Earthwork backfill construction backfill quality detection system based on image recognition

Abstract

The application relates to the technical field of image recognition, and discloses a soil backfilling construction backfilling soil quality detection system based on image recognition; the system comprises an image tiling module, which is used for constructing outer tiling images and inner tiling images; a first extraction module, which is used for extracting common features; a second extraction module, which is used for collecting outer unique features and inner unique features; a parameter identification module, which is used for identifying quality parameters; and a quality analysis module, which is used for determining the backfilling grades and prompt information of detection samples; the application realizes the effect of full-range and seamless splicing imaging of the external vertical surface and the internal cutting surface of the backfilling soil sample, avoids the information loss problem caused by improper shooting angles or splicing, avoids the misjudgment phenomenon that the internal quality problem is ignored due to the flat surface of the detection sample, can provide diversified data support for backfilling soil quality detection from two different dimensions of common dimensions and individual dimensions, and improves the accuracy of backfilling soil quality detection results.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically, to an image recognition-based system for detecting the quality of backfill soil during earthwork backfilling construction. Background Technology

[0002] During earthwork backfilling construction, the quality of the backfill soil directly affects the bearing capacity of the foundation, settlement control, and the safety of subsequent structures. Commonly used backfill soil quality evaluation indicators include compaction degree, moisture content, and uniformity. With the rapid development of smart construction sites and digital construction management, the quality control of earthwork backfilling projects is gradually shifting towards intelligence and data-driven approaches. Therefore, solutions for backfill soil quality inspection based on image recognition technology are also beginning to be widely used.

[0003] Reference patent application CN117670764A discloses a method, apparatus, and equipment for backfill soil quality inspection based on image data analysis. The method includes acquiring an image of backfill soil corresponding to a gas pipeline and determining whether the backfill soil image contains location data corresponding to the gas pipeline. When the backfill soil image contains location data corresponding to the gas pipeline, a longitudinal section of the backfill soil is cropped from the backfill soil image of the gas pipeline based on the location data of the gas pipeline to obtain a longitudinal section image of the backfill soil. Based on threshold binarization image processing, the longitudinal section image of the backfill soil is segmented to obtain the segmented image and corresponding image data. A connectivity threshold is determined for the segmented image, and the connectivity threshold is compared with a preset threshold. Based on the comparison result, it is determined whether the segmented image is a qualified image, so that the quality of the backfill soil for the gas pipeline can be inspected based on the comparison result. In current methods of quality testing for backfill soil, soil samples are typically collected and images of the exterior facade are taken from a single perspective. The quality of the backfill soil is then analyzed based on the information in the facade images. This single-perspective image recognition and testing method is prone to the loss of some effective information due to the limitations of the shooting angle, which limits subsequent quality testing operations. Furthermore, it is difficult to detect internal quality defects in the tested samples, which can easily lead to misjudgments where the surface quality is acceptable but the internal quality is unacceptable, thereby reducing the accuracy of the backfill soil quality testing results.

[0004] In view of this, the present invention proposes an image recognition-based backfill soil quality detection system to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a backfill soil quality detection system based on image recognition for earthwork backfilling construction, comprising: The image tiling module is used to collect test samples of backfill soil, construct an outer tiling image using the ring tiling rule, and construct an inner tiling image using the axial tiling rule. The first extraction module is used to extract common features from the outer and inner tiling images. The common features include thickness compliance rate, crack area ratio, brightness abrupt change rate, and particle synchronization rate. The second extraction module is used to determine the outer feature region and the inner feature region in the outer tiled image and the inner tiled image, and to collect the outer unique features of the outer feature region and the inner unique features of the inner feature region respectively. The parameter identification module is used to combine common features, external unique features and internal unique features into comprehensive quality data, and to identify quality parameters through the backfill quality identification model. The quality parameters include compaction degree, moisture content and uniformity. The quality analysis module is used to perform combined analysis of quality parameters through a dual analysis mechanism to determine the backfill level of the test sample. The backfill level includes a high quality level and a low quality level, and corresponding prompts are generated.

[0006] Furthermore, the circular tiling rule is: first determine the baseline, then construct the tiling surface; The steps for constructing an outer tiled image are as follows: S11: Place the test sample perpendicular to the horizontal plane and determine the vertically downward axis of the test sample on the outer facade; S12: Adjust the scan line of the line scan camera to be parallel to the axis of the sample being tested, and control the line scan camera to circle around the outer facade of the sample once to obtain the video of the outer facade. S13: Using the calibrated overlap rate as the standard, extract A frames from the facade video and mark the overlapping areas where the overlap rate is located in the frames; S14: Using the first frame image as the reference image, delete the overlapping areas in the remaining A-1 frame images, and stitch and merge the A frame images in sequence according to the shooting time to generate an outer tiled image.

[0007] Furthermore, the axial tiling rule states that the same subsample is only spliced ​​and merged once; The steps for constructing an inner tiled image are as follows: S21: Divide the test sample into four equal parts along the axial direction into four sub-samples, and in a clockwise direction, record the two axial cutting surfaces of the sub-samples as the mother cutting surface and the sub-cutting surface, respectively. S22: Take pictures of the parent cutting surface of the four sub-samples one by one using a line scan camera to obtain four sub-sample images, and number the four sub-sample images in ascending order in a clockwise direction. S23: Align the edges of adjacent sub-sample images according to their numbers in ascending order, and then merge the aligned sub-sample images to generate an inner tiled image.

