Multi-feature fusion-based training method of mine selection identification model

Abstract

The invention relates to the technical field of mineral image recognition, in particular to a training method of a mineral selection recognition model based on multi-feature fusion. The method comprises the following steps: acquiring a mineral image; mineral crystal structure information of the mineral image is identified, geometric type division is carried out on the mineral crystal structure information, and mineral crystal system types are obtained; performing mineral crystal symbiosis identification on the mineral image according to the type of the mineral crystal system, evaluating a mineral crystal symbiosis relationship, and generating mineral crystal symbiosis data; mineral crystal grain connectivity analysis is conducted on the mineral crystal symbiosis data, and the mineral porosity is detected. Performing point-line-plane defect feature detection on the mineral image according to the mineral crystal system type to generate mineral point-line-plane defect features; according to the method, the ore selection recognition model is constructed through the multi-modal feature fusion technology, so that comprehensive recognition and analysis of the crystal structure, the symbiotic relationship, the morphological features and the surface abrasion of the minerals are achieved, and therefore the accuracy and efficiency of ore selection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mineral image recognition, and in particular to a training method for a mineral separation recognition model based on multi-feature fusion. Background Art

[0002] From the perspective of the crystal structure characteristics of minerals, the diversity of the crystal structures of minerals has not been fully utilized. Crystal defects (such as point defects, line defects, plane defects, etc.) will affect the physical properties and appearance characteristics of minerals, but these defects are not fully considered in the existing models during the recognition process, which easily leads to misrecognition or missed recognition. In terms of microscopic structure characteristics, the microscopic structures of minerals (such as crystal morphology, grain size, mineral symbiotic relationship, etc.) are difficult to fully reflect in macroscopic images, and the existing models have insufficient extraction and utilization of microscopic features, resulting in low accuracy when identifying minerals with similar appearances but different microscopic structures. From the perspective of physical property characteristics, the physical properties of minerals (such as hardness, density, luster, etc.) are important recognition bases, but most of the existing models rely on image data and consider less the physical property characteristics of ores. During the mineral separation process, the surface characteristics of ores will change due to operations such as mining, crushing, and transportation (such as surface wear, color change, etc.), and the existing models have insufficient adaptability to these dynamic change characteristics and are difficult to maintain stable recognition performance at different stages. Summary of the Invention

[0003] Based on this, it is necessary to provide a training method for a mineral separation recognition model based on multi-feature fusion to solve at least one of the above technical problems.

[0004] To achieve the above object, a training method for a mineral separation recognition model based on multi-feature fusion, the method includes the following steps: Step S1: Collect mineral images; identify the mineral crystal structure information of the mineral images, perform geometric type division on the mineral crystal structure information to obtain the mineral crystal system type; perform point-line-plane defect feature detection on the mineral images according to the mineral crystal system type to generate mineral point-line-plane defect features; Step S2: Perform mineral crystal symbiosis recognition on the mineral images according to the mineral crystal system type, and evaluate the mineral crystal symbiosis relationship to generate mineral crystal symbiosis data; perform mineral grain connectivity analysis on the mineral crystal symbiosis data, and detect the mineral porosity; determine the mineral grain shape characteristics based on the mineral porosity; Step S3: Identify the mineral morphology characteristics of the mineral images, determine the mineral luster area of the mineral morphology characteristics; perform luster density detection on the mineral luster area to obtain regional luster density data; perform color change gradient division on the mineral morphology characteristics according to the regional luster density data, and determine the mineral surface wear characteristics based on the color change gradient; Step S4: Perform mineral multi-modal feature fusion on the mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features to generate mineral multi-modal fusion features; construct a preset ore separation recognition model through deep learning technology, train the preset ore separation recognition model according to the mineral multi-modal fusion features, and verify the model performance indicators; optimize the model parameters according to the model performance indicators to obtain the ore separation recognition model.

[0005] Through the synergistic effect of step S1 and step S2, the present invention can accurately detect the point-line-plane defect characteristics of mineral images and accurately determine the mineral grain shape characteristics based on the mineral porosity. This accurate detection and determination method provides detailed and accurate microstructure information for subsequent ore selection and identification, thereby effectively distinguishing minerals of different quality grades during the ore selection process, improving the accuracy of ore selection, avoiding ore selection errors caused by internal structural defects or grain shape non-conformities of minerals, and further enhancing the overall quality of ore selection. Step S2 evaluates the symbiotic relationship of mineral crystals, detects the mineral porosity, and generates mineral crystal symbiotic data. This process can comprehensively understand the symbiotic situation of minerals during the natural formation process and the key physical property of porosity. In the ore selection scenario, the symbiotic relationship of minerals is of great significance for judging the formation environment of minerals and the value of associated minerals, while the porosity directly affects the physical properties of minerals and subsequent processing and utilization. Through the comprehensive evaluation and detection of these two aspects, more comprehensive basis can be provided for ore selection, which helps to formulate more reasonable ore selection strategies and improve the scientificity and effectiveness of ore selection. Step S3 can accurately reflect the wear condition of the mineral surface during mining, transportation or processing by detecting the gloss density of the mineral gloss area and determining the mineral surface wear characteristics based on the color change gradient. During the ore selection process, the degree of mineral surface wear affects its subsequent processing performance and application value. Accurately determining the mineral surface wear characteristics provides reliable information about the surface quality of minerals for ore selection, so that minerals with lower surface wear degree and better quality can be reasonably screened during ore selection, improving the subsequent processing utilization rate of minerals after ore selection and reducing the processing loss caused by surface wear. Step S4 performs multi-modal fusion on the mineral point-line-plane defect characteristics, mineral grain shape characteristics and mineral surface wear characteristics to generate mineral multi-modal fusion characteristics, and constructs an ore selection and identification model through deep learning technology, and optimizes the parameters according to the model performance indicators. This multi-modal feature fusion method can make full use of the information of different feature dimensions, enabling the ore selection and identification model to learn various characteristics of minerals more comprehensively. At the same time, the parameter optimization process based on the model performance indicators can ensure that the ore selection and identification model has a high recognition accuracy and stability in practical applications. In the actual operation of ore selection, this method can quickly and accurately identify minerals, improve the ore selection efficiency, reduce the errors and labor intensity of manual identification, and realize the automation and intelligence of the ore selection process. The present invention constructs an ore selection and identification model through multi-modal feature fusion technology to achieve comprehensive identification and analysis of the crystal structure, symbiotic relationship, morphological characteristics and surface wear of minerals, thereby improving the accuracy and efficiency of ore selection.

[0006] Preferably, in step S1, the mineral crystal structure information of the mineral image is identified, and the geometric type division of the mineral crystal structure information includes: The mineral image is scanned line by line, and the starting and ending point coordinates of the crystal structure in the horizontal and vertical directions in the image are recorded to form a mineral crystal structure coordinate data set; Based on the mineral crystal structure coordinate data set, the center point position of each crystal structure is calculated. Taking the center point as a reference, the crystal structure is divided into four quadrant regions, and the pixel point distribution in each quadrant is recorded; For the crystal structure in each quadrant region, its crystal shape factor is calculated, where the shape factor is determined by the ratio of the perimeter to the area of the crystal structure, and the shape factor value of each quadrant is generated; The crystal structures are classified into regular crystals and irregular crystals according to the shape factor values of each quadrant; For the structure of irregular crystals, the curvature change value of its edge points is calculated; the degree of concavity and convexity is determined according to the curvature change value; Based on the degree of concavity and convexity, the geometric type of the mineral crystal structure information is divided to obtain the mineral crystal system type.

[0007] In the present invention, the mineral image is scanned line by line to form a mineral crystal structure coordinate data set, which can accurately record the starting and ending point coordinates of the crystal structure in the horizontal and vertical directions in the image, providing an accurate positioning basis for subsequent analysis. Based on this coordinate data set, the center point position of each crystal structure is calculated, and four quadrant regions are divided based on this, and the pixel point distribution in each quadrant is recorded, which makes the analysis of the crystal structure more detailed and comprehensive, and can extract the characteristics of the crystal from different directions and regions. Further, the shape factor value of each quadrant is calculated, and based on this, the crystal structures are classified into regular crystals and irregular crystals. This classification method can effectively distinguish the morphological characteristics of the crystals, providing a clear basis for subsequent analysis. For irregular crystals, the curvature change value of its edge points is calculated and the degree of concavity and convexity is determined, and then based on the degree of concavity and convexity, the geometric type of the mineral crystal structure information is divided, and finally the mineral crystal system type is obtained. This series of steps can accurately start from the morphological characteristics of the crystal structure, accurately divide the mineral crystal system type, provide high-quality and high-precision mineral crystal structure information for the training of the ore selection recognition model, enable the model to more accurately identify the crystal characteristics of different minerals, thereby improving the accuracy and reliability of ore selection recognition, and effectively improving the ore selection efficiency and quality control level.

[0008] Preferably, in step S1, the detection of point, line and surface defect features of the mineral image according to the mineral crystal system type includes: The gradient of the mineral image is calculated according to the mineral crystal system type. The Sobel operator is used to calculate the gradients of the image in the horizontal and vertical directions respectively, and a gradient magnitude image and a gradient direction image are generated; Based on the gradient magnitude image, local extreme points of the gradient magnitude are identified, and the local extreme points are used as potential point defect candidate points; calculate the ratio of the gradient magnitude of each candidate point to the gradient magnitudes of its surrounding 8 neighborhood pixels, and retain the candidate points with a ratio greater than 1.2 as mineral crystal system defect points, and record the positions and gradient magnitudes of the mineral crystal system defect points; Starting from the upper left corner of the gradient direction image, track along the gradient direction until a point where the gradient direction changes by more than 45 degrees is encountered or the image boundary is reached; For each extracted line, calculate its length and curvature, where the curvature is determined by calculating the average value of the angular changes between every two adjacent points on the line; Mark the lines with a length greater than 10 pixels and a curvature greater than 0.1 as mineral crystal system defect point lines, and record the positions, lengths, and curvatures of the mineral crystal system defect point lines; Perform multi-region segmentation on the gradient magnitude image, with each region corresponding to a part of the mineral crystal; Calculate the standard deviation of the gradient magnitudes of each region, and mark the regions with a standard deviation greater than 0.5 as mineral crystal system defect surfaces, and record the positions and areas of the mineral crystal system defect surfaces.

