Rockfill material grading intelligent detection method and system based on ai vision

Abstract

The present application relates to a kind of rockfill material gradation intelligent detection method and system based on AI vision, belong to image processing technical field.The sample to be measured is divided into multiple groups according to particle size, and a differentiated preprocessing strategy is used, the ultra-fine particle group is directly weighed, the main particle group is collected by dynamic video stream free-fall image, and the super-large particle group is multi-angle static shooting.AI vision model is used to segment and identify the image, extract the particle pixel area, and convert it into physical projection area by calibrating the background plate, establish the empirical relationship between pixel area and particle volume, calculate the mass of single particle, and calibrate it by combining the mass conversion coefficient, fuse the data of each group, sort all particles according to equivalent particle size, calculate the cumulative mass percentage, and automatically generate a complete particle grading curve.The present application realizes efficient and high-precision detection in the full particle size range, significantly improves the sample representativeness and detection automation level, and can be directly used for engineering quality control.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an intelligent detection method and system for riprap gradation based on AI vision. Background Technology

[0002] With the continuous advancement of intelligent construction and engineering testing technologies, major infrastructure projects such as earth-rock dams and roadbeds are placing higher demands on the precise control of particle size distribution of fill materials. As a core parameter determining the compactness, permeability, and stability of the fill, particle size distribution directly impacts project quality and safety. Current standards generally employ sieving for gradation determination. This method requires the stepwise sieving and weighing of representative samples exceeding 1000 kg, resulting in long testing cycles, high reliance on manual labor, and limited sample representativeness. It is ill-suited to the demands of modern, efficient construction processes and real-time quality control. Especially in large-scale rockfill dam projects, where the fill particle size range is wide (0–800 mm) and the volume of a single test is large, traditional methods are severely inadequate in terms of timeliness and coverage integrity.

[0003] Among these technologies, image recognition-based particle size analysis has been introduced into the field of engineering inspection due to its potential for non-contact and high-throughput processing. However, its direct application to the gradation detection of riprap faces several technical bottlenecks: Firstly, riprap has a very wide particle size range, with small particles easily obscured and large particles difficult to image completely within a limited field of view; secondly, the large volume of on-site samples makes it difficult to ensure that all stones are clearly captured, leading to statistical bias. Furthermore, riprap is generally dark in color and highly irregular in shape. Under complex lighting, shadow interference, and particle adhesion conditions, traditional image segmentation algorithms exhibit poor robustness and high parameter sensitivity, making it difficult to achieve stable, universal, and high-precision identification.

[0004] Existing technologies are either limited to small-particle-size scenarios for fine-grained materials such as sand and soil, or rely on static single-point photography to infer overall gradation from a single point, failing to accurately reflect the actual distribution of large-scale, wide-particle-size rockfill. The test results have significant shortcomings in terms of accuracy, representativeness, and engineering applicability, making it difficult to meet the stringent requirements for gradation data in critical projects such as high rockfill dams. Therefore, there is an urgent need for an intelligent detection method for rockfill gradation that can achieve full particle size coverage, high representativeness, high efficiency, and direct applicability to engineering quality inspection. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent detection method for gradation of rockfill based on AI vision, which can effectively solve the problems mentioned in the background art, such as long detection cycle, high dependence on manual labor, and limited sample representativeness of traditional sieving methods, as well as the problems of incomplete imaging, severe occlusion, poor anti-interference ability, and difficulty in achieving full particle size coverage and direct engineering application of existing image recognition technology in rockfill scenarios with wide particle size range and large sample size.

[0006] To achieve the above objectives, this application provides the following technical solution:

[0007] A method for intelligent detection of gradation of riprap based on AI vision includes the following specific steps:

[0008] S1: The rockfill sample to be tested is divided into multiple particle size groups according to particle size, and different strategies are used to preprocess the multiple particle size groups.

[0009] S2: Based on the dynamic video stream of the multiple particle size groups, image features are extracted to obtain the pixel area of ​​the particle size groups;

[0010] S3: Based on the calibration background plate, convert the pixel area of ​​each particle into the physical projection area, establish the empirical function relationship between the pixel area and the particle volume, calculate the uncalibrated mass of a single particle, and calculate the calibrated mass of all particles in the multiple particle size groups.

[0011] S4: Perform data fusion on the particle mass of the multiple particle size groups, and sort all identified particles from largest to smallest according to their equivalent particle size;

[0012] S5: Calculate the total calibration mass based on the total number of identified particles and the calibration mass of each particle, calculate the mass percentage of particles with a preset particle size, and plot the complete particle gradation curve based on the mass percentage of particles with the preset particle size.

[0013] Furthermore, the method includes:

[0014] The rockfill samples to be tested were divided into ultrafine particle group, main particle group and ultra-large particle group according to particle size.

[0015] The ultrafine particle group with a particle size smaller than the first set threshold is separated by a pre-screening device and directly weighed to obtain the particle mass of the ultrafine particle group.

