A method of grading superimposed granular gravel

Abstract

The application discloses a superimposed particle size gravel grading method, and relates to the technical field of image data processing and automatic grading of industrial bulk materials. A conveying gravel image is acquired, and an initial gravel connected region is obtained through image enhancement and region segmentation. Adhesion connected regions are identified according to the texture homogeneity change quantity on both sides of a candidate segmentation line and are secondarily segmented. The area, length-width ratio and centroid coordinates of the gravel separation region are extracted, the equivalent particle size is obtained from the area, the fuzzy membership degree is corrected according to the length-width ratio, and a particle size grade label is formed. Parallel shunt control instructions are generated according to the label and the centroid coordinates, and multi-particle size parallel landing is completed. Verification data are read to form a particle size boundary directional grading error, and particle size boundary parameters are updated for subsequent determination. The application can reduce the cross-class shunt error caused by the adhesion of the gravel area, the misjudgment of the irregular particle shape boundary and the long-term drift of the particle size boundary.

Description

Technical Field

[0001] This invention relates to the fields of image data processing and automatic classification of industrial bulk materials, specifically to a method for classifying crushed stone with superimposed particle size. Background Technology

[0002] Crushed stone and coarse aggregates are widely used in road base courses, concrete structures, railway ballast, mining processing, and other civil engineering and industrial material production processes. Different engineering applications typically correspond to different particle size grades and gradation requirements, and the particle size classification affects subsequent batching, conveying, mixing, and construction quality control processes.

[0003] Traditional crushed stone particle size classification mainly relies on mechanical screening structures. Mechanical screening can separate particles according to the size of the screen openings. However, under continuous production conditions, screen wear, clogging, changes in the dust content of the material, and changes in production line conditions all increase the workload of equipment maintenance and parameter adjustment. With the development of industrial image processing technology, acquiring images of crushed stone in the conveyor state through industrial cameras, using image processing algorithms to extract information such as crushed stone contour, area, major axis, minor axis, particle size, and particle shape, and then performing classification control based on the image analysis results has become a way to realize intelligent detection and automatic sorting in continuous production lines.

[0004] Images of industrial crushed stone exhibit characteristics distinct from those of regular industrial parts. The surface texture of crushed stone is uneven, with random edge morphology. Dust can create isolated bright spots, localized dark spots, and continuous grayscale disturbances. During transport, multiple pieces of crushed stone may also come into contact, be adjacent, or have partially overlapping projections. When two or more pieces of crushed stone form the same connected region in a binary image, directly calculating the area of ​​this connected region will merge the projected areas of multiple pieces, leading to area magnification errors in subsequent equivalent particle size calculations. Even for correctly separated individual crushed stone regions, simply converting the equivalent circle diameter based on the projected area may still result in misjudgment of particle size boundaries. Blocky and elongated crushed stone may have similar projected areas, but their spatial dimensions and particle shapes differ. Using a fixed hard threshold can cause misjudgment of crushed stone with equivalent particle sizes near adjacent grade boundaries.

[0005] Continuous production lines also suffer from long-term parameter drift issues. Changes in ore source can alter particle shape distribution; variations in camera operating temperature and installation status can slowly change image scale calibration relationships; and aging of the lighting system and dust adhesion can alter image edge quality. If all particle size boundaries are adjusted uniformly based solely on overall classification accuracy, it's impossible to determine which adjacent grade boundary the error originates from, or whether the boundary needs adjustment towards larger or smaller particle sizes. Therefore, a superimposed particle size classification method is needed that can process images of adhered gravel, correct critical particle size assignments using particle shape information, apply image determination results to parallel physical sorting, and selectively update corresponding particle size boundary parameters using silo quality verification results. Summary of the Invention

[0006] The purpose of this invention is to provide a superimposed particle size classification method for solving the problems of image adhesion causing area magnification during continuous conveying, misjudgment of critical particle size boundaries due to irregular particle shapes, and drift of particle size boundary parameters due to long-term operation.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for classifying aggregates by superimposed particle size distribution includes the following steps: Acquiring multi-directional raw images of crushed stone: Crushed stone is conveyed through the image acquisition area via an industrial conveyor path. Multi-directional image acquisition units simultaneously acquire raw images of the crushed stone from different directions, resulting in multi-directional raw images of the crushed stone. Each multi-directional raw image carries a unified timestamp. The image processing server establishes a correspondence between the same crushed stone target from different viewpoints based on the timestamp and spatial calibration relationship. In some embodiments, an ambient light detection element is configured in the image acquisition area. The ambient light detection results, along with the image's high-brightness saturation ratio and low-brightness underexposure ratio, are used to adjust the exposure time of subsequent images.

[0008] Image enhancement and initial debris region segmentation are performed: The image processing server establishes a background model of the transport path and obtains the background residual image between the current debris image and the background model. In some embodiments, the background residual image is divided into multiple background residual image blocks, and the normalized continuous noise variance and impulse noise density are calculated for each block. Gaussian filter fusion weights are calculated based on the two noise states, and the Gaussian filter result and the median filter result are fused. After background subtraction, automatic thresholding, and contour restoration, the fused and denoised image forms the initial debris connected region.

[0009] Identifying Adhesive Connected Regions and Performing Secondary Segmentation: Closed contour chains are extracted from the initial gravel connected regions. Local contour corners and concavity depths are analyzed to obtain contour concave points. Concave point combinations satisfying distance and orientation constraints form candidate segmentation lines. Texture neighborhoods are established on both sides of each candidate segmentation line's normal direction. The homogeneity of the gray-level co-occurrence matrix of the texture neighborhood is calculated, and the change in texture homogeneity on both sides is calculated. When the change in texture homogeneity reaches the homogeneity abrupt change threshold, the corresponding candidate segmentation line is retained. Region segmentation is performed according to the retained candidate segmentation lines, and connectivity and minimum region area checks are performed on the segmented regions to form gravel separation regions. In some embodiments, candidate segmentation lines formed from the main viewpoint are also confirmed for consistency through an auxiliary viewpoint. The candidate segmentation lines are projected onto the auxiliary viewpoint image, and the edge response and texture homogeneity change within the projection search neighborhood are calculated to form cross-viewpoint support. Only candidate segmentation lines whose cross-viewpoint support reaches the segmentation line confirmation threshold enter secondary segmentation.

[0010] Extracting area, aspect ratio, and centroid coordinates: For each fragmentation separation region, the pixel area is statistically analyzed, the long and short sides of the minimum bounding rectangle are calculated to obtain the aspect ratio, and the centroid coordinates are obtained based on the region's pixel coordinates. Pixel area and pixel physical scale calibration parameters are used to calculate the equivalent particle size. For the... For each gravel separation zone, the equivalent particle size is calculated using the following formula:

[0011] in, For the first Equivalent particle size of each crushed stone separation zone; Parameters for pixel physical scale calibration; For the first The pixel area of ​​each gravel separation region; Pi is the mathematical constant of a circle.

[0012] The aspect ratio is calculated using the following formula:

[0013] in, For the first The aspect ratio of each gravel separation zone; For the first The length of the longest side of the minimum bounding rectangle of the gravel separation region; For the first The length of the short side of the smallest circumscribed rectangle of the gravel separation region.

[0014] Perform particle shape coupling fuzzy particle size level determination: Establish fuzzy overlap intervals between adjacent particle size levels based on the current particle size boundary parameters. When the equivalent particle size falls into the fuzzy overlap interval, calculate the initial fuzzy membership degree of the adjacent particle size levels. (The text then repeats the process of determining particle size levels based on particle shape coupling.) The shape penalty coefficient corresponding to each particle size level is used to correct the initial fuzzy membership degree:

[0015] in, For the first The first gravel separation area is related to the first Corrected fuzzy membership degree for each particle size level; The initial fuzzy membership degree; This is the shape penalty coefficient; The aspect ratio is 1.