[0008] Furthermore, the steps for extracting the thickness compliance rate are as follows: S31: Convert the outer tiled image into a grayscale image, and set the scaling factor of the grayscale image according to the ratio of the image height of the grayscale image to the actual height of the detection sample; S32: Based on the global grayscale histogram of the grayscale image, the low threshold and high threshold are calculated by the Otsu algorithm, and all grayscale edge lines are extracted from the grayscale image by combining the Canny edge detection technology. S33: Identify straight lines from grayscale edge lines using Hough line transform, set the baseline length and baseline angle range, and record the straight lines whose length is greater than the baseline length and whose tilt angle is within the baseline angle range as valid straight lines. Draw a horizontal line through the midpoint of the valid straight lines to create candidate lines. S34: The regions located one height cell above the candidate line and one height cell below the candidate line are respectively called the upper region and the lower region. The gray values ​​of all points in the upper region and the lower region are summed and the average is calculated. The difference between the two averages is then calculated to obtain the gray value difference. S35: Candidate lines with a grayscale difference greater than the calibrated grayscale difference threshold are recorded as layer lines. The absolute value of the difference between the ordinates of two adjacent layer lines is multiplied by the scaling factor to calculate the layer thickness value. S36: Divide the layer thickness value by the calibrated thickness threshold to calculate the sub-compliance rate, and then sum all the sub-compliance rates and average them to calculate the thickness compliance rate.

[0009] Furthermore, the extraction steps for particle synchronization rate are as follows: S41: Convert the grayscale image into a binary image using an adaptive thresholding method, and denote the points in the binary image as white pixels and black pixels respectively; S42: Through Euclidean distance transformation, calculate the Euclidean distance from white pixel to black pixel one by one, summarize the minimum value of all Euclidean distances to generate a distance map, and record the areas in the distance map where the Euclidean distance is greater than the calibrated distance threshold as the confidence area. S43: The confidence region is subjected to a combination operation of first dilation and then erosion using the morphological closing operation method. The confidence regions where adjacent particles are located are removed, and the remaining confidence regions are recorded as particle regions. The particle contours are extracted from the particle regions using the watershed algorithm. S44: Draw the smallest circumcircle on each particle outline one by one, sum the areas of all the smallest circumcircles and calculate the average area, and record the particle outlines whose smallest circumcircle area is greater than the average area as large particle entities. S45: Measure the point-to-point distance between any two points on the large particle entity one by one, and connect the two points where the maximum point-to-point distance value is located to generate the long axis line; S46: Using the horizontal direction as the reference plane, measure the angle between the long axis and the reference plane. Record the long axis with the angle within the synchronization angle range as the target line. Divide the number of target lines by the number of long axes to calculate the particle synchronization rate.

[0010] Furthermore, when determining the outer feature region, two auxiliary lines are drawn vertically in the outer tiled image. The area between the two auxiliary lines is recorded as the first inner region, and the area outside the two auxiliary lines is recorded as the first outer region. The span of the first inner region and the first outer region is measured horizontally. The positions of the two auxiliary lines are moved horizontally until the ratio of the span of the first inner region to the span of the two first outer regions is 6:2:2. The first inner region is then recorded as the outer feature region.

[0011] Furthermore, when determining the inner feature region, two auxiliary lines are drawn horizontally in the inner tiled image. The area between the two auxiliary lines is recorded as the second inner region, and the area outside the two auxiliary lines is recorded as the second outer region. The span of the second inner region and the second outer region is measured vertically. The positions of the two auxiliary lines are moved vertically until the ratio of the span of the second inner region to the span of the two second outer regions is 6:2:2. The second inner region is then recorded as the inner feature region.

[0012] Furthermore, the unique characteristics include the proportion of large particles and the proportion of particle area. The unique features include the proportion of compliant circumcircle and the proportion of circumcircle area.

[0013] Furthermore, the dual analysis mechanism is as follows: first, a single parameter analysis is performed, and then a summary analysis of the parameters is performed; The steps for determining the backfill level are as follows: S51: Subtract the compaction degree, moisture content, and uniformity from the preset compaction threshold, preset moisture content threshold, and preset uniformity threshold respectively, and take the absolute value to obtain the compaction difference, moisture content difference, and uniformity difference. S52: When the compaction difference is greater than the upper limit of the compaction difference, the compaction degree is recorded as a low-quality parameter; when the moisture content difference is greater than the upper limit of the moisture content difference, the moisture content is recorded as a low-quality parameter; when the uniformity difference is greater than the upper limit of the uniformity difference, the uniformity is recorded as a low-quality parameter. S53: After dividing the compaction degree, moisture content and uniformity by the preset compaction threshold, preset moisture content threshold and preset uniformity threshold respectively, calculate the compaction ratio, moisture content ratio and uniformity ratio, and then add the compaction ratio, moisture content ratio and uniformity ratio together and calculate the average to obtain the mass ratio. S54: If the number of low-quality parameters in the test sample is less than or equal to 1 and the quality ratio is greater than or equal to 0.9, the backfill quality will be determined as high quality; if the number of low-quality parameters in the test sample is greater than 1 or the quality ratio is less than 0.9, the backfill quality will be determined as low quality.

[0014] Furthermore, if the backfill grade is high quality, no prompt message will be provided; if the backfill grade is low quality, the compaction ratio, moisture content ratio, and uniformity ratio will be compared. When the compaction ratio is greater than 1, information on reducing compaction degree is formulated; when the compaction ratio is less than 1, information on increasing compaction degree is formulated. When the moisture content is greater than 1, information on reducing the moisture content is formulated; when the moisture content is less than 1, information on increasing the moisture content is formulated. When the uniformity ratio is greater than 1, information to decrease uniformity is formulated; when the uniformity ratio is less than 1, information to increase uniformity is formulated.