[0009] In the process of mineral image analysis of the present invention, gradient calculations are performed according to the type of mineral crystal system to generate a gradient magnitude image and a gradient direction image, which can accurately capture the structural change characteristics of mineral crystals from the image perspective. The local extreme points identified based on the gradient magnitude image are used as potential point defect candidate points, and mineral crystal system defect points are screened out by calculating the ratio. This process can accurately identify point defects in mineral crystals and provide accurate information at the microscopic structure level for mineral separation. Extracting lines from the gradient direction image and calculating their lengths and curvatures, and then marking the mineral crystal system defect point lines, as well as performing multi-region segmentation on the gradient magnitude image and marking the mineral crystal system defect surfaces, these steps can comprehensively identify line defects and surface defects in mineral crystals. The accurate identification and recording of these defect characteristics provide rich and accurate mineral crystal defect data for the mineral separation recognition model, enabling the model to more comprehensively evaluate the quality and integrity of mineral crystals, thereby more accurately screening out minerals that meet the quality requirements in the mineral separation process, improving the accuracy and efficiency of mineral separation, ensuring the reliability of the mineral separation results, and further enhancing the overall quality control level of mineral separation.

[0010] Preferably, in step S2, mineral crystal symbiosis recognition is performed on the mineral image according to the type of mineral crystal system, and the evaluation of the mineral crystal symbiosis relationship includes: Perform edge detection on the mineral image, and extract the edge information of the mineral crystal to generate a mineral edge image; Identify the contours of the mineral edge image according to the type of mineral crystal system, where each contour corresponds to a mineral crystal, to obtain mineral edge contour information; Calculate the geometric center of the mineral edge contour information to determine the geometric center position of each mineral crystal; Calculate the Euclidean distance between the geometric centers of each pair of mineral crystals, and identify the pairs of mineral crystals with a distance less than 10% of the image width as symbiotic relationships; For each pair of mineral crystals identified as having a symbiotic relationship, calculate the overlapping area ratio of their contours, where the overlapping area ratio is the ratio of the area of the overlapping region of the two contours to the total area of the two contours; Mark the pairs of mineral crystals with an overlapping area ratio greater than 0.1 as closely symbiotic relationships; mark the pairs of mineral crystals with an overlapping area ratio between 0.05 and 0.1 as moderately symbiotic relationships; mark the pairs of mineral crystals with an overlapping area ratio less than 0.05 as loosely symbiotic relationships.

[0011] By performing edge detection on the mineral image and extracting the edge information of the mineral crystals to generate a mineral edge image, the present invention can clearly outline the outer contour of the mineral crystals. Identify the contours of the mineral edge image according to the type of mineral crystal system, with each contour corresponding to a mineral crystal, thereby obtaining the mineral edge contour information, providing accurate mineral crystal boundary data for subsequent analysis. Calculate the geometric center of the mineral edge contour information to determine the geometric center position of each mineral crystal, and then calculate the Euclidean distance between the geometric centers of each pair of mineral crystals, and identify the pairs of mineral crystals with a distance less than 10% of the image width as symbiotic relationships. This process can accurately judge the spatial correlation between mineral crystals. For each pair of mineral crystals identified as having a symbiotic relationship, calculate the overlapping area ratio of their contours, and mark the pairs of mineral crystals as closely symbiotic relationships, moderately symbiotic relationships, and loosely symbiotic relationships respectively according to the overlapping area ratio. This classification method can describe in detail the degree of symbiotic tightness between mineral crystals. The implementation of these steps can provide accurate data on the symbiotic relationship of mineral crystals for the ore separation recognition model, enabling the model to more comprehensively understand the symbiotic characteristics of mineral crystals in the natural formation process, thereby more accurately evaluating the formation environment of minerals and the value of associated minerals during the ore separation process, providing a scientific basis for formulating ore separation strategies, improving the scientificity and effectiveness of ore separation, ensuring the reliability of ore separation results, and further enhancing the overall quality control level of ore separation.

[0012] Preferably, in step S2, the mineral crystal symbiotic data is analyzed for mineral grain connectivity and the mineral porosity is detected, including: Analyze the mineral crystal symbiotic data to extract the geometric center positions of each pair of symbiotic mineral crystals; Based on the geometric center positions, construct a connectivity graph of mineral crystals, where each mineral crystal is represented as a node and the connectivity relationship between each pair of symbiotic mineral crystals is represented as an edge; Perform a depth-first search on the connectivity graph, starting from each unvisited node, and record the number of connected components and the size of each connected component; Segment the mineral image and mark the mineral background area; By counting the number of background pixels after segmentation and multiplying by the area of each pixel, the area of the mineral background area is obtained; the area of the mineral background area is determined as the mineral pore area; Take the ratio of the mineral pore area to the area of the entire mineral image to generate the mineral porosity.

[0013] In the present invention, by analyzing the mineral crystal symbiotic data and extracting the geometric center positions of each pair of symbiotic mineral crystals, precise positioning information can be provided for subsequent analysis. Based on the geometric center positions, a connectivity graph of mineral crystals is constructed, where each mineral crystal is represented as a node and the connectivity relationship between each pair of symbiotic mineral crystals is represented as an edge. Then, by performing a depth-first search on the connectivity graph, starting from each unvisited node, the number of connected components and the size of each connected component are recorded. This process can accurately analyze the spatial connectivity between mineral crystals. At the same time, the mineral image is segmented and the mineral background area is marked. The area of the mineral background area is obtained by counting the number of background pixels after segmentation and multiplying by the area of each pixel, and it is determined as the mineral pore area. Furthermore, by taking the ratio of the mineral pore area to the area of the entire mineral image to generate the mineral porosity, the pore characteristics of the mineral can be precisely quantified. The implementation of these steps provides quantitative data on the connectivity and porosity of mineral crystals for the ore selection recognition model, enabling the model to more comprehensively evaluate the physical structure characteristics of minerals, thus more accurately screening out minerals that meet the quality requirements during the ore selection process, improving the scientificity and effectiveness of ore selection, ensuring the reliability of the ore selection results, and further enhancing the overall quality control level of ore selection.

[0014] Preferably, determining the mineral grain shape characteristics based on the mineral porosity in step S2 includes: Preliminarily classify the mineral grains according to the mineral porosity to obtain the mineral grain shape characteristics; If the porosity is less than 0.1, mark the grain as a low-porosity grain; If the porosity is between 0.1 and 0.3, mark the grain as a medium-porosity grain; If the porosity is greater than 0.3, mark the grain as a high-porosity grain.

[0015] The present invention preliminarily classifies mineral grains according to the porosity of minerals, and can divide mineral grains into low-porosity grains, medium-porosity grains, and high-porosity grains according to the size of the porosity, so as to obtain the shape characteristics of mineral grains. This classification process is based on clear porosity thresholds, ensuring the objectivity and accuracy of the classification. Through this classification method, the ore separation recognition model can quickly identify mineral grains with different porosity levels according to the porosity characteristics of mineral grains, providing an important reference basis for the subsequent ore separation process. In the ore separation application scenario, this porosity-based classification method can effectively distinguish the compactness and structural integrity of mineral grains, and then help to more accurately screen out mineral grains that meet specific quality requirements in the ore separation process, improving the ore separation efficiency and quality control level.

[0016] Preferably, in step S3, to identify the mineral morphology characteristics of the mineral image and determine the mineral luster area of the mineral morphology characteristics includes: Perform color space conversion on the mineral image and convert the image to the HSV color space; Divide the mineral image into multiple regions in the HSV color space, and calculate the brightness mean and brightness standard deviation of each region; Determine the mineral morphology characteristics according to the brightness mean and brightness standard deviation; If the brightness mean of a region is greater than 0.7 and the brightness standard deviation is less than 0.1, then this region is marked as a high-brightness region; if the brightness mean of a region is between 0.5 and 0.7 and the brightness standard deviation is between 0.1 and 0.3, then this region is marked as a medium-brightness region; if the brightness mean of a region is less than 0.5 and the brightness standard deviation is greater than 0.3, then this region is marked as a low-brightness region; Calculate the texture contrast of each region, and determine the mineral luster area according to the texture characteristics; For the high-brightness region, if the contrast is less than 0.2, then this region is marked as a high-luster region; for the medium-brightness region, if the contrast is between 0.2 and 0.5, then this region is marked as a medium-luster region; for the low-brightness region, if the contrast is greater than 0.5, then this region is marked as a low-luster region.

[0017] The present invention performs color space conversion on mineral images to the HSV color space, divides multiple regions to calculate the brightness mean and standard deviation, and can accurately quantify the brightness characteristics of the mineral morphology. According to the brightness mean and standard deviation, the regions are marked as high, medium, and low brightness regions, and the texture contrast is further calculated to determine the mineral luster regions, and based on this, the regions are marked as high, medium, and low luster regions, realizing the accurate classification of the mineral luster characteristics. This process provides the mineral morphology and luster quantification data for the ore selection recognition model, and the model can accurately identify the appearance characteristics of the minerals based on this, so as to effectively distinguish minerals with different lusters and brightness in ore selection, improve the accuracy and efficiency of ore selection, and ensure that the ore selection results meet the quality requirements.

[0018] Preferably, the luster density detection of the mineral luster region in step S3 includes: Extract the brightness mean and brightness standard deviation of the mineral luster region; For each luster region, divide the brightness mean by the brightness standard deviation to evaluate the brightness uniformity of the luster region; determine the luster density according to the brightness uniformity; Divide the luster density into the maximum luster density, the minimum luster density, and the average luster density; For each luster region, calculate the difference between the maximum luster density and the minimum luster density within the luster region, and compare it with the average luster density to obtain the regional luster density data.

[0019] The present invention extracts the brightness mean and brightness standard deviation of the mineral luster region, evaluates the brightness uniformity of the luster region by dividing the brightness mean by the brightness standard deviation, and then determines the luster density. This process can accurately quantify the brightness characteristics of the mineral luster region. Divide the luster density into the maximum luster density, the minimum luster density, and the average luster density, and calculate the difference between the maximum luster density and the minimum luster density within each luster region, and compare it with the average luster density to obtain the regional luster density data, further refining the feature description of the luster region. These steps provide the detailed quantification data of the mineral luster region for the ore selection recognition model, enabling the model to more accurately identify and distinguish minerals with different luster densities, so as to more accurately screen out minerals meeting specific luster requirements in the ore selection process, improve the accuracy and reliability of ore selection, and ensure the quality of the ore selection results.

