A method for testing the apparent quality of a mirror-finished fair-faced concrete

Abstract

The application relates to the technical field of building engineering quality detection, and discloses a surface quality testing method for mirror-finished fair-faced concrete. The method comprises the following steps: acquiring a multi-band multi-spectrum image of a component surface; obtaining independent mirror gloss distribution maps, micro relief topological maps and chroma coordinate mapping maps through image fusion and feature decoupling; registering and correlating the three maps to construct a three-dimensional full-factor surface feature field; grid scanning the feature field to extract the gloss uniformity, texture roughness and chroma stability indexes of each grid; inputting a pre-trained surface quality analysis network to output the local quality evaluation value and the defect type label of each grid; and integrating the overall quality index and generating a repair guidance atlas according to the defects. The method realizes multi-dimensional, full-field and intelligent fine detection of the surface quality, can independently quantitatively analyze the gloss, texture and color, and accurately locates the defects.

Description

Technical Field

[0001] This invention relates to the field of building engineering quality testing technology, specifically a method for testing the appearance quality of mirror-finish fair-faced concrete. Background Technology

[0002] In architectural decoration and bridge engineering, mirror-finished exposed concrete is widely used due to its unique gloss and decorative effect; its appearance quality directly determines the overall aesthetic value of the project. Currently, the assessment of the appearance quality of such components mainly relies on manual visual inspection, sampling measurements using point-contact gloss meters and colorimeters, or image analysis based on ordinary imaging techniques. These methods mostly only obtain local, isolated parameters, or can only qualitatively and holistically identify appearance defects. They cannot accurately isolate and quantify the independent contributions of different attributes such as gloss, texture, and color on a spatial scale, and are insufficient to reveal the precise correspondence between appearance defects and specific physical characteristics.

[0003] Existing technologies have shortcomings. Manual visual inspection and point-based measurements are inefficient, highly subjective, and have sampling blind spots, failing to achieve full coverage and objective evaluation of component surfaces. While conventional two-dimensional imaging analysis can acquire spatial information, the image data is mixed with coupled information from illumination, color, and texture, making it impossible to separate the gloss distribution, which purely characterizes specular reflection intensity, or to remove the microscopic morphology influenced by color. Furthermore, existing methods lack a digital model that can fuse and correlate multi-dimensional appearance attributes under a unified spatial coordinate system. This prevents refined diagnosis of any tiny area of ​​a component based on multiple indicators, and further hinders the automated output of precisely located defect types and repair guidance. A detection technology is needed that can decouple appearance elements and achieve refined intelligent diagnosis across the entire field. Summary of the Invention

[0004] The purpose of this invention is to provide a method for testing the appearance quality of mirror-finished concrete to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for testing the apparent quality of mirror-finish fair-faced concrete, the method comprising:

[0006] Acquire surface image data of the mirror-finished fair-faced concrete component under set illumination conditions. The surface image data includes multi-band imaging information and multi-spectral reflectance information.

[0007] Image fusion and feature decoupling operations are performed on the surface image data to decouple the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map, which independently characterize the surface gloss features of the component, the surface texture features of the component, and the surface color features of the component.

[0008] Spatial registration and feature association were performed on the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map to construct a three-dimensional full-element appearance feature field of the tested mirror-finished fair-faced concrete component.

[0009] The three-dimensional full-element appearance feature field is scanned in a partitioned grid, and the gloss uniformity index, texture roughness index and color stability index are extracted in each grid cell.

[0010] The gloss uniformity index, texture roughness index, and color stability index of each grid cell are input into a pre-constructed appearance quality analysis network, which outputs the local quality evaluation value and defect type label of each grid cell.

[0011] The local quality evaluation values ​​of all grid cells are integrated, and the overall apparent quality index of the fair-faced concrete component under test is generated by weighted fusion.

[0012] Based on the defect type labels and spatial locations of all grid cells, a visual repair guidance map corresponding to the defect type is generated.

[0013] Preferably, the step of performing image fusion and feature decoupling operations on the surface image data includes:

[0014] An image fusion strategy based on multi-scale decomposition is adopted to fuse the multi-band imaging information and the multi-spectral reflectance information to generate a high-resolution fused surface image.

[0015] The fused surface image is subjected to feature separation operations, the feature separation operations including:

[0016] By using directional filtering and illumination model inversion, the reflection component image, which is mainly affected by the surface geometry, is separated from the fused surface image.

[0017] Through texture frequency analysis and morphological reconstruction, the intrinsic texture image, which is mainly affected by the material microstructure, is separated from the fused surface image;

[0018] By performing color space transformation and spectral demixing, the intrinsic color image, mainly affected by material composition and contamination, is separated from the fused surface image.

[0019] The reflection component image is normalized and specular reflection enhancement is performed to generate the specular gloss distribution map;

[0020] The texture intrinsic image is subjected to three-dimensional topography restoration and mesh refinement processing to generate the micro-undulation topology map;

[0021] The intrinsic color image is subjected to colorimetric transformation and noise suppression processing to generate the colorimetric coordinate mapping map.

[0022] Preferably, the step of spatially registering and feature-correlateding the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map to construct a three-dimensional full-element appearance feature field of the tested specular fair-faced concrete component includes:

[0023] Using the surface vertices of the micro-undulation topology map as a spatial reference, the gloss value of each pixel in the specular gloss distribution map is mapped to the corresponding surface vertex, and a gloss attribute is assigned to each vertex.

[0024] Map the chromaticity coordinates of each pixel in the chromaticity coordinate mapping map to the surface vertex corresponding to the micro-undulation topology map, and assign chromaticity attributes to each vertex.

[0025] Based on the glossiness attribute, the chromaticity attribute associated with each surface vertex, and the three-dimensional coordinates and normal vector of the surface vertex itself, a vertex feature vector containing multi-dimensional attributes is generated.

[0026] The vertex feature vectors of all surface vertices in the micro-undulation topology map are organized according to the topological connection relationship between vertices to construct a spatially continuous three-dimensional full-element appearance feature field with vertices as the basic feature carriers.

[0027] Preferably, the step of performing a partitioned gridded traversal scan of the three-dimensional full-element appearance feature field and extracting the gloss uniformity index, texture roughness index, and color stability index within each grid cell includes:

[0028] The three-dimensional full-element apparent feature field is divided into equal-area grids on a two-dimensional plane, with each grid corresponding to a surface region.

[0029] Traverse each grid and read the vertex feature vectors of all surface vertices covered by that grid;

[0030] The glossiness attribute values ​​of all surface vertices belonging to the same grid are analyzed for dispersion and trend change, and the glossiness uniformity index of the grid is calculated.

[0031] Statistical analysis is performed on the three-dimensional coordinates and normal vectors of all surface vertices belonging to the same mesh, and the root mean square deviation and slope distribution of the surface height within the mesh are calculated as the texture roughness index of the mesh.

[0032] The degree of clustering and offset distance of the chromaticity attribute values ​​of all surface vertices belonging to the same mesh in the chromaticity space are analyzed to calculate the chromaticity stability index of the mesh.

[0033] Preferably, the step of inputting the gloss uniformity index, texture roughness index, and color stability index of each grid cell into a pre-constructed appearance quality analysis network includes:

[0034] The appearance quality analysis network includes a feature encoding layer, a cross-attention fusion layer, and a multi-task decision layer;

[0035] The feature encoding layer performs deep feature extraction on the input gloss uniformity index sequence, texture roughness index sequence, and color stability index sequence to generate high-dimensional feature vectors.

