A mine belt intelligent foreign matter identification method and system based on machine vision

Abstract

The present application relates to the technical field of image recognition, in particular to a mine belt intelligent foreign matter identification method and system based on machine vision, wherein the mine belt intelligent foreign matter identification method based on machine vision comprises: acquiring an image of coal material transported by a mine belt, identifying and segmenting a foreign matter area; determining the material type of the foreign matter according to the brightness distribution parameters of the foreign matter area; for the stick-shaped foreign matter prone to misidentification, calculating the alignment deviation index of the reflection area and the boundary of the foreign matter area, and calculating the dissimilarity of the reflection area boundary of the stick-shaped foreign matter in the front and rear frame images; determining the stick-shaped foreign matter discrimination index based on the alignment deviation index and the dissimilarity, and distinguishing the material type of the stick-shaped foreign matter; and performing corresponding processing on the foreign matter according to the distinguished material type. The above method constructs a lightweight identification system independent of deep neural networks, reduces the dependence on large-scale labeled data and hardware computing power requirements, and realizes accurate material classification of mine belt foreign matters.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a machine vision-based intelligent foreign object recognition method and system for mine conveyor belts. Background Technology

[0002] Belt conveyors are specialized equipment used for the continuous transport of materials in mineral mining, production, and transfer processes. They offer advantages such as large transport capacity, strong environmental adaptability, and long transport distances, and are widely used in major mining areas. For example... Figure 1 As shown, during coal mining, belt conveyors often encounter various foreign objects while transporting crushed coal, including coal gangue, discarded steel pipes and iron bars, scrap wood, and abandoned tools. These foreign objects vary in shape and material, and if they are not identified and handled promptly, they can easily cause blockages, scratches, or tears in the transportation system, seriously affecting the safe operation of the mine belt conveyor. Therefore, rapid and accurate material identification and classification of foreign objects in belt-transported coal is a crucial step in achieving safe production in mines.

[0003] Currently, existing methods for identifying foreign objects on mine conveyor belts mainly rely on deep learning-based object detection methods, which train deep neural network models to identify and classify foreign objects. However, this approach has certain limitations: on the one hand, deep learning methods rely on large-scale labeled data for model training, and the inference process consumes enormous computing power, resulting in high-performance hardware support for system deployment, leading to high costs and making it difficult to widely promote in mine sites; on the other hand, for foreign objects with similar surface reflective features but completely different materials, such as wooden tool handles with waxed surfaces and smooth metal rods, existing methods struggle to accurately distinguish their material types, easily leading to incorrect processing strategies (e.g., magnetic separators cannot remove wooden foreign objects), resulting in missed foreign objects and potentially causing safety accidents. Summary of the Invention

[0004] This invention provides a machine vision-based intelligent foreign object identification method and system for mine conveyor belts to solve existing problems.

[0005] The present invention provides a machine vision-based intelligent foreign object identification method and system for mine conveyor belts, which adopts the following technical solution:

[0006] In a first aspect, one embodiment of the present invention provides a machine vision-based intelligent foreign object identification method for mine conveyor belts. The method includes: acquiring images of coal transported by a mine conveyor belt, identifying and segmenting foreign object regions; determining the material type of the foreign object based on brightness distribution parameters of the foreign object region; for easily misidentified rod-shaped foreign objects, calculating the alignment deviation index between the reflective area and the boundary of the foreign object region, and calculating the dissimilarity of the reflective area boundaries of the rod-shaped foreign object in consecutive frames; determining a rod-shaped foreign object discrimination index based on the alignment deviation index and the dissimilarity, and distinguishing the material type of the rod-shaped foreign object based on the rod-shaped foreign object discrimination index; and processing the foreign object accordingly based on the distinguished material type.

[0007] Furthermore, the method further includes: obtaining the circumscribed geometry of the foreign object region, determining the longitudinal and lateral dimensions of the circumscribed geometry; calculating the ratio between the longitudinal and lateral dimensions, and the filling ratio between the area of ​​the foreign object region and the area of ​​the circumscribed geometry; and identifying the rod-shaped foreign object from the foreign object region based on the ratio and the matching of the filling ratio with a preset threshold.

[0008] Furthermore, the method further includes: acquiring optical reflection characteristics at various locations within the region of the rod-shaped foreign object; determining a strong reflection region inside the rod-shaped foreign object based on the distribution of the optical reflection characteristics; and extracting the boundary of the strong reflection region as a reflective area.

[0009] Further, the calculation of the alignment deviation index between the reflective area and the foreign object area boundary includes: geometrically registering the reflective area and the foreign object area to determine the morphological scale difference index of the reflective area relative to the foreign object area; determining the deviation metric between the boundary of the reflective area and the boundary of the foreign object area after registration; and determining the alignment deviation index between the reflective area and the foreign object area based on the morphological scale difference index and the deviation metric; wherein the alignment deviation index is used to characterize the degree of geometric deviation between the reflective area and the foreign object area.

[0010] Further, calculating the dissimilarity of the reflective region boundaries of the stick-shaped foreign object in consecutive frame images includes: establishing a temporal correspondence of the stick-shaped foreign object in adjacent frame images based on belt running parameters and imaging timing, and determining the associated regions of the stick-shaped foreign object in consecutive frames; obtaining the reflective region boundaries of the associated regions and extracting the geometric morphological features of the reflective regions; comparing and analyzing the geometric morphological features in consecutive frames to determine the degree of morphological difference of the reflective region boundaries; and determining the dissimilarity based on the degree of morphological difference; wherein the dissimilarity is used to quantify the temporal variation characteristics of the reflective region morphology.

[0011] Further, the step of determining the rod-shaped foreign object discrimination index based on the alignment deviation index and the dissimilarity, and distinguishing the material type of the rod-shaped foreign object according to the rod-shaped foreign object discrimination index, includes: performing a nonlinear transformation on the alignment deviation index to obtain a geometric deviation correction factor; performing a fusion operation on the dissimilarity and the geometric deviation correction factor to obtain the rod-shaped foreign object discrimination index; and determining the material type of the rod-shaped foreign object according to the numerical relationship between the rod-shaped foreign object discrimination index and a preset threshold.

[0012] Furthermore, the identification and segmentation of the foreign object region includes: extracting the belt region from the image and removing the background region outside the belt region; performing semantic segmentation on the belt region to identify and extract the coal region; and removing the remaining region outside the coal region from the belt region to obtain the foreign object region.

[0013] Further, determining the material type of the foreign object based on the brightness distribution parameters of the foreign object region includes: converting the image data of the foreign object region to a color space containing a brightness channel, and extracting the brightness value of each pixel in the foreign object region; constructing a brightness distribution histogram and a reflection distribution map of the foreign object region based on the distribution of the brightness values; and determining the material type of the foreign object based on the statistical characteristics of the brightness distribution histogram and the reflection distribution map.

[0014] Furthermore, the rod-shaped foreign object includes a surface-treated wooden tool handle and a metal rod-shaped object;

[0015] The process of handling the foreign objects according to their material type includes: for rod-shaped foreign objects identified as being made of metal, controlling a magnetic suction device to perform a rejection operation; for rod-shaped foreign objects identified as being made of non-metallic material, controlling a vacuum adsorption device or a mechanical actuator to perform a rejection operation; and for foreign objects exceeding the system's preset processing capacity, triggering an alarm signal and controlling the conveyor belt to slow down.

[0016] Secondly, another embodiment of the present invention provides a machine vision-based intelligent foreign object identification system for mine conveyor belts, including a host computer and an image acquisition device communicatively connected to the host computer, wherein:

[0017] The image acquisition device is used to acquire images of coal transported on the mine conveyor belt and send the images to the host computer.

[0018] The host computer is used to acquire the image, identify and segment the foreign object region; determine the material type of the foreign object based on the brightness distribution parameters of the foreign object region; for easily misidentified rod-shaped foreign objects, calculate the alignment deviation index between the reflective area and the boundary of the foreign object region, and calculate the dissimilarity of the reflective area boundary of the rod-shaped foreign object in the preceding and following frames; determine the rod-shaped foreign object discrimination index based on the alignment deviation index and the dissimilarity, and distinguish the material type of the rod-shaped foreign object based on the rod-shaped foreign object discrimination index; and perform corresponding processing on the foreign object according to the distinguished material type.

[0019] The beneficial effects of the technical solution of the present invention are:

[0020] In this embodiment of the invention, an image of coal transported by a mine conveyor belt is acquired, and the foreign object region is identified and segmented. The material type of the foreign object is determined based on the brightness distribution parameters of the foreign object region. For rod-shaped foreign objects that are easily misidentified, the alignment deviation index between the reflective area and the boundary of the foreign object region is calculated, and the dissimilarity of the reflective area boundary of the rod-shaped foreign object in the previous and next frame images is calculated. The rod-shaped foreign object discrimination index is determined based on the alignment deviation index and the dissimilarity, and the material type of the rod-shaped foreign object is distinguished based on the rod-shaped foreign object discrimination index. The foreign object is then processed accordingly based on the distinguished material type.

[0021] Thus, this invention establishes a lightweight identification system that does not rely on deep neural networks by constructing a three-level progressive discrimination mechanism: initial material judgment based on brightness distribution parameters, geometric fine judgment based on alignment deviation index, and temporal correction based on dissimilarity. This reduces reliance on large-scale labeled data and hardware computing power requirements while achieving accurate material classification of foreign objects on mine conveyor belts. Furthermore, it uses traditional image processing methods to extract brightness distribution features of the foreign object region to infer material type, replacing deep learning object detection models that require massive training samples and high-performance computing resources. This significantly reduces deployment costs and improves applicability in mine environments. Additionally, it addresses the issue of surfaces treated with anti-corrosion or... Wooden tool handles treated with slippage enhancement and metal rods, which have similar reflective features but drastically different physical properties, are easily misidentified foreign objects. By calculating the alignment deviation index between the reflective area and the foreign object area boundary, the geometric deviation between the reflective area and the foreign object outline is quantified, effectively distinguishing rod-shaped foreign objects of different materials and avoiding the failure of the rejection strategy due to material misjudgment. On the other hand, by combining the temporal variation characteristics of the reflective area boundary with the analysis of the previous and next frame images, the dissimilarity is calculated to characterize the temporal stability of the reflective morphology. The physical difference between the stable reflective morphology of metal materials and the volatile reflective morphology of treated wooden surfaces is utilized to further verify the material recognition results and improve the recognition robustness under noise interference such as cinder adhesion. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A schematic diagram of a mining dataset provided in an embodiment of this application;

[0024] Figure 2 A flowchart illustrating the intelligent foreign object identification method for mine conveyor belts based on machine vision provided in this application embodiment;

[0025] Figure 3 This is a schematic diagram of the architecture of a machine vision-based intelligent foreign object recognition system for mine conveyor belts, provided in an embodiment of this application. Detailed Implementation

[0026] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, provides a specific implementation method, structure, features, and effects of the invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] The following description, in conjunction with the accompanying drawings, details a specific solution provided by the present invention.