[0016] Large particle groups with a particle size greater than the second set threshold are manually or mechanically sorted and guided to the static image acquisition area;

[0017] The remaining particle groups with particle sizes between the first and second set thresholds are then transported to the dynamic image acquisition system.

[0018] Furthermore, the method also includes:

[0019] For the ultra-large particle group, the ultra-large particle group particles within the group are laid in a single layer on a calibration background plate with known physical dimensions, and static images of the ultra-large particle group particles are captured from multiple angles using a high-resolution camera.

[0020] The static image of the ultra-large particle group is input into a pre-trained AI semantic segmentation model to extract the maximum projected contour pixel area of ​​each particle.

[0021] The main particle group is controlled by a conveyor belt and a vibrating screen to form a single layer of material flow that is separated from each other. When it enters the front of the high-contrast background plate, it undergoes free fall motion, and a high-speed camera continuously records a video sequence of its falling process.

[0022] Furthermore, the method also includes:

[0023] Based on the effective height of the background panel, the initial velocity of the stone when it enters the background panel, the camera frame rate, and the gravitational acceleration, the time required for the stone to pass through the background panel is calculated, and then the total number of frames is obtained.

[0024] Based on the principles of avoiding duplicate counting and ensuring capture integrity, the frame extraction interval N is determined, and one frame is extracted every N frames starting from the first frame.

[0025] All extracted frames and ultra-large particle static images are input into the locally deployed AI vision model to generate a binarized image with stone as the white foreground and black background.

[0026] Subsequently, connected component analysis was performed on the binarized image to extract the pixel area and principal axis size of each particle.

[0027] Furthermore, the method also includes:

[0028] The calibration coefficients are calculated based on the actual size of the calibration background and its pixel size ratio in the image;

[0029] Convert the pixel area of ​​each particle into its physical projected area;

[0030] An empirical functional relationship between experimental data and particle volume was established by fitting the data.

[0031] Calculate the uncalibrated mass of a single particle based on the premise that the density of the stone is constant;

[0032] The mass conversion factor is obtained based on the measured mass of sieving and the total mass calculated by the system identification. The mass conversion factor is obtained by parallel testing of samples from the same batch and the standard sieving method. The calculated mass of all particles in the main particle group and the ultra-large particle group is multiplied by K to obtain the calibrated mass.

[0033] Furthermore, the method also includes:

[0034] Perform data normalization and synthesis to include the total weighing mass of the ultrafine particle group in the cumulative mass;

[0035] The calibrated mass of each particle identified in the main particle group and the super-large particle group was calculated;

[0036] Sort all identified particles in descending order of equivalent particle size.

[0037] Furthermore, the method also includes:

[0038] Real-time monitoring of the coverage rate, which is the ratio of the total area of ​​all particles projected within a unit frame of the dynamic video stream to the total area of ​​the background image.

[0039] When the coverage rate is higher than the preset ratio, the feeding rate will be automatically reduced or the conveying speed will be increased.

[0040] When the coverage rate is lower than the preset ratio, the feeding rate is increased while ensuring recognition accuracy, so as to achieve a dynamic balance between detection efficiency and recognition accuracy.

[0041] Furthermore, the method also includes:

[0042] The first set threshold d1 is 5 mm, and the second set threshold d2 is 400 mm;

[0043] The pre-screening device uses a vibrating screen with an aperture of 5 mm;

[0044] The main particle group is transported by a belt conveyor, and the ultra-large particle group is placed in the static image acquisition area after being sorted manually or mechanically.

[0045] The high-contrast background panel is made of matte red or blue material;

[0046] The light source system consists of two sets of high-brightness LED diffused lights, which are arranged on both sides of the background panel.

[0047] The initial velocity was calculated using the free fall of the stone before it entered the background plate.

[0048] Furthermore, the method also includes:

[0049] The AI ​​vision big model is a locally deployed variant of the Segment Anything Model, and the prompting mechanism is set to mark all falling rock targets in the image as foreground.

[0050] For particles with slight adhesion, secondary segmentation is performed by combining morphological opening operation and watershed algorithm after the model output mask;

[0051] The empirical function relationship is based on material-related parameters and uses a power function form.

[0052] Once the mass conversion factor is determined through a standard sieve calibration experiment, it can be reused under the same working conditions and the same batch of raw materials.

[0053] The quality conversion factor is updated based on a triggered recalibration according to changes in the construction section or significant environmental changes.

[0054] The equivalent particle size is based on the diameter of the equal-area circle of the particle projection profile as a unified benchmark.

[0055] The static image acquisition module acquires images of the ultra-large particle group from at least the top and two side angles, and then fuses them using a three-dimensional reconstruction algorithm to generate the maximum projection surface.

[0056] Image stitching employs feature point matching and perspective transformation techniques;

[0057] The real-time coverage analysis module samples one frame at preset intervals for statistical analysis.

