Coal gangue aggregate particle shape image intelligent recognition and sorting quality evaluation method and system

Abstract

This invention relates to the field of intelligent image recognition and sorting quality assessment technology for the utilization of coal gangue as recycled aggregate, specifically to a method and system for intelligent image recognition and sorting quality assessment of coal gangue aggregate particle shape. The method includes: acquiring and preprocessing images of coal gangue aggregate particles; detecting inversion points of particle edge concavity / convexity attributes and locating particle contact surfaces; segmenting the image into a set of single particle outlines and labeling contact surface signals; calculating a set of particle shape parameters (aspect ratio, roundness, and angularity) for each particle to distinguish between needle-like, flaky, angular, and rounded particles; measuring particle size and statistically analyzing particle size distribution; comprehensively assessing sorting quality based on needle-like / flaky particle content, particle shape distribution, particle size distribution, and contact surface signals, and providing alerts for exceeding standards; and inverting the particle contact surface and multi-scale edges from geometric artifacts to be eliminated into dual signal carriers of sorting process feedback and lithological attribution, thereby achieving simultaneous quantification of particle shape and size and comprehensive assessment of sorting quality including upstream process feedback.

Description

Technical Field

[0001] This invention relates to the field of image intelligent recognition and sorting quality assessment technology for the utilization of coal gangue as recycled aggregate, specifically to a method and system for image intelligent recognition and sorting quality assessment of coal gangue aggregate particle shape. Background Technology

[0002] Coal gangue is a solid waste generated during coal mining and washing. Long-term stockpiling not only occupies significant land resources but also causes serious environmental pollution. With the advancement of solid waste resource utilization technologies, processing coal gangue through crushing, screening, and sorting into recycled aggregates for use in roadbed filling, concrete aggregate replacement, and mine backfilling has become a major approach for large-scale utilization of coal gangue. However, when used as recycled aggregate, the particle shape of coal gangue significantly affects aggregate quality: excessively high content of needle-like and flaky particles significantly reduces the workability and compressive strength of concrete, while excessively low content of angular particles weakens the interlocking skeleton strength of asphalt mixtures, and an unreasonable particle size distribution affects the material's density and stability. Therefore, quantitatively testing the particle shape and size of coal gangue aggregates and evaluating sorting quality accordingly is a crucial step in the high-value utilization of coal gangue.

[0003] Traditional methods for detecting the particle shape of coal gangue aggregates mainly rely on manual screening, standardization, or visual judgment using vernier calipers. These methods suffer from low efficiency, high subjectivity, and lack of batch data traceability, making them unsuitable for the online detection requirements of modern coal gangue sorting lines. Therefore, aggregate particle shape recognition technology based on digital image analysis has become a research hotspot in the industry. Chinese invention patent application CN106442231A discloses a method for evaluating the angularity of coarse aggregates based on digital image analysis technology. This method obtains coarse aggregates of a single particle size through water washing and screening. Particle images are acquired in a backlight box using digital image acquisition equipment. Image-proPlus image processing software is used to extract basic parameters such as area, perimeter, major axis, and minor axis, and then quantifies angularity indicators such as roundness, angularity, average minimum radius of curvature, and circularity to evaluate the angularity of the coarse aggregates. Chinese invention patent application CN105136622A discloses a method for batch testing the flatness ratio of coarse aggregates. This method involves supporting coarse aggregate particles on a paper surface with modeling clay, taking multi-directional photos with a camera, calculating the ratio of the major and minor axes of the equivalent ellipse for each particle using Image-ProPlus software, and taking the maximum ratio after ten flips as the flatness ratio of each particle, thereby statistically analyzing the content of needle-like and flaky particles. Chinese invention patent CN109116002B discloses a method for quantifying the angularity of road coarse aggregates. This method performs preprocessing on the aggregate image, including grayscale conversion, median filtering for noise reduction, image entropy segmentation, and target extraction using a closing operation. It then uses M-structuring elements to extract the morphological aggregate skeleton, statistically analyzes the average number of angular points, and combines this with area weights to quantify the angularity of the entire batch of aggregates. These methods have promoted the transformation of aggregate particle shape quantification from manual visual inspection to digital image recognition in the fields of asphalt pavement and road engineering.

[0004] However, all the aforementioned schemes treat the contact surfaces between particles as topological artifacts that must be completely separated and eliminated by the algorithm in image segmentation, and the multi-scale micro-serrations at the particle edges as image noise that must be suppressed by smoothing filters. They also equate the sorting quality assessment with an open-loop binary judgment of whether the content of needle-like and flaky particles exceeds a preset threshold. This cognitive framework is applicable in offline detection scenarios for natural crushed stone or artificial aggregates, but it fundamentally contradicts the characteristics of coal gangue material itself in the continuous online detection scenario of coal gangue production lines: coal gangue is a multi-lithological brittle solid waste mixed with coal-bearing shale, symbiotic rocks, and coal gangue. The dense stacking and contact of particles on the production line conveyor belt is unavoidable. The concave geometric distribution of the particle contact surfaces actually carries key process signals of the matching state between the upstream sorting screen aperture and the coal gangue gradation. The density of multi-scale micro-serrations at the particle edges actually encodes the brittle fracture of coal gangue and the plasticity of symbiotic rocks. The fingerprint information of three types of lithological fracturing modes: crushed coal gangue, shale bedding fracturing, and shale bedding fracturing. When existing schemes eliminate these two types of signals as artifacts and noise, the sorting quality assessment can only stop at the isolated conclusion of whether the needle-like and flaky particles exceed the standard. It is impossible to attribute the problem to lithology, because it is unknown whether it is due to excessive brittle fracturing products of coal gangue or the intrusion of shale bedding fracturing products. Furthermore, it is impossible to reverse the particle shape distribution at the detection end to the quantitative adjustment suggestions for the deviation of the upstream sorting screen aperture. As a result, the core bottleneck of the lack of a key feedback link between coal gangue sorting quality control and upstream sorting process is achieved. Summary of the Invention

[0005] Addressing the core bottleneck of existing coal gangue aggregate particle shape image recognition and sorting quality assessment technologies—the lack of a key feedback link between sorting quality control and upstream sorting processes—this invention provides a method and system for intelligent recognition and sorting quality assessment of coal gangue aggregate particle shape images. By reversing the multi-scale morphology of particle contact surfaces and edges from geometric artifacts and image noise to dual signal carriers of sorting process feedback and lithological attribution, a cognitive reversal strategy is constructed. This establishes a deeply coupled closed loop of particle segmentation, particle shape recognition, particle size quantification, and sorting quality assessment. Without introducing additional hardware sensors or disrupting existing production line cycles, this approach combines image geometric features with the crushing mechanism of coal gangue materials to achieve objective particle-by-particle quantification of coal gangue aggregate particle shape and size, comprehensive multi-index assessment of sorting quality, and a feedback loop from the detection end to the sorting process end.

[0006] The technical solution of this invention is: an intelligent recognition and sorting quality assessment method for coal gangue aggregate particle shape images, comprising the following steps: acquiring images of coal gangue aggregate particles flowing through the image acquisition area on a sorting conveyor belt; performing illumination equalization and contrast enhancement on the particle images to obtain a preprocessed image; detecting the particle edge concavity / convexity attribute reversal points in the preprocessed image, locating the particle contact surface based on the concavity / convexity attribute reversal points, segmenting the preprocessed image into a set of single particle contours and labeling the contact surface signal, wherein the contact surface signal is the particle contact surface on the single particle wheel. The geometric distribution in the profile set is characterized; the particle shape parameters of each single particle profile in the single particle profile set are calculated, including aspect ratio, roundness, and angularity; based on the particle shape parameters, the particles corresponding to each single particle profile are identified as needle-like particles, angular particles, or rounded particles; the particle size of each single particle is measured based on the single particle profile and the particle size distribution is statistically obtained; the sorting quality of the coal gangue aggregate is comprehensively evaluated based on the needle-like particle content, particle shape distribution, particle size distribution, and the contact surface signal, and the sorting quality evaluation result is output. When the needle-like particle content exceeds a preset content threshold, a needle-like particle over-limit prompt is output.