[0036] The cross-attention fusion layer receives the high-dimensional feature vector and calculates the interdependence and influence weights between glossiness features, texture features and chroma features through the cross-attention mechanism to generate a fused comprehensive feature vector.

[0037] The multi-task decision layer executes the quality regression task and the defect classification task in parallel based on the comprehensive feature vector;

[0038] The quality regression task outputs a local quality evaluation value that characterizes the quality of the grid cell, and the defect classification task outputs a defect type label that characterizes the main defect categories of the grid cell.

[0039] Preferably, the integration of the local quality evaluation values ​​of all grid cells, and the generation of the overall apparent quality index of the fair-faced concrete component under test through weighted fusion, includes:

[0040] Based on the spatial position of each grid cell on the surface of the component, a position weight coefficient is assigned to the local quality evaluation value of each grid cell. The position weight coefficient is determined based on the salience of the grid cell in the field of view.

[0041] Calculate the weighted average of the local quality evaluation value of all grid cells and their corresponding location weight coefficients, and use the weighted average as the preliminary overall quality index;

[0042] Historical curing environment data of the fair-faced concrete component to be tested are introduced, and environmental factors are used to correct the preliminary overall quality index according to the severity of the environment, so as to obtain the overall apparent quality index.

[0043] Preferably, the step of generating an appearance repair guidance map corresponding to the defect type based on the defect type labels and their spatial locations of all mesh cells includes:

[0044] All mesh cells are clustered according to the defect type label to identify continuous surface regions with the same or similar defect types and mark them as blocks to be repaired.

[0045] For each type of defect, the corresponding standard repair process and key control parameters are retrieved from the repair strategy library;

[0046] By combining the spatial dimensions, geometry, and micro-undulation topology information of each block to be repaired in the three-dimensional full-element apparent feature field, the standard repair process is adaptively adjusted to generate a customized repair plan for the block to be repaired.

[0047] Customized repair plans for all areas to be repaired are superimposed and marked on the surface development diagram of the component according to their spatial location, forming an appearance repair guidance map that includes the boundaries of the repair area, repair process steps and parameters.

[0048] Preferably, the method for constructing the pre-built appearance quality analysis network includes:

[0049] A large number of mirror-finished concrete samples with known apparent quality grades were collected. The three-dimensional full-element apparent feature field of each sample was obtained and partitioned into grids to obtain the gloss uniformity index, texture roughness index, color stability index of all grid units of each sample, as well as the real quality label and real defect label of each grid unit.

[0050] Using the gloss uniformity index, texture roughness index, and color stability index of the grid cells as input features, and the real quality label and real defect label of the grid cells as supervision signals, a multi-task deep learning network model is constructed.

[0051] The multi-task deep learning network model is trained using the sample data, and the network parameters are optimized through backpropagation until the loss function between the model's predicted output and the true label converges, thus obtaining the trained appearance quality analysis network.

[0052] Preferably, the step of introducing historical curing environment data of the fair-faced concrete component to be tested, and correcting the preliminary overall quality index for environmental factors according to the severity of the environment, includes:

[0053] The historical maintenance environment data includes the temperature fluctuation range, humidity change curve, and cumulative light radiation during the maintenance period;

[0054] An environmental severity assessment model is established, and the temperature fluctuation range, humidity change curve and cumulative solar radiation are used as inputs to calculate an environmental severity coefficient.

[0055] Construct a mapping function between the environmental severity coefficient and the degree of apparent quality degradation;

[0056] Using the mapping function, the preliminary overall quality index is compensated based on the calculated environmental severity coefficient. The purpose of the compensation calculation is to eliminate the apparent quality evaluation bias caused by differences in maintenance environment, thereby obtaining a more comparable overall apparent quality index.

[0057] Preferably, after generating the overall appearance quality index and appearance repair guidance map, the method further includes a long-term quality tracking step:

[0058] At a preset time interval after the tested mirror-finished fair-faced concrete component is put into use, the steps of acquiring surface image data and generating overall appearance quality index are repeated to obtain a series of overall appearance quality indices over time.

[0059] Analyze the changing trend and rate of change of the overall apparent quality index over the time series to predict the future evolution path of the apparent quality of components.

[0060] The future evolution path is compared with a preset quality degradation warning threshold. If the warning threshold is predicted to be reached, the appearance repair guidance map is updated, and preventive maintenance measures are added to the repair plan.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By acquiring surface image data with multi-band and multi-spectral information and performing image fusion and feature decoupling operations, it is possible to accurately separate from a composite image a specular gloss distribution map reflecting only the surface's specular reflectivity, a microscopic undulation topology map reflecting only the physical morphology, and a chromaticity coordinate mapping map reflecting only the material's spectral characteristics. This overcomes the problem of mutual interference of multi-attribute information in traditional images, enabling the three core aesthetic indicators of gloss, roughness, and color to be presented independently, quantitatively, and visually. This provides a clean and dimensional data foundation for subsequent precise analysis, realizing the digital analysis and element extraction of the complex surface structure of concrete.

[0063] The decoupled multidimensional feature maps are spatially registered and correlated to construct a three-dimensional full-element appearance feature field encompassing gloss, texture, and color information. This feature field is then subjected to a partitioned, gridded traversal scan, and a pre-trained appearance quality analysis network processes the multi-index data of each grid unit. This transforms the overall evaluation of the component into parallel intelligent diagnosis of countless micro-units. The network model can directly output quantitative scores and qualitative defect labels for each unit based on the fusion analysis of multidimensional data within the grid. This achieves 100% coverage and blind-spot-free detection of appearance quality, automatically identifying defect types and accurately locating their spatial distribution, generating defect maps that guide construction and repair, and advancing quality assessment from macroscopic qualitative to microscopic quantitative and intelligent decision-making stages. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the surface quality testing method for mirror-finish fair-faced concrete described in this invention.

[0065] Figure 2 A flowchart for image fusion and feature decoupling;

[0066] Figure 3 A flowchart for constructing a three-dimensional full-element appearance feature field;

[0067] Figure 4 A bar chart comparing the quality scores and location weights of grid cells in different areas of a mirror-finished fair-faced concrete wall.

[0068] Figure 5 A pie chart showing the area percentage of surface repair types for mirror-finish fair-faced concrete. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides a method for testing the appearance quality of mirror-finished fair-faced concrete. The method includes: acquiring surface image data of the mirror-finished fair-faced concrete component under set illumination conditions, wherein the surface image data includes multi-band imaging information and multispectral reflectance information; performing image fusion and feature decoupling operations on the surface image data to decouple and extract a specular gloss distribution map, a micro-undulation topology map, and a chromaticity coordinate mapping map, each independently representing the surface gloss characteristics of the component; spatially registering and feature associating the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map to construct a three-dimensional full-element appearance feature field of the mirror-finished fair-faced concrete component; performing a partitioned gridded traversal scan of the three-dimensional full-element appearance feature field to extract gloss uniformity, texture roughness, and chromaticity stability indices within each grid cell; inputting the gloss uniformity, texture roughness, and chromaticity stability indices of each grid cell into a pre-constructed appearance quality analysis network, which outputs a local quality evaluation value and defect type label for each grid cell. The local quality evaluation values ​​of all grid cells are integrated, and a weighted fusion is used to generate the overall apparent quality index of the fair-faced concrete component under test. Based on the defect type labels and spatial locations of all grid cells, an appearance repair guidance map corresponding to the defect type is generated.