[0029] like Figure 2 As shown in the figure, this application provides a machine vision-based intelligent foreign object identification method for mine conveyor belts, including:

[0030] Step S110: Acquire images of coal transported by conveyor belt in the mine, and identify and segment the foreign object area.

[0031] The aforementioned images of coal transported by conveyor belts in mines refer to visual information data collected by optical imaging equipment, containing the material carried on the conveyor belt during operation. This image data primarily characterizes the accumulation morphology and spatial distribution of the continuously transported coal on the belt, while also encompassing various foreign objects mixed in the coal flow, including but not limited to coal gangue, metal components, wooden debris, and tools. The pixel information in the images reflects the optical reflection characteristics of different materials under specific lighting conditions. The coal area exhibits a relatively uniform texture and grayscale distribution, while the foreign object area displays optical characteristics distinct from the coal due to material differences, providing a data foundation for subsequent material identification based on brightness distribution and reflectivity.

[0032] The aforementioned images can be acquired using an industrial-grade high-resolution image acquisition device deployed above the belt conveyor system. During image acquisition, an industrial camera with high sensitivity and wide dynamic range can be mounted at a predetermined height above the conveyor belt, ensuring the optical axis forms a specific angle (e.g., 60°) with the belt plane and faces the direction of coal delivery. This ensures the imaging field of view covers the effective width of the belt and reduces motion blur. During continuous conveyor operation, the image acquisition device performs periodic exposures at preset time intervals, or achieves synchronous acquisition with the belt speed based on encoder trigger signals, thereby obtaining a high-definition image sequence of the current coal on the conveyor belt. The acquired raw image data undergoes preprocessing such as denoising, enhancement, and grayscale conversion to form standardized image data suitable for subsequent foreign object recognition algorithms.

[0033] In this application embodiment, optional implementations for identifying and segmenting the foreign object region include:

[0034] The first implementation method: based on semantic segmentation;

[0035] Optionally, the above-mentioned identification and segmentation of the foreign object region includes: extracting the belt region from the image and removing the background region outside the belt range; performing semantic segmentation on the belt region to identify and extract the coal region; and removing the remaining region outside the coal region from the belt region to obtain the foreign object region.

[0036] The conveyor belt area refers to the image region within the camera's field of view corresponding to the surface of the conveyor belt that actually carries and transports coal. Because industrial cameras are installed in relatively fixed positions in mines, their imaging field of view typically includes background interference information such as supports on both sides of the belt, tunnel walls, or other fixed facilities. These areas do not participate in material transport but may introduce false detections in subsequent image processing. Therefore, after acquiring the original image, pixel areas outside the effective width of the conveyor belt can be removed based on camera calibration parameters or pre-set mask templates, retaining only the area of ​​the conveyor belt surface and the space above it as the processing target. This allows computational resources to be focused on the effective material monitoring area.

[0037] Semantic segmentation is a pixel-level image understanding technique that accurately divides regions of different semantic categories by classifying and labeling each pixel in an image. In the scenario of conveyor belt foreign object recognition, given that coal deposits typically exhibit relatively uniform color and texture features, contrasting with foreign objects of varying shapes and characteristics, a lightweight semantic segmentation model can be used to classify pixels within the conveyor belt area, effectively identifying and extracting the coal region. This process does not rely on extensive labeling of foreign object samples; instead, it utilizes the relatively stable features of coal as a background category, generating a pixel mask of the coal region through model inference, thus marking the portion of the conveyor belt surface covered by coal from the image.

[0038] After obtaining the coal area mask through semantic segmentation, the foreign object area is acquired based on the principle of logical subtraction. Since foreign objects constitute only a small portion of the total material volume during actual transportation and are typically scattered among coal piles or on their surface, the remaining pixel areas within the conveyor belt region that are not marked as coal are identified as candidate foreign object areas. Each independent connected region corresponds to a potential foreign object target. Considering the sparsity of foreign objects in the material flow, it is generally assumed that each segmented foreign object area contains only a single foreign object entity. Based on this, the minimum bounding rectangle or precise contour boundary for each foreign object can be generated as input for subsequent material discrimination algorithms.

[0039] Furthermore, given the continuous motion characteristics of belt conveyor systems, adjacent frames overlap and continue in time and space. To prevent fast-moving foreign objects from being truncated at the boundaries of the image sequence, causing the same object to appear as an incomplete local shape in different frames and thus affecting recognition accuracy, a buffer zone can be set at the boundary between the already identified coal area and the area to be identified. The buffer zone is usually calculated based on the belt speed and the camera shooting time interval. For example, a certain proportion (approximately 10%) of the strip-shaped area at the boundary can be included in the current processing range as a temporal overlap buffer. Through this sliding window-like region management strategy, it is ensured that moving foreign objects can be completely captured even at the inter-frame boundary, maintaining the continuity of foreign object tracking and recognition.

[0040] The second implementation method: segmentation based on traditional image features;

[0041] In this embodiment, the distinguishability of coal and foreign objects in terms of grayscale distribution, texture complexity, or edge gradient can be utilized. Traditional image processing methods such as adaptive threshold segmentation, edge detection operators, or region growing algorithms can be used to directly extract the connected regions of foreign objects that differ from the coal background from the original image or the preprocessed image. The boundaries can be optimized through morphological operations to achieve the localization and segmentation of the foreign object region.

[0042] The third implementation method: an end-to-end segmentation method based on deep learning;

[0043] In this embodiment, a pre-trained deep neural network model can be used to perform end-to-end instance segmentation or panoramic segmentation on the input image. The pixel-level mask and bounding box of the foreign object target can be directly generated through forward inference of the model, so as to realize the automatic identification and accurate segmentation of the foreign object region.

[0044] The fourth implementation method: segmentation based on spectral feature enhancement;

[0045] In this embodiment, the diagnostic absorption characteristics or reflectance differences of different materials in specific spectral bands can be utilized to enhance the spectral contrast between foreign objects and coal through algorithms such as spectral angle mapping, normalized difference index, or principal component analysis. Then, based on threshold segmentation or classification algorithms, the foreign object region can be accurately extracted.

[0046] Step S120: Determine the material type of the foreign object based on the brightness distribution parameters of the foreign object area.

[0047] In this application embodiment, optional implementation methods for determining the material type of the foreign object include:

[0048] The first implementation method: brightness distribution analysis based on optical reflection characteristics;

[0049] Optionally, step S120 may include: converting the image data of the foreign object region to a color space that includes a brightness channel, and extracting the brightness value of each pixel in the foreign object region; constructing a brightness distribution histogram and a reflection distribution map of the foreign object region based on the distribution of brightness values; and determining the material type of the foreign object based on the statistical characteristics of the brightness distribution histogram and the reflection distribution map.

[0050] Optical reflection characteristics characterize the combined intensity of specular and diffuse reflection response of a material surface to incident electromagnetic radiation. This response depends on the inherent physical properties of the material surface, such as microscopic roughness, refractive index, and absorption coefficient. In mine conveyor belt scenarios, metallic foreign objects, due to their smooth surfaces and high conductivity, typically exhibit high reflectivity, forming continuous high-brightness reflective areas in images. Mineral foreign objects, with relatively rough surfaces and complex mineral compositions, have moderate reflectivity and a more uniform brightness distribution. In contrast, wooden foreign objects, with their loose fibrous structure and strong light absorption, typically exhibit low reflectivity; even after surface waxing, their reflective areas show a fragmented and discontinuous spatial morphology. These differences based on physical optical mechanisms constitute the theoretical basis for distinguishing different materials through image brightness analysis.

[0051] To achieve objective quantification of optical reflectance characteristics, the raw RGB data acquired by the image acquisition device can be converted to a device-independent color space containing an independent luminance channel. The raw RGB image is affected by the spectral response characteristics of the camera sensor and the color temperature variations of the ambient lighting source, making it difficult for the pixel values ​​to directly reflect the true reflective physical quantities of the material surface. By introducing a standard white point reference (such as the standard light source D65), the RGB data is mapped to the Lab color space, allowing the separation of the L channel, which represents brightness information, and the a and b channels, which represent chromaticity information. This conversion process effectively eliminates the interference of light source color and camera characteristics on luminance analysis, ensuring that the extracted L value reflects only the objective reflectance intensity of the incident light from the material surface, providing a standardized physical quantity basis for subsequent statistical analysis.

[0052] After obtaining the luminance channel data, the luminance values ​​of all pixels within the foreign object region can be statistically summarized to construct a luminance distribution histogram. The luminance distribution histogram uses discrete luminance levels as the x-axis and the frequency of pixels corresponding to each luminance level as the y-axis, forming a statistical distribution curve characterizing the overall reflectivity of the foreign object. By analyzing the peak position of this distribution curve, the typical luminance level of the foreign object can be determined. The uniformity of the reflective area can be assessed by calculating the dispersion of the distribution, while the skewness of the distribution reflects the ratio between high-reflectivity and low-reflectivity areas. These statistics collectively constitute a numerical feature vector describing the optical properties of the foreign object, which can be used for comparison with preset thresholds or training samples.

[0053] Besides statistical distribution, the spatial distribution pattern of highly reflective areas on the surface of foreign objects also contains important material discrimination information. Based on the brightness distribution analysis results, by setting an adaptive threshold or using an inter-class variance maximization algorithm, the foreign object region can be binary classified into high-reflectivity and low-reflectivity regions, thereby generating a reflectivity distribution map that only marks the spatial location of highly reflective pixels. This distribution map retains the boundary contours, geometric shapes, and relative positional information of the high-reflectivity regions within the overall scope of the foreign object, reflecting the regularity of the material surface microstructure. Metallic materials typically exhibit axially extending, regular bright bands, while surface-treated wood materials exhibit scattered or irregular bright spots. This difference in geometric morphology provides a spatial dimension for material subdivision.

[0054] Based on the aforementioned statistical characteristics of brightness distribution and spatial characteristics of reflectivity distribution, the material type of foreign objects can be determined through rule inference or pattern matching. The rule inference path uses empirical differences in the average brightness, proportion of reflective areas, and distribution concentration of metals, ores, and wood to set grading thresholds for logical discrimination. The pattern matching path fuses the brightness histogram features and reflectivity distribution geometric features into a feature vector, which is then used to calculate similarity with a sample library of known material types or input into a lightweight classification model for inference. Neither path relies on large-scale labeled data to train deep neural networks; material classification can be rapidly determined using only traditional image processing algorithms, significantly reducing computational resource requirements and deployment costs.