[0007] This invention also provides an intelligent recognition and sorting quality assessment system for coal gangue aggregate particle shape images, comprising: an image acquisition unit for acquiring images of coal gangue aggregate particles flowing through the image acquisition area on a sorting conveyor belt; an image preprocessing module for performing illumination equalization and contrast enhancement on the particle images to obtain preprocessed images; a particle segmentation module for detecting particle edge concavity / convexity attribute reversal points in the preprocessed images, locating particle contact surfaces based on the concavity / convexity attribute reversal points, segmenting the preprocessed images into a set of single particle contours, and labeling contact surface signals; and a particle shape recognition module for analyzing the single particle contour sets. The system calculates particle shape parameters for each individual particle profile and classifies the corresponding particles as needle-like, angular, or rounded particles. A particle size quantification module measures the particle size of each individual particle based on its profile and statistically obtains the particle size distribution. A sorting quality assessment module comprehensively evaluates the sorting quality of the coal gangue aggregate based on the needle-like particle content, particle shape distribution, particle size distribution, and the contact surface signal, outputs the sorting quality assessment results, and outputs a needle-like particle exceeding the preset content threshold as a warning. A statistical traceability module summarizes the particle shape distribution and particle size distribution to form statistical traceability data, providing a basis for adjusting the sorting process.

[0008] The beneficial effects of this invention are as follows: First, by detecting the reversal point of the concave-convex attribute of the particle edge and locating the particle contact surface based on the reversal point, the coal gangue aggregate particles in contact with each other in the preprocessed image no longer rely on the premise of artificial laying where the particles do not contact each other beforehand, and can be effectively segmented into the set of single particle outlines. This allows the invention to directly address the real working condition of continuous and densely stacked coal gangue aggregate flowing through the sorting conveyor belt. The mechanism is that the local curvature of the coal gangue aggregate particle edge at the contact position between two particles will reverse from convex to concave and cross zero. This zero-crossing point is the intrinsic marker of the geometric boundary of the coal gangue aggregate particle rather than external image noise. Compared with the offline detection route that relies on backlighting box shooting or separate placement supported by clay, which relies on particles not contacting each other beforehand, this invention fundamentally breaks through the detection premise limitation of offline manual laying.

[0009] Secondly, by simultaneously incorporating the contact surface signal as an evaluation input when comprehensively assessing the sorting quality of the coal gangue aggregate, this invention upgrades the sorting quality assessment from an open-loop binary judgment relying solely on the threshold of needle-like and flaky particle content to a multi-index comprehensive assessment including the contact surface distribution signal. The mechanism lies in the physical relationship of the conservation of packing entropy between the geometric distribution of the contact surface of the coal gangue aggregate particles and the deviation of the upstream sorting screen aperture. That is, when the aperture of the upstream sorting screen is too small, the coal gangue aggregate particles are forced to stack into a deep contact surface, and when the aperture is appropriate, they are spread out into a shallow contact surface. Thus, the contact surface signal becomes a mirror carrier of the upstream sorting process status. Compared with the open-loop route that only evaluates the angularity of the entire batch of aggregate based on the statistics of skeleton corner points, this invention achieves a feedback closed loop from the detection end to the sorting process end without adding additional sensors.

[0010] Furthermore, by utilizing the three-piece quantization of particle shape parameters, including aspect ratio, roundness, and angularity, and the volume equivalent diameter calculation based on the true three-dimensional equivalent size, this invention enables the synchronous quantization and mutual verification of particle shape and particle size on the single particle profile set. The mechanism lies in the fact that the rolling posture change of coal gangue particles on the sorting conveyor belt causes a single-frame two-dimensional projection deviation. The true three-dimensional equivalent size obtained by multi-frame multi-posture trajectory association and multi-view joint inversion of the minimum bounding rectangle overcomes the posture deviation of single-frame measurement, and the misjudgment rate of needle-like and flaky particles is significantly reduced compared with single-frame measurement.

[0011] In summary, this invention employs a deeply coupled closed-loop architecture that combines particle segmentation, particle shape recognition, particle size quantification, and sorting quality assessment. This transforms the particle shape detection and sorting quality evaluation of coal gangue aggregates from manual screening and visual inspection to intelligent image recognition. While objectively quantifying particle shape and size on a particle-by-particle basis, it provides a comprehensive evaluation of sorting quality including feedback from upstream processes, thus offering technical support for quality control and sorting process optimization of recycled coal gangue aggregates. Attached Figure Description

[0012] Figure 1 This is a flowchart of the intelligent recognition and sorting quality assessment method for coal gangue aggregate particle shape image provided in the embodiments of the present invention.

[0013] Figure 2 This is an architecture diagram of the intelligent recognition and sorting quality assessment system for coal gangue aggregate particle shape image provided in this embodiment of the invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the specific embodiments of this invention will be described in further detail below with reference to the accompanying drawings. The listed embodiments are intended to provide illustrative examples of this invention and do not constitute a limitation on the scope of protection of this invention.

[0015] This embodiment describes in detail the implementation of the intelligent image recognition and sorting quality assessment method for coal gangue aggregates described in this invention, based on the actual production line of a coal gangue recycled aggregate sorting workshop in a coal washing and processing enterprise. The sorting workshop is equipped with a coal gangue aggregate sorting production line with a particle size range of 5mm to 40mm. The sorting conveyor belt is 800mm wide and operates at a linear speed of 0.6m / s to 1.2m / s. After being classified by an upstream vibrating screen and air classifier, the coal gangue aggregate particles are densely stacked on the sorting conveyor belt and flow through the image acquisition area. The implementation flow of this embodiment is as follows: Figure 1 As shown, it includes steps S1 to S10.

[0016] Step S1: Acquire images of the coal gangue aggregate particles flowing through the conveyor belt. An image acquisition unit is installed above the sorting conveyor belt. The image acquisition unit includes an industrial area array camera with a resolution of 2448 pixels × 2048 pixels and an LED array light source that is synchronously triggered with the industrial area array camera. The frame rate of the image acquisition unit is 100fps, the object distance of the industrial area array camera is 800mm, and the actual size of a single pixel is 0.3mm. The LED array light source uses a combination of 4 strip LED lights with a color temperature of 5500K and an illuminance uniformity of 90% calibrated on-site. The LED array light source is evenly arranged along the width of the sorting conveyor belt to avoid lateral shadows on the particles. The image acquisition unit takes a vertical downward shot from a position 800mm above the surface of the sorting conveyor belt, with the field of view covering the entire width of the sorting conveyor belt and a 600mm length along the direction of conveyor belt movement. This field of view is the image acquisition area. The industrial area array camera acquires images of coal gangue aggregate particles flowing through the image acquisition area on the sorting conveyor belt in a continuous trigger mode, and outputs the original particle image. In this embodiment, the particle image is a grayscale image with a resolution of 2448 pixels × 2048 pixels and 8-bit grayscale depth, and the single frame acquisition time interval is 10ms. After being generated by the image acquisition unit, the particle image is transmitted to the processing server via gigabit Ethernet for subsequent processing.

[0017] Step S2: Image Preprocessing. The particle image obtained in Step S1 is subjected to illumination equalization and contrast enhancement to obtain the preprocessed image. In this embodiment, contrast-limited adaptive histogram equalization is used to perform illumination equalization on the particle image. During processing, the particle image is divided into an 8×8 grid, and histogram equalization is performed independently within each grid. The contrast limit value is set to 2.0, and bilinear interpolation is used between adjacent grids to avoid obvious seams at grid boundaries. This processing can effectively suppress the grayscale gradient difference between the overall shadow area formed by the dense stacking of coal gangue aggregate particles on the sorting conveyor belt and the overall bright spot area illuminated by the LED array light source, making the local contrast of each region in the particle image more consistent. Based on illumination equalization, gamma correction is performed on the particle image to enhance contrast, with a gamma value of 0.85. This processing can increase the grayscale contrast of the coal gangue aggregate particle edges relative to the background of the sorting conveyor belt by approximately 20%, laying the foundation for edge extraction in subsequent steps. The preprocessed image and the grain image maintain the same resolution and grayscale depth, with the only difference being the local redistribution of pixel grayscale values.

