Ore identification method and system based on large model algorithm

Abstract

The invention provides an ore recognition method and system based on a large model algorithm, and is applied to the technical field of computer vision and data processing, and the method comprises the steps: obtaining an ore microscopic image, and generating a feature point location candidate set through a convolutional neural network model; calculating a spectral value and a morphological parameter of each feature point location to determine a multi-dimensional data point; clustering the multi-dimensional data points to obtain a feature point cluster; and when the spatial distance between the adjacent feature point position clusters is smaller than a first preset threshold value and the spectral similarity is higher than a second preset threshold value, combining the adjacent feature point position clusters into the same mineral region. The segmentation precision and the type identification accuracy of the complex symbiotic mineral area can be obviously improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and data processing, and particularly relates to a mineral identification method and system based on a large model algorithm. BACKGROUND

[0002] In the industrial processes of geological exploration, mineral development and metallurgical beneficiation, the accurate identification and analysis of the mineral types, content, intergrowth relationship and spatial distribution in the ore are the links that determine the accuracy of resource assessment, the rationality of beneficiation scheme development and the final economic benefit.

[0003] The traditional mineral analysis method mainly relies on artificial microscopic identification and chemical phase analysis. The artificial microscopic identification method is low in efficiency, high in cost and strong in subjectivity. For ore samples with complex composition, small crystals and close intergrowth relationship, the consistency and accuracy of the identification results are difficult to guarantee. Although the chemical phase analysis can provide accurate element content, it cannot provide mineral spatial distribution information and crucial intergrowth structure information, and the intergrowth structure directly determines the selectivity and smelting property of the ore.

[0004] With the development of computer technology, some automatic mineral identification methods based on image processing have appeared. However, the existing methods mostly rely on single-dimensional visual features such as color, brightness or simple morphology of minerals. The same mineral may present significant differences in appearance under different ore-forming environments, weathering degrees or lighting conditions (i.e. "same object different spectrum"); while different types of minerals may also exhibit similar visual features (i.e. "different objects same spectrum"). This leads to insufficient accuracy and robustness of the recognition method based on single feature in the face of complex geological samples. In the complex scene of multiple minerals closely intergrown and interlocked, the existing algorithms are difficult to accurately define the physical boundaries between different mineral phases. The fuzziness of this boundary recognition directly leads to inaccurate calculation of mineral content, and even more cannot effectively quantify the intergrowth structure which determines the difficulty of ore processing technology.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] Therefore, the present application provides a mineral identification method and system based on a large model algorithm, aiming to solve the problems of low accuracy and poor robustness of mineral identification in the prior art. The method generates a feature point candidate set by processing a micrograph with a convolutional neural network model, calculates spectral values and morphological parameters for the feature points, groups the multi-dimensional data points using a clustering algorithm, and finally merges the feature point clusters based on spatial distance and spectral similarity, thereby improving the accuracy and robustness of mineral identification.

[0007] The embodiment of the present application provides a mineral identification method based on a large model algorithm, comprising: An image of a to-be-tested mineral sample is acquired, and a convolutional neural network model is used to process the image to generate a feature point candidate set; For each feature point in the feature point candidate set, a spectrum value and a morphological parameter are calculated to determine a multi-dimensional data point containing spatial coordinates, the spectrum value and the morphological parameter of each feature point; A clustering algorithm is used to group the multi-dimensional data points to obtain a feature point cluster; The spatial distance between any two adjacent feature point clusters in space is calculated, and the spectrum similarity of the two is evaluated; When the spatial distance is less than a first preset threshold and the spectrum similarity is higher than a second preset threshold, the two adjacent feature point clusters are merged into a mineral region.

[0008] In some optional embodiments, acquiring the image of the to-be-tested mineral sample comprises: collecting the mineral sample by a hyperspectral microscopic imaging system to obtain hyperspectral image data containing a spectrum curve on each spatial pixel point.

[0009] In some optional embodiments, calculating the spectrum value of each feature point comprises: extracting the spectrum curve of the feature point in the hyperspectral image data, and performing dimension reduction processing on the spectrum curve to generate a spectrum feature vector, and the spectrum feature vector is the spectrum value.

[0010] In some optional embodiments, the dimension reduction processing on the spectrum curve comprises: extracting core feature components of the spectrum curve by using a principal component analysis method or a wavelet transform method.

[0011] In some optional embodiments, calculating the morphological parameter of each feature point comprises: calculating the Euclidean distance and the spatial angle between the feature point and one or more adjacent feature points of the feature point.

[0012] In some optional embodiments, the clustering algorithm is a K-means clustering algorithm, and before the K-means clustering algorithm is used for grouping, the method further comprises: automatically determining the optimal clustering number K required by the K-means clustering algorithm by performing contour coefficient analysis or elbow rule analysis on the multi-dimensional data points.

[0013] In some optional embodiments, the first preset threshold and the second preset threshold are dynamic thresholds, and before the merging, the method further comprises: based on the dispersion degree of the intra-cluster data point distribution of the feature point cluster and the inter-cluster distance distribution, calculating to generate a dynamic threshold as the first preset threshold and a dynamic threshold as the second preset threshold.

[0014] In some optional embodiments, after being merged into the mineral region, further comprising: further extracting a texture correlation parameter and a color transition parameter of the mineral region, and determining the specific mineral species of the mineral region based on the texture correlation parameter and the color transition parameter.

[0015] In some optional embodiments, after being merged into the mineral region, further comprising: when it is identified that at least two different mineral regions coexist, performing image analysis on the contact boundary between the different mineral regions to quantify the physical intergrowth relationship between the different mineral regions.

[0016] In some optional embodiments, the quantification of the physical intergrowth relationship comprises: calculating the length proportion of the boundary of one mineral region being surrounded by the boundary of another mineral region to obtain an inclusion degree parameter.

[0017] In some optional embodiments, the quantification of the physical intergrowth relationship comprises: calculating the curve complexity of the contact boundary to obtain a parameter representing the complexity degree of the intergrowth morphology.

[0018] In some optional embodiments, the calculation of the curve complexity of the contact boundary comprises: calculating the fractal dimension of the contact boundary.

[0019] In some optional embodiments, after the quantification of the physical intergrowth relationship, further comprising: inputting the quantified physical intergrowth relationship parameter and the area proportion data of each mineral region into a resource assessment model to output a resource assessment level of the ore sample.

[0020] In some optional embodiments, the resource assessment model is a machine learning model trained by physical intergrowth relationship parameters, area proportion data and corresponding known metallurgical experimental results of historical ore samples.

[0021] In some optional embodiments, the processing of the micro image by the convolutional neural network model to generate the candidate set of feature point positions comprises: extracting a multi-scale texture feature map of the micro image by the convolutional neural network model; processing the multi-scale texture feature map by a Harris corner point detection algorithm to identify corner points with a corner point response value higher than a preset response threshold value; and performing non-maximum suppression processing on the corner points to generate the candidate set of feature point positions.

[0022] Embodiments of the present application provide a mineral identification system based on a large model algorithm, comprising: a data processing module configured to acquire a micro image of an ore sample to be tested, and process the micro image by a convolutional neural network model to generate a candidate set of feature point positions; The feature calculation module is configured to calculate the spectral value and the morphological parameter of each feature point in the feature point candidate set, so as to determine a multi-dimensional data point containing the spatial coordinates, the spectral value and the morphological parameter of each feature point. The clustering module is configured to group the multi-dimensional data points by using a clustering algorithm, so as to obtain a feature point cluster. The region merging module is configured to calculate the spatial distance between any two adjacent feature point clusters in space, evaluate the spectral similarity of the two feature point clusters, and merge the two adjacent feature point clusters into a mineral region when the spatial distance is less than a first preset threshold and the spectral similarity is higher than a second preset threshold.

[0023] In some optional embodiments, the data processing module is specifically configured to: acquire the ore sample by using the hyperspectral microscopic imaging system, so as to obtain hyperspectral image data containing a spectral curve at each spatial pixel point; and the feature calculation module is specifically configured to: extract the spectral curve of the feature point in the hyperspectral image data, and perform dimension reduction processing on the spectral curve to generate a spectral feature vector, the spectral feature vector being the spectral value.