[0055] It is understandable that the homogeneous crystalline structure of metals makes their surfaces smooth and isotropically reflective, forming regular and continuous strong reflective bands. In contrast, even with a wax coating, the inherent pits and depressions in the fiber structure of wood still cause microscopic irregularities at the edges of reflective areas, and the temporal sequence of reflective patterns is unstable due to changes in viewing angle during belt movement. The aforementioned solution, by eliminating ambient light interference through Lab color space conversion, amplifies subtle differences in brightness statistics and spatial geometric patterns into quantifiable classification features. This allows for accurate differentiation of foreign objects with similar surface reflective features but drastically different physical properties without employing high-performance deep learning inference.

[0056] The second implementation method: spectral identification based on spectral diagnostic features;

[0057] Different mineral compositions, metals, and organic materials exhibit diagnostic absorption valleys or reflection peaks in specific spectral bands. This embodiment extracts the spectral curves of the foreign object region in the visible to short-wave infrared bands, calculates the spectral angular distance or spectral characteristic ratio, and identifies the material category by matching it with a preset spectral database or classification model. For example, ferrous metals have high reflectivity in specific bands, while lignocellulose exhibits obvious absorption characteristics in the short-wave infrared band, thereby distinguishing between seemingly similar metal rods and wooden handles.

[0058] The third implementation method: texture analysis based on surface microtexture;

[0059] Metallic materials typically exhibit smooth or regularly processed textures, wood surfaces have fibrous or pitted microstructures, and mineral surfaces exhibit crystalline or irregularly broken textures. This implementation method extracts texture parameters such as contrast, roughness, and directionality by calculating the gray-level co-occurrence matrix, local binary pattern, or texture energy features of the foreign object region. It then determines the material type of the foreign object by matching the texture feature vector with the similarity of training samples or by classifier inference.

[0060] The fourth implementation method: polarization characteristic analysis based on polarization imaging;

[0061] Smooth metal surfaces exhibit strong polarization retention capabilities, while rough wooden surfaces cause significant depolarization effects. This implementation method acquires images of foreign objects at different polarization angles within the same scene, calculates the degree of polarization image or polarization angle distribution, and extracts polarization feature parameters. Based on the physical principle that metal materials typically exhibit high polarization and wood materials exhibit low polarization, a polarization threshold is set or a polarization feature classification model is constructed to distinguish foreign objects with similar reflectivity but different materials. This method demonstrates particularly good recognition stability for wooden tools with wax or paint coatings.

[0062] Step S130: For easily misidentified stick-shaped foreign objects, calculate the alignment deviation index between the boundary of the reflective area and the boundary of the foreign object area, and calculate the dissimilarity of the boundary of the reflective area of ​​the stick-shaped foreign object in the images of the previous and next frames.

[0063] It is understandable that in mine belt conveyor systems, rod-shaped foreign objects pose a unique risk of mechanical damage due to their slender geometry and rigid physical properties. Their sharp ends or edges can easily scratch or even tear the belt surface during the tumbling process with coal, causing serious production accidents. However, wooden tool handles (such as shovel handles with anti-corrosion or anti-slip treatment) and metal rod-shaped objects (such as discarded steel pipes or iron bars) often exhibit similar reflective characteristics in initial optical identification: the former has localized specular reflection due to its surface wax coating, while the latter produces high-brightness reflection due to its metallic luster. The two tend to fall into similar ranges in terms of brightness distribution statistics, making it difficult to reliably distinguish them based on a preliminary judgment mechanism based solely on reflective intensity. Such material misjudgment will directly cause the processing strategy to fail. If the system misjudges the wooden stick as metal and activates the magnetic separator, or if it misjudges it as ordinary coal and ignores the processing, such high-risk foreign objects will be missed in the detection process and enter the downstream equipment. Therefore, it is necessary to start a two-level discrimination process based on geometric alignment analysis and spatiotemporal consistency verification for stick-shaped foreign objects to accurately distinguish the material categories with similar surface optical properties but completely different physical properties, and ensure that the subsequent physical separation methods are accurately matched with the actual material of the foreign object.

[0064] Optionally, the above-mentioned machine vision-based intelligent foreign object recognition method for mine conveyor belts further includes: acquiring the circumscribed geometry of the foreign object region, determining the longitudinal and lateral dimensions of the circumscribed geometry; calculating the ratio between the longitudinal and lateral dimensions, and the filling ratio between the area of ​​the foreign object region and the area of ​​the circumscribed geometry; and identifying the rod-shaped foreign object from the foreign object region based on the matching of the ratio and the filling ratio with a preset threshold.

[0065] The aforementioned circumscribed geometry refers to the smallest closed shape that can completely enclose all pixels of the foreign object region and satisfy specific geometric constraints. Its function is to abstract the irregular outline of the foreign object into a standard geometric shape with clear longitudinal and lateral directions, thereby extracting quantitative indicators that can be used for shape classification. In the scenario of foreign object recognition in mine conveyor belts, since the foreign objects mixed into the coal flow have various shapes, directly judging the shape based on the original outline lacks a stable comparison benchmark. However, by constructing a circumscribed geometry, complex boundaries can be normalized into regular shapes such as rectangles, ellipses, or convex hulls, thus obtaining physically meaningful length and width measurements. The construction of this geometry must ensure that it completely contains all pixels of the foreign object region, and that the tangent points or support points of its boundary with the foreign object outline satisfy specific optimization criteria, such as minimizing the area or perimeter, to ensure the effectiveness of shape approximation.

[0066] The implementation method for determining the longitudinal and lateral dimensions of the circumscribed constraint geometry can be selected based on the characteristics of the chosen geometric shape. For example, when using the minimum bounding rectangle as the constraint geometry, the set of convex hull vertices of the object's contour can be extracted using the convex hull algorithm. Then, the rotating caliper method can be used to find the envelope rectangle with the smallest area among the convex hull vertices. The long side of this rectangle corresponds to the longitudinal dimension, and the short side corresponds to the lateral dimension. When using the minimum bounding ellipse, the major and minor axes of the equivalent ellipse can be calculated based on the second moment of the object's contour as the longitudinal and lateral dimensions. When using a directional bounding box, the principal axis direction of the object's contour can be determined through principal component analysis. The maximum span along the principal axis direction is used as the longitudinal dimension, and the maximum span along the perpendicular direction is used as the lateral dimension. These dimensional measurements reflect the degree of anisotropy of the object in spatial extension, providing a geometric basis for distinguishing between rod-shaped objects with significant length and width differences and approximately equiaxial block-shaped objects.

[0067] After obtaining the longitudinal and lateral dimensions, their proportional relationship can be calculated to quantify the elongation of the foreign object. This ratio is typically defined as the quotient of the longitudinal and lateral dimensions. When this value exceeds a certain threshold, it indicates that the foreign object has significant elongation characteristics in a certain direction, consistent with the morphological characteristics of a rod-shaped object. Simultaneously, the fill ratio of the foreign object's area to the area of ​​its circumscribed constraint geometry is calculated. This ratio reflects the proportion of the actual pixel area of ​​the foreign object to its theoretical envelope space. For solid and regular rod-shaped foreign objects, this ratio is usually high and concentrated, while for hollow or extremely irregular objects, the ratio is relatively low. By combining the length-to-width ratio and the fill ratio, it is possible to effectively distinguish between slender solid rod-shaped objects and flat or lumpy foreign objects, avoiding misjudgments that may arise from a single indicator.

[0068] In the above scheme, by pre-setting aspect ratio thresholds and fill ratio thresholds, when the aspect ratio of a foreign object region exceeds the elongation threshold and the fill ratio meets the solid object threshold range, the foreign object is determined to belong to the rod category, triggering the subsequent fine material discrimination process. The geometric intuition of this recognition mechanism lies in the fact that metal rods and wooden tool handles both present slender cylindrical or prismatic shapes, with aspect ratios significantly larger than those of ordinary coal blocks or gangue. As rigid tools, they also have a high solid fill ratio. Therefore, through the joint judgment of geometric constraints, target objects that require further analysis of alignment deviation index and dissimilarity can be reliably screened from foreign object regions in complex backgrounds, achieving targeted allocation of recognition resources and optimization of computational efficiency.

[0069] In addition to the above schemes, the embodiments of this application can also identify rod-shaped foreign objects through one of the following methods: (1) A linear structure detection method based on the Hough transform spatial domain voting mechanism, which identifies candidate regions with significant linear geometric features in the image by accumulating sine curve voting on edge pixels in the parameter space. The Hough transform is applied to the preprocessed edge image of the foreign object region to map the collinear edge points in the Cartesian coordinate system to the intersection points in the parameter space, and the orientation and position of the straight line segment are determined by detecting the local extreme points in the parameter space. When there is a significant straight line segment or approximately straight line structure with a length exceeding a preset threshold in the foreign object region, it is determined that the region has the axial extension characteristics of a rod-shaped foreign object, which is especially suitable for detecting metal rods and wooden tool handles with obvious edges or cylindrical generatrices. (2) A skeleton topology analysis method based on morphological thinning algorithm, which extracts the central skeleton structure that maintains the topological connectivity of the original region by performing iterative erosion and opening operations on the binarized foreign object region. The skeleton is represented by a continuous curve with a single pixel width. Its total length reflects the axial extension of the foreign object, while the number of branch points reflects the shape complexity. When the ratio of the extracted skeleton length to the long side of the minimum bounding rectangle of the region exceeds a certain threshold, and the number of skeleton branch points is less than the preset value, it indicates that the region presents a slender, single-connected topological structure, which conforms to the geometric characteristics of low branching and high extension of a rod-shaped object. It can effectively distinguish rod-shaped foreign objects from coal gangue with complex branching or clumping morphology. (3) Based on the second-order central moment inertial principal axis analysis method, by calculating the second-order moment matrix of the pixel point of the foreign object region relative to the centroid, the direction of the principal axis and the direction of the secondary axis of the region are determined, and then the ratio of the inertial moments describing the anisotropy of the region is obtained. For rod-shaped foreign objects, their mass distribution shows a significant directional aggregation. The inertial moment along the long axis is much greater than that along the short axis, resulting in a significantly larger ratio of the long and short axis inertial moments than that of general block-shaped foreign objects. By setting the discrimination threshold of this ratio and combining it with the auxiliary verification of the centroid offset, the stick-shaped target with obvious directional extension can be quantitatively identified. This method has good directional invariance for stick-shaped foreign objects that are tilted or partially obscured. (4) Morphological discrimination method based on boundary contour curvature distribution. By calculating the curvature value of each sampling point on the boundary curve of the foreign object region, the frequency of curvature sign change and the distribution characteristics of curvature amplitude are analyzed. The boundary contour of the stick-shaped foreign object is usually composed of two approximately parallel long sides and two short sides. The curvature value of the long side region is close to zero and the sign is stable. The short side region shows a limited curvature extreme value. The overall curvature distribution shows the characteristics of low frequency and low amplitude. The boundary curvature of the clump-shaped foreign object changes frequently and irregularly. By statistically analyzing the concentration of the curvature histogram or calculating the bending energy of the boundary curve, the stick-shaped object with regular long strip boundary and other foreign object categories with irregular boundaries can be effectively distinguished.