[0016] Particle size classification labels are determined based on the maximum corrected fuzzy membership degree:

[0017] in, For particle size classification labels; This is a set of particle size level indexes.

[0018] In some embodiments, the width of the blurred overlap region is dynamically adjusted according to the image grain size uncertainty, which can be formed according to the following formula:

[0019] in, For the first Image particle size uncertainty of each gravel separation region; Confidence level for the split boundary; This is the normalized result of the pixel physical scale drift. This is the normalized result of the multi-view particle size discretization. , and These are the uncertainty weights corresponding to boundary confidence, scale drift, and multi-view dispersion, respectively. All three weights are non-negative and their sum is 1.

[0020] The dynamic overlap half-width of the corresponding particle size boundary is:

[0021] in, For the first The first gravel separation zone in the first Dynamic overlap half-width near the particle size boundary; For the first The basic overlap half-width of each particle size boundary; For uncertainty width mapping coefficients.

[0022] Forming diversion control commands: In the basic implementation, the target grading path is determined based on the particle size classification label and centroid coordinates, and the execution timing is determined based on the transmission speed and the distance between the image acquisition area and the grading execution area. In some embodiments, the initial motion velocity is obtained by associating targets with consecutive image frames, and the wind resistance acceleration coefficient is formed using equivalent particle size. The wind resistance acceleration coefficient is:

[0023] in, For the first Each gravel separation zone corresponds to the wind resistance acceleration coefficient of the gravel; air density; For the first The resistance coefficient of the crushed stone corresponds to the crushed stone in each crushed stone separation zone; Density of crushed stone material; For the first The equivalent particle size of each gravel separation zone.

[0024] Discrete trajectory state update is as follows:

[0025] in, For the first The gravel in the first The horizontal velocity at each discrete moment of the trajectory; For the first The gravel in the first The vertical downward velocity at each discrete moment of the trajectory; The discrete time step of the trajectory; It is the acceleration due to gravity; For the first The gravel in the first The physical x-coordinate of a discrete moment in a trajectory; For the first The gravel in the first The physical vertical coordinates of a trajectory at discrete moments; This is an index for discrete time points of the trajectory.

[0026] When the trajectory reaches the hierarchical execution plane, the predicted landing point coordinates and predicted arrival time are obtained. In some embodiments, the target association residual, centroid coordinate error, and initial velocity error are used to form the trajectory state covariance matrix, which is then recursively derived according to the state transition Jacobian matrix:

[0027] in, For the first The gravel in the first The trajectory state covariance matrix at discrete moments of the trajectory; The state equation of the discrete trajectory in the th The state transition Jacobian matrix for each discrete moment of the trajectory; The noise covariance matrix of the trajectory process; This indicates the matrix transpose.

[0028] The earliest and latest arrival times are determined using the estimated arrival times and the standard deviation of arrival times:

[0029] in, The earliest arrival time; The latest arrival time; To predict the arrival time; This is the coverage factor for time uncertainty; This represents the standard deviation of the arrival time.

[0030] The time window for outputting the shunt control command is:

[0031] in, For the first The output time window for the diversion control command corresponding to each piece of crushed stone; The delay in the response of the implementing agency; To control communication delays.

[0032] Parallel grading and collection are performed: The grading execution end receives the diversion control command, selects the target collection path according to the particle size classification label, and controls the corresponding grading execution channel according to the diversion execution time window. The superimposed grading collection mechanism can adopt a pneumatic diversion, servo diversion plate, switchable guide slot, or a combination of multiple execution methods. The core method does not limit the grading execution end to using a certain execution medium.

[0033] Perform quality verification and generate directional graded errors: The quality verification window can be divided according to a fixed time, a fixed processing quality, or a fixed amount of crushed stone. For each quality verification window, the predicted particle size distribution and the particle size verification distribution corresponding to different collection paths are recorded, and a correspondence between the predicted distribution and the verification distribution is established according to the quality verification window identifier and the collection path identifier. For the first... For each particle size boundary, the number of erroneous shunting flows to higher particle size classes and lower particle size classes are counted respectively. The directional classification error is:

[0034] in, For the first The first quality verification window Directional grading error corresponding to each particle size boundary; For the first Within the quality verification window, crossing the first The number of errors that are assigned to higher particle size levels at each particle size boundary; For the first Within the quality verification window, crossing the first The number of errors that are assigned to lower particle size levels at each particle size boundary; For the first Within the quality verification window and the first The number of validation samples associated with each particle size boundary; Index for quality verification window.

[0035] The particle size boundary parameters are updated as follows:

[0036] in, For the first The quality verification window uses the first One particle size boundary parameter; The updated particle size boundary parameters; For the first Boundary update coefficients for each particle size boundary.

[0037] The closed-loop stopping condition is:

[0038] and:

[0039] Both conditions are maintained continuously After one quality verification window, pause the updating of particle size boundary parameters; among which... For the first Graded error rate of each quality verification window; The threshold for the allowable grading error rate; This refers to the allowable variation threshold for particle size boundary parameters; The number of quality verification windows that continuously meet the stopping conditions.

[0040] In some embodiments, image particle size uncertainty is used to dynamically adjust the width of the blurred overlap interval. When the image boundary quality deteriorates, scale calibration drift increases, or the dispersion of multi-view particle size estimation increases, the soft decision region expands accordingly; when the image state is stable, the soft decision region shrinks accordingly. In some embodiments, the equivalent particle size continues to be used as the input for wind resistance trajectory calculation, and the target association residual in consecutive frames is further propagated into landing point and arrival time uncertainty. The hierarchical execution end no longer relies solely on fixed delay actions, but controls the hierarchical path switching based on the split execution time window.

[0041] Compared with the prior art, the present invention has the following beneficial effects: This invention identifies the segmentation location of connected regions by measuring the change in texture homogeneity in the texture neighborhood on both sides of the candidate segmentation line. This allows texture parameters to directly enter the region segmentation process. Compared to only calculating the area of ​​complete connected regions, this can prevent the data propagation path of multiple contacting gravel areas from merging and continuing into the equivalent particle size calculation. This invention obtains the area and aspect ratio from the same gravel separation region. The area is used to form the equivalent particle size, and the aspect ratio is used to correct the fuzzy membership of adjacent particle size levels. These two types of geometric features converge at the particle size level label after entering different calculation stages, enabling differentiated boundary determination for critical gravel with similar equivalent particle sizes but different particle shapes. This invention transmits the particle size level label and centroid coordinates to the physical separation process and returns the grain size verification results to the image processing process. The grading error is statistically analyzed according to the specific particle size boundary and error direction, updating only the particle size boundaries that have experienced directional shifts. This avoids synchronous disturbances to all particle size boundaries caused by changes in overall accuracy.

[0042] This invention identifies candidate segmentation lines for adhesion regions by varying the amount of texture homogeneity, forms equivalent particle sizes using the area of ​​the gravel separation region, corrects the fuzzy membership of adjacent particle size levels using the aspect ratio of the same gravel separation region, and then converts the particle size level labels and centroid coordinates into parallel flow control commands. After the grain is dropped into the bin, the particle size verification data corresponding to the quality verification window is returned to the image processing server to form directional grading errors corresponding to specific particle size boundaries, and selectively updates the corresponding particle size boundary parameters. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 is a schematic diagram of the overall architecture of the hierarchical system of the present invention.

[0045] Figure 2 is a schematic diagram of the main process of the superimposed particle size crushed stone classification method of the present invention.

[0046] Figure 3 is a schematic diagram of the hardware entity architecture of the hierarchical system of the present invention. Detailed Implementation

[0047] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0048] The following is in conjunction with the appendix Figures 1-3 The embodiments of the present invention will be described in detail below.

