Magnetite beneficiation processing method and system fusing visual recognition

Abstract

The application provides a magnetite beneficiation processing method and system fusing visual recognition, relates to the AI visual technical field, and comprises the following steps: extracting macro visual features of the current batch of rough concentrates, and determining the dominant ore category to which the current batch of rough concentrates belongs; matching the dominant ore category with an ore category-optimal imaging parameter scheme mapping library to obtain an optimal imaging parameter scheme; based on the optimal imaging parameter scheme, performing high-definition image collection on the passing rough concentrates at an imaging station to obtain a high-definition image sequence; inputting the high-definition image sequence into a magnetite image recognition model array to output the mineral composition category label and the corresponding image coordinate system position coordinates of each ore particle in each high-definition image; and based on the image coordinate system position coordinates, the mineral composition category label and the ore conveying parameters, controlling a sorting execution mechanism to perform differential sorting actions on the corresponding ore particles. The beneficiation precision is improved.

Description

Technical Field

[0001] This invention relates to the field of AI vision technology, and in particular to a method and system for magnetite beneficiation that integrates visual recognition. Background Technology

[0002] In the fine selection stage of the magnetite beneficiation process, it is necessary to accurately identify the mineral composition and particle location in order to ensure the separation accuracy and efficiency of magnetite from gangue and other impurities. Visual recognition technology can effectively assist in achieving this precise identification requirement.

[0003] However, traditional mineral processing methods often use fixed imaging parameter configurations without dynamically adapting and adjusting them according to actual production conditions. This makes it difficult to cope with changes in visual characteristics caused by the different characteristics of different batches of ore, which can easily lead to insufficient refining accuracy. Summary of the Invention

[0004] This invention addresses the technical problem in the existing magnetite beneficiation and refining stage where fixed imaging parameters are difficult to adapt to the differences in characteristics of different batches of ore, resulting in insufficient visual recognition accuracy and thus failing to meet the requirements of fine sorting. The invention provides a magnetite beneficiation processing method and system that integrates visual recognition.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a magnetite beneficiation method incorporating visual recognition, comprising:

[0007] Macroscopic visual features are extracted from the pre-scanned image data of the current batch of rough and concentrate, and the dominant ore category of the current batch of rough and concentrate is determined based on the macroscopic visual features.

[0008] The dominant ore category is matched with a pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is called. If the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features.

[0009] At the preset imaging station, the visual imaging system is controlled to continuously acquire high-definition images of the passing coarse and concentrate ore based on the optimal imaging parameter scheme, thereby obtaining a high-definition image sequence.

[0010] The high-definition image sequence is input into a pre-trained magnetite image recognition model array, and the output is the mineral composition category label of each ore particle in each frame of high-definition image and the corresponding image coordinate system position coordinates.

[0011] Based on the image coordinate system position coordinates, mineral composition category labels, and ore transport parameters, the sorting execution mechanism is controlled to perform differentiated sorting actions on the corresponding ore particles.

[0012] Secondly, the present invention provides a magnetite beneficiation system integrating visual recognition, comprising:

[0013] The category determination module is used to extract macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and determine the dominant ore category to which the current batch of rough and concentrate belongs based on the macroscopic visual features.

[0014] The imaging parameter configuration module is used to match the dominant ore category with the pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is called. If the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features.

[0015] The image acquisition module is used to control the visual imaging system to continuously acquire high-definition images of the rough and concentrate ore at a preset imaging station based on the optimal imaging parameter scheme, so as to obtain a high-definition image sequence.

[0016] The sorting and recognition module is used to input the high-definition image sequence into a pre-trained magnetite image recognition model array and output the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each frame of high-definition image.

[0017] The sorting execution module is used to control the sorting execution mechanism to perform differentiated sorting actions on the corresponding ore particles based on the image coordinate system position coordinates, mineral composition category labels and ore transport parameters.

[0018] The beneficial effects of this invention are:

[0019] Compared to existing technologies, this application achieves precise adaptation to the characteristics of different batches of ore by extracting the macroscopic visual features of the current batch of rough and concentrate and determining its dominant ore category, providing a targeted basis for subsequent imaging and identification. By dynamically generating the optimal imaging parameter scheme through matching the dominant ore category with the ore category-optimal imaging parameter scheme mapping library, it effectively solves the problem of poor adaptability of traditional fixed imaging parameters, ensuring the clarity and effectiveness of high-definition image acquisition. Simultaneously, it achieves dynamic updating and optimization of the ore category-optimal imaging parameter scheme mapping library, improving the adaptability to different working conditions. High-definition image sequences acquired based on optimal imaging parameters provide high-quality data support for accurate identification of mineral particles. By processing these high-definition image sequences using a pre-trained magnetite image recognition model array, the system can accurately output the mineral composition category label and image coordinate system position coordinates of each ore particle, improving the accuracy and efficiency of mineral identification while balancing recognition performance and computing power utilization. Finally, combined with the control of conveying parameters, the sorting actuator performs differentiated sorting actions, effectively improving the sorting accuracy between magnetite and impurities, reducing mis-sorting and missed-sorting rates, adapting to real-time conveying conditions, and ensuring sorting efficiency.