[0018] Step S3: Detect the contact surface between the inversion point of the particle edge's concavity / convexity attribute and the positioning particle. This step is one of the core innovative aspects of this invention.

[0019] First, edge extraction is performed on the preprocessed image to obtain a particle edge pixel sequence. In this embodiment, the Canny edge detection operator is used to perform edge extraction on the preprocessed image. The low threshold of the Canny operator is taken as the 30th percentile of the gray-level gradient magnitude distribution, and the high threshold is taken as the 70th percentile of the gray-level gradient magnitude distribution. The standard deviation of Gaussian smoothing is 1.0. After the Canny operator outputs a binary edge image, eight-neighbor connected component labeling is performed on the binary edge image. Each connected component is taken as a candidate particle edge and sorted counterclockwise along the contour direction to obtain the corresponding particle edge pixel sequence. In this embodiment, the typical particle edge pixel sequence length is between 80 pixels and 500 pixels.

[0020] Subsequently, the local curvature directed value of each edge pixel is calculated along the pixel sequence at the particle edge. The local curvature directed value is calculated using the following formula: ,in: The first pixel in the particle edge pixel sequence edge pixels The local curvature at a given point is a directed value, which is a scalar and takes values ​​ranging from 0 to 1. rad / pixel, calculated by this formula, characterizes the edge of the particle within a pixel. The degree of curvature and the orientation of the concavity and convexity at the point, with positive values ​​representing convex orientation and negative values ​​representing concave orientation; The first pixel in the particle edge pixel sequence Each edge pixel is a two-dimensional coordinate vector, measured in pixels, and is extracted from the edges. The pixel number of the particle edge pixel sequence is a positive integer. , The total number of pixels in the particle edge pixel sequence is a positive integer, ranging from 80 to 500, and is measured in pixels. It is obtained by counting the length of the particle edge pixel sequence. and The pixel sequence at the edge of the particle is respectively in pixel The first derivative components along the contour parameter direction are all scalars, with units of pixels per step, and their values ​​range from... The value is calculated from the pixel sequence at the particle edge using the three-point difference method, representing the particle edge at the pixel level. Tangential direction at the location; and The pixel sequence at the edge of the particle is respectively in pixel The second derivative components along the contour parameter direction are all scalars, with units of . The range of values ​​is The value is calculated from the pixel sequence at the particle edge using the three-point difference method, representing the particle edge at the pixel level. The magnitude of curvature at that point; It is a symbolic function, with a scalar as input and an output of... One of the three values ​​is used to decouple the absolute value of curvature from the concave-convex orientation. For the particle edge pixel sequence in pixels The normal vector at point is a two-dimensional unit vector, dimensionless, with a constant magnitude of 1. Normalization yields a value that characterizes the particle edge at the pixel level. Located in a direction perpendicular to the tangential direction; For pixels The relative position vector to the centroid of the pixel sequence at the particle edge is a two-dimensional vector, with units of pixels and a value range of [value range missing]. ,Depend on The calculation yielded, where The centroid of the pixel sequence representing the grain edge indicates the position of the grain edge within the pixel. The orientation relative to the overall particle size distribution is considered. The inversion point of the concavity / convexity attribute is detected based on the directional value of the local curvature. The criterion for determining the inversion point of the concavity / convexity attribute is that the zero-crossing point where the directional value of the local curvature changes from positive to negative or from negative to positive simultaneously satisfies the significant curvature amplitude condition. .

[0021] in: For pixels The indicator function for whether a point is marked as a reversal point of the concavity / convexity attribute is a scalar with a value range of [value range missing]. Dimensionless, calculated by this formula, a value of 1 represents a pixel. Marked as the point where the concave-convex attribute is reversed; and pixels The local curvature directed values ​​at the preceding and following pixels along the pixel sequence at the edge of the particle are defined as in the aforementioned formula; The preset significance threshold for curvature amplitude is denoted as , which is a scalar and has a value range of . rad / pixel, in this embodiment 0.08 rad / pixel, is determined by adding 0.5 times the standard deviation to the mean of the statistical distribution of edge curvature of typical particles in the particle image. A value that is too small may cause edge jagged noise to be misjudged as the concavity / convexity attribute reversal point, while a value that is too large may cause some zero-crossing points corresponding to shallow contact surfaces to be missed. This value characterizes the sensitivity in distinguishing between real contact surface concavity and edge jagged noise. The units on both sides of the inequality in this formula are... The symbol indicates that it is dimensionless, and the overall unit is strictly consistent.

[0022] Finally, based on the ratio of the edge segment curvature valley depth to the edge segment length between adjacent pairs of concave-convex attribute inversion points, the particle contact surface is classified into three types: V-shaped contact, U-shaped contact, and linear contact. Let the two adjacent pairs of concave-convex attribute inversion points along the pixel sequence of the particle edge be... and The edge segment length ratio is defined as: ,in: The value is the ratio of the depth of the curvature valley of the edge segment to the length of the edge segment; it is a scalar and its value range is... , dimensionless, is calculated by this formula, and characterizes the concave morphology of the particle contact surface; The depth of the curvature valley value of the edge segment is defined as the pixel value. and The maximum perpendicular distance between the connecting line and the edge segment is a scalar value in pixels, and its range is [value missing]. The depth of the indentation on the particle contact surface is obtained by taking the maximum value of the vertical distance from each pixel on the edge segment to the connecting line; The length of the edge segment is defined in pixels. to The arc length along the pixel sequence at the edge of the particle is a scalar, measured in pixels, and its value ranges from [value missing]. The lateral span of the particle contact surface is represented by the number of pixels between two inversion points in the pixel sequence at the particle edge.

[0023] Based on the edge segment length ratio, the particle contact surfaces are classified into three types:

[0024] ,in: The classification type of the particle contact surface is a category variable with a value range of [value range missing]. V-shaped contact, U-shaped contact, linear contact , dimensionless, obtained by this formula, characterizes the geometric morphology of the particle contact surface; The threshold for distinguishing between V-shaped and U-shaped contacts is denoted by , which is a scalar and has a value range of . , dimensionless, is taken as 0.30 in this embodiment, and is obtained by statistical empirical calibration of the stacking angle of the coal gangue aggregate particles, characterizing the boundary between the V-shaped sharp depression and the U-shaped smooth depression; Let be the threshold for distinguishing between U-shaped and linear contact, and be a scalar with a value range of . The dimensionless value, taken as 0.08 in this embodiment, is obtained through statistical experience calibration of the flat contact of coal gangue aggregate particles, characterizing the boundary between shallow U-shaped depressions and near-flat straight-line contacts. This step outputs intermediate results containing all the positions of the concave-convex attribute reversal points and their corresponding particle contact surface classifications, and performs pixel-level segmentation on the preprocessed image based on the concave-convex attribute reversal points to obtain the single-particle contour set; for each single-particle contour, the edge segments belonging to the particle contact surface in its contour are marked to form the contact surface signal, which includes the quantity and depth distribution of the three types of contact—V-shaped contact, U-shaped contact, and linear contact—in the single-particle contour set.

[0025] Step S4: Temporal Stabilization Processing. Considering the unavoidable mechanical vibration of the sorting conveyor belt during production line operation, which causes sub-pixel-level positional jitter at the particle edges and positional drift of some of the concave-convex attribute reversal points in single-frame acquisition, this embodiment further introduces a multi-frame temporal sequence of particle images to perform temporal stabilization processing. The particle image sequence is continuously acquired by the image acquisition unit at 100fps. In this embodiment, each temporal stabilization process calls 5 temporally adjacent particle images from the particle image sequence. Inter-frame matching is performed on the contours of each particle in the particle image sequence to obtain the instantaneous displacement field of the particle's centroid. The inter-frame matching uses the following cost function: .in: For the first The first frame of the grain image Each particle With the The first frame of the grain image Each particle The inter-frame matching cost is a scalar with a value range of 1. , dimensionless, is calculated by this formula and represents the dissimilarity between two frames of particles that are the same physical particle; and Particles and The two-dimensional coordinates of the centroid are all two-dimensional vectors, with units of pixels and values ​​ranging from [value range missing]. The zero-order moment and the first-order moment of the corresponding contour in the single particle contour set are calculated to characterize the position of the particle in the particle image. This is the L2 norm operator for two-dimensional vectors, outputting a scalar in pixels. and Particles and The outline area is a scalar, and the unit is . The range of values ​​is The size of the particles is determined by the total number of pixels enclosed by the corresponding contour. and Particles and The perimeter of the outline is a scalar, with the unit being pixels, and the value range is [value range missing]. , is obtained by counting the total number of pixels in the corresponding contour, and represents the length of the particle edge; , , are the weighting coefficients of the centroid distance, area difference, and perimeter difference in the cost function, respectively. All are scalars, dimensionless, and their ranges are respectively... , , In this embodiment, values ​​of 1.0, 0.005, and 0.05 are used respectively, obtained by prior empirical tuning of the sorting conveyor belt speed and particle size, representing the weighting of the importance of the three similarity indicators. After conversion of the corresponding units for each weighting coefficient (…),… carry , carry , carry (Post-normalization and dimensionlessization).