[0058] The feedback control system connects the electric gate of the hopper to the frequency converter of the conveyor belt, forming a closed-loop regulation circuit.

[0059] According to a second aspect of the present invention, the present invention claims protection for an AI vision-based intelligent detection system for rockfill gradation, used to implement the AI ​​vision-based intelligent detection method for rockfill gradation, the system comprising:

[0060] The particle size separation module is used to separate ultrafine particles and ultralarge particles.

[0061] The dynamic image acquisition module includes a conveying and paving mechanism that enables the main aggregate to form a separated material flow, a high-contrast background plate, and a high-speed camera for capturing the free fall process of the aggregate.

[0062] The static image acquisition module is used to capture images of flat, ultra-large particle groups;

[0063] The computing and control center is used to control the operation of each module, execute algorithms for video frame extraction, AI image recognition, geometric information extraction, quality conversion, data fusion and gradation curve generation, and can dynamically adjust feeding or conveying parameters based on image feedback.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] This invention, through the construction of a three-in-one detection architecture integrating pre-screening, dynamic video AI recognition, and static image AI recognition, achieves for the first time high-precision, high-efficiency, and highly representative intelligent detection of gradation in aggregates across the entire particle size range of 0 to 800 mm. For ultrafine particles, physical sieving and direct weighing are employed to ensure the accuracy of the microparticle mass data. For the main particle group, an innovative free-fall dynamic imaging and kinematics-based intelligent frame extraction mechanism are designed to effectively solve the problems of particle occlusion and duplicate counting. Furthermore, a large AI vision model is used to achieve image segmentation with strong anti-interference capabilities, significantly improving the robustness of identifying dark and irregular aggregates. For ultralarge particles, multi-angle static imaging and maximum projection surface extraction technology are employed to overcome the shortcomings of incomplete imaging from a single viewpoint. By introducing a mass conversion coefficient K based on the principle of mass conservation, the geometric parameters obtained from AI recognition are converted into physically meaningful mass data, which is seamlessly integrated with the screening results to generate a complete gradation curve that meets engineering specifications. This curve can be directly used for engineering quality assessment without any indirect conversion. The system possesses adaptive control capabilities based on image coverage feedback, dynamically adjusting feeding parameters to maintain optimal recognition while balancing detection accuracy and throughput efficiency. The entire method compresses the traditional screening process, which takes over 4 hours, to less than 10 minutes, increasing the sample size by tens of times and ensuring high representativeness. It fundamentally changes the current situation of lagging quality inspection and insufficient sampling of fill materials in major projects such as earth-rock dams and roadbeds, providing a crucial data support tool for intelligent construction systems and demonstrating significant engineering practical value and promising prospects for widespread application. Attached Figure Description

[0066] Figure 1 The flowchart illustrates a method for intelligent detection of rockfill gradation based on AI vision, as claimed in an embodiment of the present invention.

[0067] Figure 2 A schematic diagram of the stone material measured in the pit after compaction, according to an AI vision-based intelligent detection method for riprap gradation claimed in an embodiment of the present invention.

[0068] Figure 3 The images shown are video frames representing an AI-based vision-based intelligent detection method for riprap gradation claimed in an embodiment of the present invention.

[0069] Figure 4 The image shows the AI ​​recognition effect of an AI vision-based intelligent detection method for riprap gradation, as claimed in an embodiment of the present invention.

[0070] Figure 5 A curve comparison chart of an AI vision-based intelligent detection method for riprap gradation claimed in an embodiment of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0072] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications in the embodiments of this application, such as up, down, left, right, front, back, etc., are only used to explain the relative positional relationships and movements between components in a specific orientation as shown in the accompanying drawings. If the specific orientation changes, the directional indications will change accordingly. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0073] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0074] Currently, in major infrastructure projects such as earth-rock dams and roadbeds, precise control of the particle size distribution of fill materials is crucial for project quality and safety. Existing screening methods suffer from long testing cycles, high reliance on manual labor, and limited sample representativeness, making them unsuitable for the demands of modern, efficient construction schedules and real-time quality control. This is especially true in large-scale rockfill dam projects, where the fill material particle size range is wide (0–800 mm) and the volume of material tested in a single test is large; traditional methods are severely inadequate in terms of timeliness and coverage integrity. While image recognition-based particle size analysis technology shows potential, it faces technical bottlenecks in wide-size, large-sample rockfill scenarios, including incomplete imaging, severe occlusion, and poor anti-interference capabilities, preventing full particle size coverage and direct engineering application. To address these technical problems, this invention proposes a three-in-one detection architecture combining pre-screening, dynamic video AI recognition, and static image AI recognition. This architecture enables high-precision, high-efficiency, and highly representative intelligent detection of rockfill gradation across the entire particle size range of 0 to 800 mm, and is applied to an AI vision-based intelligent detection method for rockfill gradation.