[0049] This invention discloses a method for classifying crushed stone by superimposed particle size, used for image analysis and parallel sorting of crushed stone under continuous industrial conveying conditions. After the crushed stone enters the image acquisition area, multi-directional image acquisition units simultaneously acquire crushed stone images from different directions. The image processing server forms a synchronous image frame group based on a unified timestamp and establishes the target correspondence in different viewpoints based on the spatial calibration parameters of multiple cameras.

[0050] The first-person view image is used for image enhancement and region segmentation, forming initial connected regions of gravel. For these initial connected regions where the contour structure indicates the possible contact of multiple gravels, the image processing server extracts contour concave points to form candidate segmentation lines. These candidate segmentation lines are not directly used for segmentation. The image processing server establishes texture neighborhoods on both sides of the candidate segmentation lines, calculates the homogeneity of the gray-level co-occurrence matrix, and determines whether the candidate segmentation lines correspond to real gravel contact boundaries based on the difference in homogeneity between the two sides. The confirmed candidate segmentation lines are used for secondary segmentation. The gravel separation region is then assigned area, aspect ratio, and centroid coordinates. The area and pixel physical scale calibration parameters together form the equivalent grain size. The equivalent grain size is used to locate grain size boundaries, and the aspect ratio is used to correct the fuzzy membership of adjacent grain size levels.

[0051] The particle size classification label and centroid coordinates are sent to the classification controller, which controls the classification execution channel according to the target classification path and execution sequence. After the crushed stone enters the corresponding collection path, the quality verification unit generates particle size verification data. The quality verification data corresponds to the image judgment result according to the quality verification window identifier and collection path identifier, forming the directional classification error corresponding to each particle size boundary. The particle size boundary parameters are selectively updated according to the directional classification error, and the updated result is returned to the subsequent crushed stone image judgment process.

[0052] The industrial conveying path can employ a belt conveyor, a vibrating conveyor, or other material conveying structures that allow crushed stone to continuously pass through the image acquisition area. The multi-directional image acquisition unit includes multiple industrial cameras. In the basic configuration, one main-view industrial camera and three auxiliary-view industrial cameras can be configured. The optical axis of the main-view camera faces the conveying plane and is primarily used for extracting the crushed stone's outline area and centroid coordinates. The auxiliary-view industrial cameras acquire the lateral outline and texture information of the crushed stone from different tilt directions. Each industrial camera is connected to a unified hardware trigger interface. Each hardware trigger generates a frame group identifier, and each industrial camera appends the frame group identifier and timestamp when outputting an image.

[0053] Let the main perspective be the first The timestamp of each image frame is , No. The timestamp of the image frame corresponding to each auxiliary viewpoint is The frame group time difference is:

[0054] in, The frame time difference between the auxiliary viewpoint and the main viewpoint; For image frame group index; For auxiliary perspective indexing.

[0055] When the frame time difference does not exceed the frame group time tolerance, the corresponding images are grouped into the same synchronized frame group. Each industrial camera completes intrinsic and extrinsic parameter calibration, and based on the spatial calibration parameters, projects the centroid of the main-view debris target onto the auxiliary-view image. For the main-view target and auxiliary-view candidate targets, a target matching cost can be established:

[0056] in, From the perspective of the main character The first goal and the first The first auxiliary perspective The cost of matching between targets; This represents the normalized projection centroid distance. This represents the normalized area ratio difference; To normalize the profile scale difference; , and These are the matching weights corresponding to centroid distance, area ratio difference, and contour scale difference, respectively. All three weights are non-negative and their sum is 1. This serves as an index for candidate targets from an auxiliary perspective.

[0057] Select the auxiliary viewpoint candidate target with the lowest matching cost and that does not exceed the corresponding threshold of the target to form a multi-view rockfall target group.

[0058] For example, take , , If the three normalized errors of a candidate target from an auxiliary viewpoint are 0.10, 0.08, and 0.15, respectively, then the matching cost is... If the threshold corresponding to the target is 0.20, then the target passes the matching.

[0059] In some embodiments, an ambient light detection element is configured in the image acquisition area. Since the ambient light state cannot solely reflect overexposure and underexposure of the gravel surface, the proportions of bright saturated pixels and low-brightness underexposure pixels in the gravel candidate area are simultaneously calculated. Exposure control error can be expressed as:

[0060] in, For exposure control error; The ratio of bright, saturated pixels; The proportion of low-brightness, underexposed pixels; Weights for highlight saturation penalty; The penalty weight is applied to low brightness and underexposure.

[0061] The subsequent release time has been updated as follows:

[0062] in, For the first Exposure time for each exposure control cycle; This is the updated exposure time; This is the exposure adjustment factor; To allow the shortest possible exposure time; To allow the longest possible exposure time; It is an interval cutoff function; This is an index for the exposure control cycle.

[0063] For example, if the current exposure time is 4ms, the highlight saturation pixel ratio is 0.12, the low-brightness underexposure pixel ratio is 0.03, and the two penalty weights are 1 and 0.5 respectively, then... When the exposure adjustment factor is 5ms, the exposure time of the next control cycle is... Images are acquired again after the exposure time is reduced.

[0064] Two main types of image interference exist in dusty environments: one manifests as continuous gray-level fluctuations, which can be described by local residual variance; the other manifests as isolated bright spots or isolated dark spots, which can be described by the pixel proportion deviating from the local median. The background residual image is divided into multiple background residual image blocks, the first... The continuous noise variance of each background residual image patch is formed after being normalized by the historical maximum variance. For the first For each background residual image patch, the difference between each pixel and the local median is calculated. When the difference exceeds the impulse discrimination scale, the corresponding pixel is recorded as an impulse noise candidate pixel. The number of impulse noise candidate pixels is divided by the total number of pixels in the image patch to obtain the result. The Gaussian filter fusion weights are:

[0065] For example, take , , , ,but If the Gaussian filter pixel value at a certain pixel location is 120 and the median filter pixel value is 116, then the fused denoising pixel value is... The denoised images are then fused and incorporated into the region segmentation process.

[0066] After system startup, background image sequences are acquired during periods when no debris passes through the transmission path. Pixel statistics are performed on the background image sequences to form a background model. Background subtraction is then performed between the fused denoised image and the background model.

[0067] in, The background difference pixel values; To fuse the pixel values ​​of the denoised image; These are the pixel values ​​of the background model.

[0068] Background subtraction results are automatically thresholded to form candidate foreground regions for gravel. The automatic thresholding method, such as the Otsu automatic thresholding method, can be used. After segmentation, morphological opening operations are performed to remove isolated small regions, and morphological closing operations are performed to fill small holes within the contour. The scale of the morphological structuring element is determined based on the equivalent scale calculated from the median area of ​​the current gravel candidate region. For example, if the median area of ​​the gravel candidate region in the current image corresponds to an equivalent circle radius of 50 pixels, the structuring element radius can be selected within the range of 0.04 to 0.10 of the equivalent circle radius. When 0.06 is selected, the structuring element radius is 3 pixels. After contour restoration, the initial connected gravel region is obtained.

[0069] For each initial connected region of gravel, a closed contour chain is extracted. A central contour point and two neighboring contour points are selected on the contour chain to form a local direction vector, and the local turning angle is calculated. When a local turning angle exhibits a concave structure and the concave depth reaches a concave point depth threshold, the corresponding contour point is recorded as a contour concave point. Concave points are then paired, and the pairing of contour concave points must satisfy the following conditions: the Euclidean distance between two contour concave points does not exceed the upper limit of the candidate line length; the angle between the candidate connection direction and the local normal of the two concave points satisfies the direction constraint; most segments of the candidate segmentation line are located inside the current initial connected region of gravel; and the areas of both estimated regions after the candidate segmentation line is divided are not less than the minimum effective region area. Concave point combinations that pass these four checks generate candidate segmentation lines.

