A method and system for evaluating the appearance quality of fair-faced concrete

Abstract

The application discloses a kind of fair-faced concrete appearance quality evaluation method and system, it is related to concrete appearance quality evaluation technical field, comprising: image acquisition module, for obtaining fair-faced concrete surface image, the image includes color reference standard area;Illumination normalization module is connected with the image acquisition module, for based on the color reference standard area to the surface image is carried out illumination normalization processing, generates standard observation condition image;Regular texture exclusion module is connected with the illumination normalization module, for executing regular texture pre-identification to the standard observation condition image, generates regular texture mask;The present application realizes illumination normalization pretreatment by color reference standard, effectively eliminates the interference of different shooting environment illumination conditions to chromatic aberration evaluation result, improves the stability and cross-scene comparability of evaluation result.

Description

Technical Field

[0001] This invention relates to the field of concrete appearance quality evaluation technology, specifically to a method and system for evaluating the appearance quality of fair-faced concrete. Background Technology

[0002] Fair-faced concrete differs from ordinary concrete components due to its natural and rustic decorative effect. Fair-faced concrete uses the natural surface after casting as the final finish, without secondary covering treatments such as plastering or applying materials. Therefore, its appearance quality directly determines the overall visual quality and acceptance results of the project.

[0003] Currently, the assessment of the appearance quality of fair-faced concrete mainly relies on manual visual inspection. Inspectors rely on their naked eyes to observe the concrete surface and classify defects such as cracks, bubbles, and color differences according to the qualitative descriptions given in relevant standards. This method has several prominent problems: First, it is highly subjective. Different inspectors may have significantly different evaluation results for the same surface, especially in the judgment of color difference, where the human eye's perception of color is significantly affected by factors such as ambient lighting, viewing angle, and individual differences. Second, it is inefficient. Large fair-faced concrete structures (such as bridge piers and station building exterior walls) have huge surface areas, and manual inspection of each one is time-consuming and laborious. Third, it lacks spatial quantitative indicators. Existing methods only give the area ratio or severity level of defects, and cannot characterize the spatial distribution characteristics of defects. For example, for the same 3% area ratio of bubble defects, the impact of a uniformly dispersed distribution and a localized dense cluster is completely different, but traditional methods cannot distinguish between these two situations.

[0004] Regarding the standards system, JGJ 169-2009 "Technical Specification for Application of Fair-faced Concrete" provides general requirements and inspection methods for appearance quality; DB32 / T 4184-2021 "Technical Specification for Application of Fair-faced Concrete," as a local standard of Jiangsu Province, further refines the acceptance indicators for fair-faced concrete; T / JXCRC 007-2025 "Standard for Grading and Evaluating the Appearance Quality of Fair-faced Concrete in Railway Engineering" proposes a relatively complete quality grading system for railway engineering scenarios. In the above standards, the appearance quality score usually includes automated evaluation of three dimensions: cracks, bubbles, and color difference, with cracks accounting for approximately 10% of the weight, bubbles for approximately 20%, and color difference for approximately 10%. However, the scoring methods of these standards are still mainly based on manual scoring, and the granularity of the scoring rules is relatively coarse, making it difficult to meet the needs of refined management.

[0005] In recent years, image processing-based methods for detecting defects on concrete surfaces have gradually emerged. Most existing solutions focus on improving the identification accuracy of single defect types (such as crack detection or bubble detection). However, in practical engineering, evaluating the appearance quality of fair-faced concrete is a systematic problem involving multiple comprehensive indicators. Existing technical solutions have significant shortcomings in the following aspects: First, the interference of lighting conditions on color difference evaluation has not been fully resolved. In outdoor shooting environments, different lighting conditions such as direct sunlight, shadow occlusion, and overcast day scattering can cause the same surface to exhibit significantly different color performances. If effective lighting normalization processing is not carried out, the stability and comparability of color difference evaluation results will be greatly reduced. Second, regular texture features such as cracks, visible joints, and bolt hole arrays on the surface of fair-faced concrete are easily misjudged as defects. These textures are inherent products of template splicing and construction processes and belong to "regular textures" rather than true appearance defects. However, they are often mixed with defects such as cracks and bubbles in the image processing process, causing false detection and false alarms. Third, there is a lack of effective methods to quantify the spatial distribution characteristics of color difference and bubbles. For color difference, it's not enough to simply look at the size of the area exceeding the standard; it's also crucial to determine whether the exceeding areas are scattered or large, continuous patches. Continuous areas of color difference have a far greater impact on visual perception than scattered spots. Similarly, for bubbles, it's not enough to simply look at their quantity and area; it's also necessary to consider whether there are localized dense clusters. Traditional methods are almost incapable of capturing this type of spatial distribution information. Summary of the Invention

[0006] To address the aforementioned technical problems, a method and system for evaluating the appearance quality of fair-faced concrete is provided. This technical solution solves the problems of interference from lighting conditions on color difference, the easy misjudgment of regular texture features such as cracks, visible joints, and bolt hole arrays on the surface of fair-faced concrete as defects, and the lack of effective quantitative means for the spatial distribution characteristics of color difference and air bubbles.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method and system for evaluating the appearance quality of fair-faced concrete includes: an image acquisition module for acquiring an image of the fair-faced concrete surface, wherein the image contains a color reference standard area; An illumination normalization module, connected to the image acquisition module, is used to perform illumination normalization processing on the surface image based on the color reference standard area to generate a standard observation condition image. The regular texture exclusion module, connected to the illumination normalization module, is used to perform regular texture pre-identification on the standard observation condition image and generate a regular texture mask; The defect detection module is connected to the regular texture exclusion module and is used to perform crack pixel segmentation, bubble target detection and color difference abnormal region extraction on the remaining region after excluding the regular texture region based on the regular texture mask. The spatial feature analysis module, connected to the defect detection module, is used to extract spatially perceived color difference descriptors based on the color difference spatial distribution, and to quantify the bubble spatial aggregation index based on the bubble detection results. The scoring fusion module, connected to the spatial feature analysis module and the defect detection module, is used to weight and fuse the crack area ratio, bubble area ratio, spatially perceived color difference descriptor and bubble spatial aggregation index according to preset weights, calculate a comprehensive score and map and output a quality level.