[0024] In some optional embodiments, the clustering module is further configured to: automatically determine the optimal clustering number required by the clustering algorithm by performing contour coefficient analysis or elbow rule analysis on the multi-dimensional data points before grouping by using the clustering algorithm.

[0025] In some optional embodiments, the method further comprises: The intergrowth structure analysis module is configured to perform image analysis on the contact boundary between different mineral regions when the region merging module merges to generate at least two different mineral regions, so as to quantify the physical intergrowth relationship between different mineral regions.

[0026] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present application.

[0027] The ore recognition method and system based on the large model algorithm have the following beneficial effects: The method fuses the visual features and the spectral features of the ore, constructs a high-dimensional feature description, and improves the recognition accuracy of complex ore samples. The region merging strategy based on the dual physical law constraint combines mathematical clustering and physical world laws, can accurately reconstruct the mineral crystal boundary, and realizes accurate segmentation of the mineral region. The present application also provides computable and quantifiable physical intergrowth relationship parameters, and realizes quantitative analysis of the intergrowth structure. Therefore, the method improves the accuracy and automation of ore recognition, and provides technical support for resource evaluation and beneficiation decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0028] Other features, objects, and advantages of the application will become more apparent from a detailed consideration of the following non-limiting embodiments when taken in conjunction with the drawings.

[0029] Figure 1 is a flow chart of a large model algorithm-based ore identification method according to an embodiment of the present application; Figure 2 is a structural schematic diagram of a large model algorithm-based ore identification system according to an embodiment of the present application. DETAILED DESCRIPTION

[0030] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the example implementations to those skilled in the art. The described features, structures, or characteristics can be combined in one or more implementations.

[0031] In addition, the drawings are only schematic and the dimensions of the various layers are not necessarily to scale with respect to one another. Like reference numerals designate corresponding parts throughout the several views. Some of the blocks in the drawings represent functional entities that can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0032] The flow diagrams depicted in the drawings are merely exemplary and do not necessarily include all of the steps that can be implemented by the actual implementation. For example, one or more of the steps can be combined or partially combined, and some steps can be omitted, altered, or rearranged, depending on the actual implementation. Thus, the actual implementation can vary from the depiction in the drawings.

[0033] The K-means clustering algorithm divides data into different clusters based on the distance between data points, and the idea is to make the data points in the cluster as similar as possible and the data points between the clusters as different as possible. In the ore recognition method, the algorithm is used to preliminarily distinguish feature points with different spectral and morphological characteristics. Spatial distance provides a measure of the proximity of feature points in physical space, while spectral similarity quantifies the degree of similarity of feature points in physical properties. Ore is a collection of multiple minerals, each mineral has a certain spatial continuity and specific spectral characteristics at the microscopic scale. Therefore, by setting the threshold of spatial distance and spectral similarity, the feature point clusters that are adjacent in space and similar in spectral characteristics can be merged, thereby effectively segmenting different mineral regions in the ore and restoring the distribution of minerals in the real physical space as much as possible. Such processing can improve the accuracy of mineral identification and lay the foundation for subsequent mineral species judgment and dissemination structure analysis.

[0034] As shown in Figure 1 The embodiment of the present application provides an ore recognition method based on a large model algorithm, which realizes automatic recognition of mineral composition in an ore sample through multi-stage processing, including the following steps: Step S100: Obtain a microscopic image of the ore sample to be measured, and process the microscopic image using a convolutional neural network model to generate a feature point candidate set.

[0035] In one example, a polarized microscope image of the ore sample to be measured is obtained using an imaging system combining an optical microscope and an image sensor. The image sensor can be a CCD (Charge-Coupled Device) or CMOS (Complementary Metal Oxide Semiconductor) image sensor. Then, the obtained microscopic image is input into a pre-trained convolutional neural network model. The convolutional neural network model is, for example, a VGG16 network, a ResNet network, or a DenseNet network. The convolutional neural network model is configured to extract features such as edges, corner points, and textures in the image. After convolution, pooling, and activation operations, the feature map output by the model is used to generate a feature point candidate set. The way to generate the feature point candidate set can be to set a threshold to filter high response points in the feature map, or to use a non-maximum suppression algorithm to select local maximum points. These high response points or local maximum points are used as the feature point candidate set for subsequent analysis.

[0036] Step S200: For each feature point in the feature point candidate set, calculate its spectral value and morphological parameter to determine a multi-dimensional data point containing the spatial coordinates, spectral value, and morphological parameter of each feature point.

[0037] For the feature point candidate set generated in step S100, the spectral value and the morphological parameter of each feature point are calculated. The calculation of the spectral value can be realized in the following way: if the micrograph is a hyperspectral image, the spectral data of the pixel corresponding to the feature point is directly extracted as the spectral value. If the micrograph is an RGB image, the RGB value is converted into the Lab value of the CIELab color space, and the Lab value is taken as the approximation of the spectral value. The calculation of the morphological parameter can include: calculating the Histogram of Oriented Gradients (HOG) feature in the neighborhood around the feature point, or calculating the Local Binary Pattern (LBP) feature. The spatial coordinates can be directly obtained from the micrograph. The spatial coordinates, the spectral value and the morphological parameter of each feature point are combined into a multi-dimensional vector, and the vector is a multi-dimensional data point containing the information of the feature point.

[0038] Step S300: A clustering algorithm is used to group the multi-dimensional data points to obtain feature point clusters.

[0039] All multi-dimensional data points generated in step S200 are taken as input, and a clustering algorithm is used for grouping. The clustering algorithm can be a K-means clustering algorithm, a DBSCAN clustering algorithm or a Gaussian mixture model clustering algorithm. Taking the K-means clustering algorithm as an example, K data points are first randomly selected as initial cluster centers, then the distance of each data point to each cluster center is calculated, and the data point is divided into the cluster where the nearest cluster center is located. Next, the cluster center of each cluster is recalculated, and the above process is repeated until the cluster center no longer changes significantly or the maximum iteration number is reached. The clustering result is a plurality of feature point clusters, and each cluster contains feature points with similar features.

[0040] Step S400: The spatial distance between any two spatially adjacent feature point clusters is calculated, and the spectral similarity of the two is evaluated.

[0041] For the feature point clusters obtained in step S300, the spatial distance and the spectral similarity between any two spatially adjacent clusters are calculated. The judgment of spatial adjacency can be: if the distance between the centroids of two clusters is less than a preset spatial distance threshold, it is considered that the two clusters are spatially adjacent. The centroid of the cluster can be obtained by calculating the average value of the spatial coordinates of all data points in the cluster. The spatial distance can be calculated using the Euclidean distance or the Manhattan distance. The evaluation of the spectral similarity can be realized in the following way: the average value of the spectral values of the two clusters is calculated to obtain the average spectral vector of the two clusters. Then, the cosine similarity, the Euclidean distance or the correlation coefficient between the two average spectral vectors are calculated. These values can be used as a measure of the spectral similarity.

[0042] Step S500: When the spatial distance is less than the first preset threshold and the spectral similarity is higher than the second preset threshold, merging the two adjacent feature point clusters into a mineral region.

[0043] According to the spatial distance and the spectral similarity calculated in step S400, it is determined whether the adjacent feature point clusters need to be merged. If the spatial distance between two clusters is less than the first preset threshold and the spectral similarity is higher than the second preset threshold, the two clusters are merged into a new cluster, and the new cluster represents a mineral region. The first preset threshold and the second preset threshold can be adjusted according to the actual application scene and data characteristics. For example, a fixed value can be set according to experience, or it can be automatically determined through statistical analysis.

[0044] The above steps S100 to S500 cooperate with each other. First, the feature points of the microscopic image are extracted by the convolutional neural network. Then, the spectral values and morphological parameters of the feature point positions are calculated. Next, the feature point positions are grouped into feature point clusters by the clustering algorithm. Finally, the adjacent feature point clusters are merged into a mineral region based on the spatial distance and the spectral similarity. Through this multi-stage processing, the automatic identification of the mineral composition in the ore sample is finally realized.