[0070] For each candidate segmentation line, a texture neighborhood is established on both sides of the normal direction. The texture neighborhood only counts pixels located within the initial connected region of the gravel. The grayscale of the texture neighborhood is then compressed to... Each gray level, for texture direction The homogeneity of the normalized gray-level co-occurrence matrix is:

[0071] in, For the first The first initial gravel connected region The texture neighborhood on the first side of the candidate segmentation line in the texture direction Homogeneity; The homogeneity corresponds to the texture neighborhood on the second side; and These are the normalized gray-level co-occurrence matrix elements of the two texture neighborhoods, respectively; and These are grayscale indexes; This represents the normalized number of gray levels. Index for candidate dividing lines; This refers to the texture direction.

[0072] Averaging across multiple texture directions yields the overall homogeneity on both sides of the candidate segmentation line:

[0073] in, and These represent the combined homogeneity of two texture neighborhoods; For texture direction set; This represents the number of texture directions.

[0074] The variation in texture homogeneity is:

[0075] when At that time, candidate dividing lines are retained, among which The homogeneity mutation threshold is used. The homogeneity mutation threshold can be determined based on historical manually confirmed images of the adhesion area. The amount of texture homogeneity change is statistically analyzed for both the real contact boundary candidate lines and the false candidate lines inside the intact gravel, and the threshold that minimizes the total number of misclassifications for the two types of candidate lines is selected.

[0076] For example, if the overall homogeneity on both sides of a candidate dividing line is 0.82 and 0.41 respectively, then... If the homogeneity mutation threshold is set to 0.25, then the candidate segmentation line will proceed to the subsequent secondary segmentation.

[0077] Furthermore, cross-view projection is performed on the candidate segmentation lines from the main viewpoint: the two endpoints of the candidate segmentation lines from the main viewpoint are mapped to the auxiliary viewpoint image through camera spatial transformation. Due to depth and calibration errors, a single pixel position is not directly used in the auxiliary viewpoint; instead, a projection search neighborhood is established around the projection endpoints. For each auxiliary viewpoint, the normalized edge response in the projection search neighborhood is calculated. The normalized texture homogeneity variation on both sides of the projection position The cross-view support of the candidate dividing line is:

[0078] in, For the first The first gravel target Cross-view support of candidate dividing lines; A collection of effective auxiliary perspectives; For the first Effective weights for each auxiliary perspective; For edge response weights; This represents the texture variation weights. For example, if the effective weights for the two auxiliary viewpoints are 0.6 and 0.4 respectively, the edge response of the first auxiliary viewpoint is 0.8 and the texture variation is 0.7, while the edge response of the second auxiliary viewpoint is 0.5 and the texture variation is 0.6. , ,but If the threshold for confirming the dividing line is 0.60, then the candidate dividing line is confirmed through multi-view consistency.

[0079] The connected regions are segmented according to the confirmed candidate segmentation lines. When multiple candidate segmentation lines exist for the same initial gravel connected region, a ranking score is generated based on the change in texture homogeneity and cross-view support. The candidate segmentation line with the higher ranking score is executed first. After executing each candidate segmentation line, the region connectivity is recalculated. If the segmentation forms two independent connected regions and the areas of both regions are greater than the minimum effective region area, the segmentation result is accepted. If the area of ​​one segmented region is less than the minimum effective region area, the current segmentation is cancelled. If a subsequent candidate segmentation line conflicts with an already accepted segmentation line, the segmentation line with the higher ranking score is retained, and the gravel separated region is finally output.

[0080] For the Each fragmentation region, pixel area The number of pixels in the region is equal to the number of pixels in the region, and the centroid pixel coordinates are:

[0081] in, For the first The horizontal pixel coordinates of the centroid of the gravel separation region; The vertical pixel coordinates of the centroid; For the first A set of pixels representing the fragmentation region.

[0082] Calculate the minimum bounding rectangle for the region outline to obtain and Aspect Ratio equivalent particle size For example, take Pixels ,but If the longest side of the smallest bounding rectangle is 66 pixels and the shortest side is 30 pixels, then The two results are then used to determine the fuzzy particle size level.

[0083] The shape penalty coefficient is calibrated separately for each grain size level. Historical quality validation samples include the image's equivalent grain size, image aspect ratio, validation grain size level, and current grain size boundary parameters. Historical samples near the grain size boundaries are selected as boundary training samples. For the ... For each particle size level, the corrected fuzzy membership degree and particle size level label of the boundary training sample are calculated one by one in the candidate shape penalty coefficient set, defining the _th _th_ ... The boundary misclassification loss for each particle size class is:

[0084] in, For the first The shape penalty coefficient calibration loss for each particle size class; For the first The number of boundary training samples corresponding to each particle size level; For the first Predicted particle size levels for each boundary training sample; To correspond to the particle size classification; For indicator functions; Index the training samples for the boundary. Select a shape penalty coefficient that minimizes the loss from boundary misclassification; different particle size levels allow for different shape penalty coefficients.

[0085] For a particle size boundary parameter The basic overlap half width is When using a fixed overlap half-width, the fuzzy overlap range is: .

[0086] Furthermore, a dynamic overlap half-width is formed based on the image particle size uncertainty. The segmentation boundary confidence can be obtained from the edge gradient integrity and the region contour closure. The pixel physical scale drift can be obtained by periodically scaling the target or fixing the geometric reference in the transmission path. The multi-view particle size discretization is the degree of dispersion of the particle size estimates from different effective viewpoints.

[0087] For example, , , ,Pick , , ,but If the basic overlap half-width is 2mm and the uncertainty width mapping coefficient is 1.5, then The current fuzzy overlap region of the crushed stone target is expanded relative to the base width.

[0088] Taking a 20mm boundary as an example, with a current dynamic overlap half-width of 2mm, the overlap range is 18mm to 22mm. For the equivalent particle size... Crushed stone, low particle size class, initial fuzzy membership degree Initial fuzzy membership degree of high particle size order Take the aspect ratio. Low particle size class shape penalty coefficient High particle size class shape penalty coefficient After correction , Therefore, the particle size grade label should be a low particle size grade.

[0089] Furthermore, in some preferred embodiments, the same rock fragment target is associated in consecutive image frames, and the target association uses centroid distance, area ratio, and contour similarity. After the association is established, the centroid pixel coordinates at two time points are converted into physical coordinates, and the initial horizontal velocity is... Vertical initial velocity ,in and The physical x-coordinate of the centroid of two adjacent frames. and For the corresponding physical vertical coordinates, The frame interval.

[0090] For example, if the centroid x-coordinates of two frames are 0.100m and 0.117m respectively, and the frame interval is 0.033s, then... The initial velocity is then used in the impact point prediction process. , , , Then the drag acceleration coefficient Take the discrete time step of the trajectory. Initial horizontal velocity Initial vertical velocity 0, first step horizontal velocity First step vertical speed Continue iterating until the trajectory reaches the hierarchical execution plane.

[0091] Trajectory state covariance by Recursion. After the trajectory reaches the hierarchical execution plane, the uncertainty related to the arrival time in the covariance is used to form... .

[0092] For example, , , The earliest arrival time Latest arrival time If the actuator response is delayed Controlling communication delay The control command output time window is... The hierarchical controller completes the output of the diversion control signal within this window.

[0093] During continuous conveying, the time windows for diverting adjacent crushed stones may overlap. When two crushed stones enter the same particle size class, the corresponding grading channel can be kept in the target state to reduce repetitive actions; when two crushed stones belong to different particle size classes, the overlap ratio of the time windows should be calculated. When the overlap ratio of time windows is low, two actions are completed by refining the execution timing. When the overlap ratio of time windows is high and there are redundant channels of the same level, the redundant channel is selected. When there are no available redundant channels and the conflict cannot be resolved, the gravel with low confidence is transferred to the review path.