[0070] Optionally, the above-mentioned intelligent foreign object recognition method for mine conveyor belts based on machine vision further includes: acquiring optical reflection features at various locations within the region of the rod-shaped foreign object; determining the strong reflection region inside the rod-shaped foreign object based on the distribution of the optical reflection features; and extracting the boundary of the strong reflection region as the reflective area.

[0071] The aforementioned optical reflection characteristics are physical quantities characterizing the intensity of a material surface's reflection response to incident electromagnetic radiation of a specific wavelength. Under visible light imaging conditions, they typically correspond to the numerical representation of the image's brightness channel, and their magnitude depends on factors such as the surface roughness, refractive index, and incident angle. In the scenario of identifying rod-shaped foreign objects, methods for obtaining these characteristics include directly extracting the brightness channel values ​​after color space conversion, calculating the reflectance ratio of a specific wavelength in a multispectral image, or obtaining polarization information through polarization imaging as a comprehensive representation of the reflection characteristics. These characteristics form a continuous or discrete spatial distribution field within the rod-shaped foreign object region, where high-value regions correspond to smooth surfaces or local areas with high reflectivity, while low-value regions correspond to rough surfaces or shadowed areas, collectively forming the data basis for distinguishing the optical properties of different material surfaces.

[0072] Determining strong reflection regions based on the distribution of optical reflection characteristics is essentially a process of classifying pixels in the feature space to extract high-response areas. Adaptive threshold segmentation based on statistical distribution can be used, automatically determining the segmentation threshold by analyzing the extreme values ​​or mean-variance characteristics of the feature histogram. Alternatively, an optimization method maximizing inter-class variance can be employed, dividing the foreign object region into two classes to maximize the difference in feature values ​​between classes, identifying the class with the higher mean as the strong reflection region. Clustering algorithms can also be used for unsupervised classification along the feature dimension, aggregating pixels with similar reflection characteristics into clusters and selecting high-value clusters as strong reflection regions. The common goal of these methods is to separate the set of pixels with high reflectivity from the complex surface reflection field. This set typically manifests spatially as continuous or fragmented bright spots or bands, and its geometric distribution is closely related to the microscopic physical structure of the material surface.

[0073] The implementation of extracting the boundary of a strongly reflective region involves contour detection and boundary tracking of the binarized region. Specifically, edge detection operators can be used to identify gradient abrupt changes between the strongly reflective region and the surrounding low-reflective region, generating a boundary map composed of edge points; alternatively, a chain code-based contour tracking algorithm can be used to traverse all boundary points along adjacency relationships, starting from the boundary pixels of the strongly reflective region, to form a closed boundary chain; or morphological boundary extraction operations can be used to obtain the set of boundary pixels located at the outer edge of the strongly reflective region by utilizing the difference between dilation and erosion operations. The extracted boundary is represented as a continuous or segmented spatial curve, and its geometric features (such as smoothness, curvature variation, and axial extensibility) contain rich information about the regularity of the material surface, providing necessary geometric elements for subsequent calculation of the alignment deviation index.

[0074] It is understandable that the smooth, homogeneous surface of a metal rod causes incident light to undergo primarily specular reflection, forming a highly reflective band that extends continuously along the axial direction on the imaging plane. The boundaries of this highly reflective region are regular and highly parallel to the outer contour of the object. In contrast, while the treated wooden tool handle exhibits localized specular reflection due to the coating, the microscopic pits and depressions in the underlying fiber structure result in a scattered distribution of the reflected light field, with broken and irregular boundaries of the highly reflective region. By extracting the boundaries of the highly reflective region and performing geometric matching analysis with the overall boundary of the object, this macroscopic geometric manifestation stemming from the microscopic physical differences in the materials can be quantified. This provides a criterion independent of brightness statistics for distinguishing rod-shaped objects with similar surface optical properties but fundamentally different characteristics.

[0075] In addition to the above schemes, the embodiments of this application can also use one of the following methods to extract reflective regions: (1) Adaptive threshold segmentation method based on pixel brightness statistical distribution: By analyzing the bimodal or multimodal characteristics of the brightness histogram in the foreign object region, the optimal segmentation threshold is automatically calculated to distinguish between high reflective and low reflective pixels. Based on the inter-class variance maximization criterion or the entropy maximization criterion, without the need to preset a fixed brightness threshold, the method dynamically determines the segmentation boundary according to the current image illumination conditions and material reflective characteristics, and marks the set of pixels with brightness values ​​higher than the adaptive threshold as candidate reflective regions, effectively adapting to the overall image brightness shift caused by the fluctuation of illumination at the mine site, and ensuring the robustness of reflective region extraction. (2) Region growth method based on seed point and neighborhood similarity constraints: Local extreme points of optical reflection features can be located in the foreign object region as growth seeds, and then iteratively expanded to the eight-neighbor or four-neighbor directions according to the predetermined brightness similarity criteria (such as the brightness difference between the current pixel and the seed point or the already grown region being less than a specific tolerance) until the stopping condition is met (such as edge gradient abrupt change or brightness difference exceeding the limit). (3) Bright detail extraction method based on mathematical morphology top-hat transformation: Morphological opening operation can be performed on the image of the foreign object region using structuring elements of specific shapes and sizes to eliminate bright reflective details and retain the background trend. Then, by the difference operation between the original image and the opening operation result, the reflective area that is significantly brighter than the surrounding local background is highlighted. (4) Unsupervised classification method based on feature space clustering analysis: The brightness value of each pixel in the foreign object region is mapped to sample points in a one-dimensional feature space. Using algorithms such as K-means clustering, mean shift, or Gaussian mixture model, the pixels are divided into multiple categories such as high reflectivity, medium reflectivity, and low reflectivity based on the natural clustering characteristics of brightness distribution. This method does not rely on a preset brightness threshold, but automatically discovers the brightness center and distribution range of the high reflectivity region through a data-driven approach, and aggregates all pixels belonging to the high reflectivity category into reflective regions.

[0076] The aforementioned alignment deviation index is a comprehensive evaluation indicator used to quantify the degree of geometric deviation between the boundaries of the reflective area and the foreign object area. It reflects the modulation effect of the microscopic physical structure of the foreign object surface on the spatial distribution of optical reflection by measuring the difference in matching between the two areas in terms of shape regularity and spatial scale. This index integrates information on the fit between the reflective area boundary and the foreign object contour, as well as information on the degree of expansion of the reflective area during the scale registration process. The former characterizes the morphological consistency between the reflective boundary and the foreign object boundary, while the latter characterizes the fullness ratio of the original reflective area relative to the overall scale of the foreign object. For smooth and uniform metal rods, the reflective area typically forms a continuous and regular band-like distribution along the axial direction, with a boundary morphology that highly matches the outer contour of the foreign object. During scale registration, only a slight expansion is needed to coincide with the foreign object area, corresponding to a smaller alignment deviation index value. However, for wooden tool handles with treated surfaces but still retaining microscopic pitted structures, the reflective area exhibits a discrete and fragmented spatial distribution, with a significant deviation from the regular outer contour. During scale registration, a substantial expansion in the width direction is required to match the foreign object area, corresponding to a larger alignment deviation index value. Therefore, through a quantitative description at the geometric level, this index effectively captures the essential differences in the spatial distribution patterns of reflective light between materials with similar surface optical properties but drastically different microstructures, providing a geometric basis for distinguishing easily confused metallic and non-metallic rod-shaped foreign objects, independent of brightness statistics.

[0077] In this application embodiment, optional implementations for calculating the alignment deviation index include:

[0078] The first implementation method: morphological analysis based on registration expansion;

[0079] Optionally, the above calculation of the alignment deviation index between the reflective area and the foreign object area includes: geometrically registering the reflective area and the foreign object area to determine the morphological scale difference index between the reflective area and the foreign object area; determining the deviation measure between the boundary of the reflective area and the boundary of the foreign object area after registration; and determining the alignment deviation index between the reflective area and the foreign object area based on the morphological scale difference index and the deviation measure. The alignment deviation index is used to characterize the degree of geometric deviation between the reflective area and the foreign object area.

[0080] Geometric registration is a crucial process for establishing the spatial correspondence between the reflective area and the foreign object area. Its purpose is to eliminate the incomparability caused by scale differences and positional offsets between the two areas, placing them within a unified geometric reference framework through segmented boundary expansion and deformation operations. Specifically, the foreign object area is first fitted as a reference rectangle with clearly defined length and width directions. This fitting process uses the convex hull vertex set and a rotational caliper method to determine the minimum bounding rectangle, using the two long sides of this rectangle as the axial reference and the two short sides as the width reference. Then, the internal reflective area is also fitted as a rectangle, constrained by the length and width directions parallel to the fitted rectangle of the foreign object area. Operations are performed on the four sides of the fitted rectangle of the reflective area, translating the long and short sides outwards until they coincide with the corresponding sides of the fitted rectangle of the foreign object area. Simultaneously, stretching is performed along the edges to align the corner points. After registration, the two region rectangles completely overlap, forming comparable geometric scales, and the displacement parameters during the translation process are recorded.

[0081] The morphological scale difference index characterizes the extent to which the reflective area expands relative to the overall size of the foreign object during registration. It is achieved by quantifying the displacement during the aforementioned edge translation operation. In the width-sensitive direction, the translation distance in the width direction of the two long sides of the fitted rectangle for the reflective area is recorded as the area is translated outwards until it coincides with the long side of the fitted rectangle for the foreign object region. This distance is denoted as... and These two parameters directly reflect the fullness and expansion requirements of the original reflective area in the width direction; based on this translation distance and the fitted rectangle width of the foreign object area... The ratio is used to calculate the width expansion ratio. ,Right now: The larger the ratio, the greater the expansion of the reflective area needs to match the boundary of the foreign object. This corresponds to a relatively smaller and more fragmented original reflective area, revealing the physical characteristics of light field dispersion caused by microscopic pits on the material surface.