[0020] Preferably, the color change gradient division of the mineral morphology characteristics according to the regional luster density data and the determination of the mineral surface wear characteristics based on the color change gradient in step S3 include: For each pixel of the mineral morphology characteristics according to the regional luster density data, calculate the difference between its brightness mean and the brightness means of the surrounding 8 neighboring pixels; For each pixel of the mineral morphology characteristics according to the regional luster density data, calculate the difference between its brightness standard deviation and the brightness standard deviations of the surrounding 8 neighboring pixels; Add the difference in average brightness and the difference in standard deviation of brightness to obtain the color change gradient value; For all pixels within each region, calculate the average value of their color change gradient values; For all pixels within each region, calculate the standard deviation of their color change gradient values; If the average value of the color change gradient is greater than 0.5 and the standard deviation is less than 0.1, then the region is marked as a high wear region; If the average value of the color change gradient is between 0.3 and 0.5 and the standard deviation is between 0.1 and 0.3, then the region is marked as a medium wear region; If the average value of the color change gradient is less than 0.3 and the standard deviation is greater than 0.3, then the region is marked as a low wear region.

[0021] According to the regional gloss density data, the present invention calculates the difference in average brightness and the difference in standard deviation of brightness between each pixel of the mineral morphology feature and its surrounding 8 neighboring pixels, and adds the two to obtain the color change gradient value. This process can accurately quantify the color change characteristics of each pixel. Further, calculate the average value and standard deviation of the color change gradient values for all pixels within each region, and mark the regions as high wear, medium wear, and low wear regions based on these values, achieving an accurate classification of the wear degree of the mineral surface. These steps provide quantitative data on the wear characteristics of the mineral surface for the ore dressing recognition model, enabling the model to more accurately identify and distinguish minerals with different wear degrees, thereby more precisely screening out minerals that meet specific wear requirements during the ore dressing process, improving the accuracy and reliability of ore dressing, and ensuring the quality of the ore dressing results.

[0022] Preferably, step S4 includes the following steps: Step S41: Perform mineral multi-modal feature fusion on the mineral point-line-plane defect feature, the mineral grain shape feature, and the mineral surface wear feature to generate a mineral multi-modal fusion feature; Step S42: Select a deep learning framework to construct a preset ore dressing recognition model; Step S43: Define the network structure of the ore dressing recognition model, including an input layer, multiple hidden layers, and an output layer; among them, the dimension of the input layer is determined according to the data dimension of the mineral multi-modal fusion feature, and the dimension of the output layer is determined according to the number of target categories of the ore dressing recognition task; Step S44: Divide the mineral multi-modal fusion feature into a mineral feature training data set and a mineral feature validation data set; Step S45: Train the ore dressing recognition model based on the mineral feature training data set to obtain an ore dressing recognition training model; in each training epoch, divide the training data set into multiple batches, and input them into the model batch by batch for forward propagation and backward propagation to update the weight parameters of the model; Step S46: Verify the ore separation recognition training model based on the mineral feature verification dataset, and calculate the performance metrics of the model, including accuracy, recall rate, and F1-score, to obtain the model performance metrics. Step S47: Optimize the model parameters according to the model performance metrics to obtain the ore separation recognition model.

[0023] In step S41 of the present invention, by fusing the mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features, multi-modal fusion features of minerals are generated, providing comprehensive and rich input data for the ore separation recognition model, ensuring that the model can learn the features of minerals from multiple dimensions. Step S42 selects a deep learning framework to build the ore separation recognition model, providing powerful computing and learning capabilities for model training. Step S43 defines the network structure of the model, determines the dimensions of the input layer and output layer according to the data dimensions of the multi-modal fusion features of minerals and the number of target categories of the ore separation task, ensuring that the model can effectively process the input data and output accurate classification results. Step S44 divides the multi-modal fusion features of minerals into a training dataset and a verification dataset, providing a data basis for model training and verification. Step S45 verifies the model based on the verification dataset, calculates performance metrics such as accuracy, recall rate, and F1-score, ensuring the reliability and effectiveness of the model in practical applications. Step S46 optimizes the model parameters according to the performance metrics, further improving the performance of the model. The finally obtained ore separation recognition model can accurately identify the multi-modal features of minerals during the ore separation process, improve the accuracy and efficiency of ore separation, and ensure the quality of ore separation results. Description of the Drawings

[0024] Figure 1 It is a schematic diagram of the step flow of a training method for an ore separation recognition model based on multi-feature fusion; Figure 2 For Figure 1 the detailed implementation step flow diagram of step S4 in The realization, functional characteristics, and advantages of the object of the present invention will be further described in conjunction with the embodiments with reference to the drawings. Specific Embodiments

[0025] The technical method of the present invention will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0026] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0027] It should be understood that although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0028] To achieve the above object, please refer to Figures 1 to 2 , a training method for a mineral selection recognition model based on multi-feature fusion, the method comprising the following steps: Step S1: Collect mineral images; identify the mineral crystal structure information of the mineral images, perform geometric type division on the mineral crystal structure information to obtain the mineral crystal system type; perform point-line-plane defect feature detection on the mineral images according to the mineral crystal system type to generate mineral point-line-plane defect features; Step S2: Perform mineral crystal symbiosis recognition on the mineral images according to the mineral crystal system type, evaluate the mineral crystal symbiosis relationship, and generate mineral crystal symbiosis data; perform mineral grain connectivity analysis on the mineral crystal symbiosis data, and detect the mineral porosity; determine the mineral grain shape characteristics based on the mineral porosity; Step S3: Identify the mineral morphology characteristics of the mineral images, determine the mineral luster area of the mineral morphology characteristics; perform luster density detection on the mineral luster area to obtain area luster density data; perform color change gradient division on the mineral morphology characteristics according to the area luster density data, and determine the mineral surface wear characteristics based on the color change gradient; Step S4: Perform mineral multi-modal feature fusion on the mineral point-line-plane defect features, mineral grain shape characteristics, and mineral surface wear characteristics to generate mineral multi-modal fusion features; construct a preset mineral selection recognition model through deep learning technology, perform model training on the preset mineral selection recognition model according to the mineral multi-modal fusion features, and verify the model performance indicators; optimize the model parameters according to the model performance indicators to obtain the mineral selection recognition model.

[0029] In the embodiments of the present invention, refer toFigure 1 As shown in the figure, it is a schematic flowchart of the steps of a training method for a mineral separation recognition model based on multi-feature fusion according to the present invention. In this example, the training method for the mineral separation recognition model based on multi-feature fusion includes the following steps: Step S1: Collect mineral images; identify the mineral crystal structure information of the mineral images, perform geometric type division on the mineral crystal structure information to obtain the mineral crystal system type; perform point-line-plane defect feature detection on the mineral images according to the mineral crystal system type to generate mineral point-line-plane defect features; In the embodiment of the present invention, a high-resolution industrial camera is used to collect images of mineral samples. The resolution of this camera is 4000×3000 pixels, the aperture is set to F8, and the exposure time is 1 / 125 second. The collected mineral image data is named mineral image data 1. Subsequently, the mineral image data 1 is input into a pre-trained deep convolutional neural network model. This model is based on the ResNet architecture and is trained with a large amount of mineral crystal image data for identifying mineral crystal structure information. The model extracts image features through operations such as convolutional layers and pooling layers, and outputs mineral crystal structure information data. Then, based on geometric feature parameters such as symmetry and the number of axes in the mineral crystal structure information data, using the crystallographic geometric division algorithm and in accordance with the crystal system division standard specified by the International Mineralogical Union, the mineral crystal structure is divided into types such as triclinic system, monoclinic system, orthorhombic system, tetragonal system, hexagonal system, and cubic system to obtain mineral crystal system type data. Finally, according to the crystal system type determined by the mineral crystal system type data 1, the corresponding point-line-plane defect feature detection algorithm is selected. For triclinic system mineral images, a point defect detection algorithm based on Fourier transform is used, and the frequency threshold is set to 0.05 pixel^-1 to detect the point defect features in the image; for monoclinic system mineral images, a line defect detection algorithm using the Hough transform is used, and the angle range is set to 0°~180° and the threshold is 100 to detect the line defect features; for other crystal system mineral images, a surface defect detection algorithm combining edge detection and region growing is used, and the gradient threshold for edge detection is set to 0.1 and the similarity threshold for region growing is set to 0.9 to generate mineral point-line-plane defect feature data, named mineral point-line-plane defect feature data.

[0030] Step S2: Perform mineral crystal paragenesis recognition on the mineral images according to the mineral crystal system type, evaluate the mineral crystal paragenesis relationship, and generate mineral crystal paragenesis data; perform mineral grain connectivity analysis on the mineral crystal paragenesis data, and detect the mineral porosity; determine the mineral grain shape characteristics based on the mineral porosity; In the embodiments of the present invention, for the co - growth recognition of mineral image data, a co - growth recognition technology based on image segmentation and feature matching is adopted. The mineral image data 1 is segmented into multiple independent mineral crystal regions by using an image segmentation algorithm. The segmentation algorithm adopts the K - means clustering method, and the number of clustering centers is set to 5 to adapt to the color and gray - scale differences of different mineral crystals. Subsequently, for each mineral crystal region, its shape, texture, and spectral features are extracted. The shape feature is obtained through a contour detection algorithm, the texture feature is calculated by using a gray - level co - occurrence matrix, and the spectral feature is obtained based on the statistical values of the RGB channels of the image. The extracted features are matched with a preset mineral crystal feature database. The matching algorithm adopts the Euclidean distance metric, and the matching threshold is set to 0.2, so as to identify the types and co - growth relationships of mineral crystals and generate mineral crystal co - growth data. For the analysis of the connectivity of mineral grains in the mineral crystal co - growth data, a connectivity analysis algorithm in graph theory is used. Each mineral crystal region is regarded as a node in the graph. If two mineral crystal regions are adjacent in space, an edge is established between the corresponding nodes. The graph structure is traversed by using a depth - first search algorithm, and the number and size of connected components are counted to evaluate the connectivity of mineral grains. At the same time, the mineral porosity is detected. The mineral region and the pore region in the mineral image data are separated by using a threshold segmentation method. The gray - level threshold is set to 120, and the proportion of the pore region in the whole image is calculated to obtain the mineral porosity data. Based on the mineral porosity data 1, the shape characteristics of mineral grains are determined. When the mineral porosity is less than 10%, it indicates that the mineral crystals are closely packed, and the grain shape tends to be a regular polyhedron. At this time, a polyhedron fitting algorithm is used to fit the polyhedron shape of the mineral grains according to the crystal system symmetry determined by the mineral crystal system type data, and the normal vectors and areas of each crystal face are recorded; when the mineral porosity is greater than 10% and less than 30%, the mineral grain shape is between regular and irregular. The shape factor analysis method is used to calculate the shape factor of the mineral crystal region. The shape factor is obtained by comparing the perimeter and area of the mineral crystal region. According to the numerical range of the shape factor, the grain shape is classified as near - spherical, flat - shaped, or column - shaped; when the mineral porosity is greater than 30%, the mineral grain shape is mostly irregular. The contour fitting algorithm is used to segment - fit the contour of the mineral crystal region. The fitting curve uses a B - spline curve, and the control point coordinates of each segment of the curve are recorded to describe the shape characteristics of the mineral grains and generate the mineral grain shape characteristic data.