[0094] The quality verification unit is positioned at the detection location of each collection path. Quality verification data is not mandatory to be correlated one-to-one with each preceding gravel target; the basic implementation uses a quality verification window correspondence. For example, a quality verification window is formed every 5 minutes. The image processing server counts the number of particle size class labels, equivalent particle size distribution, aspect ratio distribution, and proportion of connected regions corresponding to each collection path within 5 minutes. The quality verification unit counts the particle size verification distribution within the same window and the same collection path, and uses the quality verification window identifier and collection path identifier for correspondence. When using a particle-by-particle re-inspection device, a particle-by-particle correspondence can be further established using the target time series.

[0095] Furthermore, Long Short-Term Memory (LSTM) neural networks do not replace real-time quality feedback. Each quality verification window forms a time-series feature vector. ,in The average area is... and It is the area quartile. The aspect ratio is the average. The third quartile of the aspect ratio. This represents the mean change in texture homogeneity. The proportion of the connected regions. This is the directional hierarchical error vector. This is the current particle size boundary parameter vector.

[0096] In one implementation, the Long Short-Term Memory (LSTM) neural network includes two cascaded LSM layers and a fully connected output layer. The first layer has a hidden state dimension of 64, and the second layer has a hidden state dimension of 32. The network gating computation includes a forgetting gate. Input gate Candidate memory states Memory state Output gate Hidden state The network outputs the candidate particle size boundary increment vector. The training loss is When applied online, if the candidate incremental direction is consistent with the directional grading error correction direction, it is truncated to within the allowable update range; if the directions are inconsistent, the update weight is reduced or the update is postponed.

[0097] Furthermore, in some preferred embodiments, a condition-triggered approach is used for three-dimensional verification. For the first... For each piece of gravel, obtain the maximum corrected fuzzy membership degree. Second largest modified fuzzy membership degree Particle size determination interval .

[0098] when Or the confidence level of the dividing boundary. At that time, a 3D verification is triggered, extracting only multi-view local images of the corresponding low-confidence targets. After image correction, disparity matching or feature matching, a 3D point cloud is formed. After outlier removal and registration, it is voxelized to determine the volume of the gravel. Three-dimensional composite particle size For example, if the voxel side length is 1 mm and the effective voxel count is 33510, then... , The three-dimensional particle size verification returns to the current particle size grade determination process, forming a verified particle size grade label.

[0099] In some embodiments, the system includes the following exception handling methods: When multi-view matching fails, the main view image continues to enter the basic image processing path; When the gravel area intersects with the image boundary, it is marked as an incomplete target and will not be included in the area and particle size determination. It will wait for subsequent complete image frames to be re-detected. If a candidate segmentation line is confirmed by texture homogeneity but fails the region area check, the current segmentation line is cancelled; when the target association confidence of consecutive image frames is too low, the landing point prediction process uses the transmission path calibration speed as a backup initial speed. When the trajectory prediction result exceeds the effective area of ​​all graded paths, the corresponding gravel target will be transferred to a safe collection path; When the execution time windows for the diversion of two different grades of crushed stone conflict and cannot be resolved using redundant channels, the crushed stone target with lower confidence in particle size determination will be transferred to the review path. When the material level in the collection path reaches the upper limit of the material level control, the grade controller prohibits further diversion to the corresponding collection path and switches to a redundant collection path of the same level. If there is no redundant collection path of the same level, the inlet feeding of the corresponding level is suspended or the material is switched to a temporary collection path. When the number of valid samples in the quality verification window is lower than the minimum number of update samples, the particle size boundary parameters are not updated, and the system waits for subsequent quality verification windows to accumulate verification data that meets the requirements.

[0100] In some embodiments, the multi-directional image acquisition unit can be arranged in a ring, top plus two sides, or segmented. Different arrangement methods establish the correspondence between multi-view targets through a unified timestamp and spatial calibration parameters.

[0101] In some embodiments, the background model can be updated using a selective moving average method, with the gravel foreground region not participating in the background update, and pixels that have been in the background state for a long time being updated gradually according to the update coefficient. In some embodiments, automatic thresholding can use an adaptive thresholding method based on the current image grayscale distribution, and the region segmentation method does not change the subsequent texture homogeneity confirmation, area and aspect ratio dual-input determination, and quality feedback closed-loop relationship.

[0102] In some embodiments, the cross-view support of the candidate segmentation line can be formed by two or more auxiliary views. When an auxiliary view is occluded, the effective weight of the corresponding view is reduced to zero, and other effective views are renormalized.

[0103] In some embodiments, the shape penalty coefficient is established as a parameter table according to particle size grade and material batch. After changing the ore source, the existing parameters of similar materials are used first, and then the quality verification window is used to gradually correct the parameters. In some embodiments, the image particle size uncertainty can also be supplemented with a target motion blur index. The motion blur index is used to increase the blur overlap range corresponding to the rapidly moving gravel targets with edge trailing.

[0104] In some embodiments, when the graded actuator uses pneumatic flow splitting, the flow splitting control command includes the air valve channel identifier, opening time, and closing time; when a servo flow splitter is used, the flow splitting control command includes the target angle, start time, and holding time.

[0105] In some embodiments, the quality verification unit uses laser diameter measurement; in other embodiments, it uses an independent re-inspection image unit; mechanical sampling measurement can also be performed according to a fixed quality cycle, and different quality verification methods can output a uniform particle size verification distribution.

[0106] In some embodiments, particle size boundary parameter updates are not performed after each crushing process, but rather according to the quality verification window, to avoid high-frequency boundary changes caused by a single abnormal detection object.

[0107] In some embodiments, the training data for the long short-term memory neural network comes from historical quality verification windows, and the temporal order between each input window and the corresponding reference boundary increment is maintained without disrupting the temporal dependency.

[0108] In some embodiments, the two-dimensional low-confidence conditional three-dimensional verification employs binocular parallax; in other embodiments, feature point matching and dense reconstruction from three or more perspectives are employed.

[0109] In some embodiments, the first layer of the dual-layer superimposed hierarchical path is responsible for the flow of coarse particle size groups, and the second layer is responsible for the flow of fine particle size levels within each coarse particle size group. The path selection of the first and second layers both originate from the same final particle size level label.

[0110] In some embodiments, the system saves the area distribution, aspect ratio distribution, texture homogeneity variation distribution, grain size level label distribution, and verification grain size distribution at the quality verification window level for parameter verification and slow timescale prediction.

[0111] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described below with reference to specific implementation examples.

[0112] Example 1: This example illustrates the core image processing, texture-assisted adhesion secondary segmentation, dual-input particle size classification based on area and aspect ratio, parallel flow splitting, and directional classification error feedback update process. An industrial conveyor path transports mixed-size gravel through the image acquisition area. Multi-directional image acquisition units simultaneously acquire gravel images. This example uses the main view image to perform core processing.

[0113] The main view image is processed by background subtraction, fusion denoising, automatic thresholding segmentation and contour restoration to form the initial gravel connected region.

[0114] For example, a processing window may contain 20 gravel objects that need to be judged. Initial region segmentation yields 16 initial gravel connected regions, 12 of which correspond to a single gravel, and the other 4 are formed by two contacting gravel. Closed contour chains are calculated for the 16 initial gravel connected regions. Each of the four adhered regions generates at least one candidate segmentation line that satisfies the contour concavity condition. The overall homogeneity on both sides of the four final candidate segmentation lines is as follows: Group 1: 0.82 and 0.41; Group 2: 0.76 and 0.40; Group 3: 0.70 and 0.39; Group 4: 0.74 and 0.45, corresponding to texture homogeneity variations of 0.41, 0.36, 0.31, and 0.29, respectively. A homogeneity abrupt change threshold of 0.25 is set, and all four candidate segmentation lines enter secondary segmentation. Each of the four adhered regions is divided into two gravel separation regions, resulting in 20 gravel separation regions.