[0075] According to a first embodiment of the present invention, the present invention claims protection for an intelligent detection method for gradation of riprap based on AI vision, referring to... Figure 1 This includes the following specific steps:

[0076] S1: The rockfill sample to be tested is divided into multiple particle size groups according to particle size, and different strategies are used to preprocess the multiple particle size groups.

[0077] S2: Based on the dynamic video stream of the multiple particle size groups, image features are extracted to obtain the pixel area of ​​the particle size groups;

[0078] S3: Based on the calibration background plate, convert the pixel area of ​​each particle into the physical projection area, establish the empirical function relationship between the pixel area and the particle volume, calculate the uncalibrated mass of a single particle, and calculate the calibrated mass of all particles in the multiple particle size groups.

[0079] S4: Perform data fusion on the particle mass of the multiple particle size groups, and sort all identified particles from largest to smallest according to their equivalent particle size;

[0080] S5: Calculate the total calibration mass based on the total number of identified particles and the calibration mass of each particle, calculate the mass percentage of particles with a preset particle size, and plot the complete particle gradation curve based on the mass percentage of particles with the preset particle size.

[0081] Furthermore, the method includes:

[0082] The rockfill samples to be tested were divided into ultrafine particle group, main particle group and ultra-large particle group according to particle size.

[0083] The ultra-fine particle group with a particle size smaller than the first set threshold is separated by a pre-screening device and directly weighed to obtain the particle mass of the ultra-fine particle group;

[0084] The oversized particle group with a particle size larger than the second set threshold is sorted manually or mechanically and guided to the static image acquisition area;

[0085] The remaining main particle group with a particle size between the first set threshold and the second set threshold is transported to the dynamic image acquisition system.

[0086] Among them, in this embodiment, sample preprocessing and size-segmented diversion:

[0087] The overall sample M_total to be detected is divided into three parts according to the particle size range:

[0088] Ultra-fine particle group d < d1, such as d1 = 5mm: mass M_fine. It is separated by a pre-screening device such as a vibrating screen and directly weighed.

[0089] Main particle group d1 ≤ d ≤ d2, such as d2 = 400mm: mass M_main. It is guided to the dynamic video shooting system.

[0090] Oversized particle group d > d2: mass M_oversize. After manual or mechanical sorting, static planar photos are taken.

[0091] Furthermore, the method further includes:

[0092] For the oversized particle group, the oversized particles in the group are laid flat in a single layer on a calibration background plate with known physical dimensions, and static images of the oversized particles are taken from multiple angles using a high-resolution camera;

[0093] The static images of the oversized particles are input into a pre-trained AI semantic segmentation model to extract the maximum projected contour pixel area of each particle;

[0094] For the main particle group, through the coordinated control of the conveyor belt and the vibrating screen, a single-layer and separated material flow is formed, and it makes a free-fall motion when entering in front of the high-contrast background plate, and a video sequence of its falling process is continuously recorded by a high-speed camera.

[0095] Among them, in this embodiment, for the oversized particle group: the stone materials are laid flat in a single layer on a calibration background plate with known dimensions, or on a site with calibration scales such as a ruler, and high-definition images are collected. An AI semantic segmentation model such as a model trained based on the U-Net architecture is used for image segmentation to extract the projected contour pixel area A_pixel_i of each particle.

[0096] Furthermore, the method further includes:

[0097] Based on the effective height of the background panel, the initial velocity of the stone when it enters the background panel, the camera frame rate, and the gravitational acceleration, the time required for the stone to pass through the background panel is calculated, and then the total number of frames is obtained.

[0098] Based on the principles of avoiding duplicate counting and ensuring capture integrity, the frame extraction interval N is determined, and one frame is extracted every N frames starting from the first frame.

[0099] All extracted frames and ultra-large particle static images are input into the locally deployed AI vision model to generate a binarized image with stone as the white foreground and black background.

[0100] Subsequently, connected component analysis was performed on the binarized image to extract the pixel area and principal axis size of each particle.

[0101] In this embodiment, for the main particle group:

[0102] Dynamic shooting: The stones are controlled in a single or near-single-layer, separated state by a conveyor belt and vibrating screen, and allowed to fall freely in front of a colored background of known size, while a high-speed camera records the video. The background color and the stone color create a high contrast, such as red or blue.

[0103] Intelligent frame skipping: To balance computational efficiency and avoid duplicate counting, the optimal frame skipping interval N is calculated.

[0104] Let the effective height of the background board be Hm, and the initial velocity of the stone entering the top of the background board be v0m / s. Based on the drop height h0, the stone falls freely before entering the top edge of the background board. The initial falling velocity of the stone is 0, which can be estimated by v0= sqrt(2gh0), where g is the acceleration due to gravity, and the camera frame rate is F fps.

[0105] The time t for the stone to pass through the background plate is t = (-v0 + sqrt(v0^2 + 2gH)) / g.