[0045] The method provided in the embodiment solves the problems of low recognition accuracy and poor robustness in the prior art, and realizes the automatic identification of the mineral composition in the ore sample by fusing various technologies such as image processing, spectral analysis and clustering algorithm, and has the advantages of high recognition accuracy and high automation degree.

[0046] In a specific implementation, on the basis of the above embodiment, the process of obtaining the microscopic image of the ore sample to be measured is as follows: first, a hyperspectral microscopic imaging system with a model of GaiaField-V10 is used to scan the ore sample to be measured. The system integrates a visible-near-infrared hyperspectral imaging module and an optical microscope, and can simultaneously obtain the microscopic image and the hyperspectral data of the sample. Specifically, an objective lens with a focal length of 20 mm is used, the magnification is 10 times, and the spatial resolution is set to 5 microns / pixel. The illumination light source is a halogen lamp, and a 45° incident angle optical fiber illuminator is configured to reduce mirror reflection.

[0047] Specifically, after the sample is placed on the stage, the automatic focusing function is used to accurately focus on the surface of the sample. Then, the hyperspectral scanning program is started, the scanning range is set to cover the entire sample area, and the scanning speed is 1 millimeter / second. The hyperspectral imaging module uses a push-scan scanning method to scan the sample line by line, and records the continuous spectral data in the wavelength range of 400 nm to 1000 nm at each pixel point, and the spectral resolution is 2.5 nm. During the data acquisition process, the system automatically performs dark current correction and spectral calibration.

[0048] Then, the system outputs a hyperspectral image datacube containing the spectral curve of each spatial pixel, which is in ENVI standard format. The dimension of the datacube is XYl, where X and Y represent the width and height (in pixels) of the image, respectively, and l represents the number of spectral channels (determined by the spectral resolution and wavelength range). Each pixel (x, y) corresponds to a spectral curve S(x, y, l) containing l spectral values. The data obtained in this way contains the microscopic morphological features and fine spectral information of the ore sample.

[0049] Alternatively, the hyperspectral microscopic imaging system described above can also use a planar array scanning method based on an acousto-optic tunable filter (AOTF) or a liquid crystal tunable filter (LCTF), which can obtain the hyperspectral image of the entire field of view at one time without line-by-line scanning. In this case, the data acquisition speed can be further improved.

[0050] In addition, a polarizing microscope equipped with a polarizing device can be used to obtain a polarized microscopic image of the sample before hyperspectral scanning. The polarized image can be used as auxiliary information and fused with the hyperspectral data to further improve the identification accuracy of complex minerals (such as minerals with optical anisotropy).

[0051] The present embodiment obtains hyperspectral image data containing spectral curves of the ore sample at each spatial pixel by using a hyperspectral microscopic imaging system, thereby enabling more accurate analysis of the composition and structure of the ore.

[0052] In a specific embodiment, on the basis of the above-described embodiment, for each feature point, the complete spectral curve S(l) corresponding to it in the hyperspectral image data is extracted. First, the spectral curve is preprocessed, including removing noise (for example, smoothing by Savitzky-Golay filtering) and correcting baseline drift (for example, estimating and subtracting the baseline using a polynomial fitting method). Specifically, a Savitzky-Golay filter with a window size of 11 and a polynomial order of 2 can be used to smooth the spectral curve, and a 3rdorder polynomial fitting can be used to estimate the baseline, which is then subtracted from the original spectrum.

[0053] Then, the preprocessed spectral curve is reduced in dimension using discrete wavelet transform (DWT). DWT can be implemented using a variety of wavelet basis functions, such as Haar wavelet, Daubechies wavelet, Symlets wavelet, or Coiflets wavelet. In one example, Daubechies 4 (db4) wavelet is chosen as the basis function, and the spectral curve is decomposed into detail coefficients (D1, D2, D3) and an approximate coefficient (A3) at 3 scales. Next, the energy of the detail coefficients and the approximate coefficient at each scale is calculated, and the energy can be calculated as the square root of the sum of the squares of the coefficients. Specifically, the energy at the i-th scale is calculated as where k runs through all wavelet coefficients at the scale. Then, the detail coefficient energy (Ed) and the approximation coefficient energy (Ea) at all scales are combined into a 4-dimensional spectral feature vector As an equivalent alternative, other wavelet basis functions or decomposition scales can be chosen, and the wavelet coefficients can be directly used as the spectral feature vector without calculating the energy values.

[0054] The embodiment extracts the energy features of the spectral curves by wavelet transform, effectively reduces the dimensionality of the spectral data, and retains the information that distinguishes different mineral spectra, thereby providing an efficient feature representation for subsequent clustering and recognition.

[0055] In a specific implementation, on the basis of the above embodiment, the process of dimensionality reduction of the extracted spectral curve specifically includes the following scheme: First, the principal component analysis (PCA) method is applied to the spectral curve S(λ) extracted from the feature points. Specifically, the covariance matrix of all feature point spectral curves is calculated, and eigenvalue decomposition is performed on it. The eigenvectors corresponding to the top N largest eigenvalues are selected, and these eigenvectors form a transformation matrix. The original spectral curve is projected into a low-dimensional space spanned by these eigenvectors to obtain the dimension-reduced spectral feature vector The selection of N can be based on the cumulative contribution rate, for example, the minimum N value that makes the cumulative contribution rate reach 95%. In an alternative, kernel principal component analysis (KPCA) can be used to map the spectral data to a high-dimensional space through a kernel function, and then perform principal component analysis to extract nonlinear features. Common kernel functions include polynomial kernel function, radial basis function (RBF) kernel function and Sigmoid kernel function.

[0056] Alternatively, as another option, the wavelet transform method is applied to the spectral curve S(λ). Specifically, a suitable wavelet basis function is selected, such as Daubechies wavelet (db4, db6, etc.), Symlets wavelet or Coiflets wavelet. The spectral curve is decomposed at multiple scales to obtain a series of approximation coefficients and detail coefficients at different scales. These coefficients reflect the feature information of the spectral curve at different frequencies. The low-frequency approximation coefficients are extracted as the dimension-reduced spectral feature vector In another implementation, some high-frequency detail coefficients can be selected to enhance the expression ability of local features of the spectral curve. In addition, the wavelet transform can be replaced by wavelet packet transform, which has finer frequency band division and can better adapt to the characteristics of different types of spectral data.

[0057] ​​​​Using principal component analysis or wavelet transform methods can effectively reduce the dimensionality of spectral data, extract the feature components of spectral curves, reduce computational load, and improve the speed of subsequent clustering and recognition.

[0058] In one specific implementation, based on the above embodiments, in the above steps, for each feature point... The process of calculating its morphological parameters is as follows: First, in the candidate set P of feature points, the search distance is calculated using Euclidean distance as the standard. The N nearest neighbors. N can be set to 3, 5, 8, or 10, depending on the required level of detail in describing the local morphology. An alternative method to accelerate the search is to use a KD-tree or Ball tree data structure to improve the efficiency of the nearest neighbor search and reduce the search complexity from... Reduce to .

[0059] Specifically, for each neighboring point ,calculate and Euclidean distance between ,in and They are and The normalized image coordinates. Simultaneously, calculate... Compared to spatial azimuth The range of values ​​for the azimuth angle is To enhance the robustness of the angle, the azimuth angle can be discretized, for example, quantized into 8 or 16 directions. Then, the calculated N distance values... Normalize the values, for example, by dividing by the maximum value of all distances, so that their values ​​fall within the range of [0, 1]. Similarly, normalize the azimuth angles. Normalization is also performed, for example, by dividing by π. Finally, the normalized distance value and azimuth angle are concatenated into a morphological feature vector. This vector represents the morphological parameters of the feature point.