[0115] The area, aspect ratio, and centroid coordinates of each debris separation region are calculated. For a debris separation region with an area of ​​8012 pixels and a pixel physical scale calibration parameter of 0.2 mm / pixel, the equivalent particle size is 20.20 mm, corresponding to a minimum bounding rectangle with a long side of 66 pixels and a short side of 30 pixels, and an aspect ratio of 2.2. Using a basic fuzzy overlap range of 18 mm to 22 mm, the initial fuzzy membership degree for the low particle size level is 0.45, and for the high particle size level it is 0.55. After applying a shape penalty coefficient of 0.05 for the low particle size level and 0.25 for the high particle size level, the corrected fuzzy membership degrees are 0.39 and 0.25, respectively, and the low particle size level label is output. The remaining debris obtains particle size level labels using the same process. The particle size level labels and centroid coordinates are sent to the grading controller, which selects the target collection path and calculates the execution time based on the particle size level labels.

[0116] Of the 20 crushed stone samples in this batch, the grading results of 18 samples were consistent with the verification grade, with a grading accuracy of 90%. In a subsequent quality verification window, for example, if 8 out of 100 verification samples related to the 20mm boundary were incorrectly categorized to higher particle size grades and 2 out of 100 were incorrectly categorized to lower particle size grades, the directional grading error would be 0.06. With a boundary update coefficient of 0.50mm, the 20.00mm boundary would be updated to 19.97mm. The updated particle size boundary parameters would then be used for fuzzy particle size determination in the next quality verification window. Once the continuous quality verification window reached the error rate condition and the boundary change condition, the current boundary parameter update would cease.

[0117] Example 2: This example is a further optimization based on Example 1, adding multi-view candidate segmentation line confirmation, image particle size uncertainty-driven dynamic fuzzy overlap interval, particle size coupling resistance trajectory, diversion execution time window, and long short-term memory neural network slow timescale prediction.

[0118] After generating candidate segmentation lines from the main view image, the candidate segmentation lines are projected onto two effective auxiliary viewpoints according to the camera spatial calibration parameters. The effective weights of the first auxiliary viewpoint are 0.6 and the second auxiliary viewpoint are 0.4. The normalized edge response of the first auxiliary viewpoint is 0.8 and the normalized texture change is 0.7, while the normalized edge response of the second auxiliary viewpoint is 0.5 and the normalized texture change is 0.6. The edge response weight is 0.4 and the texture change weight is 0.6. The cross-view support is calculated to be 0.668, and the segmentation line confirmation threshold is set to 0.60. The candidate segmentation lines are confirmed through multi-view consistency.

[0119] In the particle size determination stage, the segmentation boundary confidence score is 0.80, the normalized result of pixel physical scale drift is 0.05, and the normalized result of multi-view particle size discretization is 0.10. The three uncertainty weights are 0.5, 0.2, and 0.3, respectively. The image particle size uncertainty is 0.14, the basic overlap half-width is 2mm, the uncertainty width mapping coefficient is 1.5, and the dynamic overlap half-width is calculated to be 2.42mm. The equivalent particle size of the current gravel enters the dynamic fuzzy overlap interval for soft determination.

[0120] During the impact point prediction phase, the centroid abscissas of adjacent image frames are 0.100m and 0.117m, respectively, with a frame interval of 0.033s. The calculated initial horizontal velocity is approximately 0.515m / s, the equivalent particle size is 0.0202m, and the wind resistance acceleration coefficient is approximately... The trajectory is iterated in 0.01s increments, with the trajectory state covariance being recursively calculated. When the predicted arrival time is 0.410s, the standard deviation of the arrival time is 0.008s, and the time uncertainty coverage factor is 2, the earliest arrival time is 0.394s, the latest arrival time is 0.426s, the actuator response delay is 0.020s, and the control communication delay is 0.005s, forming a control command output time window of 0.369s to 0.421s.

[0121] In the historical trend prediction section, a network input sequence is composed of 12 consecutive quality validation windows. Each window generates area statistics, aspect ratio statistics, texture homogeneity change statistics, proportion of connected regions, directional hierarchical error vector, and grain size boundary parameter vector. Two cascaded long short-term memory layers are configured with 64-dimensional and 32-dimensional hidden states, respectively. The output layer dimension equals the number of updatable grain size boundaries. The network training objective is to predict the reference boundary increments generated by historical quality feedback. During online operation, the network output is only used as candidate boundary increments. When directional conflicts occur, the use of network candidate increments is paused, and real-time directional error updates continue.

[0122] Example 3: This example is based on the same principles as Example 1, but is applied to a high-precision crushing and grading production line, and adds condition-triggered multi-view three-dimensional verification and double-layer superimposed grading path.

[0123] The multi-directional image acquisition unit includes eight industrial cameras, forming four sets of binocular acquisition directions. Each industrial camera acquires images through a unified hardware trigger signal.

[0124] The two-dimensional processing flow remains consistent with Example 1. After determining the two-dimensional particle size level, the particle size determination interval between the maximum corrected fuzzy membership degree and the second largest corrected fuzzy membership degree is calculated. For example, if the maximum corrected fuzzy membership degree of a gravel target is 0.46 and the second largest corrected fuzzy membership degree is 0.42, then the particle size determination interval is 0.04. When the particle size determination interval threshold is 0.08, the current gravel enters the three-dimensional verification. The system only reads the multi-view local images of the corresponding gravel target, forms a local depth map after binocular correction and disparity matching, and converts it to a unified three-dimensional coordinate system. After statistical outlier removal, the point cloud is voxelized, and the voxel side length is 1mm. The effective number of voxels for a target is 33510, and the volume is 33510 mm^3. The three-dimensional verification particle size is calculated to be approximately 40mm, and the particle size level determination of the current gravel target is returned. If the two-dimensional particle size label is the third level and the three-dimensional verification result enters the fourth level, then the verified label is used to generate a diversion control command.

[0125] The grading execution end adopts a two-layer superimposed grading path. The first layer selects a primary path based on the coarse particle size group, and the second layer selects a secondary path based on the verified fine particle size level. The quality verification unit is set at the detection position of the secondary collection path. This embodiment also includes a material level status check. When a target path reaches the upper limit of the material level control, the grading controller queries redundant collection paths of the same level. If a redundant collection path exists, it switches to it; if no available redundant path exists, it controls the inlet feeding cycle or switches to a temporary collection path.

[0126] Comparative Example 1: This comparative example uses the same gravel input, the same image acquisition conditions, the same image enhancement parameters, the same initial region segmentation process, the same area calculation method, the same equivalent particle size formula, the same aspect ratio correction relationship, the same particle size boundary parameters, and the same parallel splitting process as Example 1. The only difference is that the candidate segmentation line is not confirmed by the change in texture homogeneity, and secondary segmentation is not performed on the four connected regions.

[0127] The initial image processing yielded 16 initial connected regions for gravel, with 12 regions corresponding to a single gravel and 4 adhered regions each containing 2 contacting gravels. Without secondary segmentation, the 4 pairs of contacting gravels were included in the equivalent grain size calculation based on their merged area. Since the merged area is larger than the area of ​​a single gravel, the equivalent grain size corresponding to the 4 pairs of contacting gravels shifted towards a higher grain size level. Using the same 20 object base as in Example 1, the 12 non-adhered objects maintained the correct level, while the 8 gravels corresponding to the 4 pairs of adhered objects were affected by the merged areas, resulting in a classification accuracy of 60% (12 / 20), compared to 90% (18 / 20) in Example 1. The only difference between the two processes is whether texture homogeneity variation is used to confirm candidate segmentation lines and perform secondary segmentation. This indicates that merging of adhered region areas will continue to propagate to the equivalent grain size and grain size level determination, while texture homogeneity-assisted secondary segmentation can block the corresponding error propagation path.