[0106] Then the total number of frames during the traversal period is F_total = t*F.

[0107] To ensure that each stone is clearly captured at least once in the video with minimal repetition, the frame extraction interval N should satisfy: 1 ​​≤ N ≤ floor(F_total / 2), preferably N = floor(F_total / k), where k is an empirical coefficient, typically 2-3. Starting from the frame where the stone first appears, extract one frame every N frames for subsequent analysis.

[0108] AI Image Binarization: Each extracted image frame is input into a locally deployed large-scale visual model such as SegmentAnything Model, SAM, or a dedicated AI segmentation model. The default instruction is to segment all falling stone targets in the image into white foreground and the rest into black background, resulting in a high-quality binary image sequence that effectively eliminates noise interference such as shadows and stains.

[0109] Feature extraction: Perform connected component analysis on the binary image to extract the pixel area A_pixel_j and principal axis size of each particle.

[0110] Furthermore, the method also includes:

[0111] The calibration coefficients are calculated based on the actual size of the calibration background and its pixel size ratio in the image;

[0112] Convert the pixel area of ​​each particle into its physical projected area;

[0113] An empirical functional relationship between experimental data and particle volume was established by fitting the data.

[0114] Calculate the uncalibrated mass of a single particle based on the premise that the density of the stone is constant;

[0115] The mass conversion factor is obtained based on the measured mass of sieving and the total mass calculated by the system identification. The mass conversion factor is obtained by parallel testing of samples from the same batch and the standard sieving method. The calculated mass of all particles in the main particle group and the ultra-large particle group is multiplied by K to obtain the calibrated mass.

[0116] In this embodiment, to balance the requirement of no severe obstruction or adhesion between particles and the requirement of high efficiency in processing large quantities per unit time, the system optimizes parameters through the following coupling relationship:

[0117] Let the conveyor belt feeding rate be Q_v (m³ / s), the effective width of the vibrating screen be B (m), the average spreading thickness of the stone on the screen surface be h (m), and the average forward speed of the stone on the screen surface be v_s (m / s). Then the volumetric flow rate of the stone passing through the background plate per unit time should be equal to the feeding rate, and reflected in the image as the percentage coverage of the projected area of ​​particles per unit frame, C.

[0118] Define C = (sum of projected areas of all particles) / (area of ​​background image). Experiments show that when C < 30%, the particle occlusion rate is low and the AI ​​recognition accuracy is high. C is positively correlated with Q_v, h, v_s, and the average particle size d_avg. Therefore, the control strategy is: after initial settings, analyze the C value of video frames in real time. If C is consistently higher than the threshold, reduce the feeding rate Q_v or increase v_s; if C is consistently too low, appropriately increase Q_v to improve detection efficiency while ensuring recognition accuracy.

[0119] Furthermore, the method also includes:

[0120] Perform data normalization and synthesis to include the total weighing mass of the ultrafine particle group in the cumulative mass;

[0121] The calibrated mass of each particle identified in the main particle group and the super-large particle group was calculated;

[0122] Sort all identified particles in descending order of equivalent particle size.

[0123] In this embodiment, all of M_fine is included in the cumulative mass less than d1.

[0124] For each particle identified in the main particle group and the super-large particle group, its calibrated mass m_calibrated = K*m.

[0125] Sort all particles, including the calculated bulk and super-large particles, in descending order of equivalent particle size, such as the major axis of an ellipse or the diameter of a circle with equal area.

[0126] Furthermore, the method also includes:

[0127] Real-time monitoring of the coverage rate, which is the ratio of the total area of ​​all particles projected within a unit frame of the dynamic video stream to the total area of ​​the background image.

[0128] When the coverage rate is higher than the preset ratio, the feeding rate will be automatically reduced or the conveying speed will be increased.

[0129] When the coverage rate is lower than the preset ratio, the feeding rate is increased while ensuring recognition accuracy, so as to achieve a dynamic balance between detection efficiency and recognition accuracy.

[0130] Furthermore, the method also includes:

[0131] The first set threshold d1 is 5 mm, and the second set threshold d2 is 400 mm;

[0132] The pre-screening device uses a vibrating screen with an aperture of 5 mm;

[0133] The main particle group is transported by a belt conveyor, and the ultra-large particle group is placed in the static image acquisition area after being sorted manually or mechanically.

[0134] The high-contrast background panel is made of matte red or blue material;

[0135] The light source system consists of two sets of high-brightness LED diffused lights, which are arranged on both sides of the background panel.

[0136] The initial velocity was calculated using the free fall of the stone before it entered the background plate.

[0137] Furthermore, the method also includes:

[0138] The AI ​​vision big model is a locally deployed variant of the Segment Anything Model, and the prompting mechanism is set to mark all falling rock targets in the image as foreground.