[0060] In another implementation, the calculation of the morphological parameter can further include calculating a Local Binary Pattern (LBP) histogram formed between the feature point and its neighboring points. Specifically, taking the feature point as the center, the gray values of the pixels in the neighborhood of the feature point are compared with the gray value of the center pixel. If the gray value of the pixel in the neighborhood is greater than or equal to the gray value of the center pixel, the pixel is marked as 1, otherwise, the pixel is marked as 0, thereby obtaining a binary code, which is taken as the LBP value of the center pixel. Then, the frequency histogram of the LBP values of all pixels in the neighborhood of the feature point is calculated as the morphological parameter of the feature point. The radius size of the neighborhood and the number of pixels in the neighborhood can be adjusted according to actual needs.

[0061] Then, when constructing the multi-dimensional feature vector, the morphological feature vector is spliced with other feature vectors to form a high-dimensional data point for subsequent clustering analysis.

[0062] By calculating the Euclidean distance and spatial angle between the feature point and its neighboring points, or by calculating the LBP histogram, the local morphological structure information around the feature point can be effectively described, thereby improving the accuracy and robustness of subsequent mineral identification.

[0063] In a specific embodiment, on the basis of the above embodiment, the clustering algorithm adopts a K-means clustering algorithm. Specifically, before grouping the multi-dimensional data point set by using the K-means clustering algorithm, the value of the number of clusters K needs to be determined.

[0064] In an example, the contour coefficient analysis method is used to adaptively determine the value of K. First, in a preset K value range, for example, , , , for each candidate K value , the K-means clustering algorithm is run to obtain a clustering result . Then, for each data point , the contour coefficient is calculated: wherein is the average distance from the data point to other points in the cluster to which the data point belongs, is the minimum value of the average distance from the data point to all points in other clusters. Then, the average contour coefficient of all data points is calculated as the clustering effect evaluation index corresponding to the current K value . The K value that makes the average contour coefficient maximum is selected as the number of clusters . ​​

[0065] In another example, the elbow rule analysis method is used to adaptively determine the K value. First, within a preset range of K values... Within, for each candidate K value Run the K-means clustering algorithm to obtain the clustering results. Then, calculate the total sum of squares within all clusters ( The sum of squares within the cluster (within-Cluster Sum of Squares) is used as the clustering loss corresponding to the current K value. Specifically, the formula for calculating the sum of squares within the cluster is: in, This is the current number of clusters. It is the k-th cluster. yes One of the data points, yes The center of mass, This represents the square of the Euclidean distance. Plot the K value against... Observe the shape of the curve in the graph. The curve usually resembles an arm, and the K value corresponding to the elbow (the point where the slope of the curve decreases significantly) is considered to represent the number of clusters. To automatically locate the elbow, calculations can be performed. The second derivative of the curve is used to determine the elbow by selecting the point with the largest absolute value of the second derivative.

[0066] Determine the number of clusters Afterwards, As input parameters to the K-means clustering algorithm, the multidimensional data point set is grouped to obtain the final feature point clusters.

[0067] This embodiment automatically determines the number of clusters K for the K-means clustering algorithm through contour coefficient analysis or elbow rule analysis, avoiding the subjectivity of manually setting the K value, improving the accuracy and reliability of the clustering results, and thus enhancing the overall performance of ore identification.

[0068] In one specific implementation, based on the above embodiments, the first preset threshold and the second preset threshold are dynamically adjusted values, rather than fixed values. First, for each feature point cluster obtained through the clustering algorithm, the average distance from all data points within the cluster to the centroid of the cluster is calculated as a measure of the cluster's intra-cluster dispersion. Euclidean distance, Manhattan distance, or Chebyshev distance can be used to calculate the distance.

[0069] Then, the average distance between all feature point clusters is calculated. Specifically, for any two clusters, the distance between their centroids is calculated. Next, the average distance between all these clusters is calculated.

[0070] Then, based on the calculated intra-cluster dispersion and inter-cluster distance average, a first preset threshold (spatial distance threshold) is dynamically set. In one example, the first preset threshold is set to be a weighted coefficient times the inter-cluster distance average, and the weighted coefficient ranges from 0.5 to 1.5. In another example, the first preset threshold is set to be a weighted coefficient times the average of the intra-cluster dispersions of all clusters, and the weighted coefficient ranges from 1.0 to 2.0. The first preset threshold can also be set to be a certain proportional coefficient times the sum of the inter-cluster distance average and the average of the intra-cluster dispersions.

[0071] Specifically, for each cluster, the variance of the spectral values of the data points in the cluster is calculated. Then, the average spectral variance of all clusters is calculated as a measure of spectral dispersion. For any two adjacent clusters, the cosine similarity between their average spectral vectors is calculated, and the average cosine similarity of all adjacent cluster pairs is calculated.

[0072] Then, based on the calculated spectral dispersion and average cosine similarity, a second preset threshold (spectral similarity threshold) is dynamically set. In one example, the second preset threshold is set to be a weighted coefficient times the average cosine similarity, and the weighted coefficient ranges from 0.8 to 1.0. In another example, the second preset threshold is set to be a weighted coefficient times the value of 1 minus the average spectral variance, and the weighted coefficient ranges from 0.7 to 0.9. The second preset threshold can also be set to be a certain proportional coefficient times the sum of the average cosine similarity and (1 minus the average spectral variance).

[0073] It is also feasible to use other equivalent statistical quantities (such as standard deviation, interquartile range) to measure the dispersion degree of data. The calculation method of the dynamic threshold can be replaced by other linear or nonlinear combination methods, such as using a neural network to fit the relationship between the threshold and the dispersion and similarity.

[0074] By dynamically setting the spatial distance threshold and the spectral similarity threshold, the method can automatically adjust the merging standard according to the specific characteristics of different ore samples, thereby enhancing the adaptability to different types of ore samples. By combining the intra-cluster dispersion degree and the inter-cluster distance information, the real relationship between clusters can be more accurately evaluated, unreasonable merging can be avoided, and the accuracy of mineral identification can be improved.

[0075] In a specific implementation, on the basis of the above examples, after completing the merging of the mineral regions, first, finer texture correlation parameters and color transition parameters are extracted from each merged mineral region. Specifically: 1. Texture correlation parameter extraction: For each mineral region, its corresponding pixel set in the original microscopic image is cropped out as the Region of Interest (ROI) of the region.

[0076] The ROI image is processed to be grayscale.

[0077] Texture features are extracted using Gray-Level Co-occurrence Matrix (GLCM) analysis. GLCM reflects the texture information of an image by calculating the co-occurrence probability of gray values in different directions and distances. In one implementation example, GLCMs are calculated for distances of 1, 2, and 3 pixels in four directions of 0°, 45°, 90°, and 135°, respectively.

[0078] Based on GLCM, the following texture statistics are calculated: Energy: measures the uniformity of the image gray distribution, the higher the energy value, the more regular the image texture. The calculation formula is: where is the element value of the (i, j) position in the GLCM.

[0079] Contrast: measures the local gray variation of the image, the higher the contrast, the rougher the image texture. The calculation formula is: .

[0080] Correlation: measures the similarity of the image texture, the higher the correlation, the more regular the image texture. The calculation formula is: where and are the mean values of the rows and columns in the GLCM, and are the standard deviations of the rows and columns.

[0081] Homogeneity: measures the uniformity of the image texture, the higher the homogeneity, the finer the image texture. The calculation formula is: .

[0082] The texture statistics calculated in different directions and distances are averaged to obtain the final texture feature vector.

[0083] As an alternative, Gabor filters can be employed to extract texture features. Gabor filters are linear filters that can extract texture information at different orientations and frequencies. Specifically, a Gabor filter bank with different scales and orientations can be used to convolve the ROI image, and then the energy and mean of the convolution results can be calculated as texture features.

[0084] 2. Color transition parameter extraction: The ROI image is converted from RGB color space to HSV color space. HSV color space decomposes color into hue (H), saturation (S), and value (V), which is more consistent with human visual perception characteristics.

[0085] The gradient histogram of each of the H, S, and V channels is calculated. The gradient histogram reflects the intensity of color change in the image. Specifically, the Sobel operator can be used to calculate the gradient value and direction of each pixel, and then the number of pixels in different gradient directions is counted to form the gradient histogram.

[0086] The mean, variance, and skewness of the gradient histogram are calculated as color transition parameters.

