Visual detection method and device for gradation segregation degree and rigidity distribution of large-particle-size macadam base

Abstract

The invention discloses a large-particle-size macadam base grading segregation degree and rigidity distribution visual detection method and device, a matrix type multi-view image acquisition unit is constructed, 3-5 industrial cameras are distributed according to vertexes and central positions of an equilateral triangle, and synchronous exposure of each camera within 0.01 s is realized through a synchronous trigger controller; carrying out full-coverage shooting on the large-particle-size macadam base detection area; the device is integrated with the GPS and IMU positioning module, so that accurate space anchoring of detection data can be realized; grading-rigidity result fusion visualization is completed through a red, yellow and green three-color coding system, a whole-process closed loop of field detection-data operation-result output-cloud supervision is realized in combination with PDF report generation and 4G / 5G cloud uploading functions, a construction party can obtain the accurate position of a segregation and rigidity abnormal area in real time, and the construction efficiency is improved. A supervision department can remotely trace the source of the cause of the quality problem, and the intelligence and refinement level of the road base quality management and control is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and in particular to a visual inspection method and apparatus for the segregation degree and stiffness distribution of large-particle-size crushed stone base course. Background Technology

[0002] Large-diameter crushed stone base courses are widely used in the pavement base structure of high-grade highways and heavy-duty traffic roads due to their advantages such as high strength, good permeability and strong load-bearing capacity. Their gradation uniformity and stiffness distribution directly determine the overall service life and driving safety of the highway pavement, and are the core indicators for highway construction quality control.

[0003] Current industry testing methods for large-diameter crushed stone base courses have many limitations: 1. Gradation detection: Traditional methods mainly rely on manual on-site sampling and indoor sieving. This not only results in long detection cycles (2-3 hours per sampling and sieving session) and low efficiency, but also introduces randomness in sampling, making it difficult to achieve full coverage of the detection area. While some portable sieving devices can be operated on-site, they are still contact-based, easily causing secondary disturbance to the substrate structure. Furthermore, they cannot simultaneously acquire the spatial distribution characteristics of particles, making it difficult to accurately determine the spatial range of gradation segregation. In addition, existing visual inspection solutions mostly use single-view imaging, which is prone to missed particle identification due to obstruction by gravel particles and uneven lighting. Moreover, the segmentation models lack sufficient accuracy in extracting edge features of large-diameter gravel, and the calculation error of gradation parameters generally exceeds 10%.

[0004] 2. Stiffness testing: The mainstream method is single-point testing using a falling weight deflectometer (FWD). This method can only obtain stiffness data at discrete points and cannot form a continuous stiffness distribution map. Furthermore, stiffness testing and gradation testing are independent of each other and lack a correlation model between the two. It is difficult to explain the cause of stiffness anomalies from the perspective of gradation, resulting in low efficiency in tracing the source of quality problems.

[0005] 3. Data management: Traditional testing data is mostly paper records or scattered electronic files, lacking a unified visualization and cloud-based monitoring channel. Data interaction between construction parties and regulatory departments is lagging, making it impossible to achieve real-time early warning and rectification of quality problems.

[0006] In summary, existing testing technologies are insufficient to meet the quality testing requirements of "high efficiency, precision, comprehensive coverage, and linkage" in the construction of large-diameter crushed stone base courses. There is an urgent need for a non-contact, integrated testing technology that can simultaneously acquire gradation and stiffness parameters. Summary of the Invention

[0007] In order to overcome the shortcomings of the prior art, one of the objectives of the present invention is to provide a visual inspection method and device for the segregation degree and stiffness distribution of large-particle-size crushed stone base course.