[0015] The technical advantages of the image recognition-based backfill soil quality detection system of this invention are as follows: (1): This invention constructs an outer tiling image using the ring tiling rule and an inner tiling image using the axial tiling rule, thereby achieving a comprehensive and seamless splicing imaging effect of the external facade and internal cut surface of the backfill soil sample. This dual detection mechanism from the surface to the inside effectively avoids the problem of information loss caused by improper shooting angle or splicing, and also avoids the misjudgment phenomenon of ignoring internal quality problems due to the flatness of the sample surface. Thus, it breaks through the limitations of single-view quality detection operation and improves the accuracy of backfill soil quality detection results.

[0016] (2): By combining common features, external unique features and internal unique features, this invention can provide diversified data support for backfill soil quality testing from two different dimensions: common dimension and individual dimension. This avoids the problem of inaccurate test results caused by single-dimensional data in identification and analysis, and ensures that the data involved in quality parameter identification can fully identify the multidimensional features of the test sample, thereby further improving the accuracy of quality parameter identification results. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a backfill soil quality detection system based on image recognition provided in Embodiment 1 of the present invention. Figure 2 This is a logical diagram of an image recognition-based backfill soil quality detection system for earthwork backfilling construction, provided in Embodiment 1 of the present invention. Figure 3 This is a flowchart illustrating a method for detecting the quality of backfill soil during earthwork backfilling construction based on image recognition, as provided in Embodiment 2 of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Please refer to Figures 1-2 As shown in this embodiment, an image recognition-based backfill soil quality detection system for earthwork backfilling construction includes: The image tiling module collects test samples of backfill soil, constructs an outer tiling image of the test sample using the ring tiling rule, and constructs an inner tiling image of the test sample using the axial tangent tiling rule. The test samples are backfill soil samples collected from the earthwork backfilling construction area using soil sampling equipment, and are used as entity objects for subsequent image recognition and detection. In this embodiment, when collecting test samples from the earthwork backfilling construction area, a ring cutter soil sampler is used as the collection device, and samples are collected vertically from the earthwork backfilling construction area, so that the collected test samples are cylindrical in shape.

[0020] To ensure the representativeness of the test samples and prevent limitations in the test results due to excessively small sample volume and area, the upper surface radius of the test samples must be sufficiently large; for example, the upper surface radius of the test samples is 20 cm.

[0021] As the collected test samples are physical structures, the quality of backfill soil cannot be quickly detected from the image level. In order to meet the quality detection effect of image recognition, it is necessary to take pictures of the test samples from multiple angles and positions using camera equipment to obtain the outer and inner tiling images for image recognition and quality detection.

[0022] An external tiled image refers to an image obtained by stitching together images taken along the circumferential exterior facade of the test sample, which can represent the overall circumferential exterior facade of the test sample in a planar manner; In order to ensure the orderliness of the construction process and the accuracy of the construction results when constructing an outer tiled image, it is necessary to use the circular tiling rule for construction. Specifically, the circular tiling rule is as follows: first determine the baseline, and then construct the tiling surface.

[0023] The steps for constructing an outer tiled image are as follows: S11: Place the test sample perpendicular to the horizontal plane and determine the vertically downward axis of the test sample on the outer surface; the axis is a straight line parallel to the axis of the test sample. Since the test sample is a cylindrical structure, the axis is a straight line. S12: Adjust the scan line of the line scan camera to be parallel to the axis of the sample being tested, and control the line scan camera to circle around the outer facade of the sample once to obtain the video of the outer facade. S13: Using the calibrated overlap rate as a standard, extract A frames from the facade video and mark the overlapping areas in each frame. The calibrated overlap rate refers to the proportion of the overlapping area between the next frame and the previous frame to ensure that adjacent frames remain connected and to lay the foundation for stitching together subsequent frames. For example, the calibrated overlap rate is 5%, meaning that the area of ​​the overlapping region between the next frame and the previous frame is 5%. S14: Using the first frame image as the reference image, delete the overlapping areas in the remaining A-1 frame images, and stitch and merge the A frame images in sequence according to the shooting time to generate an outer tiled image.

[0024] It should be noted that the external tiled image can provide an image basis for judging and detecting the quality of backfill soil in the test sample from the external exposed layer. By first overlapping and then deleting the overlapping area, it can be ensured that there is an overlapping image area between any two adjacent frames. This ensures that when two adjacent frames are stitched together, the stitching point can maintain a seamless stitching effect and avoid the phenomenon of image loss at the stitching point.

[0025] An inner tiled image is an image obtained by stitching together the images of the cut positions after dividing the sample into four equal parts along its axial direction. This allows for a complete image representation of the axial cut surface of the sample. When constructing an inner tiled image, it is necessary to use the axial tiling rule to ensure the integrity and continuity of the inner tiled image. Specifically, the axial tiling rule is as follows: the same subsample is only stitched and merged once; this ensures that the images participating in the axial stitching and merging come from all subsamples, guaranteeing the reliability and comprehensiveness of the inner tiling image construction results.

[0026] The steps for constructing an inner tiled image are as follows: S21: Divide the test sample into four equal parts along the axial direction into four sub-samples, and in a clockwise direction, record the two axial cutting surfaces of the sub-samples as the mother cutting surface and the sub-cutting surface, respectively. S22: Take pictures of the parent cutting surface of the four sub-samples one by one using a line scan camera to obtain four sub-sample images, and number the four sub-sample images in ascending order in a clockwise direction. S23: Align the edges of adjacent sub-sample images according to their numbers in ascending order, and then merge the aligned sub-sample images to generate an inner tiled image.