[0087] In another implementation, color gradients can also be calculated directly in the RGB color space. For example, the first and second derivatives of the image in the R, G, and B channels are calculated, and then the mean and variance of these derivatives are counted as color transition parameters.

[0088] 3. Mineral species determination: The texture feature vector and color transition parameters extracted above are concatenated to form a comprehensive feature vector.

[0089] The comprehensive feature vector is input into a pre-trained classifier model to determine the specific mineral species of the mineral region. The classifier model can be a support vector machine (SVM), a random forest, or a deep neural network (DNN).

[0090] The training data set of the classifier model contains a large number of texture features, color transition parameters, and corresponding mineral species labels of known mineral samples. Through training, the model learns the mapping relationship from texture and color features to mineral species.

[0091] The classifier outputs the mineral species and its confidence.

[0092] Through the above steps, the merged mineral region can be more accurately judged for mineral species based on the texture correlation parameters and color transition parameters.

[0093] The present embodiment realizes more accurate identification of mineral species by extracting the texture and color gradient features of the mineral region and combining the classifier model, thereby improving the accuracy and reliability of ore identification.

[0094] In a specific implementation, on the basis of the above embodiment, after the identification and segmentation of the mineral region are completed, first, the number of identified mineral regions is checked. If the number of identified mineral species is more than one, subsequent physical intergrowth relationship quantification analysis is performed. Specifically, image analysis is performed on the contact boundaries between different mineral regions.

[0095] To accurately extract the contact boundaries, the following methods can be used: first, convert the mineral region map into a binary image, where the pixel value of one mineral region is 0 and the pixel value of the other mineral region is 255. Then, use the Canny edge detection algorithm to process the binary image to extract clear edge pixel points. The high and low thresholds in the Canny algorithm can be adaptively adjusted according to the gradient intensity of the image, for example, the median of the gradient amplitude can be used as the low threshold, and the high threshold can be set to 2 to 3 times the low threshold. After edge detection, the edge pixels are thinned, for example, using the Zhang-Suen thinning algorithm, to ensure the single-pixel width of the edge.

[0096] Next, the extracted contact boundaries are vectorized. Use a contour tracing algorithm, such as the Moore neighborhood tracing algorithm, to connect consecutive edge pixel points into a vectorized contour line. Perform polygon approximation on the contour line, for example, using the Douglas-Peucker algorithm, to simplify the representation of the contour line and reduce the complexity of subsequent calculations. The threshold for polygon approximation can be set to 0.5 to 1 pixel.

[0097] Then, based on the vectorized contact boundary data, the physical interlocking relationship is quantified. For example, the degree to which a specified mineral grain is surrounded by another mineral is calculated. The specific method is that, for each mineral grain, the length of its contact boundary with the adjacent mineral is calculated, as well as the total boundary length of the grain. The ratio of the two defines the degree of wrapping. A minimum grain size threshold, for example 10 pixels, can be set to exclude too small, noise-induced grains. Alternatively, the roughness of the contact boundary can be calculated, which can be quantified using the fractal dimension. The fractal dimension is calculated using the box counting method, that is, covering the contact boundary with boxes of different sizes, counting the number of boxes needed to cover the boundary, and then calculating the fractal dimension according to the relationship between the box size and the number of boxes. The box size can be chosen as a power of 2, for example 2, 4, 8, 16 pixels.

[0098] In the above embodiments, the edge detection algorithm is not limited to the Canny edge detection algorithm, and the Sobel operator, Prewitt operator or Laplacian operator can also be used. The contour tracking algorithm is also not limited to the Moore neighborhood tracking algorithm, and the Suzuki85 algorithm or the split-and-merge algorithm of Theo Pavlidis can also be used.

[0099] Through the present embodiment, the spatial relationship between minerals can be automatically and quantitatively analyzed, and the subjectivity and inefficiency of traditional manual analysis are eliminated, providing a more reliable data basis for ore dressing and smelting evaluation.

[0100] In a specific embodiment, on the basis of the above embodiments, the image analysis process of the contact boundary of the mineral region comprises the following steps: first, for the identified symbiotic useful mineral A and gangue mineral B, a binary image containing only the two minerals is extracted from the mineral region map. The region pixels of the useful mineral A are assigned a value of 1, and the region pixels of the gangue mineral B are assigned a value of 0. Specifically, the image segmentation method based on color range can be used to convert the mineral region map into a binary image according to the color ranges of A and B calibrated in advance. Then, morphological closing operation is applied to the binary image to fill the small holes and cracks inside the mineral region, ensuring the integrity of the region. The structure element of the closing operation can be a circular structure with a radius of 3 pixels. Next, the Canny edge detection algorithm is used to extract the contact boundary between A and B in the binary image. The gradient threshold value of the Canny algorithm is set to a low threshold value of 30 and a high threshold value of 100.

[0101] The specific method for calculating the wrapping degree parameter is: performing connected component analysis on the binary image to label all the A mineral particle regions connected to each other. For each A mineral particle, it is detected whether it is completely wrapped or partially wrapped by the B mineral. The specific strategy for judging the wrapping state is: first, calculate the boundary pixel set of the A mineral particle. Then, investigate the number of pixel points adjacent to the B mineral (pixel value is 0) in the boundary pixel set. If all the boundary pixels of an A mineral particle are adjacent to the B mineral, it is determined that the particle is completely wrapped. If only part of the boundary pixels are adjacent to the B mineral, it is determined that the particle is partially wrapped. The wrapping degree of each wrapped A mineral particle is calculated as follows: wherein is the boundary length of the particle contacting the B mineral, is the complete boundary length of the particle. and The calculation of the wrapping degree of each wrapped A mineral particle can be realized by counting the number of contact boundary pixels. In addition, The calculation of the wrapping degree of each wrapped A mineral particle can be realized by counting the number of contact boundary pixels. In addition,

[0102] In this embodiment, an alternative to connected component analysis is to use a watershed algorithm to segment mineral particles. The calculation of the boundary length can be replaced by using a chain code tracking method to accurately calculate the number of boundary pixels. In addition, in addition to counting the number of pixels, the length information of the boundary can also be directly obtained using the contour finding function provided by the image processing library (such as OpenCV).

[0103] By calculating the wrapping degree parameter of the mineral particles, the intergrowth relationship between the useful minerals and the gangue minerals can be quantitatively described, thereby providing more accurate guidance for subsequent beneficiation process optimization.

[0104] In a specific embodiment, on the basis of the above embodiment, first, the pixel set representing the contact boundary between different mineral species is extracted from the mineral region map. Specifically, the Canny edge detection algorithm is used, and the high threshold is set to 80 and the low threshold is set to 40 to detect the edge pixels from the mineral region map.

[0105] Then, the detected edge pixels are curve-fitted. In one example, a B-spline curve fitting method is used. The order of the B-spline curve is set to 3, and the number of control points is adaptively adjusted according to the number of edge pixels to ensure that the average distance between the fitted curve and the original edge pixels is less than 0.5 pixels.

[0106] Next, the fractal dimension of the fitted curve is calculated. In one implementation, the box counting method is used to calculate the fractal dimension. Specifically, a series of boxes of different sizes are used to cover the fitted curve, and the number of boxes N(ε) required to cover the curve is counted, where ε is the size of the box. By linear regression analysis and The slope of the regression line is an estimate of the fractal dimension. To improve the accuracy of the fractal dimension calculation, the box size ε can be set to a range from 1 pixel to 1 / 10 of the total curve length, taking 10 different values ​​evenly.

[0107] In another implementation, the fractal dimension is calculated using the Hausdorff dimension. The curve is decomposed into a series of line segments, the length of each segment is calculated, and the p-th powers of all segment lengths are summed, denoted as [the sum of these p-th powers]. When p approaches a certain value D, If it approaches a finite nonzero value, then D is the Hausdorff dimension, which can also be used to characterize the complexity of the contact boundary.

[0108] In the above processing, the edge detection algorithm can be replaced with other sub-pixel level edge extraction algorithms, such as gradient vector field-based edge extraction algorithms. The curve fitting method can be replaced with other parametric curve fitting methods, such as Bézier curve fitting. The fractal dimension calculation method can also be replaced with other methods, such as the box-counting dimension method.