[0008] One of the objectives of this invention is achieved through the following technical solution: A visual inspection method for the segregation degree and stiffness distribution of large-particle-size crushed stone base course, the method is based on the principle of non-contact visual inspection to achieve simultaneous acquisition of gradation and stiffness parameters, and specifically includes the following steps: S1. Construct a matrix-style multi-view image acquisition unit, distributing 3-5 industrial cameras according to the vertices and center of an equilateral triangle. A synchronous trigger controller enables each camera to be synchronously exposed within 0.01s, achieving full coverage imaging of the large-diameter crushed stone base detection area. Before imaging, the camera's internal and external parameters are calibrated using the Zhang Zhengyou calibration method. Combined with GPS positioning data, latitude, longitude, and elevation information are added to each frame of image, forming multi-source image data containing precise spatial location information. The particle size range of the large-diameter crushed stone is 20-100mm, which covers the commonly used crushed stone particle size range for highway base courses, meeting the requirements of the "Technical Specifications for Highway Pavement Base Course Construction". S2. Standardize and preprocess the multi-source image data. The processing flow is as follows: highlight removal, noise filtering, and image registration. Highlight removal adopts a "global threshold filtering-local gradient correction" composite algorithm. First, the global highlight threshold is automatically determined by the Otsu algorithm to filter out the highlight areas. Then, the Sobel operator is used to calculate the local gradient of the highlight areas. The texture of the highlight areas is restored by gradient inversion correction. Noise filtering adopts a Gaussian filtering and median filtering cascade method. First, Gaussian noise in the image is filtered out by Gaussian filtering (standard deviation 1.2-1.8). Then, salt-and-pepper noise is eliminated by median filtering (5×5 pixel window). Image registration is based on the SIFT feature point matching algorithm to unify the multi-view images to the world coordinate system and obtain a standardized image set without distortion or misalignment. S3. Accurate segmentation of gravel particles is performed on a standardized image set based on an improved U-Net semantic segmentation model. The improved U-Net model introduces a dual attention module that integrates channel attention and spatial attention at the encoding end, and uses a deformable convolution module (with 9 sampling points) to optimize edge feature extraction at the decoding end. After training on 1200 sets of labeled gravel images, the F1 score of the model is stable at over 97.5%. After segmentation, the closed contour of each gravel particle is extracted by a contour tracking algorithm. The equivalent particle size (using the area equivalence method, equivalent particle size d=√(4S / π), where S is the particle contour area), particle distribution density (number of particles per unit area), and aspect ratio (length-to-width ratio of the particle's circumscribed rectangle) parameters are calculated to establish a gradation feature database containing single particle features and regional statistical features. S4. Construct a gradation segregation evaluation model based on highway engineering standards. First, determine the standard gradation sieve residue range corresponding to each sieve size (2.36mm, 4.75mm, 9.5mm, 19mm, 31.5mm, 63mm) according to the "Technical Specification for Construction of Asphalt Pavement of Highway" (JTGF40-2004). Then, convert the proportion of each particle size statistically analyzed in the gradation characteristic database into equivalent sieve residue values, and calculate the single sieve residue deviation rate and cumulative deviation rate. The single sieve residue deviation rate = (measured equivalent sieve residue value - median standard sieve residue) / median standard sieve residue × 100%, and the cumulative deviation rate is the weighted sum of the deviation rates of each sieve size (the weights are 0.3, 0.25, 0.2, 0.15, 0.07, and 0.03 in descending order of sieve particle size). Finally, combine the single sieve deviation and cumulative deviation results to determine the gradation segregation level. S5. Establish a gradation-stiffness correlation model. The model training samples are obtained through a dual approach of "indoor testing + field measurement": Indoorly, 30 sets of specimens are made according to different gradation ratios, and the dynamic rebound modulus is tested using a falling weight deflectometer (FWD); On-site, 50 typical road sections are selected, and image gradation data and FWD measured stiffness values ​​are collected simultaneously, ultimately forming 1200 sets of effective training samples; The mapping relationship is constructed using a random forest algorithm, with 100 decision trees, a maximum depth of 15 layers, and a minimum number of sample splits of 8. After 5-fold cross-validation, the model prediction error is controlled within 7.2%; The gradation feature parameters obtained in step S3 are input into the model, and combined with the GPS coordinate information attached to the image, a stiffness distribution heat map with a 1m×1m grid precision is generated through an interpolation algorithm; S6. Integrate the segregation evaluation results and stiffness distribution heatmap, and use a three-color coding system of red (severe segregation / insufficient stiffness), yellow (moderate segregation / stiffness meets the standard), and green (slight segregation / excellent stiffness) for visualization labeling. The comprehensive report should also include the test time, road segment chainage, environmental parameters (temperature 5-35℃, humidity 30%-80%), and model confidence level. Data can be stored locally in PDF format or uploaded to the highway quality supervision platform via 4G / 5G module.