[0031] Step S3: Identify the mineral morphology features of the mineral image, and determine the mineral luster region of the mineral morphology features; perform luster density detection on the mineral luster region to obtain the regional luster density data; divide the color change gradient of the mineral morphology features according to the regional luster density data, and determine the mineral surface wear features based on the color change gradient; In the embodiment of the present invention, in step S3, first, mineral morphology features of the mineral image are identified. An algorithm based on edge detection and region segmentation is adopted. The Canny edge detection operator is used to extract the edges of the mineral image. The low threshold is set to 50 and the high threshold is set to 150 to accurately outline the contour of the mineral. Subsequently, the region growing algorithm is used to segment the image after edge detection to identify different morphological feature regions of the mineral. The initial gray threshold of the seed point is set to 100, and the growth direction is 8-connected, so as to obtain the mineral morphology feature region. In the identified mineral morphology feature region, the mineral luster region is determined. A method based on gray histogram analysis is adopted to calculate the gray histogram of each morphological feature region, and the luster region is identified by analyzing the peak value and distribution of the histogram. The gray threshold range of the luster region is set to 180-255 to distinguish the luster region from other regions. The luster density of the identified mineral luster region is detected. The local mean filtering algorithm is adopted to calculate the average gray value within the neighborhood of each pixel point in the luster region, and the neighborhood size is set to 3×3 pixels. The luster density of each pixel point is evaluated through the calculated local mean value, so as to obtain the regional luster density data. The mineral morphology features are divided into color change gradients according to the regional luster density data. The Sobel operator is used to calculate the gradient of the mineral image, and the gradient values in the horizontal and vertical directions are calculated respectively. The gradient threshold is set to 120 to determine the gradient boundary of the color change. The mineral morphology feature region is divided into multiple sub-regions according to the color change gradient. Based on the color change gradient division result, the mineral surface wear feature is determined. By analyzing the position and shape of the color change gradient boundary and combining the luster density data, the worn area on the mineral surface is identified. The worn area usually shows an area with a large color change gradient and a low luster density. The erosion and dilation algorithms in morphological operations are used to refine and optimize the worn area. The structural element of the erosion operation is a 3×3 rectangle, and the structural element of the dilation operation is also a 3×3 rectangle. Through these operations, the mineral surface wear feature is finally determined.

[0032] Step S4: Fuse the mineral point-line-plane defect features, the mineral crystal grain shape features, and the mineral surface wear features to generate mineral multi-modal fusion features; construct a preset ore separation recognition model through deep learning technology, train the preset ore separation recognition model according to the mineral multi-modal fusion features, and verify the model performance indicators; optimize the model parameters according to the model performance indicators to obtain the ore separation recognition model.

[0033] In the embodiments of the present invention, mineral multi-modal feature fusion is performed on mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features. A method based on feature splicing and weighted fusion is adopted. The three features are respectively represented in vector form. The dimension of the mineral point-line-plane defect feature vector is 120, the dimension of the mineral grain shape feature vector is 80, and the dimension of the mineral surface wear feature vector is 50. These three feature vectors are spliced in sequence into a high-dimensional feature vector to obtain an initial fusion feature vector with a dimension of 250. Subsequently, according to the weight coefficients determined by pre-experiments, the initial fusion feature vector is weighted. The weight of the mineral point-line-plane defect feature is 0.4, the weight of the mineral grain shape feature is 0.3, and the weight of the mineral surface wear feature is 0.3. The mineral multi-modal fusion feature vector is obtained through weighted summation, and its dimension is still 250. Then, a preset ore selection recognition model is constructed through deep learning technology. The convolutional neural network (CNN) is selected as the model architecture, and specifically, the improved ResNet-50 model is adopted. This model contains 4 residual blocks, and each residual block consists of 3 convolutional layers. The kernel sizes of the convolutional layers are 3×3, the stride is 1, and the padding is 1. A batch normalization layer and a ReLU activation function are added after each convolutional layer. At the end of the model, a global average pooling layer and a fully connected layer are added. The output dimension of the fully connected layer is the number of ore selection categories. Assuming there are 10 types of minerals, the output dimension is 10. The loss function of the model adopts the cross-entropy loss function, the optimizer selects the Adam optimizer, the learning rate is set to 0.001, the momentum parameter is 0.9, and the weight decay coefficient is 0.0001. The preset ore selection recognition model is trained according to the mineral multi-modal fusion features. The fusion feature vector is used as the input of the model, and the corresponding mineral category label is used as the output, and the stochastic gradient descent method is used for training. The training dataset contains 10,000 samples, and the validation dataset contains 2,000 samples. The training process is set to 100 epochs, and each epoch makes a complete traversal of the training dataset. After each epoch ends, the accuracy, recall, and F1 score of the model on the validation dataset are calculated as the model performance indicators. The model parameters are optimized according to the model performance indicators. During the training process, an early stopping mechanism is adopted. When the accuracy on the validation set does not improve within 10 consecutive epochs, the training is stopped to avoid overfitting. At the same time, a learning rate decay strategy is adopted. During the training process, every 20 epochs, the learning rate decays to 0.1 times the original. After the training is completed, according to the performance indicators on the validation set, the hyperparameters of the model are fine-tuned, including adjusting the kernel size, learning rate, and weight decay coefficient, etc., and finally an optimized ore selection recognition model is obtained.

[0034] Preferably, in step S1, the mineral crystal structure information of the mineral image is identified, and the geometric type division of the mineral crystal structure information includes: The mineral image is scanned line by line, and the starting and ending point coordinates of the crystal structure in the horizontal and vertical directions in the image are recorded to form a mineral crystal structure coordinate data set; Based on the mineral crystal structure coordinate data set, the center point position of each crystal structure is calculated. Taking the center point as a reference, the crystal structure is divided into four quadrant regions, and the pixel point distribution in each quadrant is recorded; For the crystal structure in each quadrant region, its crystal shape factor is calculated, where the shape factor is determined by the ratio of the perimeter to the area of the crystal structure, and the shape factor value of each quadrant is generated; The crystal structure is divided into two categories: regular crystals and irregular crystals according to the shape factor value of each quadrant; For the structure of the irregular crystal, the curvature change value of its edge points is calculated; the degree of concavity and convexity is determined according to the curvature change value; Based on the degree of concavity and convexity, the geometric type division of the mineral crystal structure information is carried out to obtain the mineral crystal system type.

[0035] In the embodiments of the present invention, the mineral image is scanned line by line. Using the scan line algorithm in image processing, starting from the upper left pixel point of the image, it is scanned line by line along the horizontal direction, and the starting point and ending point coordinates of the crystal structure in the horizontal direction of the image are recorded; then, starting from the upper left pixel point of the image, it is scanned column by column along the vertical direction, and the starting point and ending point coordinates of the crystal structure in the vertical direction of the image are recorded, forming a mineral crystal structure coordinate data set. Based on the mineral crystal structure coordinate data set, the center point position of each crystal structure is calculated. Using the geometric center calculation formula, the starting point and ending point coordinates of the crystal structure in the horizontal direction and the starting point and ending point coordinates in the vertical direction are respectively averaged to obtain the center point coordinates. Taking the center point as the reference, the crystal structure is divided into four quadrant regions. By the coordinate comparison method, the upper left region of the center point is defined as the first quadrant, the upper right region is defined as the second quadrant, the lower right region is defined as the third quadrant, and the lower left region is defined as the fourth quadrant, and the pixel point distribution in each quadrant is recorded. Using the pixel point counting method, the number of pixel points in each quadrant is counted. For the crystal structure in each quadrant region, its crystal shape factor is calculated. The shape factor is determined by the ratio of the perimeter to the area of the crystal structure. The contour tracing algorithm is used to calculate the perimeter of the crystal structure, and the step size of the contour tracing is set to 1 pixel unit. The pixel filling algorithm is used to calculate the area of the crystal structure, and the filling threshold is set to 0.5, generating the shape factor value of each quadrant. According to the shape factor value of each quadrant, the crystal structure is divided into two categories: regular crystals and irregular crystals. The shape factor threshold is set to 1.5. When the shape factor is less than or equal to 1.5, it is determined as a regular crystal; when the shape factor is greater than 1.5, it is determined as an irregular crystal. For the structure of the irregular crystal, the curvature change value of its edge points is calculated. The differential method is used to calculate the curvature, and the coordinate difference of the adjacent points before and after each edge point is calculated to obtain the curvature change value. The degree of concavity and convexity is determined according to the curvature change value. The curvature change threshold is set to 0.2. When the curvature change value is greater than 0.2, it is determined that the degree of concavity and convexity is larger; when the curvature change value is less than or equal to 0.2, it is determined that the degree of concavity and convexity is smaller. Based on the degree of concavity and convexity, the geometric type of the mineral crystal structure information is divided to obtain the mineral crystal system type. Using the crystallographic geometric classification rules, the crystal structures with a smaller degree of concavity and convexity are classified as low-level crystal systems, such as the triclinic crystal system; the crystal structures with a larger degree of concavity and convexity are classified as high-level crystal systems, such as the cubic crystal system.