[0027] It should be noted that the inner tiled image can provide image basis for judging and detecting the quality of backfill soil in the test sample from the internal cutting layer. The combination of the outer tiled image and the inner tiled image achieves the combined internal and external quality detection effect of backfill soil quality.

[0028] The outer and inner tiling images obtained above cannot be directly used as the objects for subsequent backfill soil quality inspection. Instead, they need to undergo preprocessing operations such as denoising and filtering to eliminate the interference caused by negative effects such as external environment and camera vibration on the overall image quality, thus providing accurate image basis for image recognition.

[0029] The first extraction module extracts common features from the outer and inner tiled images. The common features include thickness compliance rate, crack area ratio, brightness abrupt change rate, and particle synchronization rate. Since both the outer and inner tiled images are image representations of the detection samples, both the outer and inner tiled images contain common features that can represent the quality of the detection samples in multiple dimensions. Specifically, common characteristics include thickness compliance rate, proportion of cracked areas, brightness abrupt change rate, and particle synchronization rate.

[0030] In this embodiment, since common features refer to common features that can represent the quality of the detected sample from multiple different dimensions from the perspective of image vision, the methods and steps for extracting common features in the outer tiled image and the inner tiled image are the same.

[0031] Thickness compliance rate refers to the proportion of soil thicknesses corresponding to different layers of backfill soil in the outer and inner tiled images that meet the preset thickness threshold. When the thickness compliance rate is closer to the preset layer thickness, it indicates that the quality of the backfill soil in the test sample is higher. The steps for extracting the thickness compliance rate are as follows: S31: Convert the outer tiled image to a grayscale image, and set the scaling factor of the grayscale image according to the ratio of the image height of the grayscale image to the actual height of the detection sample; the scaling factor refers to the scaling ratio between the grayscale image and the detection sample. S32: Based on the global grayscale histogram of the grayscale image, the low threshold and high threshold are calculated by the Otsu algorithm, and all grayscale edge lines are extracted from the grayscale image by combining the Canny edge detection technology. S33: Identify straight lines from grayscale edge lines using Hough line transform, set a baseline length and a baseline angle range, and record straight lines whose length is greater than the baseline length and whose tilt angle is within the baseline angle range as valid straight lines. Draw a horizontal line through the midpoint of the valid straight lines to create candidate lines. The baseline length refers to the minimum length of the layer line between different layers of backfill soil, and the baseline angle range refers to the tilt angle range of the layer line between different layers of backfill soil. For example, the baseline length is 60% of the actual layer line length between different layers in the test sample, and the baseline angle range is (-15 degrees, 20 degrees), that is, tilting downwards by 15 degrees to tilting upwards by 20 degrees. S34: The regions located one height cell above the candidate line and one height cell below the candidate line are respectively called the upper region and the lower region. The gray values ​​of all points in the upper region and the lower region are summed and the average is calculated. The difference between the two averages is then calculated to obtain the gray value difference. S35: Candidate lines with a grayscale difference greater than the calibrated grayscale difference threshold are designated as layer lines. The absolute value of the difference between the ordinates of two adjacent layer lines is multiplied by the scaling factor to calculate the layer thickness value. The calibrated grayscale difference threshold refers to the minimum grayscale difference of the layer lines. The calibrated grayscale difference threshold is obtained by averaging historical data. S36: Divide the layer thickness value by the calibrated thickness threshold to calculate the sub-compliance rate, and then sum all the sub-compliance rates and average them to calculate the thickness compliance rate. The calibrated thickness threshold refers to the actual thickness of each layer in the test sample, which can be found in the technical parameter table of the backfill construction operation.

[0032] The crack area ratio refers to the proportion of the area of ​​unfilled areas such as cracks, voids, and gaps in the outer and inner tiled images to the total area. The larger the crack area ratio, the lower the quality of the backfill soil in the test sample. In this embodiment, when extracting the crack region ratio, points in the grayscale image with grayscale values ​​greater than the calibrated crack grayscale threshold are recorded as crack points. A combination operation of dilation and erosion is performed on the crack points using a morphological closing operation method to connect discrete crack points in the same crack region to generate a crack region. The area and aspect ratio of each crack region are measured one by one, and crack regions with an area greater than 3% of the grayscale image area and an aspect ratio greater than 2 are recorded as valid regions. The areas of all valid regions are added together and compared with the area of ​​the grayscale image to calculate the crack region ratio.

[0033] Brightness abrupt change rate refers to the ratio between points with large changes in brightness and darkness in the outer and inner tiled images and the total number of points. The higher the brightness abrupt change rate, the lower the quality of the backfill soil in the test sample. In this embodiment, when extracting the brightness mutation rate, the horizontal and vertical gradients of all points in the grayscale image are calculated one by one using the Sobel operator. The gradient magnitude of all points is calculated using the magnitude calculation formula. Points with gradient magnitudes greater than the calibrated gradient threshold are recorded as mutation points. The brightness mutation rate is calculated by dividing the number of mutation points in the grayscale image by the total number of points.

[0034] It should be noted that the gradient threshold calibrated above is actually recorded as the minimum gradient threshold at the point of mutation, thus providing a precise numerical basis for the identification of mutation points.