[0009] As a further improvement to the above technical solution: The noise filtering in step S2 uses a cascaded Gaussian filter and a median filter. The standard deviation of the Gaussian filter is set to 1.2-1.8, and the window size of the median filter is set to 5×5 pixels.

[0010] The improved U-Net semantic segmentation model described in step S3 optimizes feature extraction by introducing a deformable convolutional module with fused attention, and the model's F1 score is no less than 97.5%.

[0011] The gradation segregation levels mentioned in step S4 are divided into slight segregation, moderate segregation and severe segregation, with the sieve residue deviation rates corresponding to ±5%, ±5%-±10% and ±10% and above, respectively.

[0012] The training samples of the gradation-stiffness correlation model in step S5 include more than 1,000 sets of different gradation parameters and corresponding measured values ​​of dynamic resilient modulus, and the model prediction error is less than 8%.

[0013] A visual inspection device for large-particle-size crushed stone base courses that implements the method of any one of claims 1-5, comprising: The image acquisition module consists of 3-5 high-definition industrial cameras and matching LED fill light units. The camera resolution is no less than 5 million pixels and the frame rate is no less than 25fps. The fill light unit supports stepless brightness adjustment. The positioning module integrates GPS and IMU inertial measurement units to achieve real-time positioning of the detected location with a positioning accuracy error of less than 0.1m; The data processing module uses an embedded processor and incorporates the image preprocessing algorithm, segmentation model, and association model described in claim 1, supporting real-time data processing. The output module, including a touch screen and a wireless communication unit, can display the test results on-site and upload the data to the cloud platform.

[0014] As a further improvement to the above technical solution: The camera lens of the image acquisition module adopts a large depth of field design with a focal length range of 8-16mm, ensuring clear imaging of gravel particles within a shooting distance of 2-5m.

[0015] The LED supplementary lighting unit adopts a ring array layout, with a color temperature adjustment range of 4500K-6500K, and can automatically switch the supplementary lighting mode according to the ambient light intensity.

[0016] The data processing module is also equipped with a storage unit with a storage capacity of not less than 128GB, which can cache at least 1000 sets of detection data and original images.

[0017] It also includes a portable stand made of carbon fiber, with a height adjustment range of 1.2-2.5m and lockable casters at the bottom.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Based on matrix-style multi-view visual acquisition and gradation-stiffness correlation model, it breaks the barrier between the traditional independent detection of gradation and stiffness. It can simultaneously acquire the gradation segregation degree and stiffness distribution data of the detection area with 1m×1m grid precision without disturbing the base structure. The detection efficiency is improved by more than 80% compared with the traditional method, and the detection range can achieve full coverage, avoiding the randomness defects of single-point sampling.

[0019] 2. In the image preprocessing stage, the "global threshold screening-local gradient correction" highlight removal algorithm and Gaussian-median cascade filtering are adopted to effectively solve the interference of strong light and noise on imaging. The improved U-Net model optimizes feature extraction through dual attention modules and deformable convolution, and the F1 score of gravel particle segmentation is stabilized above 97.5%, and the calculation error of gradation parameters is reduced to within 5%. The gradation-stiffness correlation model is trained with 1200 sets of samples, and the prediction error is controlled within 7.2%, which is far lower than the industry average, providing double accuracy guarantee for the detection data.

[0020] 3. The device integrates GPS and IMU positioning modules, enabling precise spatial anchoring of test data; it achieves visualization of gradation-stiffness results through a red, yellow, and green three-color coding system, and combined with PDF report generation and 4G / 5G cloud upload functions, it realizes a closed-loop process of "on-site testing - data processing - result output - cloud supervision". Construction parties can obtain the precise location of segregation and stiffness anomaly areas in real time, and regulatory departments can remotely trace the causes of quality problems, which greatly improves the intelligence and precision of highway base quality control.