[0139] For particles with slight adhesion, secondary segmentation is performed by combining morphological opening operation and watershed algorithm after the model output mask;

[0140] The empirical function relationship is based on material-related parameters and uses a power function form.

[0141] Once the mass conversion factor is determined through a standard sieve calibration experiment, it can be reused under the same working conditions and the same batch of raw materials.

[0142] The quality conversion factor is updated based on a triggered recalibration according to changes in the construction section or significant environmental changes.

[0143] The equivalent particle size is based on the diameter of the equal-area circle of the particle projection profile as a unified benchmark.

[0144] The static image acquisition module acquires images of the ultra-large particle group from at least the top and two side angles, and then fuses them using a three-dimensional reconstruction algorithm to generate the maximum projection surface.

[0145] Image stitching employs feature point matching and perspective transformation techniques;

[0146] The real-time coverage analysis module samples one frame at preset intervals for statistical analysis.

[0147] The feedback control system connects the electric gate of the hopper to the frequency converter of the conveyor belt, forming a closed-loop regulation circuit.

[0148] According to a second aspect of the present invention, the present invention claims protection for an AI vision-based intelligent detection system for rockfill gradation, used to implement the AI ​​vision-based intelligent detection method for rockfill gradation, the system comprising:

[0149] The particle size separation module is used to separate ultrafine particles and ultralarge particles.

[0150] The dynamic image acquisition module includes a conveying and paving mechanism that enables the main aggregate to form a separated material flow, a high-contrast background plate, and a high-speed camera for capturing the free fall process of the aggregate.

[0151] The static image acquisition module is used to capture images of flat, ultra-large particle groups;

[0152] The computing and control center is used to control the operation of each module, execute algorithms for video frame extraction, AI image recognition, geometric information extraction, quality conversion, data fusion and gradation curve generation, and can dynamically adjust feeding or conveying parameters based on image feedback.

[0153] The following is a specific example:

[0154] This embodiment was conducted during the main rockfill filling construction of a high-core rockfill dam at a hydroelectric power station. The test object was a batch of main rockfill material after on-site compaction, with a designed maximum particle size of 800mm and a total sample volume of approximately 1.5 cubic meters. Based on the core concept of this invention and the characteristics of the material source, the distribution particle sizes were set as d1=5mm and d2=400mm. The stone material tested in the pit after compaction is as follows... Figure 2 As shown:

[0155] System configuration and parameters:

[0156] Particle size separation module: The vibrating screen has a 5mm aperture; large particles larger than 400mm are manually sorted out on-site.

[0157] Dynamic image acquisition module:

[0158] Conveying and paving: The process involves feeding hopper → conveyor belt with a width of 1.0m and a speed of 0.25m / s → vibrating screen with a screen size of 5mm and a width of 1.0m, ensuring that the 5-400mm diameter stones fall in a single layer and in a separated state.

[0159] Shooting system: Employs an industrial high-speed camera with a resolution of 1936×1464 and a frame rate of 240fps. The backdrop is a bright red matte board with physical dimensions of 1.5m high × 1.2m wide, and is 1.5m horizontally from the drop point.

[0160] Light source: Two sets of high-brightness LED diffused light sources are placed on both sides of the background panel to ensure uniform lighting.

[0161] Static image acquisition module: Use the same camera to take multi-angle photos of ultra-large particle groups d>400mm laid flat on the known grid size of the calibration mat to ensure that their maximum projection surface is obtained.

[0162] Computing and Control Center: Workstations equipped with high-performance GPUs, running locally deployed AI visual segmentation models and self-developed analysis software.

[0163] Implementation steps and application of core algorithms:

[0164] Step S1: Sample triage and pretreatment. A 1.5 m³ sample was passed through a 5 mm vibrating sieve. The mass of fine particles (d < 5 mm) M_fine = 22.5 kg, accounting for approximately 1.5% of the total sample mass. Oversized particles (> 400 mm) were manually removed, weighing M_oversize = 1500 kg, accounting for approximately 10%. The remaining particles, the main group (5-400 mm), had a mass M_main ≈ 13327.5 kg, obtained from the sieving experiment, accounting for approximately 88.5%.

[0165] Multi-mode image acquisition:

[0166] Large particle size: Take a picture after laying a single layer to obtain a high-definition image.

[0167] Main Particle Group: Start the system to feed materials. Key frame extraction calculation: Given the background board height H = 1.5m, drop height h0 = 0.1m, the initial velocity v0 = sqrt(2 * 9.8 * 0.1) ≈ 1.4m / s. Using t = (-1.4 + sqrt(1.4^2 + 2 * 9.8 * 1.5)) / 9.8 ≈ 0.47s, the total number of frames F_total = 2400.47 ≈ 113 frames. To ensure each stone is clearly captured, an empirical coefficient k = 3 is used, resulting in a frame extraction interval N = floor(113 / 3) ≈ 38 frames. The system automatically identifies the first frame when a stone enters the frame and uses this as the first frame, extracting one frame every 38 frames for subsequent analysis.