[0008] Preferably, the spatial feature analysis module includes a color difference spatial perception subunit and a bubble aggregation quantification subunit: The color difference spatial perception subunit is specifically used to: perform CIELAB color space conversion on the standard observation condition image, calculate the color difference value between each pixel and the standard reference color, extract the color difference exceeding the standard area ratio, row-normalized Moran's I space autocorrelation coefficient and the area ratio of the largest connected color difference region, which together constitute the spatial perception color difference descriptor; The bubble aggregation quantification subunit is specifically used to: generate a bubble density field by estimating the kernel density based on the detected center pixel coordinates of each bubble, and calculate the bubble density variation coefficient and the ratio of the maximum local density to the average density at the grid level, which together constitute the bubble spatial aggregation index.

[0009] A method for evaluating the appearance quality of fair-faced concrete includes the following steps: S1. Obtain an image of the fair-faced concrete surface, wherein the image contains a color reference standard area; S2. Perform illumination normalization processing on the surface image based on the color reference standard area to generate a standard observation condition image; S3. Perform regular texture pre-identification on the standard observation condition image to generate a regular texture mask; S4. After excluding the regular texture region based on the regular texture mask, crack pixel segmentation, bubble target detection and color difference abnormal region extraction are performed on the remaining effective region respectively. S5. Based on the color space distribution of the standard observation condition image, extract the spatially perceived color difference descriptor; S6. Based on the bubble detection results of step S4, quantify the bubble spatial aggregation index; S7. Based on the crack area ratio and bubble area ratio obtained in step S4, the spatially perceived color difference descriptor obtained in step S5, and the bubble spatial aggregation index obtained in step S6, a comprehensive score is calculated by weighted fusion according to a preset weight, and a quality level is output based on the comprehensive score.

[0010] Preferably, the method for solving the color correction matrix based on the color reference standard in step S2 is as follows: Select in the color reference standard area A reference color block, The The chromaticity values ​​of each reference color cover at least three different hue regions in the color space; based on the deviation between the standard chromaticity values ​​of each reference color block and the actual measured chromaticity values, the 3×3 color correction matrix is ​​solved by the least squares method. When there is spatial illumination non-uniformity in the surface image, the surface image is divided into multiple overlapping sub-regions. For sub-regions containing pixels of the color reference standard region, the local color correction matrix is ​​solved independently. For sub-regions that do not contain pixels of the color reference standard region, the color correction matrix of the nearest neighbor sub-region is used. The overlapping regions are fused using bilinear interpolation.

[0011] Preferably, the regular texture pre-identification in step S3 involves preset straight line thresholds, preset minimum length thresholds, preset angle ranges, preset peak thresholds, frequency ranges, preset radius ranges, and preset circle thresholds; the method for determining each parameter is as follows: The preset straight line threshold is determined based on a preset percentage range of the total voting value of the Hough transform of edge pixels; the preset minimum length threshold is determined based on a preset percentage range of the number of pixels on the short side of the image; the preset angle range is determined based on the template splicing direction angle. The preset peak threshold is determined based on a preset multiple range of the power spectral density mean; the frequency range is determined based on the expected spacing range of bolt holes and the image pixel resolution. The preset radius range is determined based on the design diameter range of the bolt hole and the image pixel resolution; the preset circle threshold is determined based on the preset percentage range of the cumulative voting value of the Hough circle transform. When no periodic feature meeting the conditions is detected by autocorrelation analysis using Fast Fourier Transform, only the detection result of Hough Circle Transform is used as the bolt hole region; when both methods detect the bolt hole region, the union of the two is taken as the final bolt hole region.

[0012] Preferably, in step S5: the method for extracting the spatially perceived color difference descriptor is: converting the standard observation condition image to the CIELAB color space, calculating the color difference value between each pixel and the standard reference color, and generating a color difference distribution map; Pixels whose color difference value exceeds the preset color difference threshold are marked as color difference exceeding standard pixels, and the proportion of the number of color difference exceeding standard pixels to the total number of pixels in the effective detection area is calculated as the color difference exceeding standard area ratio. Using the effective pixels after excluding the regular texture mask as the node set, a spatial weight matrix is ​​constructed using the Rook adjacency rule and row normalization is performed. The row-normalized Moran's I spatial autocorrelation coefficient is then calculated. For pixels with excessive color difference, connected regions are marked based on Rook adjacency relationship. The connected region with the largest area is extracted, and the ratio of the area of ​​the largest connected region to the total area of ​​the effective detection region is calculated as the proportion of the area of ​​the largest connected color difference region.

[0013] Preferably, in step S6: the quantification method of the bubble spatial aggregation index is as follows: obtain the pixel coordinates of each bubble center detected in step S4, determine the kernel bandwidth based on the Silverman rule, and then perform Gaussian kernel density estimation with each bubble center as the kernel center to generate a bubble density field; The effective detection area is uniformly divided into multiple grids, and the average density value of each grid is calculated. The ratio of the standard deviation to the mean of all grid density values ​​is calculated as the bubble density variation coefficient, and the ratio of the maximum density value to the mean density value in all grids is calculated as the ratio of the maximum local density to the average density.

[0014] Preferably, in step S7: The crack area ratio is mapped to a crack score through a piecewise linear mapping function. The bubble area ratio and bubble density variation coefficient are mapped to a bubble score by weighted combination after being mapped by a piecewise linear mapping function. The color difference score is calculated by weighted combination of the area ratio of color difference exceeding the standard, the spatial autocorrelation coefficient, and the area ratio of the largest connected color difference region after being mapped by a piecewise linear mapping function. Each piecewise linear mapping function outputs a full score when the index value is below the qualified threshold, a zero score when it exceeds the severe threshold, and linearly interpolates between the qualified threshold and the severe threshold. The crack score, bubble score, and color difference score are weighted and fused according to weighting coefficients to calculate an automated comprehensive score. .