[0021] 4. The detection device adopts a carbon fiber portable bracket with a height adjustment of 1.2-2.5m and a lockable universal wheel for movement, which can flexibly adapt to the detection needs of different construction sections; the LED supplementary light unit supports adaptive adjustment of color temperature and brightness, and can work stably in complex environments of 5-35℃ and 30%-80% humidity; the camera's large depth-of-field lens ensures clear imaging of particles within a shooting distance of 2-5m. The device's environmental adaptability and ease of operation are significantly better than similar equipment.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0023] Figure 1 Flowchart of a visual inspection method for the segregation and stiffness distribution of large-diameter crushed stone base course gradation; Figure 2 This is a flowchart of the structure of a visual inspection device for large-diameter crushed stone base layers. Detailed Implementation

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0025] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Specific Implementation

[0027] This embodiment is applied to the construction quality inspection of large-diameter crushed stone base course in the K236-K238 section of the G107 National Highway reconstruction and expansion project in a certain province. The base course of this section uses crushed stone mixture with a particle size range of 20-100mm. It is necessary to simultaneously test its gradation segregation degree and stiffness distribution to ensure that the construction quality meets the requirements of the "Technical Specifications for Construction of Highway Pavement Base Course". The ambient temperature during the test is 25℃ and the humidity is 60%, which meets the environmental parameter requirements of 5-35℃ and 30%-80% in the method.

[0028] II. Setup of the Detection Device This embodiment uses the large-diameter crushed stone base visual inspection device as defined in claims 6-10, and the specific configuration of each module is as follows: 1. Image acquisition module Four 5-megapixel high-definition industrial cameras were selected, with a frame rate of 25fps, a large depth-of-field lens, and a focal length of 12mm to ensure clear imaging of gravel particles within a 3-meter shooting distance. The matching LED fill light unit is arranged in a ring array, with a color temperature adjusted to 5500K. It can automatically switch fill light modes according to the ambient light to ensure uniformity of image acquisition.

[0029] Four cameras are arranged in a matrix pattern with the vertices and center of an equilateral triangle. Synchronous exposure within 0.01 seconds is achieved through a synchronous trigger controller, thus achieving full coverage of the detection area.

[0030] 2. Positioning module It integrates GPS and IMU inertial measurement units, with positioning accuracy error controlled within 0.08m (below the requirement of 0.1m), and can add latitude, longitude, elevation and road segment station information to the acquired images in real time.

[0031] 3. Data Processing Module It adopts a high-performance embedded processor, with built-in image preprocessing algorithms, improved U-Net segmentation model and gradation-stiffness correlation model; it is equipped with a 256GB storage unit (greater than 128GB requirement), which can cache 1500 sets of detection data and raw images to meet the needs of real-time computing and data storage on site.

[0032] 4. Output module Equipped with a 10-inch touchscreen display, it can show test results on-site; it integrates a 5G wireless communication unit, supporting data upload to the provincial highway quality supervision platform; it also supports local storage of PDF format reports.

[0033] 5. Portable stand Made of carbon fiber, the bracket height is adjustable to 1.8m (within the range of 1.2-2.5m), and the bottom can be locked with casters to fix it to the detection point, ensuring the stability of the device.

[0034] III. Implementation of the Detection Method (corresponding to claims 1-5) Step S1: Multi-source image acquisition and calibration 1. Before testing, the intrinsic parameters (focal length, distortion coefficient) and extrinsic parameters (spatial position relationship) of the four cameras were calibrated using the Zhang Zhengyou calibration method to eliminate imaging distortion.

[0035] 2. The detection section from K236+100 to K236+200 (100m long and 12m wide) is divided into zones for shooting. One detection point is set up every 10m, and 10 frames of multi-view images are collected at each point. The positioning module adds accurate spatial information to each frame of image to form multi-source image data containing latitude, longitude, elevation and station number.