[0036] Preferably, in step S1, the detection of point, line and surface defect characteristics of the mineral image according to the mineral crystal system type includes: The gradient of the mineral image is calculated according to the mineral crystal system type. The Sobel operator is used to calculate the gradients of the image in the horizontal and vertical directions respectively, generating a gradient magnitude image and a gradient direction image; Based on the gradient magnitude image, identify the local extreme points of the gradient magnitude, and use the local extreme points as potential candidate points for point defects; calculate the ratio of the gradient magnitude of each candidate point to the gradient magnitudes of its 8 surrounding neighborhood pixels, and retain the candidate points with a ratio greater than 1.2 as mineral crystal system defect points, and record the positions and gradient magnitudes of the mineral crystal system defect points; Starting from the upper left corner of the gradient direction image, track along the gradient direction until a point where the gradient direction changes by more than 45 degrees is encountered or the image boundary is reached; For each extracted line, calculate its length and curvature, where the curvature is determined by calculating the average of the angular changes between every two adjacent points on the line; Mark the lines with a length greater than 10 pixels and a curvature greater than 0.1 as mineral crystal system defect point lines, and record the positions, lengths, and curvatures of the mineral crystal system defect point lines; Perform multi-region segmentation on the gradient magnitude image, with each region corresponding to a part of the mineral crystal; Calculate the standard deviation of the gradient magnitudes of each region, and mark the regions with a standard deviation greater than 0.5 as mineral crystal system defect surfaces, and record the positions and areas of the mineral crystal system defect surfaces.

[0037] In the embodiments of the present invention, the gradient of the mineral image is calculated according to the type of mineral crystal system. The Sobel operator is used to calculate the gradients of the image in the horizontal and vertical directions respectively. In the horizontal direction, the Sobel operator detects horizontal edges by calculating the gray difference between each pixel in the image and its adjacent pixels above and below, and enhances the gradient information in the horizontal direction of the image. In the vertical direction, the Sobel operator detects vertical edges by calculating the gray difference between each pixel in the image and its adjacent pixels on the left and right, and enhances the gradient information in the vertical direction of the image. Through the gradient calculations in these two directions, a gradient magnitude image and a gradient direction image are generated. The gradient magnitude image represents the gradient intensity of each pixel, and the gradient direction image represents the gradient direction of each pixel. Based on the gradient magnitude image, the local extreme points of the gradient magnitude are identified. The non-maximum suppression algorithm is used to scan the gradient magnitude image. For each pixel, the gradient magnitudes in its 8 neighborhoods are checked. If the gradient magnitude of this pixel is greater than the gradient magnitudes of all its neighborhood points, it is marked as a local extreme point and used as a potential candidate point for point defects. Then, the ratio of the gradient magnitude of each candidate point to the gradient magnitudes of its surrounding 8 neighborhood pixels is calculated. If the ratio is greater than 1.2, the candidate point is retained as a mineral crystal system defect point, and its position and gradient magnitude are recorded. Starting from the upper left corner of the gradient direction image, tracking is performed along the gradient direction. The gradient tracking algorithm is used, with the gradient direction of each pixel as the direction vector, and gradually move to the next pixel. During the tracking process, if a point with a gradient direction change exceeding 45 degrees is encountered or the image boundary is reached, the tracking stops. For each extracted line, its length and curvature are calculated. The line length is obtained by calculating the Euclidean distance between the starting point and the ending point of the line, and the curvature is determined by calculating the average value of the angular changes between every two adjacent points on the line. If the length of the line is greater than 10 pixels and the curvature is greater than 0.1, the line is marked as a mineral crystal system defect line, and its position, length, and curvature are recorded. The gradient magnitude image is segmented into multiple regions using a threshold-based region segmentation method. The gradient magnitude threshold is set to 0.3, and according to this threshold, the gradient magnitude image is segmented into multiple regions, each region corresponding to a part of the mineral crystal. The standard deviation of the gradient magnitudes of each region is calculated using the standard deviation calculation formula for statistical analysis of the gradient magnitudes within each region. If the standard deviation of the gradient magnitudes of a certain region is greater than 0.5, the region is marked as a mineral crystal system defect surface, and its position and area are recorded.

[0038] Preferably, in step S2, mineral crystal symbiosis recognition is performed on the mineral image according to the type of mineral crystal system, and the evaluation of the mineral crystal symbiosis relationship includes: Edge detection is performed on the mineral image, and the edge information of the mineral crystal is extracted to generate a mineral edge image; Identify the contours of the mineral edge images according to the mineral crystal system type, where each contour corresponds to a mineral crystal, and obtain the mineral edge contour information; Calculate the geometric center of the mineral edge contour information to determine the geometric center position of each mineral crystal; Calculate the Euclidean distance between the geometric centers of each pair of mineral crystals, and identify the pairs of mineral crystals with a distance less than 10% of the image width as symbiotic relationships; For each pair of mineral crystals identified as having a symbiotic relationship, calculate the overlapping area ratio of their contours, where the overlapping area ratio is the ratio of the area of the overlapping region of the two contours to the total area of the two contours; Mark the pairs of mineral crystals with an overlapping area ratio greater than 0.1 as closely symbiotic relationships; mark the pairs of mineral crystals with an overlapping area ratio between 0.05 and 0.1 as moderately symbiotic relationships; mark the pairs of mineral crystals with an overlapping area ratio less than 0.05 as loosely symbiotic relationships.

[0039] In the embodiments of the present invention, first, edge detection is performed on the mineral image using the Canny edge detection algorithm, which identifies edges by calculating the gradient magnitude and direction of the image. The low threshold of the Canny algorithm is set to 50 and the high threshold is set to 150 to extract the edge information of the mineral crystal and generate a mineral edge image. In the mineral edge image, the edges are represented by white pixels and the background is represented by black pixels. The contour of the mineral edge image is identified according to the type of mineral crystal system using a contour detection algorithm such as the findContours function in OpenCV. This algorithm can identify the closed contours in the image, and each contour corresponds to a mineral crystal, obtaining the mineral edge contour information. The mode of contour detection is set to a tree structure to obtain the hierarchical relationship between the contours. The geometric center of the mineral edge contour information is calculated using the geometric center calculation formula, that is, the coordinates of all pixel points within each contour are averaged to obtain the geometric center position of each mineral crystal. The specific calculation method is to add the abscissas and ordinates of all pixel points within the contour respectively, and then divide by the total number of pixel points within the contour. The Euclidean distance between the geometric centers of each pair of mineral crystals is calculated using the Euclidean distance formula, that is, the straight-line distance between two points. For each pair of mineral crystals, the distance between their geometric centers is calculated, and the pairs of mineral crystals with a distance less than 10% of the image width are identified as symbiotic relationships. For example, if the image width is 1000 pixels, the pairs of mineral crystals with a distance less than 100 pixels are identified as symbiotic relationships. For each pair of mineral crystals identified as having a symbiotic relationship, the overlapping area ratio of their contours is calculated. First, the area of the overlapping region between the two contours is calculated using the contour intersection algorithm, and then the total area of the two contours is calculated, that is, the sum of the areas of the two contours respectively. The overlapping area ratio is the ratio of the overlapping region area to the total area of the two contours. Classification is performed according to the overlapping area ratio: pairs of mineral crystals with an overlapping area ratio greater than 0.1 are marked as having a close symbiotic relationship; pairs of mineral crystals with an overlapping area ratio between 0.05 and 0.1 are marked as having a medium symbiotic relationship; pairs of mineral crystals with an overlapping area ratio less than 0.05 are marked as having a loose symbiotic relationship.

[0040] Preferably, in step S2, the mineral grain connectivity analysis of the mineral crystal symbiotic data is performed, and the detection of the mineral porosity includes: Analyze the mineral crystal symbiotic data and extract the geometric center positions of each pair of symbiotic mineral crystals; Based on the geometric center positions, construct a connectivity graph of the mineral crystals, where each mineral crystal is represented as a node and the connectivity relationship between each pair of symbiotic mineral crystals is represented as an edge; Perform a depth-first search on the connectivity graph, starting from each unvisited node and traversing, recording the number of connected components and the size of each connected component; Segment the mineral image and mark the mineral background area; By multiplying the number of background pixels after statistical segmentation by the area of each pixel, the area of the mineral background region is obtained; the area of the mineral background region is determined as the mineral pore area; The ratio of the mineral pore area to the area of the entire mineral image is calculated to generate the mineral porosity.

[0041] In the embodiments of the present invention, the symbiotic data of mineral crystals is analyzed to extract the geometric center positions of each pair of symbiotic mineral crystals. Using data analysis techniques, the geometric center coordinate information in the symbiotic data of mineral crystals is read, and these coordinates are calculated based on the mineral edge contour information. The geometric center position of each mineral crystal is recorded in the form of (x, y) coordinates. Based on the geometric center positions, a connectivity graph of mineral crystals is constructed. In this graph, each mineral crystal is represented as a node, and the connectivity relationship between each pair of symbiotic mineral crystals is represented as an edge. Using the concepts of nodes and edges in graph theory, a connectivity graph is established according to the extracted geometric center positions, and the edges between nodes represent the symbiotic relationship between mineral crystals. Depth-first search is performed on the connectivity graph, starting from each unvisited node and traversing along the edges, recording the number of connected components and the size of each connected component. Using the depth-first search algorithm, starting from an unvisited node in the connectivity graph, depth-first traversal is performed along the edges until all reachable nodes are visited, recording the size of this connected component, and then starting from another unvisited node, repeating the above process until all nodes are visited, so as to obtain the number of connected components and the size of each connected component. The mineral image is segmented, and the mineral background region is marked. Using image segmentation techniques, such as threshold segmentation method, according to the gray or color characteristics of the mineral image, the image is segmented into a mineral crystal region and a background region. By setting an appropriate threshold, the background region is distinguished from the mineral crystal region, and the background region is marked. By multiplying the number of background pixels after segmentation by the area of each pixel, the area of the mineral background region is obtained. After the mineral image is segmented, the number of pixels in the background region is counted, and then multiplied by the actual area represented by each pixel (determined according to the resolution of the image), so as to calculate the area of the mineral background region. The area of the mineral background region is determined as the mineral pore area. The above-calculated area of the mineral background region is regarded as the mineral pore area, because the background region usually represents the voids between mineral crystals. The ratio of the mineral pore area to the area of the entire mineral image is calculated to generate the mineral porosity. The ratio of the mineral pore area to the area of the entire mineral image is calculated to obtain the mineral porosity, which reflects the proportion of pores in the mineral image and can be used to evaluate the pore characteristics of the mineral.