[0026] According to the principle of minimizing the cost function, the first... Frame granules With the Frame best matching particles Correlation, thus obtaining particles The instantaneous displacement vector of the center of gravity: ,in: Particles In the The instantaneous displacement vector of the frame's center of gravity is a two-dimensional vector, with units of pixels per second, and its value range is... The instantaneous velocity of the particle, calculated by this formula, represents the instantaneous velocity of the particle along the direction of movement of the sorting conveyor belt and the direction of lateral vibration. and The definition is the same as the formula mentioned above; The time interval between two consecutive frames of particle images is a scalar value in seconds, with a range of values ​​of [value missing]. In this embodiment, s (corresponding to a 100fps acquisition frame rate) is determined by the reciprocal of the acquisition frame rate of the image acquisition unit. A value that is too large will result in excessive inter-frame displacement of particles and blurred matching, while a value that is too small will increase the acquisition and processing load. All particles The instantaneous displacement vector of the center of gravity constitutes the instantaneous displacement field of the particle's center of gravity.

[0027] Based on the instantaneous displacement field of the particle's center of gravity, multi-frame temporal interpolation is performed on the concavity / convexity attribute reversal points in each time frame to obtain the temporally stabilized reversal points: ,in: Particles In the Frame number The position of the temporal stabilization reversal point is a two-dimensional vector with the unit of pixel, which is calculated by this formula and represents the spatial coordinates of the concave-convex attribute reversal point stabilized after multi-frame temporal fusion. For the first The set of temporally neighboring frames centered on the frame is a set of integer indices; in this embodiment, it is taken as... A total of 5 frames, dimensionless, are determined by matching the frame rate of the image acquisition unit with the running speed of the sorting conveyor belt, representing the number of frames participating in the temporal interpolation; The index of the temporal neighbor frame is an integer. ; For the first On-frame particles The first corresponding physical particle The position of the concave / convex attribute inversion point is a two-dimensional vector, with the unit being pixels, determined by step S3 for the first... The processing result of the frame is obtained; For the first The frame-referenced particle displacement compensation term, in pixels, is used to adjust the particle displacement of the first frame. The frame reversal point position is compensated to match the first frame. In a reference frame where the grain center of gravity is aligned; Here, represents the weighting coefficients of the temporal neighborhood frames, and is a scalar. Dimensionless, with a range of values ​​of The exponentially decaying weight, which is calculated by this formula, represents the contribution of temporal neighboring frames. Let be the time constant of the time-weighted characteristic curve, be a scalar, be in seconds, and have a range of values ​​of . In this embodiment, s is determined by the reciprocal estimation of the main frequency band of the sorting conveyor belt vibration. If the value is too large, it will cause the weight of the temporal neighboring frame to be uniform and lose the temporal stabilization effect. If the value is too small, it will degenerate into single frame detection. This is the absolute value operator. The unit on the left side of the equation is pixels, and the displacement term on the right side is... The unit is pixel, and the displacement compensation item is... Units are Weighted normalization and dimensionless transformation.

[0028] The classification of V-shaped contact, U-shaped contact, and linear contact is performed based on the temporal stabilization reversal point. After multi-frame temporal stabilization, the standard deviation of the concavity / convexity attribute reversal point position is reduced from approximately 2.5 pixels in a single frame detection to approximately 0.6 pixels, significantly improving the stability of the V-shaped contact, U-shaped contact, and linear contact classification.

[0029] Step S5: Particle shape parameter calculation. The particle shape parameters are calculated for each individual particle profile in the set of individual particle profiles. These parameters include aspect ratio, roundness, and edge angle. The aspect ratio is determined based on the ratio of the longer side to the shorter side of the smallest bounding rectangle of the individual particle profile, denoted as [missing value]. ,in and These are the lengths of the longest and shortest sides of the minimum bounding rectangle of the single particle profile, calculated using the rotating caliper algorithm. In this embodiment... and The typical value ranges are as follows: pixel and pixel. The roundness is determined based on the normalized ratio of the area of ​​the single particle's outline to the square of its perimeter: ,in: Let be the roundness, and be a scalar with a value range of . , dimensionless, is calculated by this formula, and characterizes the degree to which the profile of the single particle approaches an ideal circle. When it is always equal to 1, it is an ideal circle; The total number of pixels enclosed by the contour of the single particle is a scalar, and the unit is . It is obtained by contour filling; The perimeter of the single particle outline is defined in the same formula as described above; Pi is a constant. The edge angle is determined based on the ratio of the total number of edge pixels on the single particle contour whose local curvature exceeds a preset curvature threshold to the perimeter of the single particle contour, denoted as π. ,in The total number of edge pixels on the single particle contour whose local curvature absolute value exceeds a preset curvature threshold is 0.12 rad / pixel in this embodiment.

[0030] Step S6: Multi-pose contour set association and joint inversion with true 3D equivalent size. When coal gangue aggregate particles flow through the image acquisition area on the sorting conveyor belt, they typically undergo slight rolling. The particle shape parameters measured by single-frame 2D projection are highly sensitive to the particle's pose on the sorting conveyor belt, exhibiting systematic bias. This embodiment further performs trajectory association on the single-particle contours of the same particle in multiple frames of the particle image sequence to obtain the multi-pose contour set.

[0031] The trajectory association cost function is defined as follows:

[0032] .

[0033] in: For the first Particle trajectory The associated cost is a scalar, and its value ranges from 1 to 2. Dimensionless, calculated by this formula, characterizing the particle in to The overall associated cost between frames as the same physical particle; and These are the start and end frame numbers of the particle trajectory, respectively, both positive integers, determined by the frame numbers when the particle enters and leaves the image acquisition area; The definition is the same as in step S4 formula; Particles The shape-invariant moment eigenvector of Hu is a 7-dimensional real vector, dimensionless, with a value range of [value missing]. The shape of the particle is obtained by calculating the profile moments and is characterized by the rotational scale translation invariance. For particle trajectory The mean of the Hu shape-invariant moment eigenvectors across all frames is a 7-dimensional real vector, dimensionless, derived from each frame. The arithmetic mean is obtained; The weight for shape consistency constraints is a scalar, dimensionless, and its value range is [value range missing]. In this embodiment, 5.0 is used, which is obtained by empirical calibration of the shape invariance moment drift tolerance, and represents the weight of shape consistency relative to position matching. This is the L2 norm square operator for 7-dimensional vectors.

[0034] Particles are obtained according to the principle of minimizing the trajectory association cost function. exist to The set of multi-pose contours in the frame, in this embodiment, typically has a trajectory length of 5 to 15 frames. The aspect ratio, roundness, and edge angle of each pose contour in the set of multi-pose contours are calculated respectively, and the maximum aspect ratio, minimum roundness, and maximum edge angle of the multi-pose contours are taken as the final particle shape parameters. This extreme value strategy is based on the mechanical fact that the needle-like features in the particle rolling posture change are only fully exposed in the pose frame with the largest projection direction of the particle's long axis.