[0071] In one embodiment of the present invention, see [reference] Figure 2 This study employs a multi-scale decomposition-based image fusion strategy to fuse multi-band imaging information and multi-spectral reflectance information, generating a high-resolution fused surface image. Feature separation operations are then performed on the fused surface image, involving multiple steps. Through directional filtering and illumination model inversion, the reflectance component image, primarily influenced by surface geometry, is separated from the fused surface image. Through texture frequency analysis and morphological reconstruction, the intrinsic texture image, primarily influenced by material microstructure, is separated from the fused surface image. Through color space transformation and spectral unmixing, the intrinsic color image, primarily influenced by material composition and contamination, is separated from the fused surface image. The reflectance component image is normalized and specular reflection enhanced to generate a specular gloss distribution map. The intrinsic texture image undergoes 3D topography restoration and mesh refinement to generate a micro-undulation topology map. The intrinsic color image undergoes colorimetric transformation and noise suppression to generate a chromaticity coordinate mapping map.

[0072] In practical implementation, surface image data of the mirror-finished fair-faced concrete component under test is acquired under set illumination conditions. This surface image data includes multi-band imaging information and multispectral reflectance information. The set illumination conditions refer to irradiation in a dark room environment using a parallel light source with a specific spectral distribution and incident angle. It can be understood that the multi-band imaging information is acquired by a high-resolution industrial camera with different optical filters, while the multispectral reflectance information is obtained by scanning the component surface with a spectroradiometer. In an example scenario, for a smooth mirror-finished fair-faced concrete wall, firstly, under stable light illumination, a camera is used to acquire narrowband grayscale images corresponding to the 450 nm, 550 nm, and 650 nm bands, respectively. These images together constitute the multi-band imaging information. Simultaneously, a spectroradiometer is used to measure the reflectance point by point on the wall surface in 5 mm increments, recording the reflectance values ​​every 10 nanometers in the visible light range from 380 nm to 780 nm, forming a multispectral reflectance information dataset.

[0073] In some embodiments, image fusion and feature decoupling operations are performed on the surface image data. A multi-scale decomposition-based image fusion strategy is used to fuse the multi-band imaging information and the multispectral reflectance information. During the fusion process, each single-band image in the multi-band imaging information is considered as an image layer, and the spectral curve of the multispectral reflectance information at each spatial sampling point is converted into a low-spatial-resolution feature image. A pyramid decomposition method is used to decompose the high-resolution single-band image and the low-resolution multispectral feature image into multiple scale spaces. At each scale, a fusion weight is calculated based on the structural similarity and spectral fidelity of the local region, combining detailed information from different source images. It can be understood that the fusion weight is jointly determined by the local gradient magnitude and spectral angle information. A formula for calculating the multi-scale fusion weight is expressed as:

[0074]

[0075] in: Indicates the fusion weight. This represents the image gradient magnitude of a specific band in multi-band imaging information. This represents the spectral vector at a specific location in multispectral reflectance information. Represents the reference standard spectral vector. and The adjustment coefficient is used. Finally, a fused surface image with both high spatial resolution and rich spectral detail is obtained through pyramid reconstruction.

[0076] In practice, feature separation operations are performed on the fused surface image. Through directional filtering and illumination model inversion, the reflection component image, primarily influenced by the surface geometry, is separated from the fused surface image. This process assumes that the surface reflection of the component follows a two-way reflection distribution function model. Directional filtering enhances information about specific specular reflection directions, while illumination model inversion uses an iterative optimization algorithm to estimate the specular reflection component related to the surface's micro-geometric undulations and separates it from the fused surface image to form the initial reflection component image. In the illumination model inversion process, the specific implementation of the iterative optimization algorithm is based on the assumption that the surface reflection of the component follows a two-way reflection distribution function model. The algorithm first initializes the relevant parameters of the surface's micro-geometric undulations, such as the initial estimates of the normal vectors at each point. Then, it enters an iterative loop, calculating the specular reflection component predicted by the model under the current parameters in each iteration. This predicted value is then compared with the actual observed data of the corresponding region in the fused surface image to generate residuals. The algorithm then adjusts the surface parameters according to the magnitude and direction of the residuals, gradually reducing the overall difference between the predicted and observed values ​​through repeated iterations until the residual norm is below a preset convergence threshold or the maximum number of iterations is reached. At this point, the algorithm terminates and outputs the optimized specular reflection component, thereby separating the initial reflection component image from the fused surface image. Simultaneously, through texture frequency analysis and morphological reconstruction, the intrinsic texture image, mainly influenced by the material's microstructure, is separated from the fused surface image. Texture frequency analysis employs two-dimensional Fourier transform to identify the dominant frequency components representing periodic or non-periodic textures such as template seams and air bubbles in the image. Morphological reconstruction utilizes opening and closing operations to remove grayscale gradations caused by uneven illumination while preserving the edge features of these textures, thereby extracting the intrinsic texture image. Furthermore, through color space transformation and spectral unmixing, the intrinsic color image, primarily influenced by material composition and contamination, is separated from the fused surface image. Color space transformation converts the image from RGB to Lab color space to better separate brightness and color information. Spectral unmixing uses endmember spectra obtained from multispectral reflectance information to linearly unmix each pixel in the fused surface image, separating the color contributions of different components such as cement matrix, aggregate, rust, and water stains, and synthesizing the intrinsic color image.

[0077] In some embodiments, the reflection component image is normalized and specular reflection enhancement is performed to generate the specular gloss distribution map. Normalization refers to linearly mapping the pixel value range of the reflection component image to a standard gloss unit range of 0 to 100. Specular reflection enhancement refers to applying an unsharpened mask algorithm to enhance the contrast between specular highlight areas and diffuse reflection areas, making the gloss distribution differences more obvious. The texture intrinsic image is then subjected to 3D topology restoration and mesh refinement to generate the micro-undulation topology map. 3D topology restoration employs photometric stereo vision technology, using a sequence of texture intrinsic image sequences obtained from light sources illuminating different directions to calculate the normal vector and relative height of each point on the surface. Mesh refinement refers to performing loop subdivision or edge folding simplification on the reconstructed initial triangular mesh to obtain a smoother or more suitable mesh model for analytical needs. The intrinsic color image is subjected to colorimetric transformation and noise suppression processing to generate the chromaticity coordinate mapping map. Colorimetric transformation refers to converting the intrinsic color image from the Lab color space to the CIE1931xyY color space and extracting the x and y coordinates of each pixel as chromaticity coordinates. Noise suppression processing refers to using a median filter or a nonlocal mean filter to remove isolated noise points in the chromaticity coordinate mapping map to ensure the spatial continuity of chromaticity information.

[0078] Optionally, in the feature separation operation, if there is residual cross-coupling information between the reflection component image, texture intrinsic image, and color intrinsic image, a secondary separation can be performed using iterative optimization until the mutual information between the component images is lower than a set threshold. Optionally, for mirror-finished fair-faced concrete components with large surface curvature variations, when performing three-dimensional morphology restoration using photometric stereo vision, a known macroscopic geometric model of the component needs to be introduced as a constraint to correct the depth calculation error caused by surface tilt.