[0035] Particle synchronization rate refers to the degree to which the long axis of large particles in the outer and inner tiled images maintains a consistent horizontal distribution. The higher the particle synchronization rate, the higher the quality of the backfill soil in the sample. The steps for extracting particle synchronization rate are as follows: S41: Convert the grayscale image into a binary image using an adaptive thresholding method, and denote the points in the binary image as white pixels and black pixels respectively; S42: Through Euclidean distance transformation, calculate the Euclidean distance from white pixel to black pixel one by one, summarize the minimum value of all Euclidean distances to generate a distance map, and record the regions in the distance map where the Euclidean distance is greater than the calibrated distance threshold as confidence regions; the calibrated distance threshold refers to the minimum value of Euclidean distance when recorded as a confidence region, thereby setting a threshold for the recognition of confidence regions. S43: The confidence region is subjected to a combination operation of first dilation and then erosion using the morphological closing operation method. The confidence regions where adjacent particles are located are removed, and the remaining confidence regions are recorded as particle regions. The particle contours are extracted from the particle regions using the watershed algorithm. S44: Draw the smallest circumcircle on each particle outline one by one, sum the areas of all the smallest circumcircles and calculate the average area, and record the particle outlines whose smallest circumcircle area is greater than the average area as large particle entities. S45: Measure the point-to-point distance between any two points on the large particle entity one by one, and connect the two points where the maximum point-to-point distance value is located to generate the long axis line; S46: Using the horizontal direction as the reference plane, measure the angle between the long axis and the reference plane. Record the long axes whose angle falls within the synchronization angle range as target lines. Divide the number of target lines by the number of long axes to calculate the particle synchronization rate. The synchronization angle range refers to the range of angles between the long axis recorded as the target line and the reference plane, ensuring that all target lines maintain relatively close tilt angles.

[0036] It should be noted that the above content only provides the method and steps for extracting the thickness compliance rate, crack area ratio, brightness abrupt change rate, and particle synchronization rate from the outer tiled image. When extracting the thickness compliance rate, crack area ratio, brightness abrupt change rate, and particle synchronization rate from the inner tiled image, the same steps as those in the outer tiled image are also used.

[0037] The second extraction module determines the feature regions in the outer tiled image and the inner tiled image, extracts the outer unique features from the feature regions of the outer tiled image, and extracts the inner unique features from the feature regions of the inner tiled image. Feature regions refer to regions that can capture unique features from both the outer and inner tiled images, enabling them to reflect the different characteristics of the test sample in two different image dimensions and providing a basis for subsequent backfill quality testing. Specifically, the feature region includes the outer feature region and the inner feature region.

[0038] Because the angles corresponding to the outer and inner tiled images are not the same, the specific locations of the feature regions in the outer and inner tiled images are not the same.

[0039] In this embodiment, when determining the outer feature region in the outer tiled image, firstly, two auxiliary lines are drawn vertically in the outer tiled image. The region between the two auxiliary lines is recorded as the first inner region, and the region outside the two auxiliary lines is recorded as the first outer region. Then, the region span of the first inner region and the first outer region is measured one by one along the horizontal direction. Finally, the relative positions of the two auxiliary lines are continuously moved along the horizontal direction until the ratio of the region span of the first inner region to the region span of the two first outer regions is 6:2:2, and the first inner region is recorded as the outer feature region.

[0040] In this embodiment, when determining the inner feature region in the inner tiled image, firstly, two auxiliary lines are drawn in the inner tiled image along the horizontal direction. The area between the two auxiliary lines is recorded as the second inner region, and the area outside the two auxiliary lines is recorded as the second outer region. Then, the area span of the second inner region and the second outer region are measured one by one along the vertical direction. Finally, the relative positions of the two auxiliary lines are continuously moved along the vertical direction until the ratio of the area span of the second inner region to the area span of the two second outer regions is 6:2:2, and the second inner region is recorded as the inner feature region.

[0041] It should be noted that the feature regions in both the outer and inner tiled images are rectangular structures.

[0042] External unique features are personalized features used in the external feature region to visually represent the exterior morphology and structure of the test sample. This allows external unique features to be distinguished from the universality of common features and can only provide a basis for testing the quality of backfill soil from the visual perspective of the exterior. Specifically, the unique characteristics include the proportion of large particles and the proportion of particle area.

[0043] In this embodiment, the proportion of large particles refers to the ratio between the number of large particle entities in the outer feature region and the number of all particle outlines. Specifically, based on the above-mentioned particle synchronization rate step, large particle entities and particle outlines are identified from the outer feature region, and the proportion of large particles is calculated after comparing the number of large particle entities with the number of particle outlines.

[0044] In this embodiment, the particle area ratio refers to the ratio between the area covered by large particles in the outer feature region and the area of ​​the outer feature region; specifically, the particle area ratio is calculated by comparing the area of ​​the region where the large particles are located with the area of ​​the outer feature region.

[0045] Unique features are personalized features used in the inner feature region to visually represent the morphological structure of the inner surface of the test sample. This allows unique features to be distinguished from the universality of common features and can only provide a basis for the detection of backfill quality from the visual perspective of the inner surface. Specifically, the unique features include the proportion of compliant circumcircle and the proportion of circumcircle area.

[0046] In this embodiment, the proportion of compliant circumscribed circles refers to the ratio of the number of large particle entities in the inner feature region whose smallest circumscribed circle is located inside the inner feature region and the area of ​​more than half of the total area is within the inner feature region. Specifically, the radius of the minimum circumcircle of all large particles in the inner feature region is measured one by one. The rated area of ​​the minimum circumcircle is calculated using the circle area formula. The area within the domain of the minimum circumcircle in the inner feature region is measured. The minimum circumcircle whose area within the domain exceeds half of the rated area is recorded as a compliant circumcircle. The number of compliant circumcircles is compared with the total number of minimum circumcircles to calculate the proportion of compliant circumcircles.