[0036] Step S2: Multi-source image normalization preprocessing 1. Highlight Removal: First, the global highlight threshold of 180 (grayscale value) is automatically determined by the Otsu algorithm to filter out the highlight areas on the surface of gravel in the image; then, the local gradient of the highlight area is calculated by the Sobel operator, and the gravel texture of the highlight area is restored by gradient inverse correction to eliminate the interference of strong light on particle recognition.

[0037] 2. Noise filtering: A cascaded method of "Gaussian filtering + median filtering" is adopted. First, a Gaussian filter with a standard deviation of 1.5 is set to filter out Gaussian noise, and then a median filter with a 5×5 pixel window is used to eliminate salt and pepper noise and improve image clarity.

[0038] 3. Image registration: Based on the SIFT feature point matching algorithm, the multi-view images from the four cameras are unified into the world coordinate system to obtain a standardized image set without distortion or misalignment.

[0039] Step S3: Crushed stone particle segmentation and gradation feature extraction 1. An improved U-Net semantic segmentation model is used to segment a standardized image set: a channel-space dual attention module is introduced at the encoding end of the model, and a deformable convolution module with 9 sampling points is configured at the decoding end. The model is trained on 1200 sets of labeled images and the measured F1 score reaches 97.8% (higher than the 97.5% requirement), which can achieve accurate segmentation of gravel particles.

[0040] 2. Extract the closed contour of the particles using a contour tracing algorithm, and calculate the equivalent particle size using the area equivalence method (formula: , The particle outline area is used to calculate parameters such as particle density per unit area and particle length-to-width ratio. Finally, a gradation feature database containing single particle characteristics and regional statistical characteristics is established. The measured equivalent particle size of crushed stone in this section is mainly concentrated in the range of 30-80mm.

[0041] Step S4: Evaluation of Gradation Segregation 1. According to the Technical Specification for Construction of Asphalt Pavement of Highway (JTGF40-2004), the median standard sieve residue values ​​for each sieve size are determined as follows: 2.36mm (3%), 4.75mm (5%), 9.5mm (12%), 19mm (25%), 31.5mm (45%), and 63mm (75%).

[0042] 2. Convert the proportion of each particle size in the gradation characteristic database into equivalent sieve residue values, and calculate the single sieve residue deviation rate and cumulative deviation rate: For example, at point K236+150, the measured equivalent sieve residue of the 63mm sieve is 82%, and the single sieve residue deviation rate = (82%-75%) / 75%*100% = 9.33%; the measured equivalent sieve residue of the 31.5mm sieve is 48%, and the deviation rate = (48%-45%) / 45%*100% = 6.67%.

[0043] 3. The cumulative deviation rate was calculated according to the weights (63mm:0.3, 31.5mm:0.25, 19mm:0.2, 9.5mm:0.15, 4.75mm:0.07, 2.36mm:0.03). The cumulative deviation rate at this point was 7.2%. Based on the deviation rate, the segregation level was determined (claim 4): the maximum deviation rate of a single sieve aperture at this point was 9.33% (±5%-±10%), and the cumulative deviation rate was 7.2%, which was determined to be moderate segregation; the deviation rates of all sieve apertures at point K236+120 were within ±4%, and the cumulative deviation rate was 2.1%, which was determined to be slight segregation.

[0044] Step S5: Stiffness Distribution Prediction 1. The gradation-stiffness correlation model in this embodiment has 30 sets of indoor specimen FWD test data (different gradation ratios) and 50 typical road sections on-site measured data, totaling 1200 sets of valid samples. The model uses a random forest algorithm with 100 decision trees, a maximum depth of 15 layers, and a minimum number of sample splits of 8. After 5-fold cross-validation, the prediction error is 6.8% (lower than the 7.2% requirement).

[0045] 2. Input the gradation characteristic parameters from step S3 into the model, and combine them with GPS coordinate information to generate a stiffness distribution heat map with a grid accuracy of 1m×1m using the Kriging interpolation algorithm; where the dynamic resilient modulus of the K236+150 (moderate segregation) region is 1800MPa (stiffness meets the standard), and the dynamic resilient modulus of the K236+180 region (severe segregation) is only 1200MPa (stiffness is insufficient).