[0015] Preferably, the quality grade mapping method is as follows: Based on the comprehensive score Determine the quality grade: when When the quality grade is determined to be "Excellent", when When the quality grade is determined to be "Good"; when When the quality level is determined to be "medium"; when At that time, the quality grade was determined to be "poor".

[0016] Preferably, the method includes weighting and fusing the crack score, bubble score, and color difference score according to weighting coefficients to calculate an automated comprehensive score. The method is as follows:

[0017] in, Score the crack. Rate the bubbles. Rate the color difference; , , These are the weighting coefficients for crack score, bubble score, and color difference score, respectively, and they satisfy... ; The weighting coefficient , , Using the analytic hierarchy process (AHP), a 3×3 judgment matrix was constructed with cracks, bubbles, and color difference as evaluation elements. The elements of this judgment matrix... Indicates the first The indicator relative to the first The importance of each indicator is assigned using a 1-9 scale; after normalizing each column of the judgment matrix, the weight vector is obtained by averaging by rows, and a consistency test is performed. When the consistency ratio... When the consistency test is passed, the weight vector that passes the consistency test is used as the weight coefficient. , , .

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention achieves illumination normalization preprocessing through a color reference standard, effectively eliminating the interference of different shooting environment lighting conditions on color difference evaluation results, and improving the stability and cross-scene comparability of the evaluation results.

[0019] By employing a regular texture pre-identification mechanism, regular textures such as cicada seams, visible seams, and bolt hole arrays are excluded in advance, significantly reducing the false alarm rate of defect detection. Furthermore, a spatially perceptive color difference descriptor is proposed, organically combining three dimensions: the proportion of color difference exceeding the standard area, the spatial autocorrelation coefficient, and the proportion of the largest connected region area. This allows for a comprehensive and accurate quantification of the severity of color difference defects, solving the problem that traditional methods cannot distinguish the spatial distribution pattern of color difference.

[0020] By employing a quantitative scheme for bubble spatial aggregation, and through kernel density estimation and gridded statistical analysis, the spatial distribution characteristics of bubbles are accurately quantified, making the evaluation results more consistent with actual visual perception. A multi-index weighted fusion scoring system balances comprehensiveness and objectivity, while automated indicators ensure the objectivity and efficiency of the evaluation, making it suitable for the appearance quality acceptance of various fair-faced concrete projects. Attached Figure Description

[0021] Figure 1 This is an overall architecture diagram of the fair-faced concrete appearance quality evaluation system of the present invention; Figure 2 This is an overall flowchart of the method for evaluating the appearance quality of fair-faced concrete according to the present invention; Figure 3 The flowcharts are for two implementation schemes of the illumination normalization processing in step S2 of the present invention; Figure 4The flowcharts are for two implementation schemes of the illumination normalization processing in step S3 of the present invention; Figure 5 This is a flowchart of the spatial perception color difference analysis in step S5 of the present invention; Figure 6 This is a flowchart of the bubble aggregation quantification process in step S6 of the present invention. Detailed Implementation

[0022] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0023] Reference Figure 1 As shown, a method for evaluating the appearance quality of fair-faced concrete is applied to, for example... Figure 1 The evaluation system shown is comprised of an image acquisition module, an illumination normalization module, a regularized texture exclusion module, a defect detection module, a spatial feature analysis module, and a scoring fusion module. The data flow relationships between these modules are as follows: Figure 1 As shown below. (Combined with...) Figure 2 The overall flowchart will be explained step by step.

[0024] Step S1: Acquire an image of the fair-faced concrete surface The image acquisition module acquires images of the fair-faced concrete surface. A crucial requirement for this step is that the image must contain a color reference standard area, which is essential for subsequent illumination normalization.

[0025] There are three optional forms for the color reference standard area: The first method is using a standard color chart. During photography, the standard color chart is placed next to the surface being tested and included in the shot. The standard color chart contains at least six color patches with known standard chromaticity values, and these patches must cover at least three different hue regions in the color space. This is because if all the reference colors are concentrated around a single hue, the solution to the color correction matrix will suffer from ill-conditioned problems, and the correction accuracy cannot be guaranteed. The more color patches and the more uniform the hue coverage, the better the correction effect; however, six color patches represent the lower limit for meeting basic solution accuracy.

[0026] The second type is a standard curing test block of concrete from the same batch. The test block and the component being tested use the same batch of concrete mix proportions and are cut into shape after standard curing. The surface color of the test block has a natural consistency with the surface being tested. The test block is placed near the surface being tested and photographed together, and the surface of the test block can be used as a color reference standard.

[0027] The third type is the defect-free reference color region. When it is inconvenient to set the first two reference standards, it can be determined statistically from the image itself. A region in the image known to be defect-free (no cracks, no bubbles, no contamination) is selected as the reference color. Specifically, the CIELAB chromaticity values ​​of all pixels in this region are counted, and the median of each channel is taken as the standard reference color. Compared with the mean, the median is more robust to outliers and is less likely to be skewed by individual noisy pixels.

[0028] The automatic identification method for defect-free reference color regions is as follows: First, suspected defective pixels are marked through preliminary crack detection and bubble detection. The preliminary detection uses a lightweight threshold segmentation method, eliminating the need for deep learning models. Crack detection can be based on gradient magnitude thresholds (e.g., the Sobel operator response exceeding three times the global mean), and bubble detection can be based on a joint judgment of dark spot area and roundness. After marking suspected defective pixels, K-means clustering analysis is performed on the remaining non-defective pixels in the CIELAB color space, and the number of clusters is... The value is set to 3-5. The pixel region corresponding to the largest cluster area is selected as the candidate reference color region. The area of ​​this candidate region is not less than 5% of the effective detection area. This minimum area constraint is to ensure that the candidate region has a sufficient number of pixels to support the statistically significant color consistency evaluation. If the percentage of the largest cluster area is less than 5%, the current image is determined not to be suitable for the defect-free reference color region scheme, and Retinex illumination normalization is used instead.