[0168] AI recognition and feature extraction:

[0169] All still photos and dynamic frame-by-frame images were input into the AI ​​model for segmentation and recognition to obtain accurate binary images. The pixel area of ​​each particle was extracted through connected component analysis. Based on the correspondence between the physical size of the background and the pixel size, the calibration coefficient S≈1.04mm / pixel was calculated. The video frame images before and after AI recognition are shown below. Figure 3 and Figure 4 As shown;

[0170] Quality conversion and data fusion:

[0171] Volume and mass estimation: Using an empirical conversion model of particle projected area and volume, combined with the stone density ρ, the uncalibrated mass of each identified particle is calculated.

[0172] System calibration: The uncalibrated total mass of this invention was measured to be M_calc = 13238.9 kg, while the actual mass measured by traditional sieving was M_actual = 13327.5 kg. Therefore, the mass conversion factor K = 13327.5 / 13238.9 ≈ 1.0067.

[0173] Data Synthesis: The total calibrated mass was obtained by summing the following: M_fine 22.5 kg, the calibrated mass of the ultra-large particle group (1.0067*1500≈1510.04 kg), and the calibrated mass of the main particle group (1.0067*13238.9≈13327.5 kg). All identified particles were sorted by equivalent particle size, the cumulative mass percentage was calculated, and a complete gradation curve was plotted.

[0174] Efficiency optimization control:

[0175] The system analyzes the particle coverage C in the dynamic video stream in real time and sets a threshold C_max=28%. When C approaches this threshold, the system automatically controls the feeding rate to stabilize it at about 25%, maximizing detection efficiency while ensuring recognition accuracy.

[0176] Verification results:

[0177] Reference Figure 5 The gradation curves obtained by this invention are compared with those obtained by traditional sieving methods. The results show that:

[0178] Accuracy: Within the range of particle size greater than 5mm, the two curves have extremely high agreement, and the relative errors of the key characteristic particle sizes d60, d30, and d10 are all less than 2.5%, which fully meets the engineering control standards.

[0179] Efficiency: The entire detection process of this invention, including sample processing, image acquisition, AI analysis, and report generation, takes approximately 9 minutes. Traditional screening methods take more than 4 hours to process a representative 1.5m³ sample.

[0180] Representativeness: The method of this invention processed a complete sample of 1.5 m³, while the traditional sieving method only obtained a few key control points of the gradation curve. Therefore, the results of this invention are more representative of the overall gradation of the material yard.

[0181] This embodiment strictly followed the updated parameters: maximum particle size 800mm, d2=400mm, resolution 1936*1464, etc. The results once again verified the superior performance of the method of this invention for the detection of gradation of large-sample-volume riprap with an ultra-wide particle size range of 0-800mm. Through the core strategy of three-segment processing and AI visual recognition, the imaging challenges of ultra-large-diameter stones and the identification challenges of fine-grained materials were successfully solved, achieving a balance between accuracy, efficiency, and representativeness.

[0182] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0183] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0184] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A method for intelligent detection of gradation of riprap based on AI vision, characterized in that, The specific steps include the following: S1: The rockfill sample to be tested is divided into multiple particle size groups according to particle size, and different strategies are used to preprocess the multiple particle size groups. S2: Based on the dynamic video stream of the multiple particle size groups, image features are extracted to obtain the pixel area of ​​the particle size groups; S3: Based on the calibration background plate, convert the pixel area of ​​each particle into the physical projection area, establish the empirical function relationship between the pixel area and the particle volume, calculate the uncalibrated mass of a single particle, and calculate the calibrated mass of all particles in the multiple particle size groups. S4: Perform data fusion on the particle mass of the multiple particle size groups, and sort all identified particles from largest to smallest according to their equivalent particle size; S5: Calculate the total calibration mass based on the total number of identified particles and the calibration mass of each particle, calculate the mass percentage of particles with a preset particle size, and plot the complete particle size distribution curve based on the mass percentage of particles with the preset particle size. The rockfill samples to be tested were divided into ultrafine particle group, main particle group and ultra-large particle group according to particle size. The ultrafine particle group with a particle size smaller than the first set threshold is separated by a pre-screening device and directly weighed to obtain the particle mass of the ultrafine particle group. Large particle groups with a particle size greater than the second set threshold are manually or mechanically sorted and guided to the static image acquisition area; The remaining particle groups with particle sizes between the first and second set thresholds are then transported to the dynamic image acquisition system. The first set threshold d1 is 5 mm, and the second set threshold d2 is 400 mm; The pre-screening device uses a vibrating screen with an aperture of 5 mm; The main particle group is transported by a belt conveyor, and the ultra-large particle group is placed in the static image acquisition area after being sorted manually or mechanically. High-contrast background panels use matte red or blue material; The light source system consists of two sets of high-brightness LED diffused lights, which are arranged on both sides of the background panel. The initial velocity was calculated using the free fall of the stone before it entered the background panel. The AI ​​vision big model is a locally deployed variant of the Segment Anything Model, and the cueing mechanism is set to mark all falling rock targets in the image as foreground. For particles with slight adhesion, secondary segmentation is performed by combining morphological opening operation and watershed algorithm after the model output mask; The empirical function relationship is based on material-related parameters and uses a power function form. Once the mass conversion factor is determined through a standard sieve calibration experiment, it can be reused under the same working conditions and the same batch of raw materials. The quality conversion factor is updated based on a triggered recalibration according to changes in the construction section or significant environmental changes. The equivalent particle size is based on the diameter of the equal-area circle of the particle projection profile as a unified benchmark. Images of ultra-large particle groups are acquired from at least the top and two side angles, and then fused using a 3D reconstruction algorithm to generate the maximum projection surface. Image stitching employs feature point matching and perspective transformation techniques; The real-time coverage analysis module samples one frame at preset intervals for statistical analysis; The feedback control system connects the electric gate of the hopper to the frequency converter of the conveyor belt, forming a closed-loop regulation circuit.