[0046] Step S6: Result Fusion and Visualization Output 1. The fusion results using the "red-yellow-green" three-color coding system are as follows: the K236+180 area is marked in red (severe segregation + insufficient stiffness), the K236+150 area is marked in yellow (moderate segregation + stiffness meets the standard), and the K236+120 area is marked in green (slight segregation + excellent stiffness).

[0047] 2. Generate a comprehensive testing report, which includes the testing date (month / day / year / 2025), road segment number, environmental parameters (25℃, 60% humidity), and model confidence level (98.2%). The report is also uploaded to the highway quality supervision platform to provide the construction party with a basis for rectification of segregation areas.

[0048] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A visual inspection method for the segregation degree and stiffness distribution of large-particle-size crushed stone base course, characterized in that, The method, based on the principle of non-contact visual inspection, achieves simultaneous acquisition of gradation and stiffness parameters, and specifically includes the following steps: S1. Construct a matrix-style multi-view image acquisition unit, distributing 3-5 industrial cameras according to the vertices and center of an equilateral triangle. A synchronous trigger controller enables each camera to be synchronously exposed within 0.01s, achieving full coverage imaging of the large-diameter crushed stone base detection area. Before imaging, the camera's internal and external parameters are calibrated using the Zhang Zhengyou calibration method. Combined with GPS positioning data, latitude, longitude, and elevation information are added to each frame of image, forming multi-source image data containing precise spatial location information. The particle size range of the large-diameter crushed stone is 20-100mm, which covers the commonly used crushed stone particle size range for highway base courses, meeting the requirements of the "Technical Specifications for Highway Pavement Base Course Construction". S2. Standardize and preprocess the multi-source image data. The processing flow is as follows: highlight removal, noise filtering, and image registration. Highlight removal adopts a "global threshold filtering-local gradient correction" composite algorithm. First, the global highlight threshold is automatically determined by the Otsu algorithm to filter out the highlight areas. Then, the Sobel operator is used to calculate the local gradient of the highlight areas. The texture of the highlight areas is restored by gradient inversion correction. Noise filtering adopts a Gaussian filtering and median filtering cascade method. First, Gaussian noise in the image is filtered out by Gaussian filtering (standard deviation 1.2-1.8). Then, salt-and-pepper noise is eliminated by median filtering (5×5 pixel window). Image registration is based on the SIFT feature point matching algorithm to unify the multi-view images to the world coordinate system and obtain a standardized image set without distortion or misalignment. S3. Accurate segmentation of gravel particles is performed on a standardized image set based on an improved U-Net semantic segmentation model. The improved U-Net model introduces a dual attention module that integrates channel attention and spatial attention at the encoding end, and uses a deformable convolution module (with 9 sampling points) to optimize edge feature extraction at the decoding end. After training on 1200 sets of labeled gravel images, the F1 score of the model is stable at over 97.5%. After segmentation, the closed contour of each gravel particle is extracted by a contour tracking algorithm. The equivalent particle size (using the area equivalence method, equivalent particle size d=√(4S / π), where S is the particle contour area), particle distribution density (number of particles per unit area), and aspect ratio (length-to-width ratio of the particle's circumscribed rectangle) parameters are calculated to establish a gradation feature database containing single particle features and regional statistical features. S4. Construct a gradation segregation evaluation model based on highway engineering standards. First, determine the standard gradation sieve residue range corresponding to each sieve size (2.36mm, 4.75mm, 9.5mm, 19mm, 31.5mm, 63mm) according to the "Technical Specification for Construction of Asphalt Pavement of Highway" (JTGF40-2004). Then, convert the proportion of each particle size statistically analyzed in the gradation characteristic database into equivalent sieve residue values, and calculate the single sieve residue deviation rate and cumulative deviation rate. The single sieve residue deviation rate = (measured equivalent sieve residue value - median standard sieve residue) / median standard sieve residue × 100%, and the cumulative deviation rate is the weighted sum of the deviation rates of each sieve size (the weights are 0.3, 0.25, 0.2, 0.15, 0.07, and 0.03 in descending order of sieve particle size). Finally, combine the single sieve deviation and cumulative deviation results to determine the gradation segregation level. S5. Establish a gradation-stiffness correlation model. The model training samples are obtained through a dual approach of "indoor testing + field measurement": Indoorly, 30 sets of specimens are made according to different gradation ratios, and the dynamic rebound modulus is tested using a falling weight deflectometer (FWD); On-site, 50 typical road sections are selected, and image gradation data and FWD measured stiffness values ​​are collected simultaneously, ultimately forming 1200 sets of effective training samples; The mapping relationship is constructed using a random forest algorithm, with 100 decision trees, a maximum depth of 15 layers, and a minimum number of sample splits of 8. After 5-fold cross-validation, the model prediction error is controlled within 7.2%; The gradation feature parameters obtained in step S3 are input into the model, and combined with the GPS coordinate information attached to the image, a stiffness distribution heat map with a 1m×1m grid precision is generated through an interpolation algorithm; S6. Integrate the segregation evaluation results and stiffness distribution heatmap, and use a three-color coding system of red (severe segregation / insufficient stiffness), yellow (moderate segregation / stiffness meets the standard), and green (slight segregation / excellent stiffness) for visualization labeling. The comprehensive report should also include the test time, road segment chainage, environmental parameters (temperature 5-35℃, humidity 30%-80%), and model confidence level. Data can be stored locally in PDF format or uploaded to the highway quality supervision platform via 4G / 5G module.