[0029] The identified candidate reference color regions also need to undergo quality verification to ensure that the color consistency of these regions meets the requirements for use as reference colors. The verification method involves calculating the standard deviation of the CIELAB chromaticity values ​​of each pixel within the candidate reference color region. ,when exist , , When all three channels are below the preset fluctuation threshold ( aisle , aisle , aisle If the color consistency of a region is deemed good, it can be used as a reference color region. If the standard deviation of any channel exceeds the fluctuation threshold, it indicates that the candidate region still contains pixels with inconsistent colors, which may indicate slight color difference, contamination, or texture interference. In this case, the region should be narrowed (e.g., only the core subset with the most concentrated colors in the largest cluster should be taken) or another color reference standard should be used.

[0030] Step S2, Illumination normalization processing: The illumination normalization module performs illumination normalization processing on the surface image obtained in step S1 to eliminate color deviations caused by uneven illumination conditions during shooting, generating an image under standard observation conditions. This embodiment provides two optional solutions, such as... Figure 3 As shown.

[0031] Option 1: Solve the color correction matrix based on the color reference standard In this embodiment, a reference color and an actual color are known, and the mapping relationship between them encodes information about illumination interference. By reversing the mapping relationship, illumination interference can be eliminated.

[0032] The specific steps are as follows: Select in the color reference standard area Reference color blocks ( ). For the first There are several reference color patches, and the known standard chromaticity value is... The actual chromaticity value measured in the image is Construct a 3×3 color correction matrix. , so that:

[0033] For all All reference color blocks hold true. The system of equations can be written in matrix form:

[0034] in, 3× The actual measurement value matrix, 3× The standard value matrix.

[0035] It is important to note that the chromaticity values ​​of the selected reference color patches must match the actual measured values ​​in the matrix. Number of conditions satisfied Below the preset threshold number .

[0036] Solving for the color correction matrix using the least squares method:

[0037] Solving for the results Then, for each pixel of the entire image Perform transformation That completes the color correction.

[0038] When the measured surface area is large and there is significant spatial lighting unevenness in the image (e.g., half in sunlight and half in shadow), a single global correction matrix may not be sufficient. In this case, the image is divided into... There are three overlapping sub-regions, with the overlap width between adjacent sub-regions not less than 10% of the side length of the sub-region. For a sub-region containing pixels of the color reference standard region, the local color correction matrix is ​​solved independently based on the reference color within the sub-region; for a sub-region that does not contain a reference color, the color correction matrix of the nearest sub-region containing the reference color is used.

[0039] After each sub-region is calibrated individually, the calibration results of adjacent sub-regions are fused using bilinear interpolation in overlapping areas to avoid seams.

[0040] Option 2: Illumination normalization based on Retinex theory When there is no suitable color reference standard in the image, the Retinex method can be used. The basic assumption of Retinex theory is that an image can be decomposed into the product of an illumination component and a reflection component. The illumination component reflects the ambient lighting conditions, and the reflection component reflects the intrinsic color of the object's surface. Removing the illumination component yields the color of the object under "standard white light".

[0041] Specifically, take the logarithm of each pixel in the image to obtain... ,in It is the light component, It is the reflection component. Then, the illumination component is estimated using a Gaussian surround filter. :

[0042] in, For Gaussian kernel function, For the filtering window, subtract the estimated illumination components from the logarithmic image: Then, the reflection component image is obtained by performing an exponential operation. The reflection component image is used as the standard observation condition image.

[0043] The Retinex method has the advantage of not requiring additional color charts or test blocks, simplifying the shooting process. Its disadvantage is that the selection of the Gaussian kernel parameter has a significant impact on the results, and its correction accuracy is lower than that of the first method for scenes with large color deviations. In practical applications, the choice can be made flexibly based on the site conditions and accuracy requirements.

[0044] Step S3, regular texture pre-identification step, in this step, regular texture features that are not defects, such as cicada seams, visible seams, and bolt holes, are identified and excluded in subsequent defect detection. Specifically, a three-channel detection method is used.

[0045] like Figure 4 As shown, each of the three channels controls a type of regular texture, and finally the output mask is merged.

[0046] Channel 1: Hough line detection is used to locate crinkle seams and visible seams. After converting the standard observation condition image to a grayscale image, edge detection is performed (using the Canny operator in this embodiment) to obtain a set of edge pixels. A Hough transform is then applied to the edge pixels in the parameter space. Voting is accumulated during the process. Votes exceeding a preset threshold are extracted. The straight line segment.

[0047] The parameter determination method is as follows: Take 5% to 15% of the total votes for the Hough transform of edge pixels. If the threshold is too low, noisy edges will be treated as straight lines, and if it is too high, actual gaps may be missed.

[0048] Preset minimum length threshold Take 5% to 15% of the number of pixels on the shorter side of the image. Cracks and visible seams are usually continuous straight lines that run through the entire or most of the surface of a component, and their length is much greater than that of ordinary cracks or noise lines. This length threshold can filter out short interference lines.

[0049] Preset angle range For the expected direction angles of the crinkle seam and the visible seam, when the template splicing direction is known, take the template splicing direction angle. When the template splicing direction is unknown, take... It covers both horizontal and vertical directions.

[0050] Line segments that meet the above three conditions are marked as candidate areas for cicada seams or open seams.

[0051] Channel 2: FFT autocorrelation analysis for locating bolt hole arrays Bolt holes are typically arranged in a regular rectangular array. This periodic arrangement will produce a noticeable spike response in the frequency domain.