[0128] Comparative Example 2: This comparative example uses the same gravel input, the same image acquisition process, the same image enhancement process, the same texture homogeneity-assisted secondary segmentation, the same area calculation, the same equivalent particle size calculation, the same particle size boundary parameters, and the same parallel splitting process as Example 1. The only difference is that the aspect ratio correction for the initial fuzzy membership is removed, and the particle size level is determined only based on the equivalent particle size and the initial fuzzy membership.

[0129] Using the same 20 objects as in Example 1, 6 of these objects are located in the fuzzy overlap region of adjacent particle size levels and have a large aspect ratio. Without aspect ratio correction, 4 irregular boundary objects enter a higher particle size level different from the verification level, while the other 2 ordinary boundary objects still exhibit level differences. In this comparative example, 14 objects are consistent with the verification level, resulting in a classification accuracy of 70% (14 / 20), while in Example 1, the classification accuracy was 90% (18 / 20). The only difference between the two processes lies in whether the aspect ratio is included in the initial fuzzy membership correction. This demonstrates that when the equivalent particle size is near the boundary of adjacent levels, the calculation relationship between the aspect ratio and the corresponding shape penalty coefficient can alter the level competition results of irregularly shaped gravel.

[0130] To further enable those skilled in the art to understand the data transfer between the formulas in the continuous processing of this invention, the following examples are provided for further illustration: The following example illustrates the data transfer between formulas using a continuous processing procedure for a rock fragment target, from image acquisition to quality feedback.

[0131] An industrial camera captures the current image of the broken stones. The normalized continuous noise variance of the background residual image block is set to 0.12, the impulse noise density to 0.08, the noise balance coefficient to 1.5, the positive stability factor to 0.01, and the Gaussian filter fusion weight to 0.48. For a given pixel, the Gaussian filter pixel value is 120, the median filter pixel value is 116, and the fused denoised pixel value is 117.92. The fused denoised image, after background subtraction and region segmentation, forms an initial connected region of broken stones. A candidate segmentation line exists within this initial connected region, with comprehensive homogeneity of 0.82 and 0.41 on either side, a texture homogeneity change of 0.41, and a homogeneity abrupt change threshold of 0.25. The candidate segmentation line is confirmed through texture analysis.

[0132] Two auxiliary perspectives were used to further confirm the candidate segmentation line. The first auxiliary perspective had an effective weight of 0.6, an edge response weight of 0.8, and a texture change weight of 0.7. The second auxiliary perspective had an effective weight of 0.4, an edge response weight of 0.5, and a texture change weight of 0.6. The cross-view support was 0.668, and the segmentation line confirmation threshold was 0.60. The candidate segmentation line was confirmed. After performing secondary segmentation, one of the fragment separation regions was tracked. The pixel area was 8012 pixels, the pixel physical scale calibration parameter was 0.2 mm / pixel, the equivalent particle size was approximately 20.20 mm, the minimum bounding rectangle had a long side of 66 pixels and a short side of 30 pixels, and an aspect ratio of 2.2.

[0133] The segmentation boundary confidence is 0.80, scale drift is 0.05, multi-view particle size discretization is 0.10, uncertainty weights are 0.5, 0.2, and 0.3, image particle size uncertainty is 0.14, basic overlap half-width is 2mm, uncertainty width mapping coefficient is 1.5, and dynamic overlap half-width is 2.42mm. Within the current boundary soft decision interval, the initial fuzzy membership degree for low particle size level is 0.45, and for high particle size level it is 0.55. A shape penalty coefficient of 0.05 for low particle size level and 0.25 for high particle size level are applied. After correction, the low-level membership degree is 0.39 and the high-level membership degree is 0.25, and the low-particle-size level label is output.

[0134] In consecutive image frames, the centroid's horizontal coordinate changes from 0.100m to 0.117m over a time interval of 0.033s. The initial horizontal velocity is approximately 0.515m / s, the air density is 1.2kg / m³, the drag coefficient is 0.9, the density of the gravel material is 2700kg / m³, the equivalent particle size is 0.0202m, and the wind resistance acceleration coefficient is approximately... The trajectory distance from the walking distance is 0.01s. The horizontal velocity of the first step is approximately 0.51496m / s, and the vertical velocity of the first step is 0.0981m / s. Calculations continue to the grading execution plane. The predicted arrival time is 0.410s, the standard deviation of the arrival time is 0.008s, the time uncertainty coverage factor is 2, the earliest arrival time is 0.394s, the latest arrival time is 0.426s, the actuator response delay is 0.020s, the control communication delay is 0.005s, and the control command output time window is 0.369s to 0.421s. The grading controller selects the corresponding grading path based on the low particle size level label and completes the execution control within the window, allowing the crushed stone to enter the target collection path.

[0135] The subsequent quality verification window counts 100 verification objects related to the 20mm boundary. There are 8 errors at the higher level and 2 errors at the lower level. The directional grading error is 0.06. The current particle size boundary is 20.00mm, the boundary update coefficient is 0.50mm, and the updated boundary is 19.97mm. The fuzzy particle size level is then determined in the next quality verification window.

[0136] When the continuous quality verification window simultaneously meets the classification error rate condition and the particle size boundary change condition and continuously maintains the specified number of windows, the particle size boundary update is paused.

[0137] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for classifying aggregates by superimposed particle size distribution, characterized in that, include: Acquire original images of crushed stone from multiple directions during industrial conveying, perform image enhancement and crushed stone region segmentation on the original images of the crushed stone from multiple directions, and obtain initial connected regions of crushed stone; Candidate segmentation lines are formed based on the contour structure of the initial gravel connected region. The texture homogeneity change of the texture neighborhood on both sides of the candidate segmentation line is extracted. Adhesive connected regions are identified based on the texture homogeneity change. Secondary segmentation is performed along the candidate segmentation lines that satisfy the texture change conditions to obtain the gravel separation region. The area, aspect ratio, and centroid coordinates of the crushed stone separation area are extracted. The equivalent particle size is calculated based on the area. The initial fuzzy membership degree of adjacent particle size levels is determined based on the equivalent particle size. The initial fuzzy membership degree is corrected based on the aspect ratio to obtain the corrected fuzzy membership degree. Particle size level labels are generated based on the corrected fuzzy membership degree. Based on the particle size class label and the centroid coordinates, a diversion control command corresponding to the parallel classification path is generated and sent to the classification execution end to control the crushed stone of different particle size classes to enter the corresponding collection path, forming a multi-particle-size parallel silo result. Read particle size verification data corresponding to different collection paths within the quality verification window. Based on the particle size verification data and the particle size grade label, form a grading error feedback quantity corresponding to each particle size boundary. Selectively update the corresponding particle size boundary parameters based on the grading error feedback quantity, and use the updated particle size boundary parameters for subsequent particle size grade determination corresponding to the original images of crushed stone in multiple directions.