[0079] In one embodiment of the present invention, see [reference] Figure 3 To construct a three-dimensional full-element appearance feature field for the tested mirror-finished fair-faced concrete component, spatial registration and feature association were performed on the specular gloss distribution map, micro-undulation topology map, and chromaticity coordinate mapping map. The process is as follows: Using the surface vertices of the micro-undulation topology map as spatial references, the gloss value of each pixel in the specular gloss distribution map is mapped to the corresponding surface vertex, assigning a gloss attribute to each vertex. The chromaticity coordinate value of each pixel in the chromaticity coordinate mapping map is mapped to the corresponding surface vertex in the micro-undulation topology map, assigning a chromaticity attribute to each vertex. Based on the gloss attribute, chromaticity attribute, and the three-dimensional coordinates and normal vector of each surface vertex, a vertex feature vector containing multi-dimensional attributes is generated. The vertex feature vectors of all surface vertices in the micro-undulation topology map are organized according to the topological connection relationship between vertices to construct a spatially continuous three-dimensional full-element appearance feature field with vertices as the basic feature carriers.

[0080] In practice, after generating the specular gloss distribution map, micro-undulation topology map, and chromaticity coordinate mapping map, spatial registration and feature association operations are performed to construct a three-dimensional full-element appearance feature field. Spatial registration aims to unify data from different images or models into a single geometric coordinate system, ensuring a precise correspondence between gloss, color, and three-dimensional morphology information at a surface location. This process uses the surface vertices of the micro-undulation topology map as the spatial reference, as the micro-undulation topology map directly provides the three-dimensional coordinates and connectivity of each sampling point on the surface, forming the basic framework for describing the surface geometry of the component. In an example scenario, the micro-undulation topology map consists of one million triangular faces and five hundred thousand vertices, each vertex storing its X, Y, and Z coordinates and normal vector in the world coordinate system. , , Components. The specular gloss distribution map and chromaticity coordinate mapping map are two-dimensional images with a resolution of 2000 pixels by 3000 pixels, and each pixel contains its corresponding gloss value or chromaticity coordinate value.

[0081] In practice, the gloss value of each pixel in the specular gloss distribution map is mapped to the corresponding surface vertex in the micro-undulation topology map. This step requires establishing a precise correspondence between the pixel coordinates of the two-dimensional image and the vertex coordinates of the three-dimensional model. This correspondence is understood to be determined through the camera imaging model and pre-calibrated system. The system calibration determines the intrinsic and extrinsic parameters of the camera and scanning equipment used to acquire surface image data relative to the component's world coordinate system. For each surface vertex in the micro-undulation topology map, using its three-dimensional coordinates and the calibrated camera parameters, the specular gloss distribution of that vertex is calculated through perspective projection transformation. Figure 2 The ideal projected coordinates on the 3D image plane are used. Since projected coordinates are usually subpixel accurate, a bilinear interpolation algorithm is needed to calculate the precise gloss value corresponding to the vertex based on the gloss values ​​of the four neighboring pixels around the projected point. Finally, each surface vertex in the micro-undulation topology map is assigned a gloss attribute G mapped from the specular gloss distribution map.

[0082] In some embodiments, the chromaticity coordinates of each pixel in the chromaticity coordinate map are mapped to the corresponding surface vertex in the micro-undulation topology map, and the mapping principle is the same as that for glossiness attributes. Using the same camera projection model and calibration parameters, each 3D surface vertex is projected onto the image plane of the chromaticity coordinate map. The chromaticity coordinates at the projection location of that vertex are obtained from the chromaticity coordinate map using the same bilinear interpolation method. Finally, each surface vertex in the micro-undulation topology map, in addition to its existing glossiness attribute G, is further assigned chromaticity attributes x and y. At this point, each surface vertex is associated with its own 3D geometric information (coordinates X, Y, Z and normal vector). , , ), optical information, and color information.

[0083] In practical implementation, based on the glossiness and chromaticity attributes associated with each surface vertex, as well as the vertex's own 3D coordinates and normal vector, a vertex feature vector containing multi-dimensional attributes is generated. This vertex feature vector is a data structure that integrates multiple appearance features at the same spatial point. A direct implementation is to arrange all attributes in order to form a multi-dimensional vector. The vector used to characterize the comprehensive attributes of the vertex can be expressed as:

[0084]

[0085] in: Represents the vertex eigenvector. Represents the three-dimensional coordinates of the vertex. This represents the normal vector component of the surface at the vertex. This represents the gloss value obtained by mapping from the specular gloss distribution map. This represents the chromaticity coordinates obtained from the chromaticity coordinate mapping diagram. This nine-dimensional vertex feature vector fully encapsulates all the key appearance features at this point in space.

[0086] Optionally, before constructing vertex feature vectors, attribute data of different dimensions and magnitudes can be normalized to eliminate the impact of numerical scale differences on subsequent analysis. For example, the 3D coordinates X, Y, Z can be transformed to a local coordinate system with the center of the component bounding box as the origin and scaled to the range [0,1]. The normal vector components are kept to unit length, and the glossiness G and chromaticity x, y are also linearly mapped to the range [0,1]. In some embodiments, when the number of patches in the micro-undulation topology map is extremely large, storing a complete vertex feature vector for each vertex will consume a lot of memory. To improve processing efficiency, a hierarchical data organization method can be adopted, storing complete vertex feature vectors only for key areas used for refined analysis, and storing the regional statistical feature values ​​as representatives for other uniform areas.

[0087] In practical implementation, the vertex feature vectors of all surface vertices in the micro-undulation topology map are organized according to the topological connections between vertices to construct a spatially continuous three-dimensional full-element appearance feature field with vertices as the basic feature carriers. The topological connections are defined by triangular facets in the micro-undulation topology map, with each facet referencing three vertex feature vectors through the index numbers of its three vertices. This organization means that the three-dimensional full-element appearance feature field is not only a collection of discrete feature points, but also a network structure that maintains the continuous geometric topology of the surface. For each triangular facet in the network structure, the appearance attribute of any point within it can be obtained through centroid interpolation of the attribute values ​​of its three vertices, thus achieving a continuous and seamless fusion expression of gloss, color, and three-dimensional shape across the entire surface of the component. The final generated three-dimensional full-element appearance feature field is a complete data model containing all vertex feature vectors and their triangular connections, providing a unified and rich data foundation for subsequent partitioned mesh traversal scanning and feature extraction.

[0088] In one embodiment of the present invention, a partitioned gridded traversal scan of the three-dimensional full-element appearance feature field is performed to extract the gloss uniformity index, texture roughness index, and chromaticity stability index within each grid cell. Specifically, the three-dimensional full-element appearance feature field is divided into equal-area grids on a two-dimensional plane, with each grid corresponding to a surface region. Each grid is traversed, and the vertex feature vectors of all surface vertices covered by the grid are read. The dispersion and trend change of the gloss attribute values ​​of all surface vertices belonging to the same grid are analyzed to calculate the gloss uniformity index of the grid. Statistical analysis is performed on the three-dimensional coordinates and normal vectors of all surface vertices belonging to the same grid to calculate the root mean square deviation and slope distribution of the surface height within the grid, which serve as the texture roughness index of the grid. The clustering degree and offset distance of the chromaticity attribute values ​​of all surface vertices belonging to the same grid in the chromaticity space are analyzed to calculate the chromaticity stability index of the grid. The index of each grid cell is input into a pre-constructed appearance quality analysis network, which includes a feature encoding layer, a cross-attention fusion layer, and a multi-task decision layer. The feature encoding layer performs deep feature extraction on the input gloss uniformity index sequence, texture roughness index sequence, and color stability index sequence to generate high-dimensional feature vectors.

