An intelligent inspection method and system for a construction site

Abstract

The application discloses an intelligent inspection method and system for a construction site, and belongs to the technical field of image processing. The application firstly processes a remote sensing image of a cement pavement of the construction site, and positions a pixel area of the pavement; then, relative difference values of each pixel point are calculated according to a pixel mean value of the area, and a difference value distribution image is generated; then, maximum values and minimum values of each window area of the distribution image are extracted, and high difference value and low difference value compression images are obtained respectively. After that, non-pavement pixel areas are marked as abnormal points, and the two types of compression images are divided into grids, difference characteristic values are calculated based on the abnormal points in the grids, and high and low difference characteristic matrices are formed; then, joint deviation degrees of the difference characteristic values of each grid are calculated, and high and low joint deviation matrices are obtained; finally, the above matrices are processed through a pavement defect characteristic evaluation network, and a pavement quality score is output. The application effectively improves the precision of pavement evaluation through multi-dimensional feature extraction and intelligent evaluation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to an intelligent inspection method and system for construction sites. Background Technology

[0002] In the field of modern infrastructure construction, quality control of cement pavement at construction sites is a crucial link in ensuring the long-term stable operation of projects. In recent years, with the rapid development of urbanization and transportation construction, the application of cement pavement in projects such as roads, bridges, and squares has continued to expand. Its construction quality not only affects the service life of the project, but also directly impacts traffic safety and the safety of public life and property.

[0003] Faced with the large-scale demand for road surface inspection, traditional manual inspection methods are no longer sufficient to meet the requirements of high-precision quality inspection due to their low efficiency, strong subjectivity, and limited detection range. In this context, image-based detection technologies have gradually become a research hotspot, attempting to compensate for the shortcomings of manual inspection. However, existing image-based algorithms lack the ability to deeply analyze subtle road surface defects (such as micro-cracks and localized wear). When processing images, they are prone to misclassifying normal texture differences as defects or overlooking truly existing defects, resulting in low accuracy in road surface assessment. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides an intelligent inspection method and system for construction sites, which solves the problem of low road surface assessment accuracy in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: an intelligent inspection method for construction sites, comprising the following steps:

[0006] Locate the pixel area of ​​the cement road surface in the remote sensing image of the construction site;

[0007] Based on the pixel mean of the road surface pixel region, calculate the relative pixel difference value of each pixel to obtain the difference value distribution image;

[0008] Extract the maximum and minimum values ​​from each window region in the difference value distribution image to obtain high difference value compressed image and low difference value compressed image;

[0009] All pixels except those in the road surface pixel region are marked as outliers. Both the high-difference-value compressed image and the low-difference-value compressed image are divided into grids. Based on the outliers in the grid, the difference feature value is calculated for each grid to obtain the high-difference feature matrix and the low-difference feature matrix.

[0010] The joint bias is calculated for the difference eigenvalues ​​of each grid in the high-difference-value compressed image and the low-difference-value compressed image, respectively, to obtain the high joint bias matrix and the low joint bias matrix.

[0011] A pavement defect feature assessment network is used to process the high-discrepancy feature matrix, low-discrepancy feature matrix, high-joint-bias matrix, and low-joint-bias matrix to obtain the pavement quality score.

[0012] Furthermore, the process of locating the road surface pixel region includes:

[0013] Take any pixel value from the remote sensing image of the cement pavement at the construction site that has not been compared before as the comparison value;

[0014] Calculate the distance between other pixel values ​​and the comparison value. When the distance is less than the distance threshold, classify the pixels corresponding to the pixel value and the comparison value into one category, and select a new comparison value. Repeat the classification process until all pixels are classified and multiple categories are obtained.

[0015] The class containing the most pixels among multiple classes is selected as the road surface pixel region.

[0016] Furthermore, the formula for calculating the relative pixel difference value for each pixel is:

[0017] ,

[0018] Among them, P re,i Let P be the relative pixel difference value of the i-th pixel, μ be the pixel mean value of the road surface pixel region, and P be the pixel relative difference value of the i-th pixel. i Let be the pixel value of the i-th pixel, where i is a positive integer and sign is the sign function.

[0019] Furthermore, the process of obtaining high-difference-value compressed images and low-difference-value compressed images includes: dividing the difference-value distribution image into multiple window regions of size 3×3;

[0020] Extract the maximum relative pixel difference value in each window region, and arrange the maximum relative pixel difference values ​​according to the window region position to obtain a high difference value compressed image;

[0021] Extract the minimum relative pixel difference value in each window region, and arrange the minimum relative pixel difference values ​​according to the window region position to obtain a low difference value compressed image.

[0022] Furthermore, the calculation of difference feature values ​​for each grid in both the high-difference-value compressed image and the low-difference-value compressed image includes the following process:

[0023] The total difference value is obtained by summing the relative pixel differences of all outlier points within the grid.