[0109] After obtaining the parameters characterizing the complexity of the mosaic morphology, as in the previous embodiment, they can be used for subsequent resource assessment and mining decisions.

[0110] This embodiment achieves a quantitative characterization of the complexity of the intermineral mosaic morphology by calculating the fractal dimension or Hausdorff dimension of the contact boundary, providing more accurate structural parameters for subsequent resource assessment.

[0111] In one specific implementation, based on the above embodiments, for calculating the curve complexity of the contact boundary in the mineral region, firstly, the contact boundary curve extracted by the edge detection algorithm (e.g., Canny algorithm, Sobel algorithm, or Laplacian algorithm) is converted into a binary image, where the pixel value of the contact boundary is 1 and the pixel value of the non-contact boundary is 0. Specifically, the fractal dimension of the contact boundary in the binary image can be calculated using box counting.

[0112] The box counting method includes the following steps: 1. Overlay the binary image containing the contact boundary curve using a series of square boxes with side length ε. The value of ε can be adjusted according to the actual image resolution, for example, from 2 pixels to 1 / 4 of the maximum image size.

[0113] 2. Count the number N(ε) of boxes that cover the contact boundary curve (i.e., the pixel value is 1).

[0114] 3. Change the side length ε of the box, and repeat steps 1 and 2 to obtain a series of (ε, N(ε)) data pairs.

[0115] 4. Perform a double-log linear regression on the above (ε, N(ε)) data pairs, i.e. linearly fit and to obtain the slope k.

[0116] 5. The fractal dimension D is equal to the inverse of the slope k, i.e. D = -k. The value of the fractal dimension D is between 1 and 2, and the larger the value is, the more complex the contact boundary curve is, and the more complex the mineral dissemination structure is. In another example, the fractal dimension of the contact boundary curve can also be obtained by equivalent algorithms known to those skilled in the art, such as Hausdorff dimension calculation, box-counting dimension calculation, etc.

[0117] By calculating the fractal dimension of the contact boundary, the complexity of the mineral dissemination structure can be more accurately quantified, and more accurate data support can be provided for subsequent beneficiation process optimization.

[0118] In a specific embodiment, on the basis of the above embodiment, first, after quantifying the physical dissemination relationship, the inclusion degree parameter, the contact boundary fractal dimension, the intergrowth ratio, and the area ratio data of each mineral region are input into the resource assessment model.

[0119] Specifically, the resource assessment model can take various forms, such as a neural network model, a support vector machine (SVM) model, or a random forest model. The resource assessment model is trained by a training data set. The training data set contains the following data of a plurality of historical ore samples: physical dissemination relationship parameters (including but not limited to inclusion degree, fractal dimension, intergrowth ratio) quantified by image analysis, area ratio data of each mineral region obtained by image segmentation, and corresponding beneficiation smelting indexes measured by actual beneficiation experiments. The beneficiation smelting indexes include but are not limited to concentrate recovery rate, concentrate grade, and beneficiation cost. The training process adjusts the weights of the neural network model using, for example, a backpropagation algorithm, or trains the SVM model using a sequential minimal optimization algorithm.

[0120] Then, the trained resource assessment model is used to evaluate a new ore sample. The model receives the input physical dissemination relationship parameters and area ratio data, and outputs a resource assessment grade of the ore sample. The resource assessment grade can be divided into multiple levels, such as “high selectivity”, “medium selectivity”, and “difficult to select”. In another implementation, the resource assessment model outputs a continuous beneficiation smelting score. The higher the score value is, the better the beneficiation smelting of the ore sample is.

[0121] Then, the resource evaluation level or the optional metallurgical score can be compared with a preset threshold to assist decision-making. For example, the optional metallurgical score threshold is set to 0.8, and if the score of an ore sample is higher than 0.8, it is considered that the ore sample has high economic value, and it is recommended to be mined preferentially.

[0122] The beneficial effect is that the objective evaluation of the value of ore resources is realized, the subjectivity of judgment only by experience is avoided, and the investment risk is reduced.

[0123] In a specific embodiment, on the basis of the above-mentioned embodiment, first, a data set containing historical ore samples is collected and sorted. Each record of the data set needs to contain: sample microscopic image data obtained by a hyperspectral microscopic imaging system, a complete set of physical dissemination relationship parameters (inclusion degree, fractal dimension, intergrowth ratio) calculated by the method of the above-mentioned embodiment, area ratio data of each mineral region, and concentrate recovery rate and concentrate grade data obtained after standard beneficiation experiments on the sample.

[0124] Specifically, the physical dissemination relationship parameters include but are not limited to: inclusion degree, defined as the degree of useful minerals being wrapped by gangue minerals, calculated as the ratio of the boundary length of a useful mineral grain in contact with gangue minerals to the total perimeter of the grain; contact boundary fractal dimension, used to quantify the complexity of the contact interface between minerals; intergrowth ratio, defined as the proportion of useful mineral grains that coexist with gangue minerals. The area ratio data refers to the area percentage of each mineral on the cross section of the ore sample, which can be calculated by image segmentation algorithm. The beneficiation experiment results include but are not limited to: concentrate grade, referring to the percentage of the content of target elements in the concentrate; concentrate recovery rate, referring to the proportion of target elements from raw ore into concentrate. The data set needs to contain at least 1000 different types of ore samples to ensure the training effect of the model.

[0125] Then, a suitable machine learning model is selected. The selectable models include but are not limited to: neural network models (such as multilayer perceptron, convolutional neural network), support vector machine (SVM), random forest, gradient boosting decision tree (GBDT), etc. In this embodiment, the gradient boosting decision tree model is selected, and the five-fold cross-validation method is used for model training and parameter optimization. The gradient boosting decision tree is composed of multiple decision trees, each decision tree learns the residual error of the previous decision tree, and the overall prediction performance is improved by integrating multiple weak learners.

[0126] Next, the dataset is divided into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the generalization ability of the model. The physical intergrowth relationship parameters and area ratio data are used as input features, and the concentrate recovery rate and concentrate grade are used as output targets. The input features are normalized to eliminate the influence of dimensional differences. The gradient boosting decision tree model is trained using the training set data, and the model parameters selected by grid search or Bayesian optimization method, such as the number of decision trees, the maximum depth of the tree, the learning rate, etc. After training, the test set data is used to evaluate the prediction accuracy of the model. Common evaluation indicators include mean square error (MSE), mean absolute error (MAE), and R-squared value.

[0127] In addition, the resource evaluation model is not limited to the above machine learning model, but can also use a rule-based reasoning system based on expert knowledge. The system pre-sets a series of rules, such as when the package degree is higher than a certain threshold, the concentrate recovery rate will decrease; when the contact boundary fractal dimension is higher, the beneficiation difficulty will increase. By inputting the physical intergrowth relationship parameters and area ratio data into the rule-based reasoning system, the resource evaluation grade of the ore sample can also be obtained.

[0128] In addition, in order to improve the robustness of the model, noise can be added to the training data. For example, random perturbations are added to the physical intergrowth relationship parameters and area ratio data to simulate errors that may exist in actual ore sample analysis. Ensemble learning methods can also be used to train multiple different machine learning models and perform weighted averaging of their prediction results to reduce model variance.

[0129] By constructing the above resource evaluation model, accurate prediction of ore selectivity and metallurgy can be realized, providing scientific basis for subsequent beneficiation process optimization and resource comprehensive utilization. The model can predict the concentrate recovery rate and concentrate grade of the ore according to the input physical intergrowth relationship parameters and area ratio data, thereby providing decision support for enterprises.

[0130] The present embodiment realizes the quantification and intelligentization of ore resource evaluation by constructing a machine learning model trained based on historical data, improves the accuracy and efficiency of evaluation, and provides technical support for the rational development and utilization of mineral resources.