[0089] The cross-attention fusion layer receives high-dimensional feature vectors and calculates the interdependencies and influence weights among gloss, texture, and chromaticity features through a cross-attention mechanism, generating a fused comprehensive feature vector. The multi-task decision layer, based on this comprehensive feature vector, executes quality regression and defect classification tasks in parallel. The quality regression task outputs local quality evaluation values ​​characterizing the quality of grid cells, while the defect classification task outputs defect type labels characterizing the main defect categories of grid cells. The construction method of this appearance quality analysis network includes the following steps: A large number of mirror-finished concrete samples with known appearance quality levels are collected. The three-dimensional full-element appearance feature field of each sample is obtained and partitioned into grids to obtain the gloss uniformity index, texture roughness index, and chromaticity stability index of all grid cells for each sample, as well as the true quality label and true defect label for each grid cell. Using the gloss uniformity index, texture roughness index, and chromaticity stability index of the grid cells as input features, and the true quality label and true defect label of the grid cells as supervision signals, a multi-task deep learning network model is constructed. The multi-task deep learning network model is trained using sample data, and the network parameters are optimized through backpropagation until the loss function between the model's predicted output and the true label converges, thus obtaining the trained appearance quality analysis network.

[0090] In practical implementation, after constructing the three-dimensional full-element apparent feature field, a partitioned meshing traversal scan is performed to extract local feature indicators. The three-dimensional full-element apparent feature field contains a spatial continuous network structure with vertices as the basic feature carriers. It can be understood that the purpose of partitioning the continuous surface feature field is to divide it into discrete units that are easy to analyze and statistically process independently. The three-dimensional full-element apparent feature field is divided into equal-area meshes on a two-dimensional plane, with each mesh corresponding to a surface region. This division is usually based on the parametric unfolded diagram of the component surface. For a flat wall surface, the two-dimensional plane is the actual projection plane of the wall surface; for a curved surface with a single curvature, it needs to be unfolded into a plane. In an example scenario, for a flat mirror-finished fair-faced concrete wall surface with an area of ​​2 meters by 3 meters, a 10-millimeter square mesh is used to divide it, ultimately dividing the wall surface into 600 rows by 200 columns, totaling 120,000 mesh units. Each mesh unit physically corresponds to a 10-millimeter by 10-millimeter area on the wall surface.

[0091] In practice, each grid is traversed, and the vertex feature vectors of all surface vertices covered by the grid are read. Since the three-dimensional full-element appearance feature field is composed of triangular facets, a grid cell usually contains multiple vertices. It is necessary to determine which grid cell the projected coordinates of each vertex on the two-dimensional plane fall into, and assign the feature vector of that vertex to that grid cell. The dispersion and trend change analysis of the gloss attribute values ​​of all surface vertices belonging to the same grid are performed to calculate the gloss uniformity index of the grid. The dispersion analysis includes calculating the standard deviation and range of the gloss values ​​of all vertices in the grid. The trend change analysis is performed by fitting a two-dimensional plane or surface within the grid region and calculating the root mean square of the residuals between the vertex gloss values ​​and the fitted surface. The gloss uniformity index is a weighted combination of the standard deviation, range, and root mean square of the residuals. Statistical analysis is performed on the three-dimensional coordinates and normal vectors of all surface vertices belonging to the same grid to calculate the root mean square deviation and slope distribution of the surface height within the grid, which serves as the texture roughness index of the grid. The root mean square deviation of surface height reflects the overall amplitude of surface undulation within a region. The slope distribution describes the anisotropic or isotropic roughness of the surface by statistically analyzing the distribution characteristics of the angle between the normal vectors of all vertices and the average normal vector. The clustering and offset distance of the chromaticity attribute values ​​of all surface vertices belonging to the same mesh in the chromaticity space are analyzed to calculate the chromaticity stability index of the mesh. A formula for comprehensively evaluating chromaticity stability is as follows:

[0092]

[0093] in: Indicates the colorimetric stability index. The area of ​​the covariance ellipse representing the chromaticity coordinate points. This represents the average Euclidean distance from a chromaticity coordinate point to its center point. and These are weighting coefficients. It's understandable that smaller ones... The value indicates that the color is more uniform and stable within the grid cell.

[0094] In some embodiments, the gloss uniformity index, texture roughness index, and chromaticity stability index of each grid cell are input into a pre-constructed appearance quality analysis network. This network includes a feature encoding layer, a cross-attention fusion layer, and a multi-task decision layer. The feature encoding layer performs deep feature extraction on the input sequences of gloss uniformity index, texture roughness index, and chromaticity stability index, generating high-dimensional feature vectors. For each index sequence, the feature encoding layer can be composed of a one-dimensional convolutional neural network or a long short-term memory network to capture local patterns and contextual dependencies within each index sequence. The cross-attention fusion layer receives high-dimensional feature vectors output from different feature encoding layers and calculates the interdependencies and influence weights between gloss, texture, and chromaticity features through a cross-attention mechanism. This cross-attention mechanism allows each element in the gloss feature vector to focus on and weight all elements in the texture and chromaticity feature vectors, thereby learning cross-feature associations such as "what defects a high-gloss region accompanied by a specific texture roughness pattern signifies," ultimately outputting a comprehensive feature vector that integrates multi-source information.

[0095] In practice, the multi-task decision layer executes the quality regression and defect classification tasks in parallel based on the comprehensive feature vector generated by the cross-attention fusion layer. The multi-task decision layer typically consists of two independent sub-network branches. The quality regression branch is usually a multilayer perceptron, which maps the comprehensive feature vector to a continuous scalar value. This value represents the local quality evaluation of the grid cell, with a range, for example, between 0 and 1, where a higher value indicates better local quality. The defect classification branch is usually a softmax classifier, which maps the comprehensive feature vector to a probability distribution vector. The category with the highest probability is the defect type label representing the main defect category of the grid cell. Defect types can be preset as "no defects," "bubbles," "color difference," "contamination," and "obvious template seams."

[0096] In some embodiments, the method for constructing a pre-built appearance quality analysis network includes data preparation, model building, and training. A large number of mirror-finished concrete samples with known appearance quality grades are collected. These samples should cover different construction techniques, curing conditions, and common defects. A three-dimensional full-element appearance feature field is obtained for each sample, and partitioned and meshed according to the aforementioned method. The gloss uniformity index, texture roughness index, and color stability index of all mesh units for each sample are calculated. Simultaneously, experienced quality control engineers manually annotate each mesh unit, providing the true quality label and true defect label for each mesh unit. Using the combination of mesh unit indices as input features and the true quality label and true defect label of the mesh units as supervision signals, a multi-task deep learning network model is constructed. The network's loss function is typically designed as a weighted sum of regression loss and classification loss. The multi-task deep learning network model is trained using sample data, and the network parameters are optimized through backpropagation until the loss function between the model's predicted output and the true label on an independent validation dataset converges to a stable value, thus obtaining the trained appearance quality analysis network that can be used for automatic evaluation.

[0097] Optionally, when extracting texture roughness indices, in addition to calculating the root mean square deviation of surface height, the autocorrelation length or fractal dimension of the surface can also be calculated to describe the complexity of the texture from different dimensions. Optionally, when constructing a multi-task deep learning network model, a position encoding module can be introduced before the feature encoding layer, using the global coordinate information of each grid cell on the component surface as an auxiliary input to help the network learn the spatial distribution patterns of defects, such as the fact that some defects are more likely to occur at the edges or bottom of the component.