[0052] Specific steps: Extract grayscale profiles along both the horizontal and vertical directions, perform a Fast Fourier Transform on each profile, and calculate the power spectral density (PSD). If the PSD is within the frequency range... The frequency exceeds the preset peak threshold at a certain point. The sharp peak indicates that there is a periodic feature in this direction, and the corresponding area is marked as a candidate area for the bolt hole array.

[0053] The parameter determination method is as follows: The value should be 2 to 4 times the average power spectral density. The peak must be a significant increase relative to the background PSD. 2 to 4 times is a reasonable range.

[0054] Frequency range Based on the expected spacing range of bolt holes in the template design. and image pixel resolution (Unit: mm / pixel) Confirm:

[0055] In other words, the larger the bolt hole spacing, the lower the corresponding frequency; the smaller the spacing, the higher the frequency. The frequency range is determined by the possible range of bolt hole spacing.

[0056] When no periodic features satisfying the conditions are detected by FFT autocorrelation analysis, only the detection result of Hough circle transform (channel 3 below) is used as the bolt hole region. When both methods detect the bolt hole region, the union of the two is taken as the final bolt hole region.

[0057] Channel 3: Hough circle transformation is used to locate a single bolt hole. Perform a Hough circle transform on the grayscale image within a preset radius. Internal detection of circular features.

[0058] The preset radius range is based on the design diameter range of the bolt holes. and image pixel resolution Sure:

[0059] Preset circular threshold Take 5% to 15% of the cumulative votes from the Hough circle transformation. If the cumulative votes exceed... The circular area is marked as a candidate area for bolt holes.

[0060] Merge to generate regular texture masks The candidate regions output by the three channels—the cicada-seam / open-seam candidate region, the bolt hole array candidate region, and the bolt hole candidate region—are merged (union) to generate a regular texture mask. All subsequent defect detection and analysis operations are performed outside the coverage area of ​​this mask.

[0061] Step S4, Defect Detection: After excluding the regular texture region based on the regular texture mask generated in step S3, three types of defect detection are performed on the remaining valid region.

[0062] Crack detection: Crack pixel segmentation is performed using a semantic segmentation network based on U-Net. U-Net is a convolutional neural network with an encoder-decoder structure. The encoder gradually extracts high-level semantic features through convolution and pooling operations, and the decoder gradually restores spatial resolution through transposed convolution, while skip connections preserve detailed information.

[0063] In this embodiment, the network structure of the U-Net crack segmentation model is as follows: The encoder contains four downsampling stages, each consisting of two 3×3 convolutions + batch normalization (BN) + ReLU activation function + 2×2 max pooling operation, with the number of feature channels being 64, 128, 256, and 512 respectively; the decoder contains four upsampling stages, each consisting of a 2×2 transposed convolution + skip connections (concatenating feature maps from the corresponding encoder layer) + two 3×3 convolutions + BN + ReLU, with the number of feature channels being 256, 128, 64, and 32 respectively. The input size is 256×256 or 512×512 pixels. Before input, the standard observation condition image is subjected to bilinear interpolation scaling and normalization (pixel values ​​scaled to...). The output is a single-channel probability map of the same size as the input. After Sigmoid activation, a binary crack mask is obtained with a threshold of 0.5. Crack pixels are labeled as 1, and non-crack pixels are labeled as 0. The model training dataset contains no fewer than 5000 manually annotated concrete surface images, with annotations in the form of pixel-level crack masks. The crack segmentation IoU is no less than 75% on the validation set.

[0064] Bubble detection: A YOLO-based target detection network is used for bubble target detection. The YOLO (You Only Look Once) series of detection networks can output the bounding box coordinates and category information of multiple targets simultaneously in a single forward propagation.

[0065] In a preferred embodiment of this application, the bubble detection model employs a YOLOv5 or YOLOv8 series network. The input size is 640×640 pixels, and adaptive scaling (maintaining aspect ratio and padding for insufficient areas) and normalization are performed before input. The output format is the bounding box coordinates of each bubble. The model training dataset contains no fewer than 8000 manually annotated concrete surface images, labeled as bubble bounding boxes. On the validation set, the bubble detection mAP@0.5 is no less than 80%. The confidence score and category label are also considered. The pixel area of ​​each bubble can be approximated by the actual segmentation result within the bounding box or by the area of ​​the bounding box itself.

[0066] Preliminary extraction of color difference anomaly regions: After converting the image under standard observation conditions to the CIELAB color space, extract regions where the color difference value exceeds a preset color difference threshold. The pixel areas are considered as areas of color difference abnormality. The CIELAB color space is a perceptually uniform color space, in which... The value corresponds well with the degree of color difference perceived by the human eye.

[0067] Step S5, Spatial Perception Color Difference Analysis: This step is the core of color difference evaluation, such as Figure 5As shown, the goal is to construct a three-dimensional "spatial-perceived color difference descriptor" to comprehensively characterize the severity of color difference defects from three dimensions: area, spatial distribution, and extreme contiguousness.

[0068] (1) CIELAB color space conversion and color difference value calculation Each pixel in the standard observation condition image is converted from the RGB color space to the CIELAB color space to obtain the individual pixel values. The conversion process involves two steps: first, converting from RGB to CIEXYZ tristimulus values, and then converting from XYZ to CIELAB.

[0069] Determine the standard reference color When the color reference standard area is a standard color card or a standard maintenance test block, the standard reference color is the known CIELAB colorimetric value of the color reference standard area.

[0070] When the color reference standard area is a defect-free baseline color area, the standard reference color is the median of the CIELAB chromaticity values ​​of all pixels within that area. The median is used instead of the mean because the median is not sensitive to individual abnormal pixels and is more representative of the color level of "most normal pixels".

[0071] Calculate the color difference between each pixel and the standard reference color:

[0072] With each pixel The values ​​are used as pixel values ​​to generate a color difference distribution map with the same size as the image under standard observation conditions. This map essentially "translates" the color deviation into a grayscale image, where the grayscale value of each pixel is the size of its color difference from the standard color.