2. The method according to claim 1, characterized in that, The noise filtering in step S2 uses a cascaded Gaussian filter and a median filter. The standard deviation of the Gaussian filter is set to 1.2-1.8, and the window size of the median filter is set to 5×5 pixels.

3. The method according to claim 1, characterized in that, The improved U-Net semantic segmentation model described in step S3 optimizes feature extraction by introducing a deformable convolutional module with fused attention, and the model's F1 score is no less than 97.5%.

4. The method according to claim 1, characterized in that, The gradation segregation levels described in step S4 are divided into slight segregation, moderate segregation, and severe segregation, with the sieve residue deviation rates corresponding to ±5%, ±5%-±10%, and ±10% and above, respectively.

5. The method according to claim 1, characterized in that, The training samples of the gradation-stiffness correlation model mentioned in step S5 include more than 1,000 sets of different gradation parameters and corresponding measured values ​​of dynamic resilient modulus, and the model prediction error is less than 8%.

6. A visual inspection device for large-particle-size crushed stone base courses that implements the method of any one of claims 1-5, characterized in that, include: The image acquisition module consists of 3-5 high-definition industrial cameras and matching LED fill light units. The camera resolution is no less than 5 million pixels and the frame rate is no less than 25fps. The fill light unit supports stepless brightness adjustment. The positioning module integrates GPS and IMU inertial measurement units to achieve real-time positioning of the detected location with a positioning accuracy error of less than 0.1m; The data processing module uses an embedded processor and incorporates the image preprocessing algorithm, segmentation model, and association model described in claim 1, supporting real-time data processing. The output module, including a touch screen and a wireless communication unit, can display the test results on-site and upload the data to the cloud platform.

7. The apparatus according to claim 6, characterized in that, The camera lens of the image acquisition module adopts a large depth of field design with a focal length range of 8-16mm, ensuring clear imaging of gravel particles within a shooting distance of 2-5m.

8. The apparatus according to claim 6, characterized in that, The LED supplementary lighting unit adopts a ring array layout, with a color temperature adjustment range of 4500K-6500K, and can automatically switch the supplementary lighting mode according to the ambient light intensity.

9. The apparatus according to claim 6, characterized in that, The data processing module is also equipped with a storage unit with a storage capacity of not less than 128GB, which can cache at least 1000 sets of detection data and original images.

10. The apparatus according to claim 6, characterized in that, It also includes a portable stand made of carbon fiber, with a height adjustment range of 1.2-2.5m and lockable casters at the bottom.

Images