[0098] In one embodiment of the present invention, the local quality evaluation values ​​of all grid cells are integrated, and a weighted fusion is used to generate the overall apparent quality index of the fair-faced concrete component under test. Based on the spatial position of each grid cell on the component surface, a positional weight coefficient is assigned to the local quality evaluation value of each grid cell. The positional weight coefficient is determined based on the salience of the grid cell in the field of view. The weighted average of the local quality evaluation values ​​of all grid cells and their corresponding positional weight coefficients is calculated, and this weighted average is used as the preliminary overall quality index. Historical curing environment data of the fair-faced concrete component under test is introduced, and the preliminary overall quality index is corrected for environmental factors according to the severity of the environment to obtain the overall apparent quality index. The historical curing environment data includes the temperature fluctuation range, humidity change curve, and cumulative solar radiation during the curing period. An environmental severity assessment model is established, using the temperature fluctuation range, humidity change curve, and cumulative solar radiation as inputs to calculate an environmental severity coefficient. A mapping function between the environmental severity coefficient and the degree of apparent quality degradation is constructed. Using a mapping function, the preliminary overall quality index is compensated based on the calculated environmental severity coefficient. The purpose of this compensation is to eliminate the appearance quality evaluation bias caused by differences in maintenance environments, thus obtaining a more comparable overall appearance quality index. Based on the defect type labels and spatial locations of all grid cells, an appearance repair guidance map corresponding to each defect type is generated. All grid cells are clustered according to the defect type labels to identify continuous surface areas with the same or similar defect types, which are marked as repair blocks. For each defect type, the corresponding standard repair process and key control parameters are retrieved from the repair strategy library. Combining the spatial dimensions, geometry, and micro-undulation topology information of each repair block in the three-dimensional full-element appearance feature field, the standard repair process is adaptively adjusted to generate customized repair schemes for the repair blocks. All customized repair schemes for the repair blocks are superimposed and annotated on the surface development diagram of the component according to their spatial locations, forming an appearance repair guidance map that includes the repair area boundaries, repair process steps, and parameters.

[0099] In practice, the local quality evaluation values ​​of all grid cells are integrated, and a weighted fusion is used to generate the overall apparent quality index of the fair-faced concrete component under test. Based on the spatial position of each grid cell on the component surface, a positional weight coefficient is assigned to the local quality evaluation value of each grid cell. This positional weight coefficient is determined based on the salience of the grid cell within the observation field of view. Salience can be quantified by calculating the angle, distance, and whether the center point of the grid cell is located in the visual center region relative to the preset observation viewpoint. In an example scenario, for a fair-faced concrete wall 3 meters high and 5 meters wide, the observation viewpoint is set 6 meters directly in front of the center. The wall surface is divided into 150,000 grid cells (300 rows x 500 columns). The solid angle of each grid cell center relative to the viewpoint and the cosine of the angle between the line of sight and the wall normal are calculated. The product of the solid angle and the cosine is normalized to obtain the positional weight coefficient for each grid cell. Grid cells located in the center of the visual field have higher weight coefficients, while those in the edge regions have lower weight coefficients. Calculate the weighted average of the local quality evaluation value of all grid cells and their corresponding location weight coefficients, and use the weighted average as the preliminary overall quality index.

[0100] In practical implementation, historical curing environment data of the fair-faced concrete components to be tested are introduced. The preliminary overall quality index is corrected for environmental factors based on the severity of the environment to obtain the overall apparent quality index. Historical curing environment data includes the temperature fluctuation range, humidity change curve, and cumulative solar radiation during the curing period. An environmental severity assessment model is established, using the temperature fluctuation range, humidity change curve, and cumulative solar radiation as inputs to calculate an environmental severity coefficient. For example, the environmental severity assessment model can be a predefined scoring function that grades and weights the temperature fluctuation amplitude, humidity change frequency, and solar radiation intensity. A mapping function between the environmental severity coefficient and the degree of apparent quality degradation is constructed. This function can be fitted by comparing the quality of a large number of similar component samples produced under different curing environments. Using the mapping function, the preliminary overall quality index is compensated based on the calculated environmental severity coefficient. The purpose of the compensation calculation is to eliminate the apparent quality evaluation bias caused by differences in curing environments, thereby obtaining a more comparable overall apparent quality index. The formula for environmental factor correction is expressed as:

[0101]

[0102] in: This represents the corrected overall apparent quality index. This indicates the preliminary overall quality index. This represents the calculated environmental severity coefficient. This represents the baseline severity coefficient corresponding to the standard maintenance environment. and The function is used to fit the model parameters determined by the sample data. Used to limit the correction range. It can be understood that when the environmental severity coefficient... Higher than the benchmark At that time, the formula applies to the preliminary overall quality index. Perform positive compensation. See Table 1.

[0103] Table 1: Local Quality Evaluation Values ​​and Location Weight Coefficients of Grid Units in Different Areas of the Wall

[0104]

[0105] In some embodiments, a surface repair guidance map corresponding to the defect type is generated based on the defect type labels and spatial locations of all mesh cells. All mesh cells are clustered according to the defect type labels to identify continuous surface regions with the same or similar defect types, which are marked as blocks to be repaired. The clustering algorithm can employ a region growing method based on spatial adjacency, merging mesh cells with the same defect type labels and adjacent locations into a continuous block. For each defect type, the corresponding standard repair process and key control parameters are retrieved from the repair strategy library. The repair strategy library is a predefined knowledge base that stores, for example, grouting filling processes and their material ratios and curing time parameters for "bubble" defects, and special colorant spraying processes and their concentration and number of passes parameters for "color difference" defects. Combining the spatial dimensions, geometry, and micro-undulation topological information of each block to be repaired in the three-dimensional full-element surface feature field, the standard repair process is adaptively adjusted to generate a customized repair plan for the block to be repaired. For example, for a large area of ​​"contaminated" surface with obvious undulations, the adjustment plan includes increasing the amount of cleaning agent and extending the contact time of physical wiping.

[0106] Optionally, when assigning positional weight coefficients, in addition to calculations based on geometric perspective, a human visual attention model can be introduced to make the weight coefficients more consistent with the actual key areas observed by the human eye. Optionally, the standard repair process flow in the repair strategy library can be associated with a list of applicable conditions. When a process is retrieved, the system automatically filters out several candidate processes that best match the size of the block to be repaired, the severity of the defect, and other information for further adjustment. In some embodiments, customized repair schemes for all blocks to be repaired are superimposed and marked on the surface unfolded diagram of the component according to their spatial location, forming an apparent repair guidance map containing the boundaries of the repair area, repair process steps, and parameters. The surface unfolded diagram is a projection or parametric unfolded diagram of the component on a two-dimensional plane. The boundary of each block to be repaired is determined by the coordinates of its outermost grid unit and is outlined and filled in the diagram with different colors or fill patterns. The interior of the block is marked with text or symbols indicating the repair process steps, required materials, and their key parameters, such as "Process: High-pressure grouting; Material: Epoxy resin; Parameter: Pressure 0.3MPa, Duration 2 minutes". This map constitutes an intuitive and visual document guiding on-site repair construction.