[0073] Color difference threshold The value is related to the color of the concrete surface; light-colored concrete ( Surface color differences are more easily perceived by the human eye. Lower values ​​can be used (e.g., 5-8); dark concrete ( The surface color difference tolerance is relatively high. The limit can be appropriately relaxed (e.g., 10-15); medium-toned concrete ( Typical values On-site selection can be adaptively based on the surface brightness of the component being measured. This makes the evaluation results more consistent with human subjective perception.

[0074] (2) Area ratio of color difference exceeding the standard Color difference value exceeds color difference threshold Pixels marked as having excessive color difference are used to calculate the area ratio of pixels with excessive color difference:

[0075] in, The number of pixels with excessive color difference. The total number of pixels in the effective detection area (excluding areas covered by regular texture masks).

[0076] (3) Row-standardized Moran's I spatial autocorrelation coefficient Moran's I is a classic indicator in spatial statistics for measuring spatial autocorrelation. In this step, it is used to measure whether pixels with excessive color difference are randomly scattered or clustered in space.

[0077] The node set is defined as all valid pixels excluding the area covered by the regular texture mask in the color difference distribution map, and the total number of nodes is... The original spatial weight matrix is ​​constructed using the Rook adjacency rule. That is, if two pixels share an edge (one of the four directions: up, down, left, right), and are adjacent to each other, their corresponding weights are... ;otherwise .

[0078] right Row standardization yields the spatial weight matrix. :

[0079] The significance of row normalization lies in ensuring that the sum of the weights of each row is 1, eliminating the difference between edge pixels and interior pixels. After row normalization, the calculation result of Moran's I is no longer affected by pixel position (edge ​​or interior), and the calculation results between images of different sizes are comparable.

[0080] Let the first The color difference value of each effective pixel is The mean of all valid pixel color difference values ​​is Calculate Moran's I:

[0081] The range of values ​​is approximately in between. A value close to 1 indicates positive spatial autocorrelation, meaning that pixels with high color difference tend to be adjacent to other pixels with high color difference, forming contiguous areas of color difference. A value close to 0 indicates a random distribution; A value close to -1 indicates negative spatial autocorrelation, with alternating high and low color differences. For evaluating the color difference of fair-faced concrete, A higher value indicates that the color difference problem is more concentrated and more serious.

[0082] (4) Percentage of the area of ​​the region with the largest connected color difference For all satisfied For pixels with excessive color difference, connected component labeling is performed based on Rook adjacency relationships (using a two-pass scanning method or an equivalent union-find algorithm). The connected component with the largest area is extracted, and its pixel area is... The total effective detection area is .calculate:

[0083] This indicator shows "how large the area with the most severe color difference is". Even if the overall area with excessive color difference is not high, if there is a large, continuous area with color difference, the visual impact will be significant.

[0084] (5) Spatial perceptual color difference descriptor The ratio of areas with excessive color difference Moran's I spatial autocorrelation coefficient and the percentage of the area of ​​the region with the largest connected color difference These three indicators together constitute the spatially perceived color difference descriptor. .

[0085] Step S6, Bubble Aggregation Quantification Step The goal of this step is to quantify the spatial distribution characteristics of the bubbles, such as... Figure 6 As shown, a bubble density field is generated by kernel density estimation, and then statistical analysis is performed at the grid level.

[0086] S6.1 Obtain the coordinates of the bubble center Obtain the pixel coordinates of the center of each bubble from the bubble detection results of step S4. , ,in This represents the total number of bubbles detected.

[0087] S6.2 Kernel Density Estimation Using the center of each bubble as the kernel center, a Gaussian kernel function is used to estimate the kernel density, with a kernel bandwidth of [missing value]. Determined based on Silverman rules:

[0088] in, Let the standard deviation be the coordinates of the center of all bubbles. The total number of bubbles. The Silverman rule is a classic bandwidth selection method widely used in kernel density estimation, which can strike a balance between "oversmoothing" and "overcoarsening".

[0089] The effective detection area is evenly divided into There are 1 grid, and the center coordinates of each grid are 1. , , Grid side length Core bandwidth 2 times, that is Number of grid cells:

[0090] in and These represent the width and height pixel counts of the effective detection area, respectively. This is for rounding up.

[0091] With core bandwidth Using the Gaussian kernel function as the parameter, the bubble density value at the center of each grid is calculated:

[0092] The physical meaning of this formula is quite intuitive: each bubble contributes a Gaussian "hill" to the surrounding space, with the contribution increasing closer to the bubble's center and decreasing further away. The sum of all the contributions from each bubble yields the overall bubble density field. High density values ​​indicate clustered bubbles, while low density values ​​indicate sparsely distributed bubbles.

[0093] Calculate the average density value of each grid. Then calculate all Mean and standard deviation of each grid density value:

[0094]

[0095] S6.3 Bubble Spatial Aggregation Index Bubble density variation coefficient:

[0096] It reflects the overall non-uniformity of bubble density in space. A value close to 0 indicates that the density is very uniform. The larger the value, the more significant the density difference, with bubbles clustered in some areas and almost none in others.

[0097] The ratio of maximum local density to average density:

[0098] This reflects the degree to which the most densely populated areas deviate from the average level. For example... This means that the bubble density in the densest area is five times the average level, indicating a significant local clustering phenomenon. and Together they constitute the index of bubble spatial aggregation.

[0099] Step S7, Score fusion and quality level determination In this step, all the indicators generated in the previous steps are summarized, weighted and integrated, and the final comprehensive score and quality level are calculated.

[0100] S7.1 Piecewise linear mapping of each indicator Each indicator is converted into a score of 0 to 100 through a piecewise linear mapping function. The shape of the mapping function is uniform: when the indicator value is below the qualified threshold, a full score of 100 is output; when it exceeds the severe threshold, a score of zero is output; and linear interpolation is performed between the qualified threshold and the severe threshold.

[0101] With crack area ratio For example, let its qualified threshold be... The severity threshold is Then the crack score for:

[0102] The mapping methods for other indicators are completely similar.