[0035] Furthermore, the true three-dimensional equivalent size of the particle is obtained by performing multi-view joint inversion on the long and short sides of the minimum bounding rectangle of each posture contour in the multi-posture contour set. Let the particle be at the... The rotation angles of the frame's orientation relative to the sorting conveyor plane are respectively (around the long axis of the particle) and (Around the short axis of the particle), the triaxial length of the true three-dimensional equivalent size , , (in (and the longest side of the minimum bounding rectangle of each frame) short side The following projection relationship exists:

[0036] .

[0037] .

[0038] in: , , These represent the major axis, minor axis, and thickness of the particle in the actual three-dimensional equivalent size, respectively. All are scalars, with units of pixels, and their value ranges are as follows: , , The formula is obtained by solving this set of formulas, which characterizes the true geometric scale of the particle in three-dimensional space. For particles in the first The length of the longest side of the minimum bounding rectangle of the frame is a scalar in pixels and is calculated by the rotating caliper algorithm. For particles in the first The length of the short side of the minimum bounding rectangle of the frame is a scalar in pixels and is calculated by the rotating caliper algorithm. For particles in the first The rotation angle of the frame around the major axis of the particle is a scalar quantity in rad, and its value ranges from [value missing]. It is obtained by azimuth angle difference estimation of the long side of the minimum bounding rectangle between frames relative to the direction of movement of the sorting conveyor belt; For particles in the first The rotation angle of the frame around the minor axis of the particle is a scalar, with units of rad and a range of values ​​of . It is estimated by inverse calculation from the inter-frame area change rate; The multi-pose frame number is a positive integer. , The number of poses in the set of multi-pose contours is 5 to 15 in this embodiment.

[0039] The above system of equations is solved using the least squares method:

[0040] .

[0041] in: , , These are the estimated three-axis dimensions of the true three-dimensional equivalent size obtained by inversion, both of which are scalars in pixel and are defined as described above; and Based on three axes With attitude angle The lengths of the longest and shortest sides of the minimum bounding rectangle predicted by the aforementioned projection equation are both scalars, with units of pixels. This is an operator that determines the parameter value of the function within parentheses. This embodiment employs Levenberg-Marquardt numerical optimization, with a single-particle solution time of approximately 2ms to 5ms. The particle size is calculated based on the volumetric equivalent diameter of the actual three-dimensional equivalent size. ,in mm / pixel is a calibration factor from pixel to actual size, obtained by calibrating the object distance and focal length of the industrial area array camera.

[0042] Step S7: Particle Shape Classification. Based on the particle shape parameters, each individual particle outline is classified into needle-like, angular, or rounded particles. When the aspect ratio is greater than 2.5, it is classified as a needle-like particle; when the aspect ratio is less than or equal to 2.5 and the angularity is greater than a preset angularity threshold, it is classified as an angular particle (in this embodiment, the preset angularity threshold is 0.18); when the aspect ratio is less than or equal to 2.5, the angularity is less than or equal to the preset angularity threshold, and the roundness is greater than a preset roundness threshold, it is classified as a rounded particle (in this embodiment, the preset roundness threshold is 0.75). It should be noted that the discrimination of needle-like and flaky particles in this step is based on the aspect ratio, which belongs to the particle shape dimension classification that describes particle morphology. The aspect ratio only reflects the elongation of the particle in the plane subtended by the major and minor axes and does not include information on the particle thickness direction. Therefore, the needle-like and flaky particles in terms of morphology include both slender needle-like particles with a thickness comparable to the minor axis and flat flaky particles with significant collapse in the thickness direction. This particle shape dimension classification and the lithological fracturing mode discrimination based on the triaxial ratio of the true three-dimensional equivalent size in step S8 are two independent classification dimensions. They are applied to the same particle in parallel. Each particle obtains a particle shape category label of needle-like, angular, or rounded, and independently obtains a lithological fracturing mode label of brittle fracturing of coal gangue, plastic fracturing of symbiotic rocks, or bedding fracturing. The particle shape distribution is obtained by counting the percentage of the three types of particles in the single particle profile set, and the needle-like and flaky particle content is obtained by counting the percentage of the number of needle-like and flaky particles in the total number of particles in the single particle profile set.

[0043] Step S8: Lithological Fragmentation Mode Identification. Coal gangue, as a type of brittle solid waste in coal-bearing strata, exhibits three typical fragmentation modes: brittle fragmentation, plastic fragmentation of associated rocks, and bedding fragmentation. The particles of these three fragmentation modes show differentiated distributions in terms of the triaxial ratio of their true three-dimensional equivalent size: brittle fragmentation products have a moderate triaxial ratio and well-developed angularity; plastic fragmentation products of associated rocks tend to have an equiaxial triaxial ratio; and bedding fragmentation products exhibit typical flat, lamellar shapes. This embodiment identifies the lithological fragmentation mode based on the triaxial ratio distribution of the true three-dimensional equivalent size. The triaxial ratio index is defined as follows: ,in: The ratio of the major axis to the minor axis is a scalar quantity with a range of values ​​of 1. , dimensionless, is calculated by this formula and characterizes the elongation of the particle on the horizontal projection plane; Let be the ratio of thickness to minor axis, and be a scalar with a range of values. , dimensionless, is calculated by this formula and characterizes the degree of flattening of the particle relative to its minor axis in the thickness direction; , , The definition is the same as in step S6.

[0044] The criteria for determining lithological fracturing mode are:

[0045] ,in: Particles The lithological fracturing mode discrimination result is a categorical variable with a value range of [value range missing]. Stratification fracturing, plastic fracturing of associated rocks, and brittle fracturing of coal gangue , dimensionless, obtained by this formula, characterizes the petrological fracturing mechanism corresponding to the particles; This is the range operator for a set of three axes, equal to the difference between the maximum and minimum values ​​in the set. It is a scalar, dimensionless, and its range is [value missing]. The result obtained from this calculation represents the degree of difference in the three-axis dimensions; The preset triaxial difference threshold is a scalar with a value range of [value range missing]. The value is dimensionless, and in this embodiment, it is taken as 0.30. It is obtained from the statistical analysis of the calibration samples of plastic fracture products of the associated rock. If the value is too large, some brittle fracture products will be misclassified as plastic fracture, and if the value is too small, plastic fracture products will be missed. It represents the tolerance of plastic fracture discrimination. The lithological mixing ratio is obtained by statistically analyzing the percentage distribution of the lithological fracture mode discrimination results in the single particle profile set. The lithological mixing ratio is the set of percentages of brittle fracture particles of coal gangue, plastic fracture particles of associated rock, and bedding fracture particles in the total number of particles. In the typical coal gangue sorting production line of this embodiment, the value range of the lithological mixing ratio is: 60% to 80% for brittle fracture of coal gangue, 10% to 25% for plastic fracture of associated rock, and 5% to 20% for bedding fracture. Furthermore, the cause of the excessive content of needle-like and flaky particles can be reliably distinguished based on the lithological fracturing mode discrimination results, and will not fail due to the partial overlap between the needle-like and flaky particle criteria and the bedding fracturing criteria in two-dimensional geometric features. As described in step S7, the discrimination of needle-like and flaky particles is based on two-dimensional aspect ratio, which only reflects the elongation in the plane of the major and minor axes and does not contain thickness information. Therefore, bedding fracturing particles with a major axis to minor axis ratio greater than 2.0 fall into the needle-like and flaky range because they are flat and flaky, and their aspect ratio is often also greater than 2.5. Some brittle fracturing particles of coal gangue with a major axis to minor axis ratio greater than 2.0 also fall into the needle-like and flaky range because they are slender and needle-like. This overlap between the two in two-dimensional aspect ratio does not hinder the differentiation of causes, because in the true three-dimensional equivalent size The thickness and minor axis ratio are independent of the two-dimensional aspect ratio and are precisely the distinguishing dimension between the two types of needle-like and flaky particles. Particles with a thickness and minor axis ratio less than 0.6 are flat, flaky particles with significant collapse in the thickness direction, classified as bedding fracture according to the discrimination criterion in step S8, and their formation is due to the intrusion of shale bedding fracture products. Particles with a thickness and minor axis ratio greater than or equal to 0.6 are slender, needle-like particles with a thickness comparable to their minor axis, and classified as brittle fracture of coal gangue according to the discrimination criterion in step S8, and their formation is due to the small aperture of the upstream sorting screen leading to the needle-like and flaky formation of brittle fracture products. Therefore, for the set of particles identified as needle-like and flaky, genetic decomposition is performed based on the lithological fracture pattern labels obtained for each particle in this step. This allows the needle-like and flaky particles to be uniquely classified into two categories: bedding fracture origin and brittle fracture origin, without any classification conflict, thus reliably achieving the lithological attribution of the needle-like and flaky particles exceeding the standard. When the excessive needle-like and flaky particles are detected, the lithological attribution for the excessive needle-like and flaky particles is simultaneously output based on the lithological mixing ratio: if the percentage of broken bedding particles is significantly higher (above the empirical threshold of 25%), it is attributed to the intrusion of shale bedding breaking products; if the proportion of needle-like and flaky particles in the brittle broken particles of coal gangue is significantly higher, it is attributed to the small aperture of the upstream sorting screen, which leads to the needle-like and flaky formation of brittle breaking products.