[0082] Deviation metrics are used to quantify the inconsistencies in position and shape between the boundaries of the two registered regions. Their implementation relies on sampling and statistical analysis of boundary deviations at a uniform scale. Within the registered geometric framework, a rectangle is fitted over the foreign object region to select... At each sampling point, the Euclidean distance between the boundary of the reflective area and the boundary of the foreign object area is measured along a direction perpendicular to the side of the rectangle, and this distance is used as the deviation distance. This distance reflects the degree of spatial separation between two boundaries at a specific location; the distance set is obtained after traversing all points on the fitted rectangle. Then, the arithmetic mean of the deviation distances of all sampling points is calculated, and the width of the fitted rectangle is used to represent the foreign object region. As a normalization factor, calculate the normalization boundary inconsistency index. Inconsistency index of normalized boundaries Used to quantify the degree of edge mismatch between reflective and foreign object areas:

[0083] ,

[0084] Normalized boundary inconsistency index The larger the value, the more significant the deviation between the boundary of the reflective area and the overall outline of the foreign object, and the more irregular the corresponding surface microstructure.

[0085] Alignment Deviation Index A comprehensive discrimination index is formed by integrating morphological scale difference indicators and deviation measures, and its mathematical expression is as follows:

[0086]

[0087] in, The width of the fitted rectangle for the foreign object region characterizes the lateral scale of the foreign object; This is the average of the aforementioned boundary deviation distances, reflecting the overall degree of boundary deviation. and These represent the translational distances of the two long sides of the reflective area in the width direction, reflecting the absolute amount of scale expansion. The above calculation formula linearly superimposes the dimensionless boundary inconsistency with the width-sensitive expansion distance, and utilizes the width... Normalization is performed to make the final index It can comprehensively reflect the degree of geometric deviation of the reflective area relative to the overall outline of the foreign object and the degree of relative scale fullness, providing a quantitative basis for distinguishing different materials.

[0088] In the above scheme, for a smooth and homogeneous metal rod, its reflective area forms a continuous and regular band-like distribution along the axial direction. The band is well-shaped and highly parallel to the outer contour of the foreign object. During geometric registration, only a small edge translation is required to achieve boundary overlap, resulting in… and The value is small, and the boundary deviation distance is small. The mean is low, and ultimately The value is relatively small; however, for wooden tool handles with a waxed surface but still retaining a fibrous pitted structure, the reflective area exhibits a discrete and fragmented spatial distribution, with irregular edges and significant deviations from the outer contour. During registration, a large-scale edge translation in the width direction is required to expand to the scale of the foreign object boundary, resulting in… and The value is large, and the boundary deviation distance is... The average value is relatively high, and ultimately The value is relatively large. By observing the difference in the value of this index, it is possible to effectively distinguish between metallic and non-metallic rod-shaped foreign objects with similar surface optical properties but completely different microscopic geometric structures.

[0089] Furthermore, it is understandable that the translation distance of the shorter side... and The extent of the reflective area's expansion along its length was recorded, but in the material discrimination index calculation of this scheme, it was only used as an auxiliary geometric parameter and did not directly participate. Value operations.

[0090] The second implementation method: a shape description method based on moment features;

[0091] In this embodiment, shape descriptors such as the central moment, principal axis of inertia, and aspect ratio of the two regions can be calculated separately. By comparing the angle between the principal axes of inertia, the difference in aspect ratio, and the distance of the Hu moment invariant, the degree of deviation of the shape regularity of the two regions can be comprehensively evaluated. For metal rods, the principal axes of inertia of the reflective area and the foreign object area are highly consistent, their aspect ratios are similar, and the difference in moment characteristics is small. However, for wooden tool handles with pitted surfaces, the fragmented distribution of the reflective area causes a deviation between its principal axis of inertia and the overall direction of the foreign object, resulting in a significant difference in aspect ratio and a large distance in moment characteristics. Based on this, a geometric deviation evaluation value equivalent to the alignment deviation index can be generated.

[0092] The third implementation method: boundary matching based on distance field;

[0093] In this embodiment, a distance transformation is first performed on the foreign object region to generate a distance field map from each pixel to the nearest boundary. Then, the boundary of the reflective region is mapped onto this distance field, and the statistical characteristics of the distance values ​​corresponding to the boundary points (such as mean, variance, or extreme values) are calculated. This statistic directly reflects the distance and uniformity of the reflective region boundary relative to the foreign object region boundary. Simultaneously, the fill ratio of the reflective region area to the foreign object region area is calculated as a scale difference index. A larger mean value in the distance field indicates that the reflective region boundary is farther from the outer contour of the foreign object, while a smaller fill ratio indicates that the reflective region is smaller relative to the overall foreign object. Combining these two factors allows for the construction of a comprehensive index characterizing geometric alignment deviation.

[0094] Fourth implementation method: Feature point-based structural alignment;

[0095] In this embodiment, curvature extrema, corner points, or zero-crossing points can be detected on the boundaries of the reflective area and the foreign object area as feature markers to establish a topological correspondence of the feature point set. The Euclidean distance deviation between corresponding points and the affine transformation residual of the feature point distribution are then calculated. For regular metallic reflective strips, the feature point distribution exhibits a linear arrangement and corresponds one-to-one with the feature points on the foreign object boundary, resulting in a smaller transformation residual. For irregular wooden reflective areas, the feature point distribution is scattered and the correspondence is ambiguous, leading to a larger transformation residual. By normalizing the residual and the difference in point set distribution, a numerical index representing the boundary alignment deviation can be generated.

[0096] The aforementioned dissimilarity is a temporal consistency index used to quantify the degree of variation in the boundary morphology of the reflective region of the same rod-shaped foreign object in consecutive time-series images. It characterizes the stability of the material's surface microstructure in response to changes in viewing angle by evaluating the geometric contour differences of corresponding reflective regions in consecutive frames. For smooth and homogeneous metal rods, due to the uniform internal crystal structure and regular surface processing, the boundary morphology of the specular reflection region on the surface remains relatively stable during changes in the relative positions of the light source, camera, and target caused by belt transport, exhibiting high geometric similarity between consecutive frames and corresponding to a small dissimilarity value. However, for wooden tool handles with treated surfaces but still retaining the micro-pitted structure of the fibrous substrate, although the local coating layer produces specular reflection, the unevenness of the substrate structure causes random changes in the light reflection path when the viewing angle changes. This results in significant edge detail jitter and contour deformation of the reflective region boundary between adjacent frames, exhibiting low temporal stability and corresponding to a large dissimilarity value. Therefore, by capturing the temporal evolution characteristics of reflective patterns, this index provides a temporal discrimination dimension independent of single-frame geometric alignment analysis, effectively distinguishing between metallic and non-metallic materials with similar surface optical features but fundamental differences in microscopic physical structure, and enhancing the robustness of discrimination against transient noise interference such as slag adhesion and local occlusion.

[0097] In this application embodiment, an optional implementation for calculating the dissimilarity of the reflective region boundaries of the stick-shaped foreign object in consecutive frame images includes:

[0098] The first implementation method is a morphological analysis method based on temporal correspondence and chain code comparison;

[0099] Optionally, the above calculation of the dissimilarity of the reflective region boundaries of the stick-shaped foreign object in consecutive frame images includes: establishing a temporal correspondence of the stick-shaped foreign object in adjacent frame images based on belt running parameters and imaging timing, and determining the associated regions of the stick-shaped foreign object in consecutive frames; obtaining the reflective region boundaries of the associated regions and extracting the geometric morphological features of the reflective regions; comparing and analyzing the geometric morphological features in consecutive frames to determine the degree of morphological difference of the reflective region boundaries; and determining the dissimilarity based on the degree of morphological difference; wherein, the dissimilarity is used to quantify the temporal variation characteristics of the reflective region morphology.

[0100] The aforementioned temporal correspondence refers to the mechanism for establishing the association of pixel coordinate mappings of the same physical object across different frames in a continuously acquired image sequence, based on the kinematic parameters of the belt conveyor system and imaging temporal constraints. Establishing this temporal correspondence relies on the precise calibration of the belt's running speed and the shooting time interval. After estimating the displacement vector of the stick-shaped object in adjacent frames using a motion prediction model and determining its approximate location, local feature descriptors such as scale-invariant feature transformation or accelerated robust features are further used to search for matching feature points in the prediction neighborhood. Geometric consistency verification and the RANSAC random sampling consistency algorithm are used to eliminate false matches, accurately locating the associated region of the stick-shaped object in the subsequent frame. This ensures the spatial consistency of the analyzed object between frames and avoids object misalignment or tracking loss due to belt movement.

[0101] Normalized differential chain code is a discrete description method for encoding the geometry of closed boundaries. It transforms continuous geometric contours into discrete directional sequences by recording the directional changes between adjacent pixels on the boundary curve. The directional code typically defines eight basic directions (0-7), corresponding to eight azimuth angles such as 0°, 45°, and 90°. The differential chain code further records the changes in adjacent chain code values ​​to eliminate the influence of the starting point. Normalization is achieved by cyclically shifting to minimize the resulting natural number, thus achieving rotation invariance. After extracting the reflective region boundary of the associated area, to avoid unequal description lengths caused by differences in the number of boundary points, the boundary needs to be resampled to determine the appropriate number of sampling points. Specifically, the curvature variance of each local window of the boundary is calculated to locate the most complex region of the edge. Points are taken in this region at different densities, and the rate of change of curvature variance is monitored. When the curvature information gain brought by increasing the number of sampling points slows down, the optimal sampling density is determined. Based on this density, n sampling points are obtained by mapping to the entire boundary, ultimately generating a normalized differential chain code sequence of equal length.

[0102] The minimum ring distance is a ring-shaped metric for measuring the difference between two directional codes. Considering that the chain code direction is circular and contains eight discrete directions from 0 to 7, the minimum ring distance is defined as follows for point a on chain code A and point b on chain code B:

[0103]

[0104] in, This is the absolute difference between the two direction codes. For ring-based compensation, this definition ensures that the difference between direction 0 and direction 7 is correctly measured as one unit rather than seven units, conforming to the topology of the direction ring. Based on this distance definition, the dissimilarity between the two normalized differential chain codes is calculated:

[0105]

[0106] in, The number of sampling points. and These represent the two chain codes at the th... The formula quantifies the overall deviation of two boundaries in their local orientation by summing the minimum loop distances at each corresponding point and taking the average. A larger value indicates a more significant difference in the geometric shape of the boundaries between the two chain codes.