[0103] S7.2 Bubble Scoring bubble area ratio and bubble density variation coefficient After being mapped using their respective piecewise linear mapping functions, the scores are weighted and combined to form the bubble score:

[0104] in, This represents the weight of the bubble area ratio.

[0105] S7.3 Color Difference Rating Area ratio with excessive color difference Spatial autocorrelation coefficient and the percentage of the area of ​​the region with the largest connected color difference The color difference score is obtained by weighting and combining the colors after mapping them separately.

[0106] in, .

[0107] S7.4 Automated Comprehensive Scoring Crack scores, bubble scores, and color difference scores are weighted and fused according to weighting coefficients to calculate an automated comprehensive score.

[0108] in, .

[0109] There are two methods for determining the weighting coefficients: The first method is based on the Analytic Hierarchy Process (AHP), using cracks, bubbles, and color difference as examples to construct a 3×3 judgment matrix. The elements of the judgment matrix... Indicates the first The indicator relative to the first The importance of each indicator is determined using a 1-9 scale (1 indicates equal importance, 3 indicates slightly important, 5 indicates significantly important, 7 indicates strongly important, 9 indicates extremely important, and 2, 4, 6, and 8 are the midpoints of adjacent scales).

[0110] Below is a specific example of an AHP judgment matrix. Taking the determination of the weights of cracks, bubbles, and color difference in the evaluation of the appearance quality of fair-faced concrete as an example, assuming that after expert review, it is determined that bubbles are significantly more important than cracks ( Color difference is slightly more important than cracks. Bubbles are just as important as color difference. The judgment matrix is ​​then:

[0111] in (Bubbles are significantly more important than cracks). (Color difference is slightly more important than cracks) (Bubbles are just as important as color difference).

[0112] Calculate the weight vector: normalize each column of the judgment matrix and then average it by row.

[0113] Sum of all columns: Column 1 ; Column 2 ; Column 3 .

[0114] First column normalization: , , Second column normalization: , , Third column normalization: , , The weight vector is obtained by averaging the rows:

[0115]

[0116]

[0117] That is, the weight of cracks is about 0.115, the weight of bubbles is about 0.480, and the weight of color difference is about 0.405, and the sum of the three is 1.

[0118] Consistency check: Calculate the largest eigenvalue of the judgment matrix. .

[0119]

[0120] First line: Second line: Third line:

[0121] Consistency Indicators For a 3-order matrix, the random consistency index (This can be found by referring to the table).

[0122] Consistency ratio when The judgment matrix is ​​considered to have passed the consistency check. In this example... The judgment matrix passed the consistency test. The crack weights were then determined. Bubble weight Color difference weight .

[0123] The second method for determining the weights is based on normalizing the weight percentages of the three indicators—cracks, bubbles, and color difference—as specified in the appearance quality scoring section of the JGJ 169-2009 or T / JXCRC 007-2025 standards. Taking T / JXCRC 007-2025 as an example, the weights for cracks are 10%, bubbles are 20%, and color difference is 10%. After normalization: , , .

[0124] After confirming the weights, an automated comprehensive score is then performed.

[0125] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A system for evaluating the appearance quality of fair-faced concrete, characterized in that, include: The image acquisition module is used to acquire images of the surface of fair-faced concrete, wherein the images contain color reference standard areas; An illumination normalization module, connected to the image acquisition module, is used to perform illumination normalization processing on the surface image based on the color reference standard area to generate a standard observation condition image. The regular texture exclusion module, connected to the illumination normalization module, is used to perform regular texture pre-identification on the standard observation condition image and generate a regular texture mask; The defect detection module is connected to the regular texture exclusion module and is used to perform crack pixel segmentation, bubble target detection and color difference abnormal region extraction on the remaining region after excluding the regular texture region based on the regular texture mask. The spatial feature analysis module, connected to the defect detection module, is used to extract spatially perceived color difference descriptors based on the color difference spatial distribution, and to quantify the bubble spatial aggregation index based on the bubble detection results. The scoring fusion module, connected to the spatial feature analysis module and the defect detection module, is used to weight and fuse the crack area ratio, bubble area ratio, spatially perceived color difference descriptor and bubble spatial aggregation index according to preset weights, calculate a comprehensive score and map and output a quality level.

2. The fair-faced concrete appearance quality evaluation system according to claim 1, characterized in that, The spatial feature analysis module includes a color difference spatial perception subunit and a bubble aggregation quantification subunit: The color difference spatial perception subunit is specifically used to: perform CIELAB color space conversion on the standard observation condition image, calculate the color difference value between each pixel and the standard reference color, extract the color difference exceeding the standard area ratio, row-normalized Moran's I space autocorrelation coefficient and the area ratio of the largest connected color difference region, which together constitute the spatial perception color difference descriptor; The bubble aggregation quantification subunit is specifically used to: generate a bubble density field by estimating the kernel density based on the detected center pixel coordinates of each bubble, and calculate the bubble density variation coefficient and the ratio of the maximum local density to the average density at the grid level, which together constitute the bubble spatial aggregation index.

3. A method for evaluating the appearance quality of fair-faced concrete, applied to the evaluation system as described in claim 1 or 2, characterized in that, Includes the following steps: S1. Obtain an image of the fair-faced concrete surface, wherein the image contains a color reference standard area; S2. Perform illumination normalization processing on the surface image based on the color reference standard area to generate a standard observation condition image; S3. Perform regular texture pre-identification on the standard observation condition image to generate a regular texture mask; S4. After excluding the regular texture region based on the regular texture mask, crack pixel segmentation, bubble target detection and color difference abnormal region extraction are performed on the remaining effective region respectively. S5. Based on the color space distribution of the standard observation condition image, extract the spatially perceived color difference descriptor; S6. Based on the bubble detection results of step S4, quantify the bubble spatial aggregation index; S7. Based on the crack area ratio and bubble area ratio obtained in step S4, the spatially perceived color difference descriptor obtained in step S5, and the bubble spatial aggregation index obtained in step S6, a comprehensive score is calculated by weighted fusion according to a preset weight, and a quality level is output based on the comprehensive score.