[0042] Preferably, in step S2, determining the mineral grain shape characteristics based on the mineral porosity includes: The mineral grains are preliminarily classified according to the mineral porosity to obtain the mineral grain shape characteristics; If the porosity is less than 0.1, mark the grain as a low-porosity grain; If the porosity is between 0.1 and 0.3, mark the grain as a medium-porosity grain; If the porosity is greater than 0.3, mark the grain as a high-porosity grain.

[0043] In an embodiment of the present invention, mineral porosity data is obtained, and this data is obtained by calculating the ratio of the area of the mineral background region to the area of the entire mineral image. According to the numerical range of the porosity, the mineral grains are classified. A porosity threshold range is set. When the porosity is less than 0.1, the grain is marked as a low-porosity grain; when the porosity is between 0.1 and 0.3, the grain is marked as a medium-porosity grain; when the porosity is greater than 0.3, the grain is marked as a high-porosity grain. Through this classification method, the mineral grains are distinguished according to the high and low porosity, so as to obtain the shape characteristics of the mineral grains.

[0044] Preferably, in step S3, the mineral morphology characteristics of the mineral image are identified, and the mineral luster regions determining the mineral morphology characteristics include: Perform color space conversion on the mineral image to convert the image to the HSV color space; Divide the mineral image into multiple regions in the HSV color space, and calculate the brightness mean and brightness standard deviation of each region; Determine the mineral morphology characteristics according to the brightness mean and brightness standard deviation; If the brightness mean of a region is greater than 0.7 and the brightness standard deviation is less than 0.1, then the region is marked as a high-brightness region; if the brightness mean of a region is between 0.5 and 0.7 and the brightness standard deviation is between 0.1 and 0.3, then the region is marked as a medium-brightness region; if the brightness mean of a region is less than 0.5 and the brightness standard deviation is greater than 0.3, then the region is marked as a low-brightness region; Calculate the texture contrast of each region, and determine the mineral luster region according to the texture characteristics; For a high-brightness region, if the contrast is less than 0.2, then the region is marked as a high-luster region; for a medium-brightness region, if the contrast is between 0.2 and 0.5, then the region is marked as a medium-luster region; for a low-brightness region, if the contrast is greater than 0.5, then the region is marked as a low-luster region.

[0045] In the embodiments of the present invention, first, color space conversion is performed on the mineral image. Using color space conversion technology, the image is converted from the RGB color space to the HSV color space. In the HSV color space, H represents hue, S represents saturation, and V represents brightness. Through color space conversion, the brightness information of the image is better separated, facilitating subsequent processing. In the HSV color space, the mineral image is divided into multiple regions. Using the uniform division method, the image is divided into several regions of equal size. For example, the image is divided into 16×16 regions. For each region, its brightness mean and brightness standard deviation are calculated. The brightness mean is obtained by calculating the average value of the V components of all pixels within the region, and the brightness standard deviation is obtained by calculating the standard deviation of the V components of all pixels within the region. The mineral morphology characteristics are determined based on the brightness mean and brightness standard deviation. The following rules are set: If the brightness mean of a region is greater than 0.7 and the brightness standard deviation is less than 0.1, then the region is marked as a high-brightness region. If the brightness mean of a region is between 0.5 and 0.7 and the brightness standard deviation is between 0.1 and 0.3, then the region is marked as a medium-brightness region. If the brightness mean of a region is less than 0.5 and the brightness standard deviation is greater than 0.3, then the region is marked as a low-brightness region. The texture contrast of each region is calculated, and the gray-level co-occurrence matrix (GLCM) method is used to calculate the texture contrast. For each region, its gray-level co-occurrence matrix is calculated, and the contrast feature is extracted from the gray-level co-occurrence matrix. The contrast feature reflects the degree of change in the pixel gray values within the region and is used to describe the texture characteristics of the region. Based on the texture characteristics, the mineral gloss regions are determined. The following rules are set: For high-brightness regions, if the contrast is less than 0.2, then the region is marked as a high-gloss region. For medium-brightness regions, if the contrast is between 0.2 and 0.5, then the region is marked as a medium-gloss region. For low-brightness regions, if the contrast is greater than 0.5, then the region is marked as a low-gloss region.

[0046] Preferably, in step S3, the gloss density detection of the mineral gloss region includes: Extracting the brightness mean and brightness standard deviation of the mineral gloss region; For each gloss region, dividing the brightness mean by the brightness standard deviation to evaluate the brightness uniformity of the gloss region; determining the gloss density according to the brightness uniformity; Dividing the gloss density into the maximum gloss density, the minimum gloss density, and the average gloss density; For each gloss region, calculating the difference between the maximum gloss density and the minimum gloss density within the gloss region and comparing it with the average gloss density to obtain the regional gloss density data.

[0047] In the embodiments of the present invention, the brightness information of the marked mineral luster regions is extracted from the mineral images. Using the image region extraction technology, according to the previously determined boundaries of the luster regions, the pixel brightness values of each luster region are extracted from the brightness channel (V channel) of the HSV color space. For each luster region, its brightness mean and brightness standard deviation are calculated. The brightness mean is obtained by adding up the brightness values of all pixels in the region and then dividing by the total number of pixels; the brightness standard deviation is obtained by calculating the sum of the squares of the differences between the brightness value of each pixel in the region and the brightness mean, then dividing by the total number of pixels and taking the square root. For each luster region, the brightness mean is divided by the brightness standard deviation to obtain the brightness uniformity of the region. The brightness uniformity is used to evaluate the brightness consistency of the luster region, and the larger the value, the more uniform the brightness. According to the magnitude of the brightness uniformity, the luster density is determined. A threshold of the brightness uniformity is set, and the luster density is divided into three levels: the maximum luster density, the minimum luster density, and the average luster density. For example, the regions with a brightness uniformity greater than 2 are defined as the maximum luster density regions, the regions with a brightness uniformity less than 1 are defined as the minimum luster density regions, and the regions with a brightness uniformity between 1 and 2 are defined as the average luster density regions. For each luster region, the difference between the maximum luster density and the minimum luster density within the region is further calculated and compared with the average luster density to obtain the luster density data of the region. The specific operation is as follows: within each luster region, find the pixel points with the maximum and minimum brightness values, corresponding to the maximum luster density and the minimum luster density respectively. Calculate the difference between these two values, and then compare this difference with the average value of all pixel brightness values within the region (i.e., the average luster density). In this way, the change of the luster density of each luster region is obtained, providing important characteristic data for subsequent ore selection and identification.

[0048] Preferably, in step S3, the color change gradient of the mineral morphology features is divided according to the regional luster density data, and the mineral surface wear features are determined based on the color change gradient, including: For each pixel of the mineral morphology features according to the regional luster density data, calculate the difference between its brightness mean and the brightness means of the surrounding 8 neighboring pixels; For each pixel of the mineral morphology features according to the regional luster density data, calculate the difference between its brightness standard deviation and the brightness standard deviations of the surrounding 8 neighboring pixels; Add the brightness mean difference and the brightness standard deviation difference to obtain the color change gradient value; For all pixels within each region, calculate the average value of their color change gradient values; For all pixels within each region, calculate the standard deviation of their color change gradient values; If the average value of the color change gradient is greater than 0.5 and the standard deviation is less than 0.1, then the region is marked as a high wear region; If the average value of the color change gradient is between 0.3 and 0.5 and the standard deviation is between 0.1 and 0.3, then the area is marked as a medium wear area; If the average value of the color change gradient is less than 0.3 and the standard deviation is greater than 0.3, then the area is marked as a low wear area.

[0049] In the embodiments of the present invention, during the mineral image analysis process, first, each pixel of the mineral morphology characteristics is analyzed based on the regional gloss density data. For each pixel point, the difference between its brightness value and the average brightness value of the surrounding 8 neighboring pixels is calculated. The specific operation is as follows: Select the target pixel point, count the brightness values of its surrounding 8 neighboring pixels, calculate the average value of these 8 brightness values, and then find the difference between the brightness value of the target pixel and this average value to obtain the brightness average difference. Next, also based on the regional gloss density data, for each pixel of the mineral morphology characteristics, the difference between its brightness standard deviation and the brightness standard deviation of the surrounding 8 neighboring pixels is calculated. First, calculate the standard deviation of the brightness values of the surrounding 8 neighboring pixels, and then find the difference between the brightness value of the target pixel and this standard deviation to obtain the brightness standard deviation difference. Add the above-obtained brightness average difference and brightness standard deviation difference to obtain the color change gradient value of this pixel. This operation is performed one by one for each pixel in the mineral image to quantify the color change degree of each pixel point. Statistical analysis is performed on the color change gradient values of all pixels in each area. First, calculate the average value of the color change gradient values of all pixels in each area, which is obtained by adding up the color change gradient values of all pixels in the area and then dividing by the total number of pixels. Second, calculate the standard deviation of the color change gradient values of all pixels in each area. Using the standard deviation calculation formula, statistical analysis is performed on the color change gradient values of each pixel in the area to evaluate the dispersion degree of the color change within the area. The wear degree of the area is marked according to the average value and standard deviation of the color change gradient. The following rules are set: If the average value of the color change gradient is greater than 0.5 and the standard deviation is less than 0.1, then the area is marked as a high wear area; If the average value of the color change gradient is between 0.3 and 0.5 and the standard deviation is between 0.1 and 0.3, then the area is marked as a medium wear area; If the average value of the color change gradient is less than 0.3 and the standard deviation is greater than 0.3, then the area is marked as a low wear area.

[0050] As an example of the present invention, refer to Figure 2 As shown, in this example, step S4 includes: Step S41: Perform mineral multi-modal feature fusion on the mineral point-line-plane defect feature, the mineral grain shape feature, and the mineral surface wear feature to generate a mineral multi-modal fusion feature; Step S42: Select a deep learning framework to construct a preset ore selection recognition model; Step S43: Define the network structure of the ore separation recognition model, including an input layer, multiple hidden layers, and an output layer; among them, the dimension of the input layer is determined according to the data dimension of the mineral multi-modal fusion features, and the dimension of the output layer is determined according to the number of target categories of the ore separation recognition task; Step S44: Divide the mineral multi-modal fusion features into a mineral feature training data set and a mineral feature validation data set; Step S45: Train the ore separation recognition model based on the mineral feature training data set to obtain an ore separation recognition training model; in each training round, divide the training data set into multiple batches, and input them into the model batch by batch for forward propagation and backward propagation to update the weight parameters of the model; Step S46: Validate the ore separation recognition training model based on the mineral feature validation data set, and calculate the performance indicators of the model, including accuracy, recall rate, and F1 score, to obtain the model performance indicators; Step S47: Optimize the model parameters according to the model performance indicators to obtain the ore separation recognition model.