[0046] Step S9: Beta Distribution Physical Constraint Inversion of Screen Aperture Deviation. There is a physical correlation between the geometric distribution of the particle contact surfaces formed by the stacking of coal gangue aggregate particles on the sorting conveyor belt and the aperture deviation of the upstream sorting screen, based on the conservation of packing entropy: when the upstream sorting screen aperture is smaller, oversized particles are forced to overlap and stack with normal particles, increasing the proportion of V-shaped contact and decreasing the proportion of linear contact; when the upstream sorting screen aperture is larger, undersized particles enter the mainstream and form a flat stack, increasing the proportion of linear contact. This embodiment establishes a Beta distribution inversion model based on the conservation of packing entropy between the depth distribution of the three types of contact (V-shaped, U-shaped, and linear) and the aperture deviation of the upstream sorting screen.

[0047] Define the normalized quantity ratio of the three types of contact: V-shaped contact, U-shaped contact, and linear contact. , , ,satisfy The constraint of conservation of accumulated entropy is expressed as: ,in: Let be the packing entropy of the contact surface distribution, and let be a scalar with a value range of . The unit is nat, which is calculated by this formula and characterizes the degree of disorder in the three types of contact distribution: V-shaped contact, U-shaped contact, and linear contact. , , These represent the normalized proportions of the three contact types—V-shaped, U-shaped, and linear—in the contact surface signal, respectively. All are scalars, dimensionless, and their values ​​range from [value missing]. The result is obtained by normalizing the statistical results of the contact surface classification in step S3; The natural logarithm operator; For the contact surface type subscript, .

[0048] Normalize the depth of the contact surface to Normalized depth variables are obtained from the interval ,Establish Physical constraints that follow a Beta distribution: ,in: For normalized depth variables The Beta probability density function is a scalar with a unit of dimensionless probability density. It is calculated by this formula and characterizes the depth distribution pattern of the contact surface. Let be the normalized depth variable of the contact surface, and let be a scalar with a range of values. Dimensionless, by The calculation yielded, where The maximum value for the calibrated depth of the curvature valley of the edge segment is 40 pixels in this embodiment; and Let be the shape parameters of the Beta distribution, and be scalars with values ​​ranging from to . and , dimensionless, is estimated by moment matching from the sample mean and variance of the measured contact surface depth distribution, and characterizes the skewness of the contact surface depth distribution; Let be the Beta function, be a scalar, and be dimensionless. It is calculated by the definition of the Beta function.

[0049] Sample mean based on measured contact surface depth distribution With variance Estimating shape parameters using moment matching: ,in: and The estimated shape parameters of the Beta distribution are all scalars, dimensionless, and are calculated using this formula. Let be the normalized depth sample mean, and be a scalar with a value range of . , dimensionless, is obtained by the arithmetic mean of the normalized values ​​of the measured contact surface depth; Let be the normalized depth sample variance, and let be a scalar with a range of values. , dimensionless, is obtained from the normalized second-order central moment of the measured contact surface depth.

[0050] There is a linear mapping relationship between the shape parameters of the Beta distribution based on the constraint of conservation of packing entropy and the aperture deviation of the upstream sorting screen: ,in: and These are the baseline values ​​of the Beta distribution shape parameters when the upstream sorting screen aperture is without deviation. Both are scalars and dimensionless, obtained in this embodiment from statistical analysis of the production line's deviation-free calibration samples. , ; and These are the sensitivity coefficients for the screen aperture deviation to the Beta distribution shape parameter, both of which are scalars, with units of . The range of values ​​is In this embodiment, we take respectively , This was obtained through regression analysis of production line deviation calibration samples. The estimated value of the aperture deviation of the upstream sorting screen is a scalar quantity in mm, and its range is [value range missing]. Positive values ​​indicate that the screen aperture is too large, and negative values ​​indicate that it is too small. This can be obtained by solving the following inversion formula. The inversion is performed by simultaneously applying the above mapping relationships. : ,in: The estimated value of the aperture deviation of the upstream sorting screen is a scalar quantity in mm, and its range is [value range missing]. The deviation of the actual aperture of the upstream sorting screen relative to its rated aperture, calculated by this formula, is the amount of deviation. Other symbols are defined as in the aforementioned formula. The estimated aperture deviation of the upstream sorting screen, combined with the lithological mixing ratio, is output to the sorting quality assessment result as a suggestion for adjusting the upstream sorting process. When the proportion of needle-like and flaky particles in the brittle crushed coal gangue is high, it is recommended to appropriately increase the mesh size of the upstream vibrating screen; when When the proportion of broken particles is high and the mesh size is small, it is recommended to appropriately reduce the screen aperture of the upstream vibrating screen and enhance the cleaning of the vibrating screen.

[0051] Step S10: Comprehensive Evaluation and Statistical Traceability of Sorting Quality. The sorting quality of the coal gangue aggregate is comprehensively evaluated based on the content of needle-like and flaky particles, particle shape distribution, particle size distribution, and the contact surface signal. This embodiment uses a weighted scoring model for comprehensive evaluation: the content of needle-like and flaky particles, the proportion of angular particles, the degree of conformity of the particle size distribution to the rated gradation, and the packing rationality reflected by the contact surface signal are weighted and summed with weights of 0.35, 0.20, 0.25, and 0.20 respectively to obtain a comprehensive sorting quality score. The scoring threshold is divided into four levels—excellent, good, qualified, and unqualified—based on the relevant provisions of the "Test Procedure for Aggregates in Highway Engineering" (JTGE42-2005) and the industry practice of recycled coal gangue aggregates. The sorting quality evaluation result is output. When the content of needle-like and flaky particles exceeds the preset content threshold of 15%, a needle-like and flaky particle exceeding the standard prompt is output, and the lithological attribution of needle-like and flaky particle exceeding the standard obtained in step S8 and the upstream sorting process adjustment suggestions obtained in step S9 are output simultaneously.

[0052] Finally, the estimated values ​​of the particle shape distribution, particle size distribution, lithological mixing ratio, and upstream sorting screen aperture deviation are summarized to form the statistical traceability data. The statistical traceability data is archived and saved according to production shift and production line section, providing a long-term historical basis for the iterative adjustment of the sorting process and quality backtracking.

[0053] In this embodiment, during a 30-day continuous production line test in the coal gangue recycled aggregate sorting workshop, the method achieved an effective particle segmentation rate of 94.2%, an accuracy rate of 92.6% for identifying needle-like and flaky particles, an average relative error of 7.8% between the volumetric equivalent diameter derived from the true three-dimensional equivalent size inversion and the offline vernier caliper measurement, an accuracy rate of 86.3% for determining the lithological fracturing mode, and an average absolute error of 1.2 mm for estimating the upstream sorting screen aperture deviation. Compared to the existing traditional scheme using single-frame two-dimensional projection measurement, the misjudgment rate of needle-like and flaky particles in this embodiment decreased from approximately 15% to approximately 5.2%, and the consistency between the sorting quality assessment results and the visual assessment results of manual screening increased from approximately 70% to approximately 91%.