[0131] In a specific embodiment, on the basis of the above embodiment, the process of processing the micrograph by the convolutional neural network model to generate the feature point candidate set includes: first, inputting the micrograph into a pre-trained convolutional neural network model. The convolutional neural network model is, for example, a VGG16 network, a ResNet50 network, or an Inception network. Specifically, the convolutional neural network model is configured to extract a plurality of texture feature maps of different scales. For example, from the layer of the VGG16 network, Layer and The layers extract shallow, middle and deep texture feature maps, respectively. In order to enhance the sensitivity to fine texture features, dilated convolution can be used to increase the receptive field of the convolution kernel while maintaining the resolution of the feature map. Specifically, in the Layer and The layer applies dilated convolution with a dilation rate of 2 and 4.

[0132] Then, the multi-scale texture feature maps are processed using the Harris corner detection algorithm to identify corners with a corner response value higher than a preset response threshold. Specifically, the Harris corner detection algorithm is applied to each scale of the texture feature map, respectively. The calculation formula of the corner response function R is where M is the structure tensor, and k is an empirical constant, for example, in the range of 0.04 to 0.06. The preset response threshold can be adaptively set according to the pixel value distribution of the texture feature map, for example, set to 1.5 to 2 times the average value of the texture feature map pixels. The detected corners on different scale texture feature maps can be fused, for example, by a non-maximum suppression algorithm, in the fused corner set, only the corner with the largest response value is retained, and the remaining corners are removed.

[0133] Next, the corners obtained after the Harris corner detection algorithm are processed by a non-maximum suppression algorithm to generate a feature point position candidate set. Specifically, within an N x N neighborhood, only the corner with the largest response value is retained, and other corners in the neighborhood are suppressed. The value of N is in the range of 3 to 7, for example. After non-maximum suppression processing, a set of sparse but representative feature point position candidates is obtained. These feature point position candidates contain information such as edges, cleavage joints and surface textures of mineral crystals. The feature point position candidate set is used for subsequent spectral value and morphological parameter calculation.

[0134] By extracting multi-scale texture feature maps of the micrograph and combining the Harris corner detection and non-maximum suppression algorithms, the feature points in the image can be more accurately identified, thereby improving the accuracy and robustness of subsequent mineral identification.

[0135] As shown in Figure 2 The embodiment of the present application provides a mineral identification system based on a large model algorithm, which is composed of a data processing module M100, a feature calculation module M200, a clustering module M300 and a region merging module M400, and is used to realize the mineral identification method based on the large model algorithm of any one of the above embodiments.

[0136] The data processing module M100 is configured to obtain microscopic images of the ore sample to be tested from an external device, and process the microscopic images using a convolutional neural network model to generate a candidate set of feature points. Specifically, the data processing module M100 can include an image acquisition unit and a preprocessing unit. The image acquisition unit is responsible for receiving ore sample image data provided by an image acquisition device such as a microscope or a scanner. The image acquisition unit can be compatible with multiple image formats, such as TIFF, JPEG, PNG, etc. The preprocessing unit is responsible for preliminary processing of the received image data, including but not limited to: image format conversion, noise filtering, brightness and contrast adjustment. The preprocessing unit can use algorithms such as median filtering and Gaussian filtering to reduce image noise. To generate the candidate set of feature points, the preprocessing unit integrates a pre-trained convolutional neural network model. The convolutional neural network model can be, but is not limited to: ResNet, VGGNet, or MobileNet, etc. The model is configured to extract visual features such as texture and edge in the image, and identify significant feature points on the feature map through corner detection algorithms (such as Harris corner detector or Shi-Tomasi corner detector). The data processing module M100 can be implemented by software, hardware or a combination of software and hardware. For example, the module can be implemented as a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC).

[0137] The feature calculation module M200 is configured to receive the candidate set of feature points from the data processing module M100, and for each feature point in the candidate set, calculate its spectral value and morphological parameters to determine a multi-dimensional data point containing the spatial coordinates, spectral value and morphological parameters of each feature point. In one implementation, the feature calculation module M200 includes a spectral analysis unit and a morphological calculation unit. The spectral analysis unit is configured to connect a hyperspectral imaging device to receive spectral data corresponding to each feature point. The spectral analysis unit can use dimensionality reduction methods such as wavelet transform or principal component analysis to extract spectral feature vectors. The morphological calculation unit is configured to calculate the spatial relationship between each feature point and its adjacent feature points, such as Euclidean distance, relative angle, etc. These spatial relationships can be used as morphological parameters. The input data of the feature calculation module M200 can be image data or spectral data. Communication between modules can be wired (e.g. through USB, Ethernet) or wireless (e.g. through Wi-Fi, Bluetooth, NFC).

[0138] The clustering module M300 is configured to group the multi-dimensional data points from the feature calculation module M200 using a clustering algorithm to obtain feature point clusters. In an embodiment, the clustering module M300 uses a K-means clustering algorithm. To determine the number of clusters K, the clustering module M300 further comprises an adaptive K value determination unit which can automatically evaluate the clustering effect under different K values by methods such as contour coefficient analysis or elbow rule, and select the optimal K value.

[0139] The region merging module M400 is configured to calculate the spatial distance between any two spatially adjacent feature point clusters, evaluate the spectral similarity of the two, and merge the two adjacent feature point clusters into a mineral region when the spatial distance is less than a first preset threshold and the spectral similarity is higher than a second preset threshold. The spatial distance can be calculated as the Euclidean distance between the two cluster centroids. The spectral similarity can be evaluated by calculating the cosine similarity between the average spectral feature vectors of the two clusters. The first preset threshold and the second preset threshold can be fixed values, or dynamic thresholds adaptively adjusted based on the data set.

[0140] The above various modules work together in cooperation, the data processing module M100 generates high-precision feature points, the feature calculation module M200 extracts the multi-dimensional feature vector of each feature point, the clustering module M300 preliminarily groups the feature points according to the feature vectors, and the region merging module M400 optimizes and adjusts the preliminary grouping result based on spatial proximity and spectral similarity, thereby realizing accurate segmentation of different mineral regions in the ore sample. Through effective cooperation between the modules, the system can solve the technical problem that the mineral boundaries are difficult to accurately divide in ore samples with complex composition, fine grains and close intergrowth, and realize high-precision region segmentation and type identification of complex intergrowth minerals.

[0141] In summary, the embodiment of the present application provides an ore identification system based on a large model algorithm, which realizes automatic processing, feature extraction, clustering analysis and region merging of ore images through modular design, can accurately identify different mineral components in ore samples, and solves the problems of low accuracy and poor robustness in traditional ore identification technology.

[0142] In a specific implementation, based on the above embodiment, the data processing module M100 comprises a hyperspectral microscopic imaging system configured to scan the ore sample to be measured to obtain hyperspectral image data. The hyperspectral microscopic imaging system comprises an optical microscope and a pushbroom hyperspectral imager integrated therewith. The optical microscope provides magnified observation capability for the ore sample to be measured. The pushbroom hyperspectral imager is used to collect spectral information at each spatial pixel point, with a spectral range covering the visible and near-infrared bands, for example, 400 nm to 1000 nm, and a spectral resolution of 5 nm. The scanning mode is line-by-line scanning, and the spatial resolution is adjustable, for example, set to 2048x2048 pixels. The data processing module M100 further comprises a data interface configured to receive hyperspectral image data from the hyperspectral microscopic imaging system and transmit the data to the feature calculation module M200. The data interface can be wired (for example, through USB, Ethernet) or wireless (for example, through Wi-Fi).

[0143] The feature calculation module M200 comprises a spectral extraction unit and a dimension reduction processing unit. The spectral extraction unit is configured to receive hyperspectral image data from the data processing module M100 and extract the complete spectral curve corresponding to each feature point according to the feature point candidate set provided by the data processing module M100. The dimension reduction processing unit is configured to receive the spectral curve from the spectral extraction unit and perform dimension reduction processing on the spectral curve using a wavelet transform algorithm to generate a spectral feature vector. The wavelet transform algorithm can select Daubechies wavelet basis, Symlets wavelet basis or Coiflets wavelet basis. The dimension reduction processing unit extracts low-frequency and high-frequency coefficients of the spectral curve at multiple scales and splices these coefficients into a spectral feature vector, which is the spectral value of the feature point. As an alternative, the dimension reduction processing unit can also use a principal component analysis (PCA) algorithm to perform dimension reduction processing on the spectral curve.