[0051] In the embodiments of the present invention, multi-modal feature fusion of minerals is performed. The point-line-plane defect features of minerals, the grain shape features of minerals, and the surface wear features of minerals are fused. The specific operation is as follows: The three features are respectively represented in vector form. Among them, the dimension of the point-line-plane defect feature vector of minerals is 120, the dimension of the grain shape feature vector of minerals is 80, and the dimension of the surface wear feature vector of minerals is 50. These three feature vectors are concatenated in sequence to form a high-dimensional feature vector, obtaining an initial fusion feature vector with a dimension of 250. Subsequently, according to the weight coefficients determined by pre-experiments, the initial fusion feature vector is weighted. Among them, the weight of the point-line-plane defect features of minerals is 0.4, the weight of the grain shape features of minerals is 0.3, and the weight of the surface wear features of minerals is 0.3. The multi-modal fusion feature vector of minerals is obtained through weighted summation, and its dimension is still 250. A deep learning framework is selected to construct a preset ore selection recognition model. The TensorFlow deep learning framework is adopted. This framework supports the construction and training of various neural network structures and has good flexibility and scalability. The network structure of the ore selection recognition model is defined. The network structure includes an input layer, multiple hidden layers, and an output layer. The dimension of the input layer is determined according to the data dimension of the multi-modal fusion features of minerals, that is, the dimension of the input layer is 250. The dimension of the output layer is determined according to the number of target categories in the ore selection recognition task. Assuming that the number of target categories in the ore selection task is 10, the dimension of the output layer is 10. The hidden layer adopts 3 fully connected layers, and the number of neurons in each layer is 128, 64, and 32 respectively. The ReLU activation function is added after each fully connected layer to introduce non-linearity. The Softmax activation function is added between the last hidden layer and the output layer to convert the output into a probability distribution. The multi-modal fusion features of minerals are divided into a mineral feature training data set and a mineral feature validation data set. The random division method is adopted to randomly divide all fusion feature data into a training data set and a validation data set, where the training data set accounts for 80% of the total data, and the validation data set accounts for 20%. Ensure the randomness and representativeness of the data so that the model can learn a wider range of data features during the training process. The ore selection recognition model is trained based on the mineral feature training data set. In each training epoch, the training data set is divided into multiple batches, and each batch contains 64 samples. The samples are input into the model batch by batch for forward propagation and backward propagation to update the weight parameters of the model. The Adam optimizer is adopted, the learning rate is set to 0.001, the momentum parameter is 0.9, and the weight decay coefficient is 0.0001. The cross-entropy loss function is used as the loss function to measure the difference between the model output and the true label. The training process is set to 100 epochs, and each epoch makes a complete traversal of the training data set. The trained ore selection recognition model is verified based on the mineral feature validation data set. The performance metrics of the model are calculated, including accuracy, recall rate, and F1 score.The accuracy rate is obtained by calculating the ratio of the number of samples correctly predicted by the model to the total number of samples; the recall rate is obtained by calculating the ratio of the number of positive samples correctly predicted by the model to the total number of actual positive samples; the F1 score is obtained by calculating the harmonic mean of the accuracy rate and the recall rate. According to these performance indicators, the parameters of the model are optimized. The early stopping mechanism (Early Stopping) is adopted. When the accuracy rate on the validation set does not improve within 10 consecutive epochs, the training is stopped to avoid overfitting. At the same time, a learning rate decay strategy is adopted. During the training process, every 20 epochs, the learning rate decays to 0.1 times the original. Through these optimization measures, the optimized ore dressing recognition model is finally obtained.

[0052] Of particular importance, step S41 includes the following steps: Step S411: Normalize the mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features, and map all feature values to the interval from 0 to 1; Step S412: Concatenate each feature vector in sequence into a long vector to form a comprehensive mineral feature vector; Step S413: Use the principal component analysis (PCA) method to reduce the dimension of the comprehensive mineral feature vector and retain the first N principal components; Step S414: Standardize the feature vector after dimension reduction so that the mean of each feature is 0 and the standard deviation is 1; Step S415: Weight the standardized feature vector, and assign different weights to different features according to the preset weight allocation scheme; Step S416: Fuse the weighted feature vectors to generate multi-modal fusion features of minerals.

[0053] In the embodiment of the present invention, in step S411, normalization processing is performed on the mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features. Using the minimum-maximum normalization technique, the feature values in each feature vector are adjusted to the interval of 0 to 1. Specifically, for each feature vector, find the minimum and maximum values therein, and then perform a linear transformation on each feature value so that it varies between 0 and 1. After such processing, the feature values of different feature vectors have a unified numerical range, which is convenient for subsequent processing. In step S412, each normalized feature vector is concatenated in sequence into a long vector to form a comprehensive mineral feature vector. Specifically, the mineral point-line-plane defect feature vector, mineral grain shape feature vector, and mineral surface wear feature vector are arranged in sequence and merged into a longer feature vector. Assuming that the dimensions of the three feature vectors are 120, 80, and 50 respectively, then the dimension of the concatenated comprehensive mineral feature vector is 250. In step S413, the principal component analysis (PCA) method is used to perform dimensionality reduction processing on the comprehensive mineral feature vector. Specifically, first calculate the covariance matrix of the comprehensive mineral feature vector, and then perform eigenvalue decomposition on this matrix to obtain the eigenvalues and corresponding eigenvectors. Sort according to the magnitude of the eigenvalues, and select the first N principal components, where N is determined according to the cumulative variance contribution rate, and usually select the principal components with a cumulative variance contribution rate reaching more than 85%. Assuming that the first 50 principal components are selected, then the dimension of the feature vector after dimensionality reduction is 50. In step S414, normalization processing is performed on the feature vector after dimensionality reduction. Using the Z-score normalization technique, the mean of each feature is 0 and the standard deviation is 1. Specifically, for each feature value in the feature vector after dimensionality reduction, calculate its mean and standard deviation in the entire dataset, and then perform normalization processing on each feature value so that its mean is 0 and the standard deviation is 1. In step S415, weighted processing is performed on the normalized feature vector. According to the preset weight assignment scheme, different weights are assigned to different features. Specifically, the normalized feature vector is divided into three parts in the order of the original features, corresponding to the point-line-plane defect features, grain shape features, and surface wear features respectively. Then, according to the preset weight values, each part of the feature vector is multiplied by the corresponding weight value. Assuming that the weight of the mineral point-line-plane defect features is 0.4, the weight of the mineral grain shape features is 0.3, and the weight of the mineral surface wear features is 0.3, then each part of the feature vector is multiplied by the corresponding weight value. In step S416, the weighted feature vectors are fused to generate a multi-modal fusion feature of the mineral. Specifically, the three weighted feature vectors are concatenated in sequence into a final fusion feature vector. This fusion feature vector is the multi-modal fusion feature of the mineral, which combines the mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features.

[0054] Particularly importantly, step S47 includes the following steps: Step S471: Optimize the learning rate of the model parameters according to the F1 score of the model performance metrics. If the F1 score of the model is lower than the preset threshold, halve the learning rate; if the F1 score of the model is higher than the preset threshold but not optimal, increase the learning rate by 10%.

[0055] Step S472: Adjust the batch size of the model according to the accuracy of the model performance metrics. If the accuracy of the model is lower than the preset threshold, halve the batch size; if the accuracy of the model is higher than the preset threshold but not optimal, increase the batch size by 20%.

[0056] Step S473: Adjust the number of training rounds of the model according to the recall rate of the model performance metrics. If the recall rate of the model is lower than the preset threshold, increase the number of training rounds by 50%; if the recall rate of the model is higher than the preset threshold but not optimal, increase the number of training rounds by 20%.

[0057] Step S474: Use the adjusted learning rate, batch size, and number of training rounds to obtain a mineral separation recognition model and retrain the model.

[0058] In the embodiments of the present invention, the learning rate is optimized according to the F1 score of the model performance index. The preset threshold of the F1 score is set to 0.85. After each training epoch ends, the F1 score of the model is evaluated. If the F1 score of the model is lower than 0.85, the current learning rate is halved. For example, if the current learning rate is 0.001, it is adjusted to 0.0005. If the F1 score of the model is higher than 0.85 but does not reach the optimal value (assuming the optimal value is 0.95), the learning rate is increased by 10%. For example, if the current learning rate is 0.001, it is adjusted to 0.0011. The batch size is adjusted according to the accuracy of the model performance index. The preset threshold of the accuracy is set to 0.90. After each training epoch ends, the accuracy of the model is evaluated. If the accuracy of the model is lower than 0.90, the current batch size is halved. For example, if the current batch size is 64, it is adjusted to 32. If the accuracy of the model is higher than 0.90 but does not reach the optimal value (assuming the optimal value is 0.98), the batch size is increased by 20%. For example, if the current batch size is 64, it is adjusted to 76. The number of training epochs is adjusted according to the recall rate of the model performance index. The preset threshold of the recall rate is set to 0.80. After each training epoch ends, the recall rate of the model is evaluated. If the recall rate of the model is lower than 0.80, the current number of training epochs is increased by 50%. For example, if the current number of training epochs is 100, it is adjusted to 150. If the recall rate of the model is higher than 0.80 but does not reach the optimal value (assuming the optimal value is 0.95), the number of training epochs is increased by 20%. For example, if the current number of training epochs is 100, it is adjusted to 120. The ore separation recognition model is retrained using the adjusted learning rate, batch size, and number of training epochs. Specifically, the adjusted learning rate, batch size, and number of training epochs are used as new training parameters to restart the model training process. During the training process, the new learning rate is used for weight update, the new batch size is used for data batch division, and iterative training is performed according to the new number of training epochs to optimize the model performance, and finally the adjusted ore separation recognition model is obtained.

[0059] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the application documents are intended to be included in the present invention.

