The defect detection data and maintenance urgency data were combined to construct a training sample set. Each sample in the training sample set contains a set of defect parameters and a corresponding maintenance urgency label.

[0108] Step S1342: Call the random forest algorithm to perform feature importance analysis on the training sample set, determine the weight coefficients of the crack length, width, extension direction, number of branches, depth, area ratio and distribution dispersion parameters, and introduce interaction terms between parameters to correct the weights.

[0109] The random forest algorithm is an ensemble learning algorithm composed of multiple decision trees. When a training set of samples is fed into the random forest algorithm, the algorithm evaluates the importance of each feature (i.e., crack length, width, extension direction, number of branches, depth, area percentage, and distribution dispersion parameter). This evaluation is performed by calculating the splitting gain of each feature in the decision tree. A larger splitting gain indicates a greater impact on the decision and a higher importance. Based on the evaluation results, a weight coefficient is determined for each feature.

[0110] However, in reality, these parameters may interact with each other. For example, the length and width of a crack may jointly affect the defect level of a wall. To account for this interaction, we introduce an interaction term. We can construct a new feature to represent the interaction between parameters, for example, by multiplying the crack length and width to create a new feature. Then, we use the random forest algorithm again to train the training set with the interaction term, and adjust the weight coefficients of each parameter based on the new training results.

[0111] Step S1343: performing nonlinear dynamic interval division on the defect parameter range according to the corrected weight coefficient, and generating the defect level mapping table corresponding to different maintenance urgency levels.

[0112] After obtaining the corrected weight coefficients, the defect parameter range needs to be divided to determine the defect levels corresponding to different maintenance urgency levels. Since the relationship between these parameters may be nonlinear, a nonlinear dynamic interval division method is adopted.

[0113] First, each defect parameter is normalized, mapping its value range to the interval [0, 1] to eliminate the dimensionality effects of different parameters. Then, a comprehensive score is calculated for each sample based on the modified weight coefficients. The comprehensive score is calculated by multiplying the normalized value of each parameter by its corresponding weight coefficient and then adding the results.

[0114] Next, the training samples are sorted based on the comprehensive scores. The sorted samples are divided into different intervals based on the maintenance urgency level. For example, if the maintenance urgency level is divided into high, medium, and low, the samples can be divided into three intervals, each corresponding to a maintenance urgency level.

[0115] Finally, based on the divided intervals, the defect parameter range corresponding to each interval is determined, and a defect level mapping table is generated. The defect level mapping table records the defect parameter ranges corresponding to different maintenance urgency levels. For example, when parameters such as crack length, width, and number of branches are within a certain range, the corresponding maintenance urgency level is high.

[0116] Step S1344: Input the crack length, width, extension direction, number of branches, depth, area ratio and distribution dispersion parameters of the current detection area into the defect level mapping table, and output the defect level label.

[0117] The crack length, width, extension direction, number of branches, depth, area percentage, and distribution dispersion parameters calculated for the current inspection area are normalized to match the dimension and value range of the parameters in the defect level mapping table. These normalized parameters are then compared with the intervals in the defect level mapping table. If the parameters of the current inspection area fall within a certain interval, the maintenance urgency level corresponding to that interval becomes the defect level label for the current inspection area.

[0118] Step S135: Associating and labeling the defect level label with the spatial coordinates of the corresponding detection area to generate the defect distribution map.

[0119] After determining the defect level labels for each inspection area, it is necessary to associate these defect level labels with the spatial coordinates of the corresponding inspection area. The spatial coordinates of the inspection area can be obtained from the positioning information in the previously collected multi-angle image set.

[0120] Specifically, a data record is created for each inspection area, containing the spatial coordinates of the inspection area and the corresponding defect level label. These records can be stored in a data structure (such as a list or dictionary). A defect distribution map is then generated based on these records. The defect distribution map can be a two-dimensional or three-dimensional image, in which different inspection areas are represented by different colors or markers based on their defect level labels, and the spatial coordinates of each inspection area are also annotated. The defect distribution map provides a visual representation of the location and severity of potential quality hazards in the inspected wall.

[0121] Step S140: generating a visual quality report according to the defect distribution map.

[0122] After generating the defect distribution map, a visual quality report needs to be generated based on the defect distribution map to provide detailed guidance information for wall maintenance. The specific process is as follows: Step S141: Divide the wall to be inspected into a plurality of maintenance priority areas according to the distribution density and severity of the defect level labels.