[0107] Since normalized differential chain codes only guarantee cyclic consistency in relative directions, the two chain codes may have an overall offset in absolute directions. Therefore, cyclic shift optimization is needed to eliminate the difference in directional reference. Specifically, one chain code sequence is fixed, while the other chain code sequence is cyclically shifted stepwise along any direction. After each shift, the dissimilarity is recalculated. This continues until the shifted chain code returns to its initial state, thus obtaining a set containing dissimilarity under each directional reference. The minimum value in this set is taken as the final dissimilarity. ,Right now This minimization operation eliminates systematic biases caused by the selection of the starting point for the boundary description or global rotation, ensuring... The value purely reflects the degree of variation in the intrinsic geometry of the boundary, rather than external directional differences.

[0108] In the above scheme, the surface of the metal rod is smooth and homogeneous, with consistent optical reflection characteristics throughout. During the relative pose changes of the light source, camera, and target caused by the belt movement, the boundary morphology of its specular reflection area only undergoes rigid translation and rotation, while the local curvature characteristics remain stable. The chain code sequences of the preceding and following frames exhibit high similarity, and the calculated dissimilarity... The value is relatively small; however, the microscopic pits and depressions formed by the base fiber structure of the treated wooden tool handles are randomly and irregularly distributed, and the wax coating only covers local areas. When the viewing angle changes, the scattering caused by the pits and the specular reflection of the coating interact, causing the boundary of the reflective area to exhibit non-rigid local deformation and edge jitter. The chain code sequences of the preceding and following frames show significant differences in local orientation, and the calculated dissimilarity is relatively small. The value is relatively large. By quantifying this temporal variation characteristic, it is possible to effectively distinguish between metallic and non-metallic rod-shaped foreign objects that have similar surface optical properties but are fundamentally different in microscopic physical structure. Moreover, this discrimination dimension based on temporal consistency has strong immunity to transient noise such as slag adhesion or local occlusion in single-frame images, thus improving the robustness and reliability of recognition.

[0109] The second implementation method: motion compensation based on optical flow tracing;

[0110] In this embodiment, the optical flow field of the rod-shaped foreign object region in the preceding and following frames can be calculated to obtain the temporal displacement vector of each pixel. Based on this, motion compensation and geometric transformation are performed on the boundary of the reflective region in the preceding frame to precisely align it with its position in the following frame. Under the registered unified coordinate system, the overlap rate or symmetrical difference area between the boundaries of the reflective regions in the two frames is directly calculated. The lower the overlap rate or the larger the difference area, the more significant the change in boundary morphology, and the higher the corresponding dissimilarity. This embodiment can effectively compensate for the translation and rotation caused by the belt movement, focusing on the variation in reflective morphology purely caused by the microstructure of the material surface.

[0111] The third implementation method: statistical comparison based on shape description moments;

[0112] In this embodiment, statistical characteristics such as the central moment, moment of inertia, and Hu moment invariant of the reflective area boundary in the preceding and following frames can be calculated to construct a feature vector characterizing the overall geometric properties of the boundary. The Euclidean distance or Mahalanobis distance between the two feature vectors is then used as a measure of morphological difference. For metallic materials, the moment feature vectors of the preceding and following frames are highly similar, resulting in a small distance value; for wooden materials, the moment features change significantly due to boundary jitter, resulting in a larger distance value.

[0113] The fourth implementation method: geometric matching based on distance metrics;

[0114] In this embodiment, after establishing the correspondence between the point sets of reflective region boundaries in consecutive frames, the Hausdorff distance metric can be used to measure the maximum mismatch between the two boundary sets, or the Frechet distance metric can be used to consider the curve similarity constrained by point order. The larger the distance value, the greater the difference in boundary morphology, and the higher the corresponding dissimilarity. This method strictly considers the geometric position and topological order of the boundaries, and can accurately capture the local deformation and overall distortion of the boundaries, providing a precise geometric quantitative basis for distinguishing between temporally stable and temporally variable reflective morphologies.

[0115] Step S140: Determine the rod-shaped foreign object discrimination index based on the alignment deviation index and dissimilarity, and distinguish the material type of the rod-shaped foreign object according to the rod-shaped foreign object discrimination index.

[0116] Optionally, step S140 may include: performing a nonlinear transformation on the alignment deviation index to obtain a geometric deviation correction factor; performing a fusion operation on the dissimilarity and the geometric deviation correction factor to obtain a rod-shaped foreign object discrimination index; and determining the material type of the rod-shaped foreign object based on the numerical relationship between the rod-shaped foreign object discrimination index and a preset threshold.

[0117] The aforementioned stick-shaped foreign object discrimination index is a composite quantitative indicator that integrates the degree of geometric alignment deviation between the reflective area and the foreign object contour in a single frame image and the variation characteristics of reflective morphology over multiple frames. Its purpose is to integrate spatial and spatiotemporal discrimination information to form a robust basis for determining the material type of the stick-shaped foreign object. By organically fusing the alignment deviation index and dissimilarity, this index reflects both the instantaneous geometric deviation characteristics caused by the microstructure of the material surface and the temporal stability characteristics of this structure under dynamic viewing angle changes. The two complement each other, enhancing the reliability of the discrimination and effectively overcoming the uncertainty of single-dimensional features under complex working conditions.

[0118] The geometric deviation correction factor is obtained by performing a nonlinear transformation on the alignment deviation index, aiming to establish a numerical scale and weight relationship adapted to the morphological temporal difference benchmark value. Since the dimensions and numerical range of the alignment deviation index may differ from the dissimilarity, direct linear superposition can easily lead to a single feature dominating the discrimination result. Therefore, a logarithmic mapping function is used for nonlinear compression and bias adjustment. Specifically, the alignment deviation index is incremented by one, and then a common logarithmic operation is performed to generate a logarithmic term. This term maps larger geometric deviation indices to a relatively gentle growth curve while preserving the discriminative sensitivity in smaller numerical ranges. This logarithmic term is then arithmetically summed with a constant of one to form the geometric deviation correction factor. This factor is greater than one and monotonically increases with the alignment deviation index, ensuring that the more significant the geometric deviation, the stronger the enhancement effect on the final discrimination index. However, its increase is suppressed by the logarithmic function to avoid excessive amplification of single-frame noise.

[0119] The morphological temporal difference baseline value is directly borne by the dissimilarity. Given that the dissimilarity of the reflective region boundaries between adjacent frames is relatively less susceptible to external random interference such as illumination flicker or momentary occlusion, exhibiting good temporal stability, it is established as the basic discriminant in the fusion operation. The fusion operation implements a multiplicative integration strategy, multiplying the morphological temporal difference baseline value by the geometric deviation correction factor to generate the rod-shaped foreign object discrimination index, whose mathematical expression is:

[0120]

[0121] in, Indicates dissimilarity, quantifying the degree of morphological difference between the boundaries of reflective areas in consecutive frames. ; The alignment deviation index represents the degree of geometric deviation in a single frame. The above calculation formula reflects the idea of ​​using temporal stability as the main criterion and geometric alignment as an auxiliary correction weight. When the dissimilarity value is large, indicating significant temporal variation, even if the alignment deviation index is moderate, the stick-shaped foreign object discrimination index will still increase significantly. Conversely, when the dissimilarity value is small, indicating temporal stability, a larger alignment deviation index can moderately increase the stick-shaped foreign object discrimination index, but it is constrained by its base value.

[0122] Material differentiation based on the discrimination index of rod-shaped foreign objects is achieved through a threshold comparison mechanism. The preset threshold is determined statistically based on the distribution of the discrimination index of known metal rod-shaped objects and wooden tool handles in historical monitoring data. Typically, the boundary value between the two types of sample clusters is selected as the discrimination limit. When the calculated discrimination index of the rod-shaped foreign object is greater than the preset threshold, it indicates that the rod-shaped foreign object exhibits significant temporal morphological variation and geometric alignment deviation, corresponding to a processed non-metallic material with irregular surface microstructure and unstable reflective areas. When the discrimination index of the rod-shaped foreign object is less than the preset threshold, it indicates that the object has a temporally stable regular reflective morphology and good geometric alignment, corresponding to a smooth and homogeneous metallic material. Through this numerical comparison, the final determination of the material type can be completed based on traditional image processing features without the need for deep neural network inference.

[0123] Step S150: Process the foreign object according to the identified material type.

[0124] Optionally, the aforementioned rod-shaped foreign objects include surface-treated wooden tool handles and metal rod-shaped objects; the aforementioned step S150 may include: processing the foreign objects according to the distinguished material type, including: for rod-shaped foreign objects determined to be of metal material, controlling the magnetic suction device to perform a rejection operation; for rod-shaped foreign objects determined to be of non-metal material, controlling the vacuum suction device or mechanical actuator to perform a rejection operation; and for foreign objects exceeding the system's preset processing capacity, triggering an alarm signal and controlling the conveyor belt to slow down.

[0125] Surface-treated wooden tool handles and metal rods are two representative types of rod-shaped foreign objects encountered during mine conveyor belt transport. The former typically refers to the wooden handles of shovels, picks, and other working tools with surfaces treated for corrosion resistance or slip resistance. Although their surfaces are coated with wax or resin to create localized specular reflections, the underlying material remains a fibrous, loose structure, and is non-conductive and non-magnetic. The latter mainly refers to discarded steel pipes, iron bars, or damaged metal tools, which have a homogeneous metallic crystalline structure and exhibit good conductivity and ferromagnetism. Both types of foreign objects are geometrically slender rods with sharp ends or edges, posing a risk of scratching or tearing the conveyor belt. However, their physical properties are distinctly different. Therefore, they can be accurately distinguished using a rod-shaped foreign object discrimination index, allowing for targeted physical separation measures and preventing rejection strategies from failing due to material misjudgment.

[0126] For rod-shaped foreign objects identified as metallic, a magnetic suction device is used to generate a strong magnetic field or electromagnetic attraction. Utilizing the physical effect of ferromagnetic materials being magnetized and attracted in a magnetic field, the metallic rod-shaped object is adsorbed and extracted from the coal flow and transferred to a collection container. For rod-shaped foreign objects identified as non-metallic, a vacuum adsorption device is used to generate negative pressure airflow or drive a mechanical actuator to perform a clamping action. The former establishes a pressure difference below atmospheric pressure at the suction cup through a vacuum generator, relying on the pressure gradient force to adsorb relatively flat wooden handles. The latter uses a servo motor to drive the end effector of the robotic arm to perform mechanical gripping of irregularly shaped foreign objects. Both non-metallic removal methods avoid the ineffective response of magnetic suction to non-ferromagnetic materials, ensuring the effectiveness and reliability of the removal action.