[0144] By using the above structure, the embodiment can obtain multispectral information of the ore sample using the hyperspectral microscopic imaging system, and efficiently extract spectral features using dimension reduction methods such as wavelet transform or PCA, providing accurate spectral data basis for subsequent ore identification.

[0145] In one specific implementation, on the basis of the above embodiment, the clustering module M300 comprises an adaptive cluster number determination unit. The adaptive cluster number determination unit is configured to: within a preset cluster number search range, for example from 2 to 20, call the K-means clustering algorithm to cluster the multi-dimensional data point set for each candidate cluster number, and calculate the corresponding cluster evaluation index. The calculation method of the cluster evaluation index can be the silhouette coefficient method or the Calinski-Harabasz index method. When the silhouette coefficient method is adopted, the adaptive cluster number determination unit calculates the silhouette coefficient of each data point, which measures the similarity of the data point to the cluster it belongs to and the dissimilarity to other clusters. Then, the average value of the silhouette coefficients of all data points is calculated as the cluster quality evaluation value under the current cluster number. The adaptive cluster number determination unit selects the cluster number that maximizes the average silhouette coefficient as the cluster number. The adaptive cluster number determination unit can also determine the cluster number using the elbow rule. The elbow rule calculates the within-cluster sum of squares (WCSS) under different K values, and plots the WCSS and K value into a curve. The K value corresponding to the point where the curve starts to significantly slow down (forming an "elbow") is considered as the cluster number. The adaptive cluster number determination unit can be implemented as a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC). The communication between the adaptive cluster number determination unit and other units in the clustering module M300 (such as the data input unit, the clustering algorithm execution unit) can be wired (for example, through USB, Ethernet) or wireless (for example, through Wi-Fi, Bluetooth, NFC).

[0146] This embodiment avoids the subjectivity and uncertainty of manual parameter setting by automatically determining the cluster number, improving the automation and adaptability of the ore recognition system.

[0147] In one specific implementation, on the basis of the above embodiment, the ore recognition system further comprises an intergrowth structure analysis module. The intergrowth structure analysis module is configured to receive the image containing mineral region segmentation information and the category labels of various minerals output by the recognition module. The intergrowth structure analysis module comprises an image preprocessing unit, a contact boundary extraction unit, and a quantitative parameter calculation unit.

[0148] The image preprocessing unit is configured to perform noise reduction processing on the received image, for example using a Gaussian filter or a median filter algorithm, to reduce the interference of image noise on subsequent boundary extraction. The unit is also configured to perform image enhancement operations, such as histogram equalization or contrast stretching, to improve the contrast of the image and make the boundaries between mineral regions clearer.

[0149] The contact boundary extraction unit is configured to extract the contact boundary between different mineral regions from the pre-processed image using an edge detection algorithm. In one implementation, the edge detection algorithm is the Canny edge detection algorithm, whose parameters include a Gaussian kernel size (e.g., 3x3 pixels) and two thresholds (e.g., a low threshold of 0.1 and a high threshold of 0.3, both normalized with respect to the maximum gray value of the image pixels). In another implementation, the Sobel operator or the Prewitt operator is also used as an alternative to the edge detection. The extracted contact boundary is represented as a set of pixel coordinate points, or as a parametric curve equation.

[0150] The quantification parameter calculation unit is configured to receive the boundary information from the contact boundary extraction unit and calculate quantification parameters that characterize the mineral intergrowth relationship. In one embodiment, the unit is configured to calculate the inclusion parameter, which is calculated by first identifying the grains of one mineral (e.g., the useful mineral) that are completely or partially wrapped by another mineral (e.g., the gangue mineral), and then calculating the proportion of the length of the boundary perimeter of these wrapped grains that is in contact with the wrapping mineral to the total perimeter. In another embodiment, the unit is configured to calculate the complexity of the contact boundary, for example by calculating the fractal dimension of the contact boundary curve. The calculation method of the fractal dimension can use the box counting method or the Hausdorff dimension method. In addition, the unit can also calculate other geometric parameters, such as the average particle size, aspect ratio, shape factor, etc. of the mineral grains, which can be further used to describe the morphological characteristics of the minerals.

[0151] The communication between the modules can be wired (e.g., through USB, Ethernet) or wireless (e.g., through Wi-Fi, Bluetooth, NFC). The intergrowth structure analysis module can be implemented as software running on a general-purpose processor, or as a dedicated hardware circuit such as a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC) to improve the computational efficiency.

[0152] Through the intergrowth structure analysis module, the system can automatically extract and quantify the intergrowth relationship between minerals in the ore, such as the inclusion, the contact boundary complexity, etc., thereby providing objective and quantitative basis for subsequent resource evaluation and beneficiation process optimization.

[0153] The above is a further detailed description of the present application in conjunction with specific preferred embodiments, and cannot be considered as limiting the specific implementation of the present application to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the protection scope of the present application.

Claims

1. A method for ore identification based on a large model algorithm, characterized in that, include: A microscopic image of the ore sample to be tested is acquired, and the microscopic image is processed using a convolutional neural network model to generate a candidate set of feature points; For each feature point in the candidate feature point set, its spectral value and morphological parameters are calculated to determine a multidimensional data point containing the spatial coordinates, spectral value and morphological parameters of each feature point. Clustering algorithms are used to group the multidimensional data points to obtain feature point clusters; Calculate the spatial distance between any two spatially adjacent feature point clusters and evaluate their spectral similarity. When the spatial distance is less than a first preset threshold and the spectral similarity is higher than a second preset threshold, the two adjacent feature point clusters are merged into a mineral region.

2. The method according to claim 1, characterized in that, The acquisition of microscopic images of the ore sample to be tested includes: acquiring the ore sample through a hyperspectral microscopic imaging system to obtain hyperspectral image data containing spectral curves at each spatial pixel.

3. The method according to claim 2, characterized in that, The calculation of the spectral value of each feature point includes: extracting the spectral curve of the feature point in the hyperspectral image data, and performing dimensionality reduction processing on the spectral curve to generate a spectral feature vector, which is the spectral value.

4. The method according to claim 3, characterized in that, The dimensionality reduction process for the spectral curve includes: extracting the core feature components of the spectral curve using principal component analysis or wavelet transform.

5. The method according to claim 1, characterized in that, After merging into the mineral region, the method further includes: when at least two different mineral regions are identified to coexist, performing image analysis on the contact boundary between the different mineral regions to quantitatively characterize the physical embedding relationship between the different mineral regions.

6. The method according to claim 5, characterized in that, The quantitative characterization of the physical embedding relationship includes: calculating the length ratio of the boundary of one mineral region being surrounded by the boundary of another mineral region to obtain the enclosedness parameter.

7. The method according to claim 5, characterized in that, The quantitative characterization of the physical embedding relationship includes: calculating the curve complexity of the contact boundary to obtain parameters characterizing the complexity of the embedding morphology.

8. The method according to claim 7, characterized in that, The calculation of the curve complexity of the contact boundary includes: calculating the fractal dimension of the contact boundary.

9. The method according to claim 5, characterized in that, After quantifying the physical embedding relationship, the method further includes: inputting the quantified physical embedding relationship parameters and the area proportion data of each mineral region into the resource assessment model to output the resource assessment level of the ore sample.

10. A mineral identification system based on a large model algorithm, characterized in that, include: The data processing module is used to acquire microscopic images of the ore sample to be tested, and to process the microscopic images using a convolutional neural network model to generate a candidate set of feature points. The feature calculation module is used to calculate the spectral value and morphological parameters of each feature point in the candidate set of feature points, so as to determine the multidimensional data points containing the spatial coordinates, spectral values ​​and morphological parameters of each feature point. The clustering module is used to group the multidimensional data points using a clustering algorithm to obtain feature point clusters; The region merging module is used to calculate the spatial distance between any two spatially adjacent feature point clusters, evaluate their spectral similarity, and merge the two adjacent feature point clusters into a mineral region when the spatial distance is less than a first preset threshold and the spectral similarity is higher than a second preset threshold.

Images