[0123] Analyze the distribution of defect level labels in the defect distribution map, considering two factors: distribution density and severity. Distribution density refers to the number of defect level labels per unit area, while severity is determined by the defect level labels themselves.

[0124] First, the wall to be inspected is divided into several small areas. For each area, the number and type of defect level labels are counted. Based on the statistical results, the defect distribution density and overall severity score for that area are calculated. The overall severity score can be calculated by assigning different weights to different defect level labels, multiplying the number of each defect level label by its corresponding weight, and finally adding the results.

[0125] Based on the calculated defect density and overall severity score, the wall to be inspected is divided into multiple maintenance priority areas. For example, areas with high defect density and high overall severity scores can be classified as high-priority maintenance areas, while areas with low defect density and low overall severity scores can be classified as low-priority maintenance areas.

[0126] Step S142: Based on a preset repair strategy knowledge base, a corresponding repair method and a required materials list are matched for each maintenance priority area.

[0127] A pre-set repair strategy knowledge base stores repair methods and required materials lists for wall defects of different types and severities. Based on the defect level label and specific conditions in the maintenance priority area, matching repair methods and required materials lists are searched from the repair strategy knowledge base.

[0128] For example, if the defect level label of a maintenance priority area indicates that there are severe cracks in the area, the knowledge base is used to search for repair methods for severe cracks, which may include grouting repair, reinforcement repair, etc., and obtain the corresponding required material list, such as grouting materials, reinforcement steel, etc.

[0129] Step S143: Generate construction cost estimation data based on the area, location and repair method of the maintenance priority area.

[0130] Factors such as the size, location, and repair method of the maintenance priority area are considered to generate construction cost estimates. Construction costs are generally higher for larger maintenance priority areas due to the need for more materials and labor. Location factors can also impact construction costs; for example, elevated or difficult-to-reach areas may require additional equipment and safety measures, increasing costs.

[0131] Different repair methods require different costs. For example, the cost of a grouting repair may be related to the type and amount of grouting material used, while the cost of a reinforcement repair may be related to the amount of reinforcement steel used and the difficulty of installation.

[0132] Taking all these factors into account, a cost estimation model is used to generate construction cost estimates. This model can be based on historical engineering data and industry standards, taking the area, location, and repair method of maintenance priority areas as input and outputting construction cost estimates.

[0133] Step S144: Integrate the maintenance priority areas, repair methods, material lists, and construction cost estimate data into the visual quality report according to a preset template.

[0134] Pre-set templates define the format and content structure of visual quality reports. Maintenance priority areas, repair methods, material lists, and construction cost estimates are integrated according to pre-set templates.

[0135] The report first provides a detailed description of the priority maintenance areas, including their location, extent, and defect severity. Next, for each priority maintenance area, the corresponding repair method and required materials list are listed. Finally, the construction cost estimate for each priority maintenance area is provided, and the total construction cost for the entire wall to be inspected is summarized.

[0136] At the same time, in order to make the report more intuitive and easy to understand, defect distribution maps and related images can be inserted into the report to show the defect conditions of the wall and the division of maintenance priority areas.

[0137] Step S145: Loading the annotation information of the defect distribution map into the three-dimensional visualization platform.

[0138] Select a suitable 3D visualization platform, such as professional architectural visualization software or an online visualization tool. Import the annotation information of the defect distribution map into the 3D visualization platform. The annotation information includes the spatial coordinates of the inspection area and the defect level label.

[0139] In the 3D visualization platform, the inspection area is marked and displayed on the 3D model based on the annotation information. Different defect level labels can be distinguished by different colors or marks to intuitively display the defect status of the wall.

[0140] Step S146: Based on the location information of the maintenance priority area and the actual wall positioning calibration data, a virtual construction layer is constructed in the three-dimensional visualization platform. The virtual construction layer is aligned with the actual wall projection through an augmented reality device, and a repair method schematic diagram and material stacking area marks are superimposed and displayed.

[0141] Utilizing the location information of maintenance priority areas and physical wall calibration data, a virtual construction layer is constructed within the 3D visualization platform. Physical wall calibration data, acquired through positioning equipment (e.g., GPS, total station, etc.), ensures accurate alignment of the virtual construction layer with the physical wall.