[0047] In this embodiment, the circumcircle area ratio refers to the ratio between the area of ​​the compliant circumcircle and the area of ​​the inner feature region. Specifically, the area of ​​all compliant circumcircles within their domains is added together and compared with the area of ​​the inner feature region to calculate the circumcircle area ratio.

[0048] The parameter identification module summarizes common features, external unique features, and internal unique features into comprehensive quality data, and inputs the comprehensive quality data into the backfill quality identification model to identify the quality parameters of the test sample; After obtaining the common features, external unique features, and internal unique features, they need to be summarized and integrated to form a comprehensive quality data with a complete structure, so that the comprehensive quality data can serve as a direct basis for subsequent testing of the backfill soil quality of the test samples. Specifically, the data is summarized in the order of common features of the outer tiled images, unique features of the outer tiles, common features of the inner tiled images, and unique features of the inner tiles to form complete comprehensive quality data.

[0049] The backfill quality identification model is an artificial intelligence model based on deep learning technology, taking comprehensive quality data as input and quality parameters as output. It ensures that the backfill quality identification model can perform deep learning on the data features identified from the images in the test samples and output the corresponding results, which can serve as the theoretical basis for backfill soil quality testing.

[0050] Quality parameters refer to the results output by the backfill quality identification model, which can provide diversified analytical basis for the quality of backfill soil in the test samples. Specifically, the quality parameters include compaction, moisture content, and uniformity. Compaction directly reflects the degree to which the backfill soil is rolled tightly; the closer the compaction is to the preset compaction, the higher the quality of the backfill soil in the test sample. Moisture content refers to the ratio of the mass of water to the mass of soil particles in the backfill soil; the closer the moisture content is to the preset moisture content, the higher the quality of the backfill soil in the test sample. Uniformity refers to the degree to which the texture and color of the backfill soil are evenly distributed; the closer the uniformity is to the preset uniformity, the higher the quality of the backfill soil in the test sample.

[0051] The backfill quality identification model is not an existing technology and needs to be trained using deep learning techniques. Specifically, the training method for the backfill quality identification model is as follows: A large amount of comprehensive quality data and corresponding quality parameters of test samples are collected in advance. A set of comprehensive quality data and quality parameters are bound into a dataset, and the data samples are divided into training set and test set in a 7:3 ratio. Set the learning rate, number of iterations, and batch size of the deep learning model, and use cross-validation to iteratively optimize the fitting performance of the deep learning model; During the iteration process, the deep learning model extracts compaction data, moisture content data, and uniformity data from the comprehensive quality data, and performs numerical extraction on the compaction data, moisture content data, and uniformity data to output compaction degree, moisture content, and uniformity. The deep learning model is trained using a training set and tested using a test set until the accuracy of the deep learning model's predictions reaches a preset accuracy threshold, thus obtaining the backfill quality identification model.

[0052] In this embodiment, the common features, unique features, common features, and unique features of the outer tiled images collected in this test sample are summarized into comprehensive quality data and input into the backfill quality recognition model, which can then output the required quality parameters.

[0053] The quality analysis module uses a dual analysis mechanism to combine and analyze quality parameters, determine the backfill grade of the test sample, and generate prompts for earthwork backfilling construction. Backfill grade is used to represent the quality of backfill soil corresponding to the detected sample in this image recognition. Different backfill soil qualities correspond to different backfill grades. Specifically, backfill grades include high-quality grades and low-quality grades.

[0054] When analyzing and determining the backfill level, it is necessary to perform a dual-combination judgment analysis on the quality parameters. This allows for a two-dimensional combined analysis of the backfill level of the test sample. Therefore, a dual analysis mechanism is required for the combined analysis. Specifically, the dual analysis mechanism involves first performing individual parameter analysis, followed by a summary analysis of the parameters. This allows for a shift from individual, independent analysis of quality parameters to a combined, holistic analysis, thereby improving the accuracy of backfill level analysis results.

[0055] The steps for determining the backfill level are as follows: S51: The compaction degree, moisture content, and uniformity are respectively subtracted from the preset compaction threshold, preset moisture content threshold, and preset uniformity threshold, and the absolute values ​​are taken to obtain the compaction difference, moisture content difference, and uniformity difference. The preset compaction threshold, preset moisture content threshold, and preset uniformity threshold refer to the compaction degree, moisture content, and uniformity set during earthwork backfilling construction, which can provide standard numerical basis for subsequent backfill quality testing. S52: When the compaction difference is greater than the upper limit of the compaction difference, the compaction degree is recorded as a low-quality parameter; when the moisture content difference is greater than the upper limit of the moisture content difference, the moisture content is recorded as a low-quality parameter; when the uniformity difference is greater than the upper limit of the uniformity difference, the uniformity is recorded as a low-quality parameter. S53: After dividing the compaction degree, moisture content and uniformity by the preset compaction threshold, preset moisture content threshold and preset uniformity threshold respectively, calculate the compaction ratio, moisture content ratio and uniformity ratio, and then add the compaction ratio, moisture content ratio and uniformity ratio together and calculate the average to obtain the mass ratio. S54: If the number of low-quality parameters in the test sample is less than or equal to 1 and the mass ratio is greater than or equal to 0.9, it indicates that the backfill soil quality of the test sample is high, and the backfill quality is determined to be of high quality grade; if the number of low-quality parameters in the test sample is greater than 1 or the mass ratio is less than 0.9, it indicates that the backfill soil quality of the test sample is low, and the backfill quality is determined to be of low quality grade.