[0054] This embodiment further describes the system supporting the method described in Embodiment 1. The architecture of the system is as follows: Figure 2 As shown, the system includes an image acquisition unit, an image preprocessing module, a particle segmentation module, a particle shape recognition module, a particle size quantification module, a sorting quality assessment module, and a statistical traceability module. The hardware deployment and functional implementation of each component are as follows.

[0055] The image acquisition unit includes an industrial area array camera, an LED array light source, and a camera-light source synchronous trigger controller. The industrial area array camera is an industrial-grade CMOS camera with a resolution of 2448 pixels × 2048 pixels, a single pixel size of 3.45μm, and a maximum acquisition frame rate of 120fps. It is equipped with a 16mm fixed-focus lens and is mounted above the sorting conveyor belt, 800mm above the conveyor belt surface, with its optical axis perpendicular to the conveyor belt plane. The LED array light source consists of four strip LEDs evenly arranged along the width of the sorting conveyor belt. Each strip LED is 600mm long, has a power of 60W, and a color temperature of 5500K. All light sources operate synchronously and continuously, with an illuminance uniformity calibrated to 90% on-site. The camera-light source synchronous trigger controller triggers the industrial area array camera with a 100fps pulse and checks the working status of the LED array light source at each trigger. The image acquisition unit acquires images of the coal gangue aggregate particles flowing through it and outputs the raw particle images to the image preprocessing module.

[0056] The image preprocessing module is deployed on the processing server and consists of a processor, memory, and an image processing submodule. In this embodiment, the processing server is configured with an Intel Xeon Silver 4314 processor (16 cores, 2.4GHz), 64GB of DDR4 memory, and an NVIDIA RTX A4000 graphics card. The image processing submodule performs contrast-limited adaptive histogram equalization and gamma correction on the particle image in software, and the processing result is output as the preprocessed image to the particle segmentation module. Preprocessing a single frame of the particle image takes approximately 6ms.

[0057] The particle segmentation module, also deployed on the processing server, includes an edge extraction submodule, a concavity / convexity attribute reversal point detection submodule, a contact surface three-type classification submodule, a temporal stabilization submodule, and a single particle contour segmentation submodule. The edge extraction submodule performs Canny edge detection on the preprocessed image to obtain a particle edge pixel sequence. The concavity / convexity attribute reversal point detection submodule calculates the local curvature directed value of each edge pixel along the particle edge pixel sequence and detects the concavity / convexity attribute reversal point based on the zero-crossing point where the local curvature directed value changes from positive to negative or from negative to positive, while simultaneously satisfying the curvature amplitude significance condition. The contact surface three-type classification submodule classifies the particle contact surface into three types—V-shaped contact, U-shaped contact, and linear contact—based on the ratio of the depth of the edge segment curvature valley between adjacent pairs of concavity / convexity attribute reversal points to the edge segment length. The temporal stabilization submodule performs inter-frame matching on the particle contours in the particle image sequence to obtain the instantaneous displacement field of the particle centroid, and performs multi-frame temporal interpolation on the concavity / convexity attribute inversion points in each time frame based on the instantaneous displacement field of the particle centroid to obtain temporal stabilization inversion points. The single particle contour segmentation submodule performs pixel-level segmentation on the preprocessed image based on the concavity / convexity attribute inversion points to obtain the single particle contour set and labels the contact surface signal. The output of the particle segmentation module includes the single particle contour set and the contact surface signal, and is sent to the particle shape recognition module, the particle size quantization module, and the sorting quality evaluation module.

[0058] The particle shape recognition module includes a particle shape parameter calculation submodule, a multi-pose contour set association submodule, and a particle shape three-class discrimination submodule. The particle shape parameter calculation submodule calculates the aspect ratio, roundness, and angularity of each single particle contour in the single particle contour set. The multi-pose contour set association submodule performs trajectory association on the single particle contours of the same particle in multiple frames of the particle image sequence to obtain the multi-pose contour set, and obtains the final value of the particle shape parameter based on the multi-pose contour set. The particle shape three-class discrimination submodule classifies the particles corresponding to each single particle contour into needle-like particles, angular particles, or rounded particles according to the particle shape parameters, and statistically analyzes the particle shape distribution and the content of needle-like particles.

[0059] The particle size quantization module includes a three-dimensional size joint inversion submodule, a particle size calculation submodule, and a particle size distribution statistics submodule. The three-dimensional size joint inversion submodule performs multi-view joint inversion on the long and short sides of the minimum bounding rectangle of each posture contour in the multi-posture contour set to obtain the true three-dimensional equivalent size. The particle size calculation submodule calculates the particle size based on the true three-dimensional equivalent size. The particle size distribution statistics submodule calculates the histogram distribution of the particle size of each particle in the single-particle contour set to obtain the particle size distribution.

[0060] The sorting quality assessment module includes a lithological fragmentation discrimination submodule, a Beta distribution inversion submodule, and a comprehensive assessment submodule. The lithological fragmentation discrimination submodule determines the lithological fragmentation mode based on the triaxial ratio distribution of the major axis, minor axis, and thickness of the true three-dimensional equivalent size and calculates the lithological mixing ratio. The Beta distribution inversion submodule executes the Beta distribution inversion model based on the three types of distributions: V-shaped contact, U-shaped contact, and linear contact, to obtain an estimated value of the upstream sorting screen aperture deviation. The comprehensive assessment submodule comprehensively assesses the sorting quality of the coal gangue aggregate based on the content of needle-like and flaky particles, particle shape distribution, particle size distribution, and the contact surface signal; outputs the sorting quality assessment result; and when the content of needle-like and flaky particles exceeds the preset content threshold, outputs a needle-like and flaky particle exceeding threshold prompt, a lithological attribution for the needle-like and flaky particle exceeding threshold, and suggestions for adjusting the upstream sorting process.

[0061] The statistical traceability module is deployed on the processing server and interfaces with the enterprise management information system via Ethernet. It summarizes the estimated particle shape distribution, particle size distribution, lithological mixing ratio, and upstream sorting screen aperture deviation to form the statistical traceability data. The statistical traceability data is archived and stored in a database according to production shift, production line section, and coal gangue source batch, providing a long-term historical basis for sorting process adjustments and product quality retrospective.

[0062] The modules exchange data at high speed via Gigabit Ethernet and shared memory. The end-to-end processing time for a single frame of particle image is approximately 25ms, which meets the requirements for online real-time detection at a 100fps acquisition frame rate. The system supports remote configuration and adjustment of preset threshold parameters for each step, enabling rapid migration and adaptation between different coal gangue sorting production lines.

[0063] The above-described embodiments are merely specific descriptions of preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A coal gangue aggregate particle shape image intelligent recognition and sorting quality evaluation method, characterized in that, include: Images of coal gangue aggregate particles flowing through the image acquisition area are collected. The particle images are images formed when the coal gangue aggregate particles flow through the image acquisition area on the sorting conveyor belt. The particle image is preprocessed by performing illumination equalization and contrast enhancement. In the preprocessed image, inversion points of particle edge concavity / convexity attributes are detected. Based on these inversion points, the particle contact surface is located. The preprocessed image is segmented into a set of single-particle contours, and contact surface signals are labeled. These contact surface signals represent the geometric distribution of the particle contact surfaces within the set of single-particle contours. Particle shape parameters are calculated for each single-particle contour in the set, including aspect ratio, roundness, and angularity. Based on these particle shape parameters, the particles corresponding to each single-particle contour are classified as needle-like, angular, or rounded particles. The particle size of each single-particle is measured based on its contour, and the particle size distribution is statistically obtained. The sorting quality of the coal gangue aggregate is comprehensively evaluated based on the needle-like particle content, particle shape distribution, particle size distribution, and the contact surface signals. The sorting quality evaluation result is output, and a needle-like particle exceeding the preset content threshold is displayed as an alert.