2. The intelligent detection method for riprap gradation based on AI vision according to claim 1, characterized in that, Also includes: For the ultra-large particle group, the ultra-large particle group particles within the group are laid in a single layer on a calibration background plate with known physical dimensions, and static images of the ultra-large particle group particles are captured from multiple angles using a high-resolution camera. The static image of the ultra-large particle group is input into a pre-trained AI semantic segmentation model to extract the maximum projected contour pixel area of ​​each particle. The main particle group is controlled by a conveyor belt and a vibrating screen to form a single layer of material flow that is separated from each other. When it enters the front of the high-contrast background plate, it undergoes free fall motion, and a high-speed camera continuously records a video sequence of its falling process.

3. The intelligent detection method for riprap gradation based on AI vision according to claim 2, characterized in that, Also includes: Based on the effective height of the background panel, the initial velocity of the stone when it enters the background panel, the camera frame rate, and the gravitational acceleration, the time required for the stone to pass through the background panel is calculated, and then the total number of frames is obtained. Based on the principles of avoiding duplicate counting and ensuring capture integrity, the frame extraction interval N is determined, and one frame is extracted every N frames starting from the first frame. All extracted frames and ultra-large particle static images are input into the locally deployed AI vision model to generate a binarized image with stone as the white foreground and black background. Subsequently, connected component analysis was performed on the binarized image to extract the pixel area and principal axis size of each particle.

4. The intelligent detection method for riprap gradation based on AI vision according to claim 3, characterized in that, Also includes: The calibration coefficients are calculated based on the actual size of the calibration background and its pixel size ratio in the image; Convert the pixel area of ​​each particle into its physical projected area; An empirical functional relationship between experimental data and particle volume was established by fitting the data. Calculate the uncalibrated mass of a single particle based on the premise that the density of the stone is constant; The mass conversion factor is obtained based on the measured mass of sieving and the total mass calculated by the system identification. The mass conversion factor is obtained by parallel testing of samples from the same batch and the standard sieving method. The calculated mass of all particles in the main particle group and the ultra-large particle group is multiplied by K to obtain the calibrated mass.

5. The intelligent detection method for riprap gradation based on AI vision according to claim 4, characterized in that, Also includes: Perform data normalization and synthesis to include the total weighing mass of the ultrafine particle group in the cumulative mass; The calibrated mass of each particle identified in the main particle group and the super-large particle group was calculated; Sort all identified particles in descending order of equivalent particle size.

6. The intelligent detection method for riprap gradation based on AI vision according to claim 5, characterized in that, Also includes: Real-time monitoring of the coverage rate, which is the ratio of the total area of ​​all particles projected within a unit frame of the dynamic video stream to the total area of ​​the background image. When the coverage rate is higher than the preset ratio, the feeding rate will be automatically reduced or the conveying speed will be increased. When the coverage rate is lower than the preset ratio, the feeding rate is increased while ensuring recognition accuracy, so as to achieve a dynamic balance between detection efficiency and recognition accuracy.

7. An AI vision-based intelligent detection system for riprap gradation, used to implement the AI ​​vision-based intelligent detection method for riprap gradation as described in any one of claims 1-6, characterized in that, The system includes: The particle size separation module is used to separate ultrafine particles and ultralarge particles. The dynamic image acquisition module includes a conveying and paving mechanism that enables the main aggregate to form a separated material flow, a high-contrast background plate, and a high-speed camera for capturing the free fall process of the aggregate. The static image acquisition module is used to capture images of flat, ultra-large particle groups; The computing and control center is used to control the operation of each module, execute algorithms for video frame extraction, AI image recognition, geometric information extraction, quality conversion, data fusion and gradation curve generation, and can dynamically adjust feeding or conveying parameters based on image feedback.

Images