[0142] In the virtual construction layer, a repair method diagram is drawn based on the information in the repair method knowledge base. This diagram can include information such as construction steps, tools, and materials used, so that construction workers can intuitively understand the repair process. Material storage areas are also marked to facilitate construction workers to arrange material storage locations.

[0143] Using augmented reality devices (such as AR glasses or tablets), the virtual construction layer is projected and aligned with the actual wall. Construction workers can use the augmented reality device to see the repair method diagram and material storage area markings superimposed on the actual wall, allowing them to perform construction operations more accurately.

[0144] Step S147: In response to user interaction, dynamically adjust the transparency and viewing angle of the virtual construction layer, and synchronously display construction cost estimation data that matches the current viewing angle.

[0145] The 3D visualization platform provides users with interactive operation functions. Users can operate the virtual construction layer through touch screen, keyboard input, etc.

[0146] The transparency and viewing angle of the virtual construction layer can be dynamically adjusted as the user interacts. For example, the user can slide their finger to adjust the transparency of the virtual construction layer, making it clearer or blurrier. They can also rotate the device or tap on an on-screen button to change the viewing angle, allowing them to observe the wall and virtual construction layer from different angles.

[0147] At the same time, the current viewing angle can be used to display construction cost estimates that match the current viewing angle. For example, when the user switches to the viewing angle of a maintenance priority area, the construction cost estimate for that area can be displayed, allowing the user to understand the construction cost situation at any time.

[0148] Step S148: When it is detected that the user has selected a specific maintenance priority area, the augmented reality device is called to project and display a repair method guidance mark for the corresponding area on the real wall.

[0149] When users select a specific maintenance priority area on the 3D visualization platform, they can use augmented reality devices to project repair instructions for that area onto the actual wall. These instructions can include text, arrows, or graphic symbols to guide construction workers through repair operations.

[0150] For example, if a user selects a priority maintenance area with cracks, the augmented reality device will project onto the actual wall the steps and precautions for repairing the cracks, as well as the location of the tools and materials required. This allows construction workers to more intuitively understand the repair method, improving work efficiency and quality.

[0151] Step S150: Transmit the visualization quality report to a terminal device to trigger a wall maintenance operation.

[0152] In order to start wall maintenance operations in a timely manner, the generated visual quality report needs to be transmitted to the terminal device. The specific process is as follows: Step S151: parsing the maintenance priority areas and repair methods in the visual quality report to generate a maintenance task list including a construction time window, a personnel allocation plan, and equipment scheduling instructions.

[0153] Parse the visual quality report to extract maintenance priority areas and repair method information. Determine the construction time window for each maintenance priority area based on the urgency of the maintenance priority area and the complexity of the repair method. The construction time window is the reasonable time frame within which maintenance work can be completed in that area.

[0154] Develop a staffing plan based on the construction timeframe and repair method requirements. This plan includes the number of required staff, types of work, and responsibilities. For example, if grouting repairs are required, staff with grouting experience should be assigned.

[0155] At the same time, equipment scheduling instructions are developed based on the repair method and construction time window. These instructions include the type, quantity, and usage time of the required construction equipment. For example, if large lifting equipment is required in an area, the equipment's arrival time and usage order must be arranged in advance.

[0156] The construction time window, personnel allocation plan and equipment scheduling instructions are combined to generate a maintenance task list. The maintenance task list records the maintenance task information for each maintenance priority area in detail, providing clear guidance for wall maintenance operations.

[0157] Step S152: sending a task reminder notification to the terminal device according to the construction time window in the maintenance task list, and confirming the maintenance personnel's task receipt status within a preset time interval.

[0158] Based on the construction time window in the maintenance task list, task reminder notifications are sent to terminal devices in advance. Terminal devices can be mobile phones, tablets, smart watches, etc., and maintenance personnel can receive task reminder notifications through these devices.

[0159] Task reminder notifications contain detailed maintenance task information, such as the location of the priority area, repair method, and construction time window. After receiving a task reminder notification, the maintenance personnel must confirm their acceptance of the task at predetermined intervals (e.g., timed intervals). This can be confirmed through feedback from the terminal device or through communication with the maintenance personnel. If the maintenance personnel do not accept the task or indicate they cannot complete it on time, prompt adjustments and coordination are required.