[0127] When a foreign object is identified that exceeds a preset processing capacity threshold in terms of size, weight, or material combination, an abnormal handling procedure is triggered to ensure equipment safety and production continuity. The abnormal handling procedure, for example, first outputs an audible and visual alarm signal to the monitoring terminal, prompting operators to intervene. Simultaneously, a deceleration command is sent to the belt drive control system to reduce the impact energy of the foreign object on downstream equipment by decreasing the conveyor belt speed, thus buying time for manual intervention or the activation of backup processing equipment. This layered handling mechanism constitutes a complete closed loop for foreign object management. It achieves automated and immediate removal of conventionally handleable foreign objects, and through safety interlock control in abnormal states, prevents mechanical damage or belt tearing accidents caused by forcibly handling foreign objects exceeding the capacity range, improving the adaptability and safety redundancy of the entire identification and processing process.

[0128] To facilitate understanding of the working principle of the machine vision-based intelligent foreign object recognition method for mine conveyor belts, this application also provides an application example of this method in a specific scenario. In this application scenario, the machine vision-based intelligent foreign object recognition method for mine conveyor belts mainly includes:

[0129] Step 1: Acquire images from the belt and perform preprocessing;

[0130] In this application scenario, the image acquisition device uses an industrial-grade high-resolution camera, mounted above the conveyor belt. Its optical axis forms an angle of approximately 60° with the belt plane and faces the direction of the coal. During the conveyor's operation, it captures high-definition images of the current coal on the conveyor belt at fixed time intervals or based on encoder trigger signals. Because the belt conveyor operates continuously, the acquired image sequences overlap in time and space. To prevent fast-moving foreign objects from being truncated at the boundary between adjacent frames, resulting in the same object appearing incomplete in different frames, the system sets a buffer area at the boundary between the processed image area and the current image area to be processed, based on the shooting time interval and the belt speed. This buffer area is incorporated into the current image area to ensure that moving foreign objects are completely captured.

[0131] After acquiring the image, the belt area is first extracted from the image based on the camera calibration parameters, and the background area outside the belt area is removed to eliminate irrelevant interference. Then, semantic segmentation is performed on the belt area to identify and extract the coal area. Considering that coal accumulation usually presents relatively uniform color and texture features, which contrasts with foreign objects of various shapes, this segmentation process marks the part of the belt surface covered by coal. Finally, the remaining area outside the coal area is removed from the belt area to obtain the foreign object area. Since foreign objects only account for a small part of the total material, it is generally assumed that each foreign object area contains only a single foreign object entity, and thus each foreign object can be processed accordingly.

[0132] Step 2: Identify and classify foreign objects, and conduct specific analysis on foreign objects that are easily misidentified;

[0133] First, a preliminary material assessment based on optical reflection characteristics is performed. Specifically, the image data of the foreign object area is converted to a color space that includes a brightness channel, with a standard white point as a reference. The brightness values ​​of each pixel in the foreign object area are extracted, and a brightness distribution histogram and a reflection distribution map of the foreign object area are constructed based on the distribution of brightness values. Based on the statistical characteristics (such as mean and variance) of the brightness distribution histogram and the reflection distribution map, a lightweight classification model is used to infer the material type of the foreign object, which is initially classified as a high-reflectivity metal material, a medium-reflectivity mineral material, or a low-reflectivity wood material.

[0134] For rod-shaped foreign objects that are easily misidentified in the initial identification, further specific analysis is required to accurately distinguish the materials. First, the circumscribed geometry of the foreign object region is obtained (such as the minimum circumscribed rectangle determined by the convex hull algorithm and the rotational caliper method). The longitudinal and transverse dimensions of this circumscribed geometry are determined, and the ratio between the longitudinal and transverse dimensions, as well as the filling ratio between the area of ​​the foreign object region and the area of ​​the circumscribed geometry, are calculated. Based on the matching of this ratio and the filling ratio with a preset threshold, the rod-shaped foreign object is identified from the foreign object region. This identification is based on the characteristics of the rod-shaped foreign object, which presents a slender geometric shape and a high solid filling rate.

[0135] For identified rod-shaped foreign objects (such as wooden tool handles with anti-corrosion or slip-enhancing surface treatments and metal rod-shaped objects), the following refined discrimination process is further performed:

[0136] First, reflective areas are extracted for geometric and spatiotemporal analysis. Optical reflection characteristics (such as brightness values ​​in the Lab color space) are obtained at various locations within the rod-shaped foreign object region. Based on the distribution of optical reflection characteristics, the foreign object region is binary-classified using the inter-class variance maximization algorithm to determine the strong reflection regions inside the foreign object. The boundaries of these strong reflection regions are then extracted as reflective areas.

[0137] Next, the alignment deviation index between the reflective area and the foreign object area is calculated. The reflective area and the foreign object area are geometrically registered to ensure they are on comparable geometric scales: specifically, the foreign object area is fitted as a reference rectangle, and the reflective area is also fitted as a rectangle. Using the length and width directions of the fitted rectangle parallel to the foreign object area as constraints, the long side of the fitted rectangle for the reflective area is translated outwards until it coincides with the long side of the fitted rectangle for the foreign object area. The translation distances of the two long sides in the width direction are recorded. and The model is then stretched along the edges to align the corner points, completing the registration. Based on this registration process, a morphological scale difference index, namely the width expansion ratio, is determined. .

[0138] Determine the deviation metric at the registered geometric scale: traverse the fitted rectangle over the foreign object region. At each sampling point, the deviation distance between the boundary of the reflective area and the boundary of the foreign object area is measured along a direction perpendicular to the side of the rectangle. Calculate the normalized boundary inconsistency index This index characterizes the deviation between the boundary of the reflective area and the boundary of the foreign object area after registration.

[0139] The alignment deviation index is determined based on the morphological scale difference index and the deviation measure. :

[0140]

[0141] Simultaneously, the dissimilarity of the reflective area boundaries of the stick-shaped foreign object in the preceding and following frames is calculated. Based on the belt running parameters and imaging timing, a temporal correspondence between the stick-shaped foreign object and adjacent frames is established: the approximate location region of the foreign object in the following frame is predicted based on the belt speed, shooting time interval, and the position coordinates of the foreign object in the previous frame. Within this location region, the associated region of the stick-shaped foreign object in the adjacent frame is determined by feature point matching. The reflective area boundary of this associated region in the preceding and following frames is obtained, and its geometric morphological features are extracted: the boundary is encoded using normalized differential chain code, and the number of sampling points of appropriate density is determined by resampling to generate a chain code sequence of equal length. The geometric morphological features in the preceding and following frames are compared and analyzed. The minimum loop distance between point a on chain code A and point b on chain code B can be defined as: Calculate the dissimilarity between two chain codes. Since the chain code direction is circular, one chain code is fixed, while the other chain code is gradually shifted cyclically along any direction, and the dissimilarity is recalculated each time. This yields the set of dissimilarity. The minimum value in this set is taken as the final dissimilarity. This dissimilarity is used to quantify the temporal variation characteristics of the morphology of reflective areas; the larger the value, the more severe the temporal variation.

[0142] Based on alignment deviation index and dissimilarity Determine the discrimination index O for rod-shaped foreign objects. Perform a nonlinear transformation on the alignment deviation index to obtain the geometric deviation correction factor. ; Dissimilarity The geometric deviation correction factor is fused together to obtain the rod-shaped foreign object discrimination index:

[0143]

[0144] The material type of the rod-shaped foreign object is determined based on the numerical relationship between the discrimination index O and the preset threshold: when the O value is greater than the preset threshold, it indicates that the rod-shaped foreign object exhibits significant temporal morphological variation and geometric alignment deviation, corresponding to a non-metallic material with a treated surface (such as a wooden tool handle); when the O value is less than the preset threshold, it corresponds to a smooth and homogeneous metallic material.

[0145] Step 3: Process the foreign object based on the foreign object identification results;

[0146] After identifying the material type of the foreign object, it is processed accordingly. For rod-shaped foreign objects identified as metallic, a magnetic suction device is used to remove them, utilizing the physical effect of ferromagnetic materials being attracted in a magnetic field to extract the metallic foreign object from the coal stream. For rod-shaped foreign objects identified as non-metallic, a vacuum adsorption device or mechanical actuator is used to remove them, removing the wooden foreign object through negative pressure airflow adsorption or mechanical clamping. For foreign objects exceeding the system's preset processing capacity (such as oversized or overweight foreign objects), an abnormal handling procedure is triggered, an alarm signal is output to remind the operator to intervene, and the conveyor belt is slowed down to reduce the impact energy of the foreign object on downstream equipment, awaiting manual handling.

[0147] like Figure 3 As shown, based on the same inventive concept, this application also provides a machine vision-based intelligent foreign object recognition system 200 for mine conveyor belts, including a host computer 210 and an image acquisition device 220 communicatively connected to the host computer, wherein:

[0148] Image acquisition device 220 is used to acquire images of coal transported on the mine conveyor belt and send the images to host computer 210;

[0149] The host computer 210 is used to acquire images, identify and segment foreign object regions; determine the material type of foreign objects based on the brightness distribution parameters of the foreign object regions; for easily misidentified rod-shaped foreign objects, calculate the alignment deviation index between the reflective area and the boundary of the foreign object region, and calculate the dissimilarity of the reflective area boundary of the rod-shaped foreign object in the previous and next frame images; determine the rod-shaped foreign object discrimination index based on the alignment deviation index and dissimilarity, and distinguish the material type of the rod-shaped foreign object based on the rod-shaped foreign object discrimination index; and process the foreign object accordingly based on the distinguished material type.

[0150] The aforementioned intelligent foreign object identification system 200 for mine conveyor belts based on machine vision achieves a complete functional closed loop from image acquisition to foreign object removal through a hardware and software collaborative architecture. The core of the system consists of a host computer 210 and an image acquisition device 220 that establishes a communication connection with it. The host computer 210 undertakes the functions of algorithm processing and decision control center, while the image acquisition device 220 is responsible for front-end visual information acquisition. The two achieve real-time interaction of image data and control commands through wired or wireless communication links, jointly building an intelligent monitoring system adapted to the complex environment of mines.

[0151] The image acquisition device 220 employs industrial-grade high-resolution imaging equipment, typically deployed at a predetermined height above the conveyor belt. Its optical axis forms an angle of approximately 60° with the belt plane and faces the direction of coal flow, ensuring that the imaging field of view covers the effective width of the belt and reduces motion blur. During continuous operation of the conveyor, the device performs periodic exposures at fixed time intervals or based on an encoder-triggered synchronization mechanism to capture high-definition image sequences of the coal flow on the belt. The acquired raw image data is transmitted in real time to the host computer 210 via a communication interface, providing a data source for subsequent algorithm processing.