2. The coal gangue aggregate grain shape image intelligent recognition and sorting quality evaluation method according to claim 1, characterized in that, The method of detecting the inversion points of concavity / convexity attributes on particle edges and locating particle contact surfaces includes: performing edge extraction on the preprocessed image to obtain a particle edge pixel sequence; calculating the local curvature directed value of each edge pixel along the particle edge pixel sequence; marking the zero-crossing points where the local curvature directed value changes from positive to negative or from negative to positive as the concavity / convexity attribute inversion points; marking the edge segments between adjacent pairs of concavity / convexity attribute inversion points along the particle edge pixel sequence as the particle contact surfaces; and classifying the particle contact surfaces into three types—V-shaped contact, U-shaped contact, and linear contact—based on the ratio of the depth of the curvature valley value of the edge segments between adjacent pairs of concavity / convexity attribute inversion points to the length of the edge segments; the contact surface signal includes the quantity distribution and depth distribution of the three types of V-shaped contact, U-shaped contact, and linear contact in the single particle contour set.

3. The coal gangue aggregate grain shape image intelligent recognition and sorting quality evaluation method according to claim 2, characterized in that, The particle images are a sequence of particle images acquired sequentially on a sorting conveyor belt, and the particle image sequence includes at least three temporally adjacent particle images; the method further includes: performing inter-frame matching on the contours of each particle in the particle image sequence to obtain the instantaneous displacement field of the particle's center of gravity, the instantaneous displacement field of the particle's center of gravity being a set of two-dimensional offset vectors of the center of gravity of the same particle between temporally adjacent frames; performing multi-frame temporal interpolation on the concavity / convexity attribute reversal points in each temporal frame based on the instantaneous displacement field of the particle's center of gravity to obtain temporally stabilized reversal points; the three-type classification of V-shaped contact, U-shaped contact, and linear contact is performed based on the temporally stabilized reversal points.

4. The coal gangue aggregate grain shape image intelligent recognition and sorting quality evaluation method according to claim 3, characterized in that, The method further includes: performing trajectory association on the single particle contours of the same particle in multiple frames of the particle image sequence to obtain a multi-pose contour set of the particle, wherein the multi-pose contour set is a set of single particle contours of the same particle under at least 3 different rolling postures; calculating the aspect ratio, roundness, and edge angle of each posture contour in the multi-pose contour set, and taking the maximum aspect ratio, minimum roundness, and maximum edge angle of the multi-pose as the particle shape parameter; performing multi-view joint inversion on the long and short sides of the minimum bounding rectangle of each posture contour in the multi-pose contour set to obtain the true three-dimensional equivalent size of the particle, wherein the true three-dimensional equivalent size includes the three-axis dimensions of the major axis, minor axis, and thickness; and calculating the particle diameter based on the volume equivalent diameter of the true three-dimensional equivalent size.

5. The coal gangue aggregate grain shape image intelligent recognition and sorting quality evaluation method according to claim 4, characterized in that, The method further includes: determining the lithological fragmentation mode of each particle in the coal gangue aggregate based on the triaxial ratio distribution of the major axis, minor axis, and thickness of the true three-dimensional equivalent size. The lithological fragmentation mode includes three categories: brittle fragmentation of coal gangue, plastic fragmentation of symbiotic rocks, and bedding fragmentation. Among these, a major axis to minor axis ratio greater than 2.0 and a thickness to minor axis ratio less than 0.6 are identified as bedding fragmentation; a difference in the triaxial ratio of major axis, minor axis, and thickness less than a preset triaxial difference threshold is identified as plastic fragmentation of symbiotic rocks; and the rest are identified as brittle fragmentation of coal gangue. The lithological mixing ratio is obtained by statistically analyzing the lithological fragmentation mode distribution. The ratio is the percentage set of brittle fractured particles of coal gangue, plastic fractured particles of symbiotic rock, and bedding fractured particles in the total number of particles; the sorting quality assessment result includes the lithological mixing ratio; when the needle-like and flaky particles exceed the standard, the set of particles identified as needle-like and flaky particles is genetically decomposed according to the lithological fracture mode corresponding to each needle-like and flaky particle, and the needle-like and flaky particles with the lithological fracture mode of bedding fracture are attributed to the cause of bedding fracture, and the needle-like and flaky particles with the lithological fracture mode of brittle fracture of coal gangue are attributed to the cause of brittle fracture, and the lithological attribution of needle-like and flaky particles exceeding the standard is output synchronously in combination with the lithological mixing ratio.

6. The intelligent recognition and sorting quality assessment method for coal gangue aggregate particle shape image according to claim 5, characterized in that, The method further includes: establishing a Beta distribution inversion model based on the constraint of accumulation entropy conservation between the depth distribution of the three types of V-shaped contact, U-shaped contact, and linear contact and the aperture deviation of the upstream sorting screen; the Beta distribution inversion model takes the quantity distribution and depth distribution of the three types of V-shaped contact, U-shaped contact, and linear contact as input and the estimated aperture deviation of the upstream sorting screen as output; executing the Beta distribution inversion model based on the distribution of the three types of V-shaped contact, U-shaped contact, and linear contact to obtain the estimated aperture deviation of the upstream sorting screen; and jointly outputting the estimated aperture deviation of the upstream sorting screen and the lithological mixing ratio to the sorting quality assessment result as a suggestion for adjusting the upstream sorting process.

7. The intelligent recognition and sorting quality assessment method for coal gangue aggregate particle shape image according to claim 1, characterized in that, The acquisition of images of coal gangue aggregate particles is performed by an image acquisition unit deployed above the sorting conveyor belt. The acquisition frame rate of the image acquisition unit is not less than 50fps. During acquisition, the image acquisition unit synchronously triggers an LED array light source to generate uniform illumination. The color temperature of the LED array light source is 5000K to 6500K, and the illuminance uniformity is not less than 85%.

8. The intelligent recognition and sorting quality assessment method for coal gangue aggregate particle shape image according to claim 1, characterized in that, The particle shape parameters are calculated as follows: the aspect ratio is determined based on the ratio of the long side to the short side of the minimum bounding rectangle of the single particle profile; the roundness is determined based on the normalized ratio of the area of ​​the single particle profile to the square of the perimeter of the single particle profile; and the angularity is determined based on the ratio of the total number of edge pixels on the single particle profile whose local curvature exceeds a preset curvature threshold to the perimeter of the single particle profile.

9. The intelligent recognition and sorting quality assessment method for coal gangue aggregate particle shape image according to claim 1, characterized in that, The distinction between needle-like, angular, and rounded particles is performed according to the following rules: particles with an aspect ratio greater than 2.5 are identified as needle-like; particles with an aspect ratio less than or equal to 2.5 and an angularity greater than a preset angularity threshold are identified as angular; particles with an aspect ratio less than or equal to 2.5, an angularity less than or equal to the preset angularity threshold, and a roundness greater than a preset roundness threshold are identified as rounded; the preset content threshold is 15%.

10. A coal gangue aggregate particle shape image intelligent recognition and sorting quality assessment system, used to implement the coal gangue aggregate particle shape image intelligent recognition and sorting quality assessment method according to any one of claims 1-9, characterized in that, include: An image acquisition unit is used to acquire images of coal gangue aggregate particles flowing through the image acquisition area on the sorting conveyor belt. The particle images are formed when the coal gangue aggregate particles flow through the image acquisition area. An image preprocessing module is used to perform illumination equalization and contrast enhancement on the particle images to obtain preprocessed images. The particle segmentation module is used to detect the inversion points of the concavity and convexity attributes of particle edges in the preprocessed image, locate the particle contact surfaces based on the inversion points, segment the preprocessed image into a set of single particle contours, and label the contact surface signals, wherein the contact surface signals are the geometric distribution representations of the particle contact surfaces in the set of single particle contours; the particle shape recognition module is used to calculate the particle shape parameters for each single particle contour in the set of single particle contours and identify the particles corresponding to each single particle contour as needle-like particles, angular particles, or rounded particles, wherein the particle shape parameters include The system includes: an aspect ratio, roundness, and angularity measurement module; a particle size quantification module for measuring the particle size of each individual particle based on its profile and statistically analyzing its particle size distribution; a sorting quality assessment module for comprehensively evaluating the sorting quality of the coal gangue aggregate based on the content of needle-like and flaky particles, particle shape distribution, particle size distribution, and the contact surface signal; outputting the sorting quality assessment results and providing an indication of excessive needle-like and flaky particle content when it exceeds a preset threshold; and a statistical traceability module for summarizing the particle shape distribution and particle size distribution to form statistical traceability data, providing a basis for adjusting the sorting process.

Images