[0060] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A training method for a mineral separation recognition model based on multi-feature fusion, characterized in that It includes the following steps: Step S1: Collect mineral images; Identify the mineral crystal structure information of the mineral image, conduct geometric type division on the mineral crystal structure information to obtain the mineral crystal system type; detect the point-line-plane defect characteristics of the mineral image according to the mineral crystal system type to generate the mineral point-line-plane defect characteristics; Step S2: Conduct mineral crystal symbiosis recognition on the mineral image according to the mineral crystal system type, evaluate the mineral crystal symbiosis relationship to generate mineral crystal symbiosis data; conduct mineral grain connectivity analysis on the mineral crystal symbiosis data and detect the mineral porosity; determine the mineral grain shape characteristics based on the mineral porosity; Step S3: Identify the mineral morphology characteristics of the mineral image, determine the mineral luster area of the mineral morphology characteristics; conduct luster density detection on the mineral luster area to obtain the area luster density data; conduct color change gradient division on the mineral morphology characteristics according to the area luster density data and determine the mineral surface wear characteristics based on the color change gradient; Step S4: Conduct mineral multi-modal feature fusion on the mineral point-line-plane defect characteristics, mineral grain shape characteristics and mineral surface wear characteristics to generate mineral multi-modal fusion characteristics; construct a preset ore selection recognition model through deep learning technology, conduct model training on the preset ore selection recognition model according to the mineral multi-modal fusion characteristics, and verify the model performance indicators; optimize the model parameters according to the model performance indicators to obtain the ore selection recognition model.

2. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that In step S1, identifying the mineral crystal structure information of the mineral image and conducting geometric type division on the mineral crystal structure information includes: Scan the mineral image line by line, record the starting point and ending point coordinates of the crystal structure in the horizontal and vertical directions in the image to form a mineral crystal structure coordinate data set; Based on the mineral crystal structure coordinate data set, calculate the center point position of each crystal structure, take the center point as the benchmark, divide the crystal structure into four quadrant areas, and record the pixel point distribution in each quadrant; Calculate the crystal shape factor of the crystal structure in each quadrant area, where the shape factor is determined by the ratio of the perimeter to the area of the crystal structure, and generate the shape factor value of each quadrant; Divide the crystal structure into two categories: regular crystals and irregular crystals according to the shape factor value of each quadrant; For the structure of irregular crystals, calculate the curvature change value of its edge points; determine the degree of concavity and convexity according to the curvature change value; Conduct geometric type division on the mineral crystal structure information based on the degree of concavity and convexity to obtain the mineral crystal system type.

3. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, wherein In step S1, detecting the point-line-plane defect characteristics of the mineral image according to the mineral crystal system type includes: Conduct gradient calculation on the mineral image according to the mineral crystal system type, use the Sobel operator to calculate the gradients of the image in the horizontal and vertical directions respectively to generate a gradient magnitude image and a gradient direction image; Based on the gradient magnitude image, identify the local extreme points of the gradient magnitude, and take the local extreme points as potential point defect candidate points; calculate the ratio of the gradient magnitude of each candidate point to the gradient magnitudes of its surrounding 8 neighborhood pixels, and retain the candidate points with a ratio greater than 1.2 as the mineral crystal system defect points, and record the positions and gradient magnitudes of the mineral crystal system defect points; Starting from the upper left corner of the gradient direction image, track along the gradient direction until a point where the gradient direction changes by more than 45 degrees is encountered or the image boundary is reached; For each extracted line, calculate its length and curvature, where the curvature is determined by calculating the average of the angular changes between every two adjacent points on the line; Mark the lines with a length greater than 10 pixels and a curvature greater than 0.1 as defect lines of the mineral crystal system, and record the positions, lengths, and curvatures of the defect lines of the mineral crystal system; Perform multi-region segmentation on the gradient magnitude image, with each region corresponding to a part of the mineral crystal; Calculate the standard deviation of the gradient magnitudes of each region, and mark the regions with a standard deviation greater than 0.5 as defect surfaces of the mineral crystal system, and record the positions and areas of the defect surfaces of the mineral crystal system.

4. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that In step S2, mineral crystal symbiosis recognition is performed on the mineral image according to the mineral crystal system type, and the evaluation of the mineral crystal symbiosis relationship includes: Perform edge detection on the mineral image, extract the edge information of the mineral crystal, and generate a mineral edge image; Identify the contours of the mineral edge image according to the mineral crystal system type, where each contour corresponds to a mineral crystal, and obtain the mineral edge contour information; Calculate the geometric center of the mineral edge contour information to determine the geometric center position of each mineral crystal; Calculate the Euclidean distance between the geometric centers of each pair of mineral crystals, and identify the pairs of mineral crystals with a distance less than 10% of the image width as symbiotic relationships; For each pair of mineral crystals identified as having a symbiotic relationship, calculate the overlapping area ratio of their contours, where the overlapping area ratio is the ratio of the area of the overlapping region of the two contours to the total area of the two contours; Mark the pairs of mineral crystals with an overlapping area ratio greater than 0.1 as a close symbiotic relationship according to the overlapping area ratio; mark the pairs of mineral crystals with an overlapping area ratio between 0.05 and 0.1 as a medium symbiotic relationship; mark the pairs of mineral crystals with an overlapping area ratio less than 0.05 as a loose symbiotic relationship.

5. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that In step S2, mineral grain connectivity analysis is performed on the mineral crystal symbiosis data, and the mineral porosity is detected, including: Analyze the mineral crystal symbiosis data, and extract the geometric center positions of each pair of symbiotic mineral crystals; Based on the geometric center positions, construct a connectivity graph of the mineral crystals, where each mineral crystal is represented as a node, and the connectivity relationship between each pair of symbiotic mineral crystals is represented as an edge; Perform a depth-first search on the connectivity graph, starting from each unvisited node, and record the number of connected components and the size of each connected component; Segment the mineral image and mark the mineral background area; By counting the number of background pixels after segmentation and multiplying by the area of each pixel, the area of the mineral background area is obtained; the area of the mineral background area is determined as the mineral pore area; Calculate the ratio of the mineral pore area to the area of the entire mineral image to generate the mineral porosity.

6. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, wherein, In step S2, determining the mineral grain shape characteristics based on the mineral porosity includes: Perform a preliminary classification of the mineral grains according to the mineral porosity to obtain the mineral grain shape characteristics; If the porosity is less than 0.1, mark the grain as a low-porosity grain; If the porosity is between 0.1 and 0.3, mark the grain as a grain with medium porosity; If the porosity is greater than 0.3, mark the grain as a grain with high porosity.

7. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that In step S3, identify the mineral morphology characteristics of the mineral image, and the mineral luster regions determined by the mineral morphology characteristics include: Perform color space conversion on the mineral image and convert the image to the HSV color space; Divide the mineral image into multiple regions in the HSV color space, and calculate the brightness mean and brightness standard deviation of each region; Determine the mineral morphology characteristics according to the brightness mean and brightness standard deviation; If the brightness mean of a region is greater than 0.7 and the brightness standard deviation is less than 0.1, then the region is marked as a high-brightness region; if the brightness mean of a region is between 0.5 and 0.7 and the brightness standard deviation is between 0.1 and 0.3, then the region is marked as a medium-brightness region; if the brightness mean of a region is less than 0.5 and the brightness standard deviation is greater than 0.3, then the region is marked as a low-brightness region; Calculate the texture contrast of each region, and determine the mineral luster region according to the texture characteristics; For a high-brightness region, if the contrast is less than 0.2, then the region is marked as a high-luster region; for a medium-brightness region, if the contrast is between 0.2 and 0.5, then the region is marked as a medium-luster region; for a low-brightness region, if the contrast is greater than 0.5, then the region is marked as a low-luster region.

8. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that In step S3, the luster density detection of the mineral luster region includes: Extract the brightness mean and brightness standard deviation of the mineral luster region; For each luster region, divide the brightness mean by the brightness standard deviation to evaluate the brightness uniformity of the luster region; determine the luster density according to the brightness uniformity; Divide the luster density into the maximum luster density, the minimum luster density, and the average luster density; For each luster region, calculate the difference between the maximum luster density and the minimum luster density within the luster region and compare it with the average luster density to obtain the regional luster density data.

9. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that In step S3, divide the mineral morphology characteristics according to the regional luster density data for color change gradient, and determine the mineral surface wear characteristics based on the color change gradient, including: For each pixel of the mineral morphology characteristics according to the regional luster density data, calculate the difference between its brightness mean and the brightness means of the surrounding 8 neighboring pixels; For each pixel of the mineral morphology characteristics according to the regional luster density data, calculate the difference between its brightness standard deviation and the brightness standard deviations of the surrounding 8 neighboring pixels; Add the brightness mean difference and the brightness standard deviation difference to obtain the color change gradient value; For all pixels within each region, calculate the average value of their color change gradient values; For all pixels within each region, calculate the standard deviation of their color change gradient values; If the average value of the color change gradient is greater than 0.5 and the standard deviation is less than 0.1, then the region is marked as a high-wear region; If the average value of the color change gradient is between 0.3 and 0.5 and the standard deviation is between 0.1 and 0.3, then the region is marked as a medium-wear region; If the average value of the color change gradient is less than 0.3 and the standard deviation is greater than 0.3, then the region is marked as a low-wear region.

10. The training method of the ore dressing recognition model based on multi-feature fusion according to claim 1, characterized in that Step S4 includes the following steps: Step S41: Perform mineral multi-modal feature fusion on mineral point-line-plane defect features, mineral grain shape features, and mineral surface wear features to generate mineral multi-modal fusion features; Step S42: Select a deep learning framework to construct a preset ore separation recognition model; Step S43: Define the network structure of the ore separation recognition model, including an input layer, multiple hidden layers, and an output layer; among them, the dimension of the input layer is determined according to the data dimension of the mineral multi-modal fusion features, and the dimension of the output layer is determined according to the number of target categories of the ore separation recognition task; Step S44: Divide the mineral multi-modal fusion features into a mineral feature training data set and a mineral feature validation data set; Step S45: Train the ore separation recognition model based on the mineral feature training data set to obtain an ore separation recognition training model; in each training epoch, divide the training data set into multiple batches, and input them into the model batch by batch for forward propagation and backward propagation to update the weight parameters of the model; Step S46: Validate the ore separation recognition training model based on the mineral feature validation data set, and calculate the performance indicators of the model, including accuracy, recall rate, and F1 score, to obtain the model performance indicators; Step S47: Optimize the model parameters according to the model performance indicators to obtain the ore separation recognition model.

Images