[0107] In practical implementation, the environmental severity assessment model can be constructed based on multivariate analysis. Temperature fluctuations in historical maintenance environmental data are described by the average and standard deviation of the difference between the daily maximum and minimum temperatures. Humidity variation curves are characterized by their coefficient of variation within the maintenance cycle, and cumulative solar radiation is measured in total joules. A specific formula for calculating the environmental severity coefficient can be a linear combination of the above-mentioned quantitative indicators after standardization. Maintenance environmental data is collected by deploying temperature and humidity recorders and solar radiation sensors on-site, continuously monitoring and recording data throughout the predetermined maintenance cycle.

[0108] See Figure 4 This is a bar chart comparing the quality scores and location weights of different grid cells in different areas of a mirror-finished fair-faced concrete wall. The central area of ​​the field of view has the highest local quality score (approximately 0.92) and location weight coefficient (approximately 1.2), making it the core area of ​​focus for quality assessment. There is a positive correlation between the local quality score and the location weight coefficient. The bottom edge of the field of view has the lowest local score (0.8) and weight coefficient (0.6), indicating a weaker impact on the overall quality index. This type of chart is a fundamental visualization tool for calculating the apparent quality index of mirror-finished fair-faced concrete, clearly defining the quality contribution of different areas and providing a basis for regional prioritization in subsequent quality index correction and repair scheme formulation.

[0109] In one embodiment of the present invention, at preset time intervals after the fair-faced concrete component under test is put into use, the steps of acquiring surface image data and generating an overall appearance quality index are repeatedly executed to obtain a series of overall appearance quality indices over time. The changing trend and rate of change of the overall appearance quality index over time are analyzed to predict the future evolution path of the component's appearance quality. The future evolution path is compared with a preset quality degradation warning threshold. If the prediction is that the warning threshold will be reached, the appearance repair guidance map is updated, and preventive maintenance measures are added to the repair plan.

[0110] In practice, after generating the overall appearance quality index and appearance repair guidance map, a long-term quality tracking step is performed. At preset time intervals after the tested fair-faced concrete component is put into use, the steps from acquiring surface image data to generating the overall appearance quality index are repeated, resulting in a series of overall appearance quality indices over time. The preset time interval is determined based on the component's usage environment and importance; for example, for public building facades in harsh outdoor environments, the preset time interval is six months; for art walls in indoor constant temperature and humidity environments, the preset time interval is two years. In an example scenario, a museum's fair-faced concrete curved exterior wall underwent four complete appearance quality tests at months 0, 6, 12, and 18 after delivery. Each test, under the same lighting conditions as the baseline test, acquired multi-band imaging information and multispectral reflectance information of the wall surface, and fully executed the entire process from image fusion and feature decoupling to generating the overall appearance quality index, thus obtaining a time series of overall appearance quality indices at four time points, assumed to be as follows: , , , .

[0111] In practical implementation, the changing trend and rate of change of the overall appearance quality index over time series are analyzed to predict the future evolution path of component appearance quality. The changing trend can be revealed by linear or nonlinear regression fitting of the time series data, and the rate of change is the first derivative of the fitted curve at the most recent time point. One method for predicting future evolution paths is to establish a time series prediction model, such as using exponential smoothing or autoregressive integral moving average models, to predict the index value several time intervals in the future based on historical overall appearance quality index sequences. It is understood that the prediction model needs to use sufficiently long historical series data to calibrate its parameters during the training phase. The formula for calculating the linear slope of the changing trend is expressed as:

[0112]

[0113] in: This represents the slope of the linear trend of the overall apparent quality index over time. Indicates the first The overall appearance quality index obtained from this test. Indicates the first The time interval between the first test and the next test. This indicates the number of tests performed. The slope is calculated using this formula. A negative value indicates a downward trend in the overall appearance quality index. Extrapolating from the current trend, we can predict the value of the overall appearance quality index at a future point in time, such as the 24th month. This allows us to depict the future evolutionary path.

[0114] In some embodiments, the predicted future evolution path is compared with a preset quality degradation warning threshold, which is a predefined critical value characterizing the apparent quality entering a state requiring attention or intervention. This threshold can be a single absolute value of the overall apparent quality index, such as 0.70; or it can be a relative degradation ratio based on the initial quality, such as relative to the initial quality index. The degradation exceeds 30%. If the predicted future evolution path indicates that the overall appearance quality index will reach or fall below the preset warning threshold at some point in the future, the previously generated appearance repair guidance map needs to be updated. The update operation adds preventative maintenance measures to the repair plan. Preventative maintenance measures are maintenance actions planned in advance to address the predicted quality degradation. For example, if the prediction shows that the warning threshold will be reached after one year due to a continuous decline in surface gloss, the updated appearance repair guidance map will, in addition to marking the existing defect repair areas, also suggest a preventative maintenance plan and its construction parameters on the entire wall surface, recommending "a protective transparent coating sprayed within six months".

[0115] Optionally, when analyzing trends, in addition to using the overall appearance quality index series, the local quality evaluation values ​​of each grid cell or the time series of each characteristic index can be analyzed simultaneously to identify which specific appearance attribute is dominating the overall quality degradation. Optionally, the preset time interval does not have to be a fixed value and can be dynamically adjusted based on the previously detected quality degradation rate. For example, if an accelerated quality degradation rate is detected, the time interval for the next detection is automatically shortened. In some embodiments, the long-term quality tracking step can be integrated with the building information model management system of the component. The time series overall appearance quality index, predicted evolution path, and updated appearance repair guidance map obtained from each test are all stored and version-managed as part of the component's digital asset, providing continuous data records for maintenance decisions throughout its entire lifecycle.

[0116] See Figure 5 This is a pie chart showing the area distribution of different types of surface repair for mirror-finished concrete. Preventative maintenance accounts for the largest share (approximately 35%), reflecting proactive control over long-term quality degradation; gloss restoration is the second largest share (approximately 25%), representing the core repair item for the appearance quality of mirror-finished concrete; crack repair and color correction have similar shares (approximately 15% each), representing targeted defect repair; texture optimization has the smallest share (approximately 10%), representing a secondary appearance optimization item. This type of chart serves as a basis for resource allocation decisions in the repair of mirror-finished concrete, reflecting the main types of appearance defects and assisting in assessing the rationality of repair plans.

[0117] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0118] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for testing the apparent quality of mirror-finish fair-faced concrete, characterized in that, Includes the following steps: Acquire surface image data of the mirror-finished fair-faced concrete component under set illumination conditions. The surface image data includes multi-band imaging information and multi-spectral reflectance information. Image fusion and feature decoupling operations are performed on the surface image data to decouple the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map, which independently characterize the surface gloss features of the component, the surface texture features of the component, and the surface color features of the component. Spatial registration and feature association were performed on the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map to construct a three-dimensional full-element appearance feature field of the tested mirror-finished fair-faced concrete component. The three-dimensional full-element appearance feature field is scanned in a partitioned grid, and the gloss uniformity index, texture roughness index and color stability index are extracted in each grid cell. The gloss uniformity index, texture roughness index, and color stability index of each grid cell are input into a pre-constructed appearance quality analysis network, which outputs the local quality evaluation value and defect type label of each grid cell. The local quality evaluation values ​​of all grid cells are integrated, and the overall apparent quality index of the fair-faced concrete component under test is generated by weighted fusion. Based on the defect type labels and spatial locations of all grid cells, a visual repair guidance map corresponding to the defect type is generated.

2. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 1, characterized in that, The step of performing image fusion and feature decoupling operations on the surface image data includes: An image fusion strategy based on multi-scale decomposition is adopted to fuse the multi-band imaging information and the multi-spectral reflectance information to generate a high-resolution fused surface image. The fused surface image is subjected to feature separation operations, the feature separation operations including: By using directional filtering and illumination model inversion, the reflection component image, which is mainly affected by the surface geometry, is separated from the fused surface image. Through texture frequency analysis and morphological reconstruction, the intrinsic texture image, which is mainly affected by the material microstructure, is separated from the fused surface image; By performing color space transformation and spectral demixing, the intrinsic color image, mainly affected by material composition and contamination, is separated from the fused surface image. The reflection component image is normalized and specular reflection enhancement is performed to generate the specular gloss distribution map; The texture intrinsic image is subjected to three-dimensional topography restoration and mesh refinement processing to generate the micro-undulation topology map; The intrinsic color image is subjected to colorimetric transformation and noise suppression processing to generate the colorimetric coordinate mapping map.

3. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 2, characterized in that, The process involves spatially registering and feature-correlatedly performing the specular gloss distribution map, the micro-undulation topology map, and the chromaticity coordinate mapping map to construct a three-dimensional, all-element appearance feature field for the tested specular fair-faced concrete component, including: Using the surface vertices of the micro-undulation topology map as a spatial reference, the gloss value of each pixel in the specular gloss distribution map is mapped to the corresponding surface vertex, and a gloss attribute is assigned to each vertex. Map the chromaticity coordinates of each pixel in the chromaticity coordinate mapping map to the surface vertex corresponding to the micro-undulation topology map, and assign chromaticity attributes to each vertex. Based on the glossiness attribute, the chromaticity attribute associated with each surface vertex, and the three-dimensional coordinates and normal vector of the surface vertex itself, a vertex feature vector containing multi-dimensional attributes is generated. The vertex feature vectors of all surface vertices in the micro-undulation topology map are organized according to the topological connection relationship between vertices to construct a spatially continuous three-dimensional full-element appearance feature field with vertices as the basic feature carriers.

4. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 3, characterized in that, The step of performing a partitioned, gridded traversal scan of the three-dimensional full-element appearance feature field to extract gloss uniformity, texture roughness, and color stability indices within each grid cell includes: The three-dimensional full-element apparent feature field is divided into equal-area grids on a two-dimensional plane, with each grid corresponding to a surface region. Traverse each grid and read the vertex feature vectors of all surface vertices covered by that grid; The glossiness attribute values ​​of all surface vertices belonging to the same grid are analyzed for dispersion and trend change, and the glossiness uniformity index of the grid is calculated. Statistical analysis is performed on the three-dimensional coordinates and normal vectors of all surface vertices belonging to the same mesh, and the root mean square deviation and slope distribution of the surface height within the mesh are calculated as the texture roughness index of the mesh. The degree of clustering and offset distance of the chromaticity attribute values ​​of all surface vertices belonging to the same mesh in the chromaticity space are analyzed to calculate the chromaticity stability index of the mesh.

5. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 1, characterized in that, The step of inputting the gloss uniformity index, texture roughness index, and color stability index of each grid cell into a pre-constructed appearance quality analysis network includes: The appearance quality analysis network includes a feature encoding layer, a cross-attention fusion layer, and a multi-task decision layer; The feature encoding layer performs deep feature extraction on the input gloss uniformity index sequence, texture roughness index sequence, and color stability index sequence to generate high-dimensional feature vectors. The cross-attention fusion layer receives the high-dimensional feature vector and calculates the interdependence and influence weights between glossiness features, texture features and chroma features through the cross-attention mechanism to generate a fused comprehensive feature vector. The multi-task decision layer executes the quality regression task and the defect classification task in parallel based on the comprehensive feature vector; The quality regression task outputs a local quality evaluation value that characterizes the quality of the grid cell, and the defect classification task outputs a defect type label that characterizes the main defect categories of the grid cell.

6. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 5, characterized in that, The integrated local quality evaluation values ​​of all grid cells are used to generate the overall apparent quality index of the tested fair-faced concrete component through weighted fusion, including: Based on the spatial position of each grid cell on the surface of the component, a position weight coefficient is assigned to the local quality evaluation value of each grid cell. The position weight coefficient is determined based on the salience of the grid cell in the field of view. Calculate the weighted average of the local quality evaluation value of all grid cells and their corresponding location weight coefficients, and use the weighted average as the preliminary overall quality index; Historical curing environment data of the fair-faced concrete component to be tested are introduced, and environmental factors are used to correct the preliminary overall quality index according to the severity of the environment, so as to obtain the overall apparent quality index.

7. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 1, characterized in that, The step of generating an appearance repair guidance map corresponding to the defect type based on the defect type labels and spatial locations of all grid cells includes: All mesh cells are clustered according to the defect type label to identify continuous surface regions with the same or similar defect types and mark them as blocks to be repaired. For each type of defect, the corresponding standard repair process and key control parameters are retrieved from the repair strategy library; By combining the spatial dimensions, geometry, and micro-undulation topology information of each block to be repaired in the three-dimensional full-element apparent feature field, the standard repair process is adaptively adjusted to generate a customized repair plan for the block to be repaired. Customized repair plans for all areas to be repaired are superimposed and marked on the surface development diagram of the component according to their spatial location, forming an appearance repair guidance map that includes the boundaries of the repair area, repair process steps and parameters.

8. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 5, characterized in that, The method for constructing the pre-built appearance quality analysis network includes: A large number of mirror-finished concrete samples with known apparent quality grades were collected. The three-dimensional full-element apparent feature field of each sample was obtained and partitioned into grids to obtain the gloss uniformity index, texture roughness index, color stability index of all grid units of each sample, as well as the real quality label and real defect label of each grid unit. Using the gloss uniformity index, texture roughness index, and color stability index of the grid cells as input features, and the real quality label and real defect label of the grid cells as supervision signals, a multi-task deep learning network model is constructed. The multi-task deep learning network model is trained using the sample data, and the network parameters are optimized through backpropagation until the loss function between the model's predicted output and the true label converges, thus obtaining the trained appearance quality analysis network.

9. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 6, characterized in that, The historical curing environment data of the fair-faced concrete component under test are introduced, and the preliminary overall quality index is corrected for environmental factors according to the severity of the environment, including: The historical maintenance environment data includes the temperature fluctuation range, humidity change curve, and cumulative light radiation during the maintenance period; An environmental severity assessment model is established, and the temperature fluctuation range, humidity change curve and cumulative solar radiation are used as inputs to calculate an environmental severity coefficient. Construct a mapping function between the environmental severity coefficient and the degree of apparent quality degradation; Using the mapping function, the preliminary overall quality index is compensated based on the calculated environmental severity coefficient. The purpose of the compensation calculation is to eliminate the apparent quality evaluation bias caused by differences in maintenance environment, thereby obtaining a more comparable overall apparent quality index.

10. The method for testing the apparent quality of mirror-finish fair-faced concrete according to claim 1, characterized in that, After generating the overall appearance quality index and appearance repair guidance map, the method also includes a long-term quality tracking step: At a preset time interval after the tested mirror-finished fair-faced concrete component is put into use, the steps of acquiring surface image data and generating overall appearance quality index are repeated to obtain a series of overall appearance quality indices over time. Analyze the changing trend and rate of change of the overall apparent quality index over the time series to predict the future evolution path of the apparent quality of components. The future evolution path is compared with a preset quality degradation warning threshold. If the warning threshold is predicted to be reached, the appearance repair guidance map is updated, and preventive maintenance measures are added to the repair plan.

Images