[0024] The ratio of the relative pixel difference value of each outlier to the total difference value is used as the weight of each outlier.

[0025] Multiply the weight of each outlier by the relative pixel difference value of that outlier;

[0026] The weighted difference value is obtained by summing the results of multiplying the outliers;

[0027] The ratio of the number of outliers in the grid to the total number of pixels in the grid is used as the density factor;

[0028] Multiplying the weighted difference value by the density factor yields the difference characteristic value of the grid.

[0029] Furthermore, the difference feature values ​​on the high difference value compressed image are arranged according to the grid order on the corresponding image to obtain the high difference feature matrix;

[0030] Arrange the differential feature values ​​on the low-difference compressed image according to the grid order on the corresponding image to obtain the low-difference feature matrix;

[0031] The calculation of joint bias for the difference feature values ​​of each grid in both the high-difference-value compressed image and the low-difference-value compressed image includes the following steps:

[0032] The global mean is obtained by averaging the difference feature values ​​of each grid belonging to an image.

[0033] Calculate the global bias based on the difference between the differential feature value of each grid and the corresponding global mean;

[0034] Taking each grid as the center, the average value of the difference characteristic values ​​of the neighboring grids of the center is taken to obtain the neighborhood mean;

[0035] The neighborhood bias is calculated based on the difference between the differential feature value of each grid and the neighborhood mean.

[0036] The global bias and the neighborhood bias are added together to obtain the joint bias.

[0037] Furthermore, the joint biases on the high-discrepancy compressed image are arranged according to the grid order on the corresponding image to obtain the high joint bias matrix;

[0038] Arrange the joint bias values ​​on the low-discrepancy compressed image according to the grid order on the corresponding image to obtain the low joint bias matrix;

[0039] The process of calculating the global deviation includes: calculating the difference between the difference feature value of each grid and the global mean, taking the absolute value of the difference to obtain the global absolute difference, and dividing the global absolute difference by the global mean to obtain the global deviation.

[0040] The process of calculating the neighborhood deviation includes: calculating the difference between the difference feature value of each grid and the neighborhood mean, taking the absolute value of the difference to obtain the absolute neighborhood difference, and dividing the absolute neighborhood difference by the neighborhood mean to obtain the neighborhood deviation.

[0041] Furthermore, the pavement defect feature evaluation network includes: a first difference deviation feature fusion unit, a second difference deviation feature fusion unit, a first maximum pooling layer, a second maximum pooling layer, a first average pooling layer, a second average pooling layer, adder A1, adder A2, a first stacked convolutional unit, a second stacked convolutional unit, and a fully connected layer;

[0042] The first input terminal of the first difference bias feature fusion unit is used to input a high difference feature matrix, its second input terminal is used to input a high joint bias matrix, and its output terminal is connected to the input terminal of the first max pooling layer and the input terminal of the first average pooling layer, respectively.

[0043] The first input of the second difference bias feature fusion unit is used to input the low difference feature matrix, the second input is used to input the low joint bias matrix, and the output is connected to the input of the second max pooling layer and the input of the second average pooling layer, respectively.

[0044] The input of adder A1 is connected to the output of the first max pooling layer and the output of the second max pooling layer, respectively, and its output is connected to the input of the first stacked convolutional unit.

[0045] The input of adder A2 is connected to the output of the first average pooling layer and the output of the second average pooling layer, and its output is connected to the input of the second stacked convolutional unit.

[0046] The input of the fully connected layer is connected to the output of the first stacked convolutional unit and the output of the second stacked convolutional unit, respectively, and its output serves as the output of the road surface defect feature evaluation network.

[0047] Furthermore, the first difference deviation feature fusion unit and the second difference deviation feature fusion unit have the same structure, both including: a first multi-scale convolutional feature extraction module, a second multi-scale convolutional feature extraction module, a first 1×1 convolutional layer, a second 1×1 convolutional layer, adder A3, adder A4 and multiplier M1;

[0048] The input of the first multi-scale convolutional feature extraction module is connected to the first input of adder A3, and serves as the first input of the first difference bias feature fusion unit and the second difference bias feature fusion unit.

[0049] The input of the second multi-scale convolutional feature extraction module is connected to the first input of adder A4, and serves as the second input of the first difference bias feature fusion unit and the second difference bias feature fusion unit.

[0050] The input of the first 1×1 convolutional layer is connected to the output of the first multi-scale convolutional feature extraction module, and its output is connected to the second input of adder A3.

[0051] The input of the second 1×1 convolutional layer is connected to the output of the second multi-scale convolutional feature extraction module, and its output is connected to the second input of adder A4.

[0052] The input terminal of multiplier M1 is connected to the output terminal of adder A3 and the output terminal of adder A4 respectively, and its output terminal serves as the output terminal of the first difference deviation feature fusion unit and the second difference deviation feature fusion unit.

[0053] An intelligent inspection system for construction sites includes: a road surface pixel acquisition subsystem, a difference value distribution image construction subsystem, a compressed image construction subsystem, a first matrix construction subsystem, a second matrix construction subsystem, and an evaluation subsystem;

[0054] The road surface pixel acquisition subsystem is used to locate the road surface pixel region on remote sensing images of cement road surfaces at construction sites;

[0055] The difference value distribution image construction subsystem is used to calculate the relative difference value of each pixel based on the pixel mean of the road surface pixel region, and obtain the difference value distribution image;

[0056] The compressed image construction subsystem is used to extract the maximum and minimum values ​​in each window region of the difference value distribution image to obtain high difference value compressed images and low difference value compressed images.

[0057] The first matrix construction subsystem is used to mark all pixels except the road surface pixel area as outliers, divide the high difference value compressed image and the low difference value compressed image into grids, and calculate the difference feature value for each grid based on the outliers in the grid to obtain the high difference feature matrix and the low difference feature matrix.

[0058] The second matrix construction subsystem is used to calculate the joint bias degree for the difference feature values ​​of each grid on the high difference value compressed image and the low difference value compressed image, respectively, to obtain the high joint bias matrix and the low joint bias matrix.

[0059] The evaluation subsystem is used to process the high-discrepancy feature matrix, low-discrepancy feature matrix, high-joint-bias matrix and low-joint-bias matrix using the pavement defect feature evaluation network to obtain the pavement quality score.

[0060] The beneficial effects of this invention are as follows:

[0061] 1. This invention calculates the relative difference value of pixels and extracts the extreme values ​​of the window to obtain high / low difference value compressed images. Then, it combines the calculation of difference feature values ​​and joint deviation degree by gridding to accurately capture the minor defects of the road surface (such as micro cracks and local wear), improves the ability to deeply analyze the minor defects of the road surface, effectively avoids misjudging normal texture differences as defects or missing real defects, and greatly improves the accuracy of road surface assessment.

[0062] 2. This invention automates the entire process from image preprocessing to evaluating the quality score output by the network using road defect features, reducing human intervention and lowering the subjectivity and uncertainty of manual inspections, thus efficiently meeting the needs of large-scale road inspections.

[0063] 3. This invention locates the pixel region of the cement pavement in remote sensing images, calculates the relative difference value of each pixel, highlights the pixel values ​​of abnormal pixels, and forms a difference value distribution image. Then, it extracts the maximum and minimum values ​​through a window to generate high difference value compressed images and low difference value compressed images, respectively. The computational workload is significantly reduced. At the same time, the maximum and minimum values ​​are retained in the window region, reflecting the two extreme value situations in the window, highlighting the minor defects of the pavement. Based on the two extreme value distribution situations, the pavement construction quality is comprehensively evaluated, improving the evaluation accuracy.

[0064] 4. This invention suppresses misjudgment caused by normal textures by calculating the difference features of gridded structures and analyzing the joint deviation degree, thus avoiding misjudgment of normal textures (such as vibration textures) within a single grid.

[0065] 5. This invention uses a pavement defect feature evaluation network to process high-discrepancy feature matrices, low-discrepancy feature matrices, high-joint-deviation matrices, and low-joint-deviation matrices, thereby achieving multi-dimensional comprehensive evaluation of defects and improving the accuracy of pavement evaluation. Attached Figure Description

[0066] Figure 1 A flowchart of an intelligent inspection method for construction sites;

[0067] Figure 2 This is a schematic diagram of the structure of a pavement defect feature assessment network;

[0068] Figure 3 This is a schematic diagram of the structure of the first difference deviation feature fusion unit and the second difference deviation feature fusion unit;

[0069] Figure 4 This is a schematic diagram of the structure of the first multi-scale convolutional feature extraction module and the second multi-scale convolutional feature extraction module;

[0070] Figure 5 This is a schematic diagram of the structure of the first stacked convolutional unit and the second stacked convolutional unit. Detailed Implementation

[0071] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0072] Example 1, such as Figure 1 As shown, an intelligent inspection method for construction sites includes the following steps:

[0073] Locate the pixel area of ​​the cement road surface in the remote sensing image of the construction site;

[0074] Based on the pixel mean of the road surface pixel region, calculate the relative pixel difference value of each pixel to obtain the difference value distribution image;

[0075] Extract the maximum and minimum values ​​from each window region in the difference value distribution image to obtain high difference value compressed image and low difference value compressed image;

[0076] All pixels except those in the road surface pixel region are marked as outliers. Both the high-difference-value compressed image and the low-difference-value compressed image are divided into grids. Based on the outliers in the grid, the difference feature value is calculated for each grid to obtain the high-difference feature matrix and the low-difference feature matrix.

[0077] The joint bias is calculated for the difference eigenvalues ​​of each grid in the high-difference-value compressed image and the low-difference-value compressed image, respectively, to obtain the high joint bias matrix and the low joint bias matrix.

[0078] A pavement defect feature assessment network is used to process the high-discrepancy feature matrix, low-discrepancy feature matrix, high-joint-bias matrix, and low-joint-bias matrix to obtain the pavement quality score.

[0079] In this embodiment, the process of finding the road surface pixel region includes:

[0080] Take any pixel value from the remote sensing image of the cement pavement at the construction site that has not been compared before as the comparison value;

[0081] Calculate the distance between other pixel values ​​and the comparison value. When the distance is less than the distance threshold, classify the pixels corresponding to the pixel value and the comparison value into one category, and select a new comparison value. Repeat the classification process until all pixels are classified and multiple categories are obtained.

[0082] The class containing the most pixels among multiple classes is selected as the road surface pixel region.

[0083] In this embodiment, the distance is calculated using the following formula: , where d i P is the distance to the i-th pixel value. i Let P be the pixel value of the i-th pixel. c For comparison.

[0084] In this embodiment, the present invention classifies pixels by calculating the distance between pixel values ​​and comparison values. This allows for precise grouping of pixels with similar characteristics based on differences in pixel values, effectively distinguishing road surface pixels from non-road surface pixels. Furthermore, it can calculate the distance between stored road surface pixel values ​​and pixel values ​​in remote sensing images of the cement road surface at the construction site. When the distance is less than a distance threshold, the pixel corresponding to that value is classified as a road surface pixel region.

[0085] In this embodiment, the pixel value is a grayscale value, and the distance threshold is a specific value set based on experience or experimentation, which can be adjusted according to actual needs.

[0086] In this embodiment, the formula for calculating the relative pixel difference value of each pixel is:

[0087] ,

[0088] Among them, P re,i Let P be the relative pixel difference value of the i-th pixel, μ be the pixel mean value of the road surface pixel region, and P be the pixel relative difference value of the i-th pixel. i Let be the pixel value of the i-th pixel, where i is a positive integer and sign is the sign function.

[0089] This invention uses the pixel mean μ of the road surface pixel region as a benchmark, and calculates each pixel P... i The difference from the mean is quantified. This invention uses the sign function sign(P) i The algorithm (−μ) can clearly distinguish whether a pixel is greater than (positive, corresponding to bright spot anomalies such as wear) or less than (negative, corresponding to dark spot anomalies such as cracks) the mean μ of the road surface pixel area, thus achieving preliminary classification and identification of different types of anomalies. Simultaneously, squaring the difference between the pixel value and the mean amplifies the magnitude of the difference, ensuring that even minute differences (such as a slight decrease in pixel value caused by a minor crack) yield a more significant relative difference after squaring, enhancing the ability to identify minute defects.

[0090] In this embodiment, the process of obtaining high difference value compressed image and low difference value compressed image includes: dividing the difference value distribution image into multiple window regions of size 3×3;

[0091] Extract the maximum relative pixel difference value in each window region, and arrange the maximum relative pixel difference values ​​according to the window region position to obtain a high difference value compressed image;

[0092] Extract the minimum relative pixel difference value in each window region, and arrange the minimum relative pixel difference values ​​according to the window region position to obtain a low difference value compressed image.

[0093] This invention divides the difference value distribution image into 3×3 window regions, and extracts only the maximum and minimum pixel relative difference values ​​for each window, reducing the amount of computation. The maximum value can focus on the significant positive difference of bright anomalies (such as wear and whitening), while the minimum value can lock in the significant negative difference of dark anomalies (such as cracks and depressions).

[0094] In this embodiment, the calculation of difference feature values ​​for each grid in both the high-difference-value compressed image and the low-difference-value compressed image includes the following process:

[0095] The total difference value is obtained by summing the relative pixel differences of all outlier points within the grid.

[0096] The ratio of the relative pixel difference value of each outlier to the total difference value is used as the weight of each outlier.

[0097] Multiply the weight of each outlier by the relative pixel difference value of that outlier;

[0098] The weighted difference value is obtained by summing the results of multiplying the outliers;

[0099] The ratio of the number of outliers in the grid to the total number of pixels in the grid is used as the density factor;

[0100] Multiplying the weighted difference value by the density factor yields the difference characteristic value of the grid.

[0101] The formula for calculating the weighted difference value is: Where r is the weighted difference value, P ab,i Let be the relative pixel difference value of the i-th outlier, and N be the number of outliers in the grid.

[0102] This invention performs weighted processing on the relative pixel differences of all anomaly points within a grid to obtain a weighted difference value, quantifying the overall degree of difference in anomalies within the grid. Simultaneously, multiplying this value by a density factor considers both the magnitude and density of anomaly differences. For anomaly areas with large and densely distributed differences, a larger difference feature value is obtained, more accurately reflecting the severity of road surface quality in that area.

[0103] In this embodiment, when the compressed image resolution is 1000×1000 and each pixel corresponds to 0.1cm of the actual road surface, each grid can be divided into 5×5 or 10×10 pixel areas. Each grid corresponds to an actual road surface area of ​​0.5cm×0.5cm-1cm×1cm, which can effectively capture subtle defects. For larger-scale defects such as localized wear areas and large potholes, larger grids can also be effectively identified. For example, for potholes with a diameter of 5cm or more, if a 20×20 or 30×30 grid is used, each grid corresponds to an actual road surface area of ​​2cm×2cm-3cm×3cm. This reduces the number of grids and improves computational efficiency without losing key information.

[0104] In this embodiment, the difference feature values ​​on the high difference value compressed image are arranged according to the grid order on the corresponding image to obtain the high difference feature matrix;

[0105] Arrange the differential feature values ​​on the low-difference compressed image according to the grid order on the corresponding image to obtain the low-difference feature matrix;

[0106] The calculation of joint bias for the difference feature values ​​of each grid in both the high-difference-value compressed image and the low-difference-value compressed image includes the following steps:

[0107] The global mean is obtained by averaging the difference feature values ​​of each grid belonging to an image.

[0108] Calculate the global bias based on the difference between the differential feature value of each grid and the corresponding global mean;

[0109] Taking each grid as the center, the average value of the difference characteristic values ​​of the neighboring grids of the center is taken to obtain the neighborhood mean;

[0110] The neighborhood bias is calculated based on the difference between the differential feature value of each grid and the neighborhood mean.

[0111] The global bias and the neighborhood bias are added together to obtain the joint bias, where one grid corresponds to one joint bias.

[0112] In this embodiment, the joint biases on the high difference value compressed image are arranged according to the grid order on the corresponding image to obtain the high joint bias matrix;

[0113] Arrange the joint bias values ​​on the low-discrepancy compressed image according to the grid order on the corresponding image to obtain the low joint bias matrix;

[0114] The process of calculating the global deviation includes: calculating the difference between the difference feature value of each grid and the global mean, taking the absolute value of the difference to obtain the global absolute difference, and dividing the global absolute difference by the global mean to obtain the global deviation.

[0115] The process of calculating the neighborhood deviation includes: calculating the difference between the difference feature value of each grid and the neighborhood mean, taking the absolute value of the difference to obtain the absolute neighborhood difference, and dividing the absolute neighborhood difference by the neighborhood mean to obtain the neighborhood deviation.

[0116] This invention calculates the global mean and global deviation to grasp the distribution of all grid difference feature values ​​in an image at a holistic level, thus understanding the overall deviation level of road surface quality. Simultaneously, it calculates the neighborhood mean and neighborhood deviation to focus on the difference features of each grid and its surrounding local areas, capturing subtle quality changes in local road surfaces. This effectively avoids misjudging normal texture differences as defects or overlooking real defects, significantly improving the accuracy of road surface assessment.

[0117] like Figure 2 As shown, the road defect feature evaluation network includes: a first difference deviation feature fusion unit, a second difference deviation feature fusion unit, a first maximum pooling layer, a second maximum pooling layer, a first average pooling layer, a second average pooling layer, adder A1, adder A2, a first stacked convolutional unit, a second stacked convolutional unit, and a fully connected layer.

[0118] The first input terminal of the first difference bias feature fusion unit is used to input a high difference feature matrix, its second input terminal is used to input a high joint bias matrix, and its output terminal is connected to the input terminal of the first max pooling layer and the input terminal of the first average pooling layer, respectively.

[0119] The first input of the second difference bias feature fusion unit is used to input the low difference feature matrix, the second input is used to input the low joint bias matrix, and the output is connected to the input of the second max pooling layer and the input of the second average pooling layer, respectively.

[0120] The input of adder A1 is connected to the output of the first max pooling layer and the output of the second max pooling layer, respectively, and its output is connected to the input of the first stacked convolutional unit.

[0121] The input of adder A2 is connected to the output of the first average pooling layer and the output of the second average pooling layer, and its output is connected to the input of the second stacked convolutional unit.

[0122] The input of the fully connected layer is connected to the output of the first stacked convolutional unit and the output of the second stacked convolutional unit, respectively, and its output serves as the output of the road surface defect feature evaluation network.

[0123] This invention uses first and second difference deviation feature fusion units to fuse high difference feature matrices with high joint deviation matrices and low difference feature matrices with low joint deviation matrices, respectively. This fusion method integrates difference features and joint deviation features, more comprehensively covering information on pavement defects in different dimensions (such as difference magnitude, deviation degree, etc.).

[0124] The max pooling layers (first and second max pooling layers) extract salient parts of the features, highlighting key features of road defects and enhancing the ability to identify obvious defects. The average pooling layers (first and second average pooling layers) retain the overall distribution information of the features. The salient features of the upper and lower roads are added element-wise, and the mean features of the upper and lower roads are added element-wise, achieving the fusion of salient and mean features, thus enhancing defect features. The first and second stacked convolutional units, through multi-layer convolution operations, can perform depth extraction on the fused features. A fully connected layer then combines the outputs of the two stacked convolutional units for classification, improving the ability to distinguish defect types and severity.

[0125] like Figure 3 As shown, the first difference deviation feature fusion unit and the second difference deviation feature fusion unit have the same structure, both including: a first multi-scale convolution feature extraction module, a second multi-scale convolution feature extraction module, a first 1×1 convolutional layer, a second 1×1 convolutional layer, adder A3, adder A4 and multiplier M1;

[0126] The input of the first multi-scale convolutional feature extraction module is connected to the first input of adder A3, and serves as the first input of the first difference bias feature fusion unit and the second difference bias feature fusion unit.

[0127] The input of the second multi-scale convolutional feature extraction module is connected to the first input of adder A4, and serves as the second input of the first difference bias feature fusion unit and the second difference bias feature fusion unit.

[0128] The input of the first 1×1 convolutional layer is connected to the output of the first multi-scale convolutional feature extraction module, and its output is connected to the second input of adder A3.

[0129] The input of the second 1×1 convolutional layer is connected to the output of the second multi-scale convolutional feature extraction module, and its output is connected to the second input of adder A4.

[0130] The input terminal of multiplier M1 is connected to the output terminal of adder A3 and the output terminal of adder A4 respectively, and its output terminal serves as the output terminal of the first difference deviation feature fusion unit and the second difference deviation feature fusion unit.

[0131] like Figure 4As shown, the first multi-scale convolutional feature extraction module and the second multi-scale convolutional feature extraction module include: a 3×3 convolutional layer, a 5×5 convolutional layer, a 7×7 convolutional layer, a first ReLU layer, a second ReLU layer, a third ReLU layer, and a Concat layer;

[0132] The input of the 3×3 convolutional layer is connected to the input of the 5×5 convolutional layer and the input of the 7×7 convolutional layer, respectively, and serves as the input of the first multi-scale convolutional feature extraction module and the second multi-scale convolutional feature extraction module.

[0133] The input of the first ReLU layer is connected to the output of the 3×3 convolutional layer; the input of the second ReLU layer is connected to the output of the 5×5 convolutional layer; the input of the third ReLU layer is connected to the output of the 7×7 convolutional layer.

[0134] The input of the Concat layer is connected to the outputs of the first ReLU layer, the second ReLU layer, and the third ReLU layer, respectively, and its output is used as the output of the first multi-scale convolutional feature extraction module and the second multi-scale convolutional feature extraction module.

[0135] In the multi-scale convolutional feature extraction module, convolutional layers of different sizes (3×3, 5×5, and 7×7) are used to capture the features of road surface defects at different scales. Small-scale convolutions (such as 3×3) can extract subtle local defect features, while large-scale convolutions (such as 7×7) can obtain a larger range of defect distribution or overall features. The multi-scale features are concatenated through a Concat layer and then mapped through a 1×1 convolutional layer. Adders A3 and A4 fuse the input with the multi-scale mapped features to enhance the features themselves. Finally, multiplier M1 fuses the features corresponding to the difference feature matrix and the features corresponding to the joint bias matrix.

[0136] like Figure 5 As shown, the first stacked convolutional unit and the second stacked convolutional unit have the same structure, both including three 3×3 convolutional layers connected in sequence.

[0137] Example 2: An intelligent inspection system for construction sites, comprising: a road surface pixel acquisition subsystem, a difference value distribution image construction subsystem, a compressed image construction subsystem, a first matrix construction subsystem, a second matrix construction subsystem, and an evaluation subsystem;

[0138] The road surface pixel acquisition subsystem is used to locate the road surface pixel region on remote sensing images of cement road surfaces at construction sites;

[0139] The difference value distribution image construction subsystem is used to calculate the relative difference value of each pixel based on the pixel mean of the road surface pixel region, and obtain the difference value distribution image;

[0140] The compressed image construction subsystem is used to extract the maximum and minimum values ​​in each window region of the difference value distribution image to obtain high difference value compressed images and low difference value compressed images.

[0141] The first matrix construction subsystem is used to mark all pixels except the road surface pixel area as outliers, divide the high difference value compressed image and the low difference value compressed image into grids, and calculate the difference feature value for each grid based on the outliers in the grid to obtain the high difference feature matrix and the low difference feature matrix.

[0142] The second matrix construction subsystem is used to calculate the joint bias degree for the difference feature values ​​of each grid on the high difference value compressed image and the low difference value compressed image, respectively, to obtain the high joint bias matrix and the low joint bias matrix.

[0143] The evaluation subsystem is used to process the high-discrepancy feature matrix, low-discrepancy feature matrix, high-joint-bias matrix and low-joint-bias matrix using the pavement defect feature evaluation network to obtain the pavement quality score.

[0144] The specific implementation methods of Example 2 and Example 1 are the same.

[0145] Based on the road quality scoring standard and corresponding situation in this embodiment: the road quality score is 90-100 points, the relative difference value of pixels in the cement road area is evenly distributed, and there are no obvious defects such as cracks or potholes.

[0146] A pavement quality score of 70-89 indicates localized fluctuations in the distribution of relative pixel differences, potentially indicating minor cracks or slight wear. A score of 50-69 indicates significant fluctuations in the distribution of relative pixel differences, potentially indicating noticeable cracks, small potholes, or exposed aggregate. A score of 30-49 indicates drastic fluctuations in the distribution of relative pixel differences, potentially indicating deep cracks, extensive potholes, or structural damage. A score of 0-29 indicates extremely uneven distribution of relative pixel differences, potentially indicating severe structural defects such as cracks or collapses.

[0147] In this embodiment, the pavement defect feature assessment network collects "high difference feature matrix, low difference feature matrix, high joint bias matrix and low joint bias matrix" of remote sensing images of cement pavement as training samples, uses pavement quality score as label, and trains the pavement defect feature assessment network using the existing gradient descent method. The training method is an existing technology.

[0148] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent inspection method for construction sites, characterized in that, Includes the following steps: Locate the pixel area of ​​the cement road surface in the remote sensing image of the construction site; Based on the pixel mean of the road surface pixel region, calculate the relative pixel difference value of each pixel to obtain the difference value distribution image; Extract the maximum and minimum values ​​from each window region in the difference value distribution image to obtain high difference value compressed image and low difference value compressed image; All pixels except those in the road surface pixel region are marked as outliers. Both the high-difference-value compressed image and the low-difference-value compressed image are divided into grids. Based on the outliers in the grid, the difference feature value is calculated for each grid to obtain the high-difference feature matrix and the low-difference feature matrix. The joint bias is calculated for the difference eigenvalues ​​of each grid in the high-difference-value compressed image and the low-difference-value compressed image, respectively, to obtain the high joint bias matrix and the low joint bias matrix. A pavement defect feature assessment network is used to process the high-discrepancy feature matrix, low-discrepancy feature matrix, high joint deviation matrix, and low joint deviation matrix to obtain the pavement quality score. The road surface defect feature evaluation network includes: a first difference deviation feature fusion unit, a second difference deviation feature fusion unit, a first maximum pooling layer, a second maximum pooling layer, a first average pooling layer, a second average pooling layer, adder A1, adder A2, a first stacked convolutional unit, a second stacked convolutional unit, and a fully connected layer; The first input terminal of the first difference bias feature fusion unit is used to input a high difference feature matrix, its second input terminal is used to input a high joint bias matrix, and its output terminal is connected to the input terminal of the first max pooling layer and the input terminal of the first average pooling layer, respectively. The first input of the second difference bias feature fusion unit is used to input the low difference feature matrix, the second input is used to input the low joint bias matrix, and the output is connected to the input of the second max pooling layer and the input of the second average pooling layer, respectively. The input of adder A1 is connected to the output of the first max pooling layer and the output of the second max pooling layer, respectively, and its output is connected to the input of the first stacked convolutional unit. The input of adder A2 is connected to the output of the first average pooling layer and the output of the second average pooling layer, and its output is connected to the input of the second stacked convolutional unit. The input of the fully connected layer is connected to the output of the first stacked convolutional unit and the output of the second stacked convolutional unit, respectively, and its output is used as the output of the road surface defect feature evaluation network. The first difference deviation feature fusion unit and the second difference deviation feature fusion unit have the same structure, both including: a first multi-scale convolution feature extraction module, a second multi-scale convolution feature extraction module, a first 1×1 convolutional layer, a second 1×1 convolutional layer, adder A3, adder A4 and multiplier M1; The input of the first multi-scale convolutional feature extraction module is connected to the first input of adder A3, and serves as the first input of the first difference bias feature fusion unit and the second difference bias feature fusion unit. The input of the second multi-scale convolutional feature extraction module is connected to the first input of adder A4, and serves as the second input of the first difference bias feature fusion unit and the second difference bias feature fusion unit. The input of the first 1×1 convolutional layer is connected to the output of the first multi-scale convolutional feature extraction module, and its output is connected to the second input of adder A3. The input of the second 1×1 convolutional layer is connected to the output of the second multi-scale convolutional feature extraction module, and its output is connected to the second input of adder A4; the input of multiplier M1 is connected to the output of adder A3 and adder A4 respectively, and its output serves as the output of the first difference deviation feature fusion unit and the second difference deviation feature fusion unit.

2. The intelligent inspection method for construction sites according to claim 1, characterized in that, The process of finding the pixel region of the road surface includes: Take any pixel value from the remote sensing image of the cement pavement at the construction site that has not been compared before as the comparison value; Calculate the distance between other pixel values ​​and the comparison value. When the distance is less than the distance threshold, classify the pixels corresponding to the pixel value and the comparison value into one category, and select a new comparison value. Repeat the classification process until all pixels are classified and multiple categories are obtained. The class containing the most pixels among multiple classes is selected as the road surface pixel region.

3. The intelligent inspection method for construction sites according to claim 1, characterized in that, The formula for calculating the relative pixel difference value of each pixel is: ， Among them, P re,i Let P be the relative pixel difference value of the i-th pixel, μ be the pixel mean value of the road surface pixel region, and P be the pixel relative difference value of the i-th pixel. i Let be the pixel value of the i-th pixel, where i is a positive integer and sign is the sign function.

4. The intelligent inspection method for construction sites according to claim 1, characterized in that, The process of obtaining high-difference-value compressed images and low-difference-value compressed images includes: dividing the difference-value distribution image into multiple window regions of size 3×3; Extract the maximum relative pixel difference value in each window region, and arrange the maximum relative pixel difference values ​​according to the window region position to obtain a high difference value compressed image; Extract the minimum relative pixel difference value in each window region, and arrange the minimum relative pixel difference values ​​according to the window region position to obtain a low difference value compressed image.

5. The intelligent inspection method for construction sites according to claim 1, characterized in that, The calculation of difference feature values ​​for each grid in both high-difference-value compressed images and low-difference-value compressed images includes the following steps: The total difference value is obtained by summing the relative pixel differences of all outlier points within the grid. The ratio of the relative pixel difference value of each outlier to the total difference value is used as the weight of each outlier. Multiply the weight of each outlier by the relative pixel difference value of that outlier; The weighted difference value is obtained by summing the results of multiplying the outliers; The ratio of the number of outliers in the grid to the total number of pixels in the grid is used as the density factor; Multiplying the weighted difference value by the density factor yields the difference characteristic value of the grid.

6. The intelligent inspection method for construction sites according to claim 1, characterized in that, Arrange the difference feature values ​​on the high difference value compressed image according to the grid order on the corresponding image to obtain the high difference feature matrix; Arrange the differential feature values ​​on the low-difference compressed image according to the grid order on the corresponding image to obtain the low-difference feature matrix; The calculation of joint bias for the difference feature values ​​of each grid in both the high-difference-value compressed image and the low-difference-value compressed image includes the following steps: The global mean is obtained by averaging the difference feature values ​​of each grid belonging to an image. Calculate the global bias based on the difference between the differential feature value of each grid and the corresponding global mean; Taking each grid as the center, the average value of the difference characteristic values ​​of the neighboring grids of the center is taken to obtain the neighborhood mean; The neighborhood bias is calculated based on the difference between the differential feature value of each grid and the neighborhood mean. The global bias and the neighborhood bias are added together to obtain the joint bias.

7. The intelligent inspection method for construction sites according to claim 6, characterized in that, Arrange the joint bias values ​​on the high-discrepancy compressed image according to the grid order on the corresponding image to obtain the high joint bias matrix; Arrange the joint bias values ​​on the low-discrepancy compressed image according to the grid order on the corresponding image to obtain the low joint bias matrix; The process of calculating the global deviation includes: calculating the difference between the difference feature value of each grid and the global mean, taking the absolute value of the difference to obtain the global absolute difference, and dividing the global absolute difference by the global mean to obtain the global deviation. The process of calculating the neighborhood deviation includes: calculating the difference between the difference feature value of each grid and the neighborhood mean, taking the absolute value of the difference to obtain the absolute neighborhood difference, and dividing the absolute neighborhood difference by the neighborhood mean to obtain the neighborhood deviation.

8. An intelligent inspection system for construction sites, characterized in that, The intelligent inspection method for construction sites based on any one of claims 1 to 7 includes: a road surface pixel acquisition subsystem, a difference value distribution image construction subsystem, a compressed image construction subsystem, a first matrix construction subsystem, a second matrix construction subsystem, and an evaluation subsystem; The road surface pixel acquisition subsystem is used to locate the road surface pixel region on remote sensing images of cement road surfaces at construction sites; The difference value distribution image construction subsystem is used to calculate the relative difference value of each pixel based on the pixel mean of the road surface pixel region, and obtain the difference value distribution image; The compressed image construction subsystem is used to extract the maximum and minimum values ​​in each window region of the difference value distribution image to obtain high difference value compressed images and low difference value compressed images. The first matrix construction subsystem is used to mark all pixels except the road surface pixel area as outliers, divide the high difference value compressed image and the low difference value compressed image into grids, and calculate the difference feature value for each grid based on the outliers in the grid to obtain the high difference feature matrix and the low difference feature matrix. The second matrix construction subsystem is used to calculate the joint bias degree for the difference feature values ​​of each grid on the high difference value compressed image and the low difference value compressed image, respectively, to obtain the high joint bias matrix and the low joint bias matrix. The evaluation subsystem is used to process the high-discrepancy feature matrix, low-discrepancy feature matrix, high-joint-bias matrix and low-joint-bias matrix using the pavement defect feature evaluation network to obtain the pavement quality score.

Images