[0160] Step S153: When the maintenance progress data uploaded by the terminal device is detected, the current progress is compared with the maintenance task list, and the construction priority is dynamically adjusted in combination with the actual defect level label of the maintenance area to generate progress deviation alarm information and adjustment suggestions based on real-time data.

[0161] During wall maintenance operations, maintenance personnel upload maintenance progress data through terminal devices. Maintenance progress data includes the amount of work completed, the amount of work remaining, and any problems encountered.

[0162] Uploaded maintenance progress data is compared with the maintenance task list to analyze the deviation between the current progress and the planned progress. The actual defect level labels of the maintenance area are combined to dynamically adjust the construction priority. For example, if the progress of a high-priority maintenance area is seriously lagging, resources will be allocated to accelerate maintenance work in that area.

[0163] Based on the comparison results and adjusted construction priorities, schedule deviation warnings and real-time data-based adjustment suggestions are generated. These warnings are sent to relevant personnel via terminal devices, alerting them to any progress deviations. These real-time data-based adjustment suggestions include adjustments to personnel assignments, equipment scheduling, and construction sequencing to ensure that maintenance work is completed on time.

[0164] Step S154: After all operations in the maintenance task list are completed, a secondary inspection instruction is triggered to verify the repair effect, and the verification result is attached to the visual quality report.

[0165] When all operations in the maintenance task list are completed, a second inspection is triggered. The second inspection can use the same method as the first inspection to obtain a multi-angle image set of the wall to be inspected, and then perform visual feature extraction and defect area location processing.

[0166] Compare the results of the second inspection with those of the first inspection to verify the effectiveness of the repair. If the repair effect meets expectations, it means that the maintenance work has met the requirements. If the repair effect is not ideal, further analysis of the cause is required and appropriate measures should be taken to correct it.

[0167] The verification results of the secondary inspection are attached to the visual quality report to form a complete wall quality inspection and maintenance report, which can serve as an important basis for wall quality assessment and provide a reference for subsequent maintenance work.

[0168] Figure 2 The following diagram illustrates exemplary hardware and software components of a construction wall quality visual inspection device 100 that can implement the concepts of the present application, as provided in some embodiments of the present application. For example, a processor 120 can be used in the construction wall quality visual inspection device 100 to perform the functions described in the present application.

[0169] The visual inspection device 100 for building wall quality can be a general-purpose server or a special-purpose server, both of which can be used to implement the visual inspection method for building wall quality of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0170] For example, the visual detection device 100 for the quality of a wall of a construction project may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the visual detection device 100 for the quality of a wall of a construction project may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application may be implemented according to the above-mentioned program instructions. The visual detection device 100 for the quality of a wall of a construction project also includes an I / O interface 150 between the computer and other input and output devices.

[0171] For ease of explanation, only one processor is described in the construction engineering wall quality visualization detection device 100. However, it should be noted that the construction engineering wall quality visualization detection device 100 in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the construction engineering wall quality visualization detection device 100 executes step A and step B, it should be understood that step A and step B may also be executed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0172] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned method for visual detection of wall quality of a construction project is implemented.

[0173] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A visual detection method for wall quality of a construction project, characterized in that: The method comprises: Acquire a multi-angle image set of the wall to be inspected, wherein the multi-angle image set includes surface images and internal structure scan images under different lighting conditions; Performing visual feature extraction on the multi-angle image set to obtain a comprehensive visual detection feature set of the wall to be inspected, wherein the comprehensive visual detection feature set includes surface integrity features, internal crack distribution features, and material uniformity features; Perform defect area positioning processing on the comprehensive visual inspection feature set based on preset quality assessment rules to generate a defect distribution map of the wall to be inspected; Generate a visual quality report based on the defect distribution map; The visual quality report is transmitted to a terminal device to trigger a wall maintenance operation.