[0020] Through the above technical solution, this application achieves intelligent and refined control of the entire process of magnetite beneficiation and selection, from ore characteristic adaptation, high-definition imaging, accurate identification to differentiated sorting. This solves the problems of poor adaptability, low identification accuracy, and unsatisfactory sorting results in traditional beneficiation methods. It improves the level of intelligence, sorting accuracy, and production efficiency of beneficiation, meeting the refined and intelligent requirements of modern magnetite beneficiation and adapting to complex and ever-changing production scenarios. Attached Figure Description

[0021] Figure 1 A schematic flowchart of a magnetite beneficiation method integrating visual recognition provided by the present invention;

[0022] Figure 2 This is a schematic diagram of a magnetite beneficiation system that integrates visual recognition, as provided by the present invention.

[0023] In the attached diagram, the components represented by each number are as follows:

[0024] Category determination module 11, imaging parameter configuration module 12, image acquisition module 13, sorting and recognition module 14, and sorting execution module 15. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0027] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0028] Example 1, as Figure 1 As shown, this embodiment of the invention provides a magnetite beneficiation method incorporating visual recognition, comprising:

[0029] S10: Extract macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and determine the dominant ore category of the current batch of rough and concentrate based on the macroscopic visual features.

[0030] In the beneficiation and refining stage of magnetite, different batches of rough concentrate have significant differences in ore source, beneficiation process, mineral composition and grade, and their dominant ore categories are different, such as high / medium / low grade magnetite, polymetallic associated minerals, etc. The macroscopic visual characteristics such as color, texture and luster of the corresponding ores also have obvious differences.

[0031] Meanwhile, the differences in visual characteristics among different dominant ore categories directly determine the varying sensitivities of the visual separability of magnetite and gangue minerals to imaging parameters. Therefore, the dominant ore category can be determined by extracting the macroscopic visual characteristics of the current batch of rough and concentrate, providing a data foundation for subsequently determining the optimal imaging parameter scheme based on the dominant ore category.

[0032] To address the aforementioned issues, this application extracts macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and determines the dominant ore category of the current batch of rough and concentrate based on the macroscopic visual features.

[0033] Specifically, step S10 in the method includes:

[0034] At the pre-scanning station, using preset standardized light sources and camera parameters, a high-speed linear array camera is used to acquire pre-scan image data of the current batch of rough and concentrate.

[0035] Calculate the average gray value and standard deviation of all pixels in the entire batch of pre-scanned image data as a color statistical indicator;

[0036] Calculate the gray-level co-occurrence matrix of the entire batch of pre-scanned image data, and calculate the average value of the contrast parameter based on the gray-level co-occurrence matrix as a texture statistical index;

[0037] The percentage of pixels with gray values ​​higher than a preset highlight threshold in the entire batch of pre-scanned image data is used as a gloss statistical indicator.

[0038] The color statistics, texture statistics, and gloss statistics are combined to form the macroscopic visual characteristics of the current batch of coarse concentrate.

[0039] In this embodiment, pre-scanning image data of the current batch of rough concentrate is first acquired at the pre-scanning station using a high-speed linear array camera with preset standardized light source and camera parameters. Specifically, the pre-scanning station is located at the front end of the conveyor belt in the refining stage, maintaining a fixed distance from the imaging station. A standardized light source and a high-speed linear array camera are configured at the pre-scanning station. The preset standardized light source can be a cool white LED light source with a color temperature of 6500K and an illumination uniformity of ≥95%. The preset camera parameters can be fixed at an exposure time of 50μs, a gain of 0dB, and a resolution of 4096 pixels. The preset standardized light source and camera parameters ensure that the acquisition conditions of pre-scanning image data from different batches are completely consistent, eliminating the interference of environmental factors on feature extraction.

[0040] Among them, the line scanning frequency of the high-speed line scan camera can be dynamically adapted to the running speed of the conveyor belt. For example, it can be adapted to the running speed of the conveyor belt from 0.5 to 2 m / s. The line scanning frequency of the high-speed line scan camera can be set to ≥20 kHz, so that the conveyed coarse concentrate can be continuously scanned with full coverage and no omissions.