4. The method for evaluating the appearance quality of fair-faced concrete according to claim 3, characterized in that, The method for solving the color correction matrix based on the color reference standard in step S2 is as follows: Select in the color reference standard area A reference color block, The The chromaticity values ​​of each reference color cover at least three different hue regions in the color space; based on the deviation between the standard chromaticity values ​​of each reference color block and the actual measured chromaticity values, the 3×3 color correction matrix is ​​solved by the least squares method. When there is spatial illumination non-uniformity in the surface image, the surface image is divided into multiple overlapping sub-regions. For sub-regions containing pixels of the color reference standard region, the local color correction matrix is ​​solved independently. For sub-regions that do not contain pixels of the color reference standard region, the color correction matrix of the nearest neighbor sub-region is used. The overlapping regions are fused using bilinear interpolation.

5. The method for evaluating the appearance quality of fair-faced concrete according to claim 3, characterized in that, Step S3, regular texture pre-identification, involves preset line thresholds, preset minimum length thresholds, preset angle ranges, preset peak thresholds, frequency ranges, preset radius ranges, and preset circle thresholds; the method for determining each parameter is as follows: The preset straight line threshold is determined based on a preset percentage range of the total voting value of the Hough transform of edge pixels; the preset minimum length threshold is determined based on a preset percentage range of the number of pixels on the short side of the image; the preset angle range is determined based on the template splicing direction angle. The preset peak threshold is determined based on a preset multiple range of the power spectral density mean; the frequency range is determined based on the expected spacing range of bolt holes and the image pixel resolution. The preset radius range is determined based on the design diameter range of the bolt hole and the image pixel resolution; the preset circle threshold is determined based on the preset percentage range of the cumulative voting value of the Hough circle transform. When no periodic feature meeting the conditions is detected by autocorrelation analysis using Fast Fourier Transform, only the detection result of Hough Circle Transform is used as the bolt hole region; when both methods detect the bolt hole region, the union of the two is taken as the final bolt hole region.

6. The method for evaluating the appearance quality of fair-faced concrete according to claim 3, characterized in that, In step S5: the method for extracting the spatially perceived color difference descriptor is as follows: convert the standard observation condition image to the CIELAB color space, calculate the color difference value between each pixel and the standard reference color, and generate a color difference distribution map; Pixels whose color difference value exceeds the preset color difference threshold are marked as color difference exceeding standard pixels, and the proportion of the number of color difference exceeding standard pixels to the total number of pixels in the effective detection area is calculated as the color difference exceeding standard area ratio. Using the effective pixels after excluding the regular texture mask as the node set, a spatial weight matrix is ​​constructed using the Rook adjacency rule and row normalization is performed. The row-normalized Moran's I spatial autocorrelation coefficient is then calculated. For pixels with excessive color difference, connected regions are marked based on Rook adjacency relationship. The connected region with the largest area is extracted, and the ratio of the area of ​​the largest connected region to the total area of ​​the effective detection region is calculated as the proportion of the area of ​​the largest connected color difference region.

7. The method for evaluating the appearance quality of fair-faced concrete according to claim 3, characterized in that, In step S6: The quantification method of the bubble spatial aggregation index is as follows: obtain the pixel coordinates of each bubble center detected in step S4, determine the kernel bandwidth based on the Silverman rule, and then perform Gaussian kernel density estimation with each bubble center as the kernel center to generate a bubble density field. The effective detection area is uniformly divided into multiple grids, and the average density value of each grid is calculated. The ratio of the standard deviation to the mean of all grid density values ​​is calculated as the bubble density variation coefficient, and the ratio of the maximum density value to the mean density value in all grids is calculated as the ratio of the maximum local density to the average density.

8. The method for evaluating the appearance quality of fair-faced concrete according to claim 3, characterized in that, In step S7: The crack area ratio is mapped to a crack score through a piecewise linear mapping function. The bubble area ratio and bubble density variation coefficient are mapped to a bubble score by weighted combination after being mapped by a piecewise linear mapping function. The color difference score is calculated by weighted combination of the area ratio of color difference exceeding the standard, the spatial autocorrelation coefficient, and the area ratio of the largest connected color difference region after being mapped by a piecewise linear mapping function. Each piecewise linear mapping function outputs a full score when the index value is below the qualified threshold, a zero score when it exceeds the severe threshold, and linearly interpolates between the qualified threshold and the severe threshold. The crack score, bubble score, and color difference score are weighted and fused according to weighting coefficients to calculate an automated comprehensive score. .

9. The method for evaluating the appearance quality of fair-faced concrete according to claim 8, characterized in that, The mapping method for quality levels is as follows: Based on the comprehensive score Determine the quality grade: when When the quality level is determined to be "excellent", when When the quality grade is determined to be "Good"; when When the quality level is determined to be "medium"; when At that time, the quality grade was determined to be "poor".

10. The method for evaluating the appearance quality of fair-faced concrete according to claim 9, characterized in that, The crack score, bubble score, and color difference score are weighted and fused according to weighting coefficients to calculate an automated comprehensive score. The method is as follows: in, Score the crack. Rate the bubbles. Rate the color difference; , , These are the weighting coefficients for crack score, bubble score, and color difference score, respectively, and they satisfy... ; The weighting coefficient , , Using the analytic hierarchy process (AHP), a 3×3 judgment matrix was constructed with cracks, bubbles, and color difference as evaluation elements. The elements of this judgment matrix... Indicates the first The indicator relative to the first The importance of each indicator is assigned using a 1-9 scale; after normalizing each column of the judgment matrix, the weight vector is obtained by averaging by rows, and a consistency test is performed. When the consistency ratio... When the consistency test is passed, the weight vector that passes the consistency test is used as the weight coefficient. , , .

Images