[0056] Once the backfill level of the test sample is obtained, reasonable and targeted prompts can be provided for earthwork backfilling operations based on the specific content corresponding to the backfill level, thereby improving the construction quality and efficiency of earthwork backfilling. When developing guidance information for earthwork backfilling construction, it should be based on the backfilling grade and quality parameters; Specifically, when formulating prompt information, if the backfill grade is high quality, then no prompt information is required. If the backfill grade is low quality grade, compare the compaction ratio, moisture content ratio and uniformity ratio. When the compaction ratio is greater than 1, information is provided to reduce the compaction degree; when the compaction ratio is less than 1, information is provided to increase the compaction degree; when it is equal to 1, no prompt information is provided. When the moisture content is greater than 1, a message is generated to decrease the moisture content; when the moisture content is less than 1, a message is generated to increase the moisture content; when the moisture content is equal to 1, no message is generated. When the uniformity ratio is greater than 1, a message to decrease the uniformity is generated; when the uniformity ratio is less than 1, a message to increase the uniformity is generated; when it is equal to 1, no message is generated.

[0057] It should be noted that once the corresponding prompts are provided, the construction personnel for earthwork backfilling can perform the corresponding construction operations according to the specific content of the prompts. The construction operations include, but are not limited to, increasing or decreasing the compaction pressure of the backfill soil, increasing or decreasing the amount of water sprayed on the backfill soil, etc., thereby improving the overall quality of the backfill soil in earthwork backfilling construction.

[0058] Example 2: Please refer to Figure 3As shown, the parts not described in detail in this embodiment are described in Embodiment 1. This embodiment provides a method for detecting the quality of backfill soil during earthwork backfilling construction based on image recognition. This method is implemented using an image recognition-based earthwork backfilling construction backfill soil quality detection system, including: Step 1: Collect test samples of backfill soil, construct an outer tiling image using the annular tiling rule, and construct an inner tiling image using the axial tiling rule; Step 2: Extract common features from the outer and inner tiled images; Step 3: Determine the outer feature region and inner feature region in the outer tiled image and the inner tiled image, and collect the outer unique features of the outer feature region and the inner unique features of the inner feature region respectively; Step 4: Combine common features, external unique features, and internal unique features into comprehensive quality data, and identify quality parameters through the backfill quality identification model; Step 5: Combine and analyze the quality parameters through a dual analysis mechanism to determine the backfill level of the test sample and formulate corresponding prompts.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A backfill soil quality inspection system based on image recognition during earthwork backfilling construction, characterized in that, include: The image tiling module is used to collect test samples of backfill soil, construct an outer tiling image using the ring tiling rule, and construct an inner tiling image using the axial tiling rule. The first extraction module is used to extract common features from the outer and inner tiling images. The common features include thickness compliance rate, crack area ratio, brightness abrupt change rate, and particle synchronization rate. The second extraction module is used to determine the outer feature region and the inner feature region in the outer tiled image and the inner tiled image, and to collect the outer unique features of the outer feature region and the inner unique features of the inner feature region respectively. The parameter identification module is used to combine common features, external unique features and internal unique features into comprehensive quality data, and to identify quality parameters through the backfill quality identification model. The quality parameters include compaction degree, moisture content and uniformity. The quality analysis module is used to perform combined analysis of quality parameters through a dual analysis mechanism to determine the backfill level of the test sample. The backfill level includes a high quality level and a low quality level, and corresponding prompts are generated.

2. The image recognition-based backfill soil quality detection system according to claim 1, characterized in that, The circular tiling rule is: first determine the baseline, then construct the tiling surface; The steps for constructing an outer tiled image are as follows: S11: Place the test sample perpendicular to the horizontal plane and determine the vertically downward axis of the test sample on the outer facade; S12: Adjust the scan line of the line scan camera to be parallel to the axis of the sample being tested, and control the line scan camera to circle around the outer facade of the sample once to obtain the video of the outer facade. S13: Using the calibrated overlap rate as the standard, extract A frames from the facade video and mark the overlapping areas where the overlap rate is located in the frames; S14: Using the first frame image as the reference image, delete the overlapping areas in the remaining A-1 frame images, and stitch and merge the A frame images in sequence according to the shooting time to generate an outer tiled image.

3. The image recognition-based backfill soil quality detection system according to claim 2, characterized in that, The axial tiling rule is: the same subsample is only spliced ​​and merged once; The steps for constructing an inner tiled image are as follows: S21: Divide the test sample into four equal parts along the axial direction into four sub-samples, and in a clockwise direction, record the two axial cutting surfaces of the sub-samples as the mother cutting surface and the sub-cutting surface, respectively. S22: Take pictures of the parent cutting surface of the four sub-samples one by one using a line scan camera to obtain four sub-sample images, and number the four sub-sample images in ascending order in a clockwise direction. S23: Align the edges of adjacent sub-sample images according to their numbers in ascending order, and then merge the aligned sub-sample images to generate an inner tiled image.