2. The method for visually detecting the quality of a building wall according to claim 1, characterized in that: The extracting visual features from the multi-angle image set to obtain a comprehensive visual detection feature set of the wall to be detected includes: performing illumination compensation processing on the surface images to obtain a set of standardized surface images, and performing noise filtering processing on the internal structure scan images to obtain a set of denoised scan images, wherein the set of standardized surface images and the set of denoised scan images are uniformly scaled to the same resolution and aligned by spatial coordinates so that pixel positions match; Calling a pre-trained deep convolutional network to perform feature extraction processing on the standardized surface image set and the denoised scan image set, respectively, to generate a primary surface feature map and a primary structure feature map with the same spatial dimension and channel dimension; Performing feature interaction processing on the primary surface feature map and the primary structure feature map to generate a fusion channel weight matrix, and performing weighted superposition on the primary surface feature map and the primary structure feature map based on the fusion channel weight matrix to obtain a fusion feature map; Multi-scale pooling processing is performed on the fused feature map to extract the surface integrity features, the internal crack distribution features and the material uniformity features, and the surface integrity features, the internal crack distribution features and the material uniformity features are associated and stored as the comprehensive visual detection feature set.

3. The method for visually detecting the quality of a building wall according to claim 2, characterized in that: The performing feature interaction processing on the primary surface feature map and the primary structure feature map to generate a fusion channel weight matrix includes: Splicing the primary surface feature map and the primary structure feature map in channel dimension to obtain a spliced feature map; Performing global average pooling processing on the spliced feature map to generate a channel statistical vector; Calling a multilayer perceptron to perform nonlinear transformation processing on the channel statistical vector to generate a channel attention weight vector; Normalizing the channel attention weight vector to obtain the fused channel weight matrix, wherein each element of the fused channel weight matrix corresponds to a fusion ratio coefficient of the primary surface feature map and the primary structure feature map on the corresponding channel; The performing multi-scale pooling processing on the fused feature map to extract the surface integrity feature, the internal crack distribution feature, and the material uniformity feature includes: Performing average pooling processing of different sizes on the fused feature map to generate a set of multi-scale pooling feature maps corresponding to the pooling window size; Splicing the pooling feature maps of each scale in the multi-scale pooling feature map set along the channel dimension to generate a multi-channel splicing feature map; Performing a channel dimensionality reduction operation on the multi-channel splicing feature map to obtain a reduced dimensionality feature map; Performing global normalization on the dimension-reduced feature map to generate a standardized multi-scale feature vector; The standardized multi-scale feature vector is divided into a surface integrity feature vector, an internal crack distribution feature vector and a material uniformity feature vector, wherein the surface integrity feature vector is composed of an edge sharpness parameter and a texture continuity parameter, the internal crack distribution feature vector is composed of a connected domain size parameter and a directional consistency parameter, and the material uniformity feature vector is composed of a grayscale variance parameter and a gradient change frequency parameter.

4. The method for visually detecting the quality of a building wall according to claim 1, wherein: The defect area positioning processing is performed on the comprehensive visual inspection feature set based on a preset quality assessment rule to generate a defect distribution map of the wall to be inspected, including: Performing edge detection processing on the surface integrity features, extracting surface crack profile data, and calculating crack length, width, and extension direction parameters based on the surface crack profile data; Performing connected domain analysis on the internal crack distribution characteristics to determine the starting position, number of branches, and depth parameters of the internal cracks; Performing regional segmentation processing on the material uniformity characteristics to divide the material density abnormality area, and calculating the area ratio and distribution dispersion parameter of the material density abnormality area; According to the crack length, width and extension direction parameters, the starting position, number of branches and depth parameters, and the area ratio and distribution dispersion parameters, a preset defect level mapping table is matched to determine the defect level label of each detection area; The defect level labels are associated with the spatial coordinates of the corresponding detection areas to generate the defect distribution map.

5. The method for visually detecting the quality of a building wall according to claim 4, characterized in that: The method of matching a preset defect level mapping table based on the crack length, width and extension direction parameters, the starting position, number of branches and depth parameters, and the area ratio and distribution dispersion parameters to determine the defect level label of each detection area includes: Obtain defect detection data and subsequent maintenance records of similar walls in historical projects to construct a training sample set that includes the correlation between defect parameter ranges and maintenance urgency; Calling the random forest algorithm to perform feature importance analysis on the training sample set, determining the weight coefficients of the crack length, width, extension direction, number of branches, depth, area proportion and distribution dispersion parameters, and introducing interaction terms between parameters to modify the weights; Performing nonlinear dynamic interval division on the defect parameter range according to the modified weight coefficient to generate the defect level mapping table corresponding to different maintenance urgency levels; Dynamically dividing the defect parameter range into intervals according to the weight coefficient, and generating the defect level mapping table corresponding to different maintenance urgency levels; The crack length, width, extension direction, number of branches, depth, area ratio and distribution dispersion parameters of the current detection area are input into the defect level mapping table, and the defect level label is output.