2. The method for classifying aggregates by superimposed particle size according to claim 1, characterized in that, The image enhancement includes: A background residual image is obtained based on the background image and the current debris image, and the background residual image is divided into background residual image blocks; The normalized continuous noise variance and impulse noise density are obtained from each background residual image block, and the Gaussian filter fusion weights are determined based on the normalized continuous noise variance and the impulse noise density. Gaussian filtering and median filtering are performed on the corresponding background residual image blocks respectively, and the Gaussian filtered pixel values ​​and median filtered pixel values ​​are fused according to the Gaussian filtering fusion weight to obtain the fused and denoised image; The Gaussian filter fusion weights and the fused denoised pixel values ​​are determined according to the following formula: in, For the first Gaussian filter fusion weights corresponding to each background residual image patch; For the first Normalized continuous noise variance of a background residual image patch; This is the noise balance factor between the continuous noise component and the impulse noise component; For the first Impulse noise density of a background residual image patch; To prevent positively stable quantities with a denominator of zero; For the first A background residual image patch at image pixel coordinates The pixel values ​​at the fused and denoised location; These are Gaussian filtered pixel values; These are the pixel values ​​filtered by median. Index for background residual image patches; The horizontal pixel coordinates of the image; These are the vertical pixel coordinates of the image.

3. The method for classifying aggregates by superimposed particle size according to claim 1, characterized in that, The step of identifying connected regions based on the amount of texture homogeneity variation and performing secondary segmentation includes: Extract the closed contour chain of the initial connected region of the gravel; Calculate the local rotation angle and concave depth of each contour point on the closed contour chain, and select the contour point that satisfies the local rotation angle condition and the concave depth condition as the contour concave point. Candidate segmentation lines are generated between contour concave points that satisfy both distance and orientation conditions. A texture neighborhood located inside the initial gravel connected region is constructed on both sides of the normal direction of each candidate segmentation line; Calculate the homogeneity of the gray-level co-occurrence matrix of two texture neighborhoods and obtain the change in texture homogeneity; Candidate segmentation lines are retained when the amount of texture homogeneity change reaches the homogeneity mutation threshold; Perform region segmentation on the connected regions according to the retained candidate dividing lines, perform connectivity checks and minimum region area checks on the region segmentation results, and output the gravel separation region.

4. The method for classifying aggregates by superimposed particle size according to claim 3, characterized in that, The original multi-directional image of the debris includes a main view image and at least two auxiliary view images. The candidate segmentation line is further confirmed through multi-view consistency verification, which includes: Based on the spatial calibration parameters of the multi-directional image acquisition unit, the endpoints of the candidate segmentation lines in the main view image are projected onto each auxiliary view image; Establish a projection search neighborhood around the projection position of the auxiliary viewpoint image; The edge response and texture homogeneity variation in the projection search neighborhood are obtained respectively; The cross-view support of candidate segmentation lines is formed based on the edge response, texture homogeneity variation, and effective view weights of each auxiliary view image. Candidate segmentation lines whose cross-view support reaches the segmentation line confirmation threshold are retained, and the confirmed candidate segmentation lines are input into the secondary segmentation process.

5. The method for classifying aggregates by superimposed particle size according to claim 1, characterized in that, Correcting the initial fuzzy membership degree based on the aspect ratio includes: The initial fuzzy membership degree corresponding to adjacent particle size levels is obtained based on the equivalent particle size and the current particle size boundary parameters. Obtain the shape penalty coefficient corresponding to each particle size grade; The initial fuzzy membership degree is corrected based on the deviation between the aspect ratio and the unit aspect ratio and the shape penalty coefficient of the corresponding particle size level. The corrected fuzzy membership degree and particle size classification label are calculated using the following formula: in, For the first The first gravel separation area is related to the first Corrected fuzzy membership degree for each particle size level; For the first The first gravel separation area is related to the first Initial fuzzy membership degree for each particle size level; For the first Shape penalty coefficient corresponding to each particle size class; For the first The aspect ratio of each gravel separation zone; For the first Particle size classification labels for each crushed stone separation zone; A set of particle size indexes; Index for the gravel separation region; This is an index for particle size levels.

6. The method for classifying aggregates by superimposed particle size according to claim 5, characterized in that, The fuzzy overlap interval between adjacent particle size levels is dynamically adjusted according to the image particle size uncertainty, which is determined by the segmentation boundary confidence, pixel physical scale drift, and multi-view particle size discretization. The image particle size uncertainty is mapped to the dynamic overlap half-width of the corresponding particle size boundary, and the lower limit of the fuzzy overlap interval is formed by subtracting the dynamic overlap half-width from the current particle size boundary parameter, and the upper limit of the fuzzy overlap interval is formed by adding the dynamic overlap half-width to the current particle size boundary parameter. Equivalent particle sizes within the fuzzy overlap region are included in the initial fuzzy membership calculation of adjacent particle size levels, while equivalent particle sizes outside the fuzzy overlap region are directly included in the corresponding particle size level.

7. The method for classifying aggregates by superimposed particle size according to claim 1, characterized in that, The formation of the diversion control command also includes a landing point prediction process, which includes: Target association is performed on the same debris separation area based on the unified timestamp of consecutive image frames; The initial motion velocity is obtained based on the change in the centroid physical coordinates of adjacent image frames and the inter-frame time interval. The wind resistance acceleration coefficient corresponding to the crushed stone is determined based on the equivalent particle size, air density, crushed stone material density and drag coefficient. Input the physical coordinates of the center of mass, the initial velocity, and the wind resistance acceleration coefficient into the discrete trajectory to calculate the predicted landing point coordinates; The drag acceleration coefficient is calculated according to the following formula: ； in, For the first Each gravel separation zone corresponds to the wind resistance acceleration coefficient of the gravel; air density; For the first The resistance coefficient of the crushed stone corresponds to the crushed stone in each crushed stone separation zone; Density of crushed stone material; For the first The equivalent particle size of each gravel separation zone.

8. The method for classifying aggregates by superimposed particle size according to claim 7, characterized in that, The trajectory state uncertainty is formed based on the target association residual, centroid coordinate error and initial motion velocity error. The trajectory state uncertainty is recursively calculated in the discrete trajectory calculation process to obtain the range of the predicted landing point uncertainty, the earliest arrival time and the latest arrival time of the crushed stone reaching the graded execution plane. The diversion execution time window is formed based on the earliest arrival time, the latest arrival time, the actuator response delay, and the control communication delay; When the execution time windows of two crushed stones overlap and the two crushed stones correspond to different grading paths, the execution path is selected based on the overlap ratio of the time windows, the current grading channel load, and the confidence status of the particle size label, including delayed execution, redundant channel switching, or verification path.

9. The method for classifying aggregates by superimposed particle size according to claim 1, characterized in that, The classification error feedback quantity is the directional classification error corresponding to the particle size boundary, and the directional classification error is: ； in For the first The first quality verification window Directional grading error corresponding to each particle size boundary; For the first Within the quality verification window, crossing the first The number of errors that are assigned to higher particle size levels at each particle size boundary; For the first Within the quality verification window, crossing the first The number of errors that are assigned to lower particle size levels at each particle size boundary; For the first Within the quality verification window and the first The number of validation samples associated with each particle size boundary; Index for quality verification window; The particle size boundary parameters are updated only based on the directional grading error, and the updating of the particle size boundary parameters stops when the grading error rate and the change in particle size boundary parameters in the continuous quality verification window simultaneously meet the convergence condition.

10. The method for classifying aggregates by superimposed particle size according to claim 9, characterized in that, The area distribution statistics, aspect ratio distribution statistics, texture homogeneity change statistics, proportion of connected regions, directional grading error, and current particle size boundary parameters corresponding to the continuous quality verification window are used to form a time series feature vector. The time series feature vector is input into a long short-term memory neural network to obtain the candidate particle size boundary increment vector. Compare the incremental direction of the candidate particle size boundary increment vector with the target correction direction corresponding to the current directional classification error; When the incremental direction is consistent with the target correction direction, the candidate particle size boundary incremental vector is limited according to the allowed update magnitude. When the incremental direction is inconsistent with the target correction direction, reduce the update weight of the candidate particle size boundary incremental vector or pause the update of the corresponding particle size boundary. The particle size boundary increments, which are constrained by direction and amplitude, are superimposed onto the current particle size boundary parameters.

Images