[0152] After receiving the image, the host computer 210 first performs preprocessing and foreign object region segmentation. It extracts the conveyor belt region using camera calibration parameters and removes background interference. Then, it uses a semantic segmentation algorithm to identify coal accumulation areas, identifying unmarked areas within the conveyor belt region as candidate foreign object regions. Simultaneously, it sets buffer zones at the boundaries of the image sequence based on the conveyor belt speed to prevent moving foreign objects from interrupting the image. After completing region segmentation, the host computer 210 converts the foreign object region image data to a color space that includes a luminance channel, extracts the luminance values ​​of each pixel to construct a luminance distribution histogram and a reflectance distribution map, and infers the material type of the foreign object based on statistical characteristics, achieving a preliminary judgment based on optical reflection characteristics.

[0153] For rod-shaped foreign objects that are easily misidentified in the initial assessment, the host computer 210 initiates a refined discrimination process. First, it identifies the rod-shaped target by analyzing the aspect ratio and fill ratio of the external constraint geometry of the foreign object region. Then, it extracts the boundary of the internal strongly reflective area as the reflective region. In the geometric analysis dimension, the host computer 210 registers the reflective region with the foreign object region, records the degree of scale expansion and boundary deviation in the width direction, and calculates the alignment deviation index to characterize the degree of geometric deviation in a single frame. In the temporal analysis dimension, based on the belt running parameters, it establishes a temporal correspondence between rod-shaped foreign objects in adjacent frames. It calculates the dissimilarity by comparing the normalized differential chain code differences of the reflective region boundaries between consecutive frames to characterize temporal variation characteristics. Finally, the host computer 210 performs a nonlinear transformation on the alignment deviation index to obtain a geometric deviation correction factor, and fuses it with the dissimilarity to generate a rod-shaped foreign object discrimination index. The material type is determined by comparing the value with a preset threshold.

[0154] Based on the judgment results, the host computer 210 generates corresponding control commands and sends them to the execution terminal. For metallic foreign objects, it triggers a magnetic attraction device to generate a magnetic field for adsorption. For non-metallic foreign objects, it drives a vacuum adsorption device or mechanical actuator to perform a gripping or air suction operation, achieving physical removal that matches the material characteristics. When an abnormal foreign object exceeding the processing capacity is detected, the host computer 210 triggers an alarm signal output and issues a belt deceleration command. By adjusting the conveyor belt's operating status, the risk is reduced, and manual intervention is awaited. This forms a complete intelligent monitoring closed loop from image acquisition and algorithm recognition to physical execution.

[0155] It is understood that the host computer 210 described above can realize any one of the functions of the intelligent foreign object identification method for mine conveyor belts based on machine vision provided in the embodiments of this application. The method embodiment section will not repeat the details in the system embodiment section.

[0156] This invention is now complete.

[0157] In summary, in this embodiment of the invention, images of coal transported by conveyor belt in a mine are acquired, and foreign object regions are identified and segmented. The material type of the foreign object is determined based on the brightness distribution parameters of the foreign object region. For easily misidentified rod-shaped foreign objects, the alignment deviation index between the reflective area and the boundary of the foreign object region is calculated, and the dissimilarity of the reflective area boundaries of the rod-shaped foreign object in the preceding and following frames is calculated. A rod-shaped foreign object discrimination index is determined based on the alignment deviation index and the dissimilarity, and the material type of the rod-shaped foreign object is distinguished based on the rod-shaped foreign object discrimination index. The foreign object is then processed accordingly based on the distinguished material type.

[0158] This invention establishes a lightweight identification system that does not rely on deep neural networks by constructing a three-level progressive discrimination mechanism: initial material judgment based on brightness distribution parameters, detailed geometric judgment based on alignment deviation index, and temporal correction based on dissimilarity. This reduces reliance on large-scale labeled data and hardware computing power requirements while achieving accurate material classification of foreign objects on mine conveyor belts. Furthermore, it uses traditional image processing methods to extract brightness distribution features of the foreign object region to infer material type, replacing deep learning object detection models that require massive training samples and high-performance computing resources. This significantly reduces deployment costs and improves applicability in mine environments. Finally, it addresses the issue of surfaces treated with anti-corrosion or enhanced... For easily misidentified foreign objects such as wooden tool handles with smooth finish and metal rods, which have similar reflective features but completely different physical properties, this study quantifies the degree of geometric deviation between the reflective area and the foreign object outline by calculating the alignment deviation index between the reflective area and the foreign object area. This effectively distinguishes rod-shaped foreign objects of different materials and avoids the failure of the rejection strategy due to material misjudgment. On the other hand, by combining the temporal variation characteristics of the reflective area boundary with the analysis of the previous and next frame images, the dissimilarity is calculated to characterize the temporal stability of the reflective morphology. The physical difference between the stable reflective morphology of metal materials and the volatile reflective morphology of processed wooden surfaces is utilized to further verify the material recognition results and improve the recognition robustness under noise interference such as cinder adhesion.

[0159] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine vision-based intelligent foreign object identification method for mine conveyor belts, characterized in that, The method includes: Acquire images of coal transported by conveyor belt in a mine, and identify and segment the foreign object area; The material type of the foreign object is determined based on the brightness distribution parameters of the foreign object region; For easily misidentified stick-shaped foreign objects, calculate the alignment deviation index between the boundary of the reflective area and the boundary of the foreign object area, and calculate the dissimilarity of the boundary of the reflective area of ​​the stick-shaped foreign object in the images of the previous and next frames. The alignment deviation index and the dissimilarity are used to determine the rod-shaped foreign object discrimination index, and the material type of the rod-shaped foreign object is distinguished according to the rod-shaped foreign object discrimination index. The foreign object is processed accordingly based on the identified material type.

2. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to claim 1, characterized in that, The method further includes: Obtain the external constraint geometry of the foreign object region, and determine the longitudinal and transverse dimensions of the external constraint geometry; Calculate the ratio between the longitudinal dimension and the lateral dimension, and the filling ratio between the area of ​​the foreign object region and the area of ​​the external constraint geometry; Based on the ratio and the matching of the filling ratio with the preset threshold, the rod-shaped foreign object is identified from the foreign object region.

3. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to claim 2, characterized in that, The method further includes: Obtain the optical reflection characteristics at each location within the region of the rod-shaped foreign object; Based on the distribution of the optical reflection characteristics, the strong reflection region inside the rod-shaped foreign object is determined; The boundary of the strongly reflective area is extracted as the reflective area.

4. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to claim 1, characterized in that, The calculation of the alignment deviation index between the reflective area and the foreign object area includes: Geometric registration is performed between the reflective area and the foreign object area to determine the morphological scale difference index between the reflective area and the foreign object area; Determine the deviation measure between the boundary of the reflective area and the boundary of the foreign object area after registration; Based on the morphological scale difference index and the deviation metric, the alignment deviation index between the reflective area and the foreign object area is determined; wherein, the alignment deviation index is used to characterize the degree of geometric deviation between the reflective area and the foreign object area.

5. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to claim 1, characterized in that, The calculation of the dissimilarity of the reflective region boundaries of the stick-shaped foreign object in the preceding and following frames includes: Based on the belt running parameters and imaging timing, the temporal correspondence of the stick-shaped foreign object is established in adjacent frame images to determine the associated region of the stick-shaped foreign object in the preceding and following frames; Obtain the boundary of the reflective area of ​​the associated region, and extract the geometric features of the reflective area; By comparing and analyzing the geometric features in the preceding and following frames, the degree of morphological difference of the boundary of the reflective area is determined. The degree of dissimilarity is determined based on the degree of morphological difference; wherein, the degree of dissimilarity is used to quantify the temporal variation characteristics of the morphology of the reflective region.

6. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to claim 1, characterized in that, The step of determining the rod-shaped foreign object discrimination index based on the alignment deviation index and the dissimilarity, and distinguishing the material type of the rod-shaped foreign object according to the rod-shaped foreign object discrimination index, includes: The alignment deviation index is subjected to a nonlinear transformation to obtain the geometric deviation correction factor; The dissimilarity is fused with the geometric deviation correction factor to obtain the stick-shaped foreign object discrimination index; The material type of the rod-shaped foreign object is determined based on the numerical relationship between the discrimination index and the preset threshold.

7. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to any one of claims 1-6, characterized in that, The identification and segmentation of the foreign object region includes: Extract the belt region from the image and remove the background region outside the belt region; The belt conveyor area is semantically segmented to identify and extract the coal material area; The area remaining outside the coal material area in the belt conveyor area is removed to obtain the foreign object area.

8. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to any one of claims 1-6, characterized in that, Determining the material type of the foreign object based on the brightness distribution parameters of the foreign object region includes: The image data of the foreign object area is converted to a color space that includes a brightness channel, and the brightness value of each pixel in the foreign object area is extracted. Based on the distribution of the brightness values, construct a brightness distribution histogram and a reflection distribution map of the foreign object area; Based on the statistical characteristics of the brightness distribution histogram and the reflectivity distribution map, the material type of the foreign object is determined.

9. The intelligent foreign object identification method for mine conveyor belts based on machine vision according to any one of claims 1-6, characterized in that, The rod-shaped foreign object includes a surface-treated wooden tool handle and a metal rod-shaped object; The step of processing the foreign object according to its material type includes: For rod-shaped foreign objects identified as being made of metal, the magnetic suction device is controlled to perform a rejection operation; For rod-shaped foreign objects identified as being made of non-metallic materials, control the vacuum adsorption device or mechanical actuator to perform a removal operation; Additionally, for foreign objects exceeding the system's preset processing capacity, an alarm signal is triggered and the conveyor belt is controlled to slow down.

10. A machine vision-based intelligent foreign object identification system for mine conveyor belts, characterized in that, It includes a host computer and an image acquisition device that is communicatively connected to the host computer, wherein: The image acquisition device is used to acquire images of coal transported on the mine conveyor belt and send the images to the host computer. The host computer is used to acquire the image, identify and segment the foreign object region; determine the material type of the foreign object based on the brightness distribution parameters of the foreign object region; for easily misidentified rod-shaped foreign objects, calculate the alignment deviation index between the reflective area and the boundary of the foreign object region, and calculate the dissimilarity of the reflective area boundary of the rod-shaped foreign object in the preceding and following frames; determine the rod-shaped foreign object discrimination index based on the alignment deviation index and the dissimilarity, and distinguish the material type of the rod-shaped foreign object based on the rod-shaped foreign object discrimination index; and perform corresponding processing on the foreign object according to the distinguished material type.

Images