6. The method for visually detecting the quality of a building wall according to claim 4, characterized in that: Generating a visual quality report according to the defect distribution map includes: Dividing the wall to be inspected into multiple maintenance priority areas according to the distribution density and severity of the defect level labels; Based on a preset repair strategy knowledge base, matching a corresponding repair method and a list of required materials for each maintenance priority area; generating construction cost estimate data based on the area, location, and repair method of the maintenance priority area; Integrating the maintenance priority areas, repair methods, material lists, and construction cost estimate data into the visual quality report according to a preset template; Loading the annotation information of the defect distribution map into the three-dimensional visualization platform; Based on the location information of the maintenance priority area and the actual wall positioning calibration data, a virtual construction layer is constructed in the three-dimensional visualization platform. The virtual construction layer is aligned with the actual wall projection through an augmented reality device, and a repair method diagram and material stacking area markings are superimposed and displayed; In response to user interaction, dynamically adjust the transparency and viewing angle of the virtual construction layer, and synchronously display construction cost estimation data that matches the current viewing angle; When it is detected that the user has selected a specific maintenance priority area, the augmented reality device is called to project the repair method guidance signs of the corresponding area on the real wall.

7. The method for visually detecting the quality of a building wall according to claim 1, characterized in that: The step of obtaining a multi-angle image set of the wall to be detected includes: The infrared camera device carried by the drone collects the surface thermal imaging image of the wall to be inspected, and the ground mobile scanning device collects the laser three-dimensional point cloud data of the wall to be inspected; performing temperature calibration processing on the surface thermal imaging image to generate a set of calibrated thermal imaging images; Performing registration and redundancy removal processing on the laser three-dimensional point cloud data to generate target structure model data; When collecting the surface thermal imaging image and the laser three-dimensional point cloud data, the positioning target coordinates of the UAV and the ground scanning equipment are synchronously recorded; According to the positioning target coordinates, the calibration thermal imaging image set is aligned with the target structure model data in three-dimensional space projection so that the surface defect position matches the internal structure coordinates to generate the multi-angle image set.

8. The method for visually detecting the quality of a building wall according to claim 7, characterized in that: The registering and de-redundancy processing of the laser three-dimensional point cloud data to generate target structure model data includes: Performing outlier filtering on the laser three-dimensional point cloud data, filtering out noise points whose spacing exceeds a preset threshold based on an adjacent point spacing threshold, to obtain preliminary point cloud data; Performing multi-frame rigid registration processing on the preliminary point cloud data by calling an iterative closest point algorithm, aligning each frame of point cloud data to a unified spatial coordinate system, and generating registered point cloud data; Performing uniform downsampling processing based on a voxel grid on the registered point cloud data, retaining the geometric center point within each voxel by setting voxel size parameters, and generating downsampled point cloud data; Performing Poisson surface reconstruction on the downsampled point cloud data, generating a closed triangular mesh surface according to the point cloud normal vector and density distribution, and retaining the non-closed topological structure of the internal crack area by adjusting the surface reconstruction parameters; The triangular mesh surface is converted into a parameterized continuous surface model to generate target structure model data containing internal defect features.

9. The method for visually detecting the quality of a building wall according to claim 1, wherein: The transmitting the visual quality report to the terminal device to trigger the wall maintenance operation includes: Analyze the maintenance priority areas and repair methods in the visual quality report to generate a maintenance task list including construction time windows, personnel allocation plans and equipment scheduling instructions; Sending a task reminder notification to the terminal device according to the construction time window in the maintenance task list, and confirming the maintenance personnel's task acceptance status within a preset time interval; When the maintenance progress data uploaded by the terminal device is detected, the current progress is compared with the maintenance task list, and the construction priority is dynamically adjusted based on the actual defect level label of the maintenance area, and progress deviation alarm information and adjustment suggestions based on real-time data are generated; After all operations in the maintenance task list are completed, a secondary inspection instruction is triggered to verify the repair effect, and the verification result is attached to the visual quality report.

10. A visual inspection device for wall quality of a construction project, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the method for visual detection of wall quality of construction projects as described in any one of claims 1 to 9.

Images