4. The image recognition-based backfill soil quality detection system according to claim 3, characterized in that, The steps for extracting the thickness compliance rate are as follows: S31: Convert the outer tiled image into a grayscale image, and set the scaling factor of the grayscale image according to the ratio of the image height of the grayscale image to the actual height of the detection sample; S32: Based on the global grayscale histogram of the grayscale image, the low threshold and high threshold are calculated by the Otsu algorithm, and all grayscale edge lines are extracted from the grayscale image by combining the Canny edge detection technology. S33: Identify straight lines from grayscale edge lines using Hough line transform, set the baseline length and baseline angle range, and record the straight lines whose length is greater than the baseline length and whose tilt angle is within the baseline angle range as valid straight lines. Draw a horizontal line through the midpoint of the valid straight lines to create candidate lines. S34: The regions located one height cell above the candidate line and one height cell below the candidate line are respectively called the upper region and the lower region. The gray values ​​of all points in the upper region and the lower region are summed and the average is calculated. The difference between the two averages is then calculated to obtain the gray value difference. S35: Candidate lines with a grayscale difference greater than the calibrated grayscale difference threshold are recorded as layer lines. The absolute value of the difference between the ordinates of two adjacent layer lines is multiplied by the scaling factor to calculate the layer thickness value. S36: Divide the layer thickness value by the calibrated thickness threshold to calculate the sub-compliance rate, and then sum all the sub-compliance rates and average them to calculate the thickness compliance rate.

5. The image recognition-based backfill soil quality detection system according to claim 4, characterized in that, The steps for extracting particle synchronization rate are as follows: S41: Convert the grayscale image into a binary image using an adaptive thresholding method, and record the points in the binary image as white pixels and black pixels respectively; S42: Through Euclidean distance transformation, calculate the Euclidean distance from white pixel to black pixel one by one, summarize the minimum value of all Euclidean distances to generate a distance map, and record the areas in the distance map where the Euclidean distance is greater than the calibrated distance threshold as the confidence area. S43: The confidence region is subjected to a combination operation of first dilation and then erosion using the morphological closing operation method. The confidence regions where adjacent particles are located are removed, and the remaining confidence regions are recorded as particle regions. The particle contours are extracted from the particle regions using the watershed algorithm. S44: Draw the smallest circumcircle on each particle outline one by one, sum the areas of all the smallest circumcircles and calculate the average area, and record the particle outlines whose smallest circumcircle area is greater than the average area as large particle entities. S45: Measure the point-to-point distance between any two points on the large particle entity one by one, and connect the two points where the maximum point-to-point distance value is located to generate the long axis line; S46: Using the horizontal direction as the reference plane, measure the angle between the long axis and the reference plane. Record the long axis with the angle within the synchronization angle range as the target line. Divide the number of target lines by the number of long axes to calculate the particle synchronization rate.

6. The image recognition-based backfill soil quality detection system according to claim 5, characterized in that, When the outer feature region is determined, two auxiliary lines are drawn vertically in the outer tiled image. The region between the two auxiliary lines is recorded as the first inner region, and the region outside the two auxiliary lines is recorded as the first outer region. The region span of the first inner region and the first outer region is measured horizontally. Move the positions of the two auxiliary lines horizontally until the ratio of the span of the first inner region to the span of the two first outer regions is 6:2:2, and then stop. The first inner region is recorded as the outer feature region.

7. The image recognition-based backfill soil quality detection system according to claim 6, characterized in that, When the inner feature region is determined, two auxiliary lines are drawn horizontally in the inner tiled image. The area between the two auxiliary lines is recorded as the second inner region, and the area outside the two auxiliary lines is recorded as the second outer region. The area span of the second inner region and the second outer region is measured vertically. Move the positions of the two auxiliary lines vertically until the ratio of the span of the second inner region to the span of the two second outer regions is 6:2:2, and then stop. The second inner region is then recorded as the inner feature region.

8. The image recognition-based backfill soil quality detection system according to claim 7, characterized in that, External characteristics include the proportion of large particles and the proportion of particle area; The unique features include the proportion of compliant circumcircle and the proportion of circumcircle area.

9. The image recognition-based backfill soil quality detection system according to claim 8, characterized in that, The dual analysis mechanism is as follows: first, a single parameter analysis is performed, and then a summary analysis of the parameters is performed. The steps for determining the backfill level are as follows: S51: Subtract the compaction degree, moisture content, and uniformity from the preset compaction threshold, preset moisture content threshold, and preset uniformity threshold respectively, and take the absolute value to obtain the compaction difference, moisture content difference, and uniformity difference. S52: When the compaction difference is greater than the upper limit of the compaction difference, the compaction degree is recorded as a low-quality parameter; when the moisture content difference is greater than the upper limit of the moisture content difference, the moisture content is recorded as a low-quality parameter; when the uniformity difference is greater than the upper limit of the uniformity difference, the uniformity is recorded as a low-quality parameter. S53: After dividing the compaction degree, moisture content and uniformity by the preset compaction threshold, preset moisture content threshold and preset uniformity threshold respectively, calculate the compaction ratio, moisture content ratio and uniformity ratio, and then add the compaction ratio, moisture content ratio and uniformity ratio together and calculate the average to obtain the mass ratio. S54: If the number of low-quality parameters in the test sample is less than or equal to 1 and the quality ratio is greater than or equal to 0.9, the backfill quality will be determined as high quality; if the number of low-quality parameters in the test sample is greater than 1 or the quality ratio is less than 0.9, the backfill quality will be determined as low quality.

10. The image recognition-based backfill soil quality detection system according to claim 9, characterized in that, If the backfill grade is high quality, no prompt message will be given; if the backfill grade is low quality, the compaction ratio, moisture content ratio and uniformity ratio will be compared. When the compaction ratio is greater than 1, information on reducing compaction degree is formulated; when the compaction ratio is less than 1, information on increasing compaction degree is formulated. When the moisture content is greater than 1, information on reducing the moisture content is formulated; when the moisture content is less than 1, information on increasing the moisture content is formulated. When the uniformity ratio is greater than 1, information to decrease uniformity is formulated; when the uniformity ratio is less than 1, information to increase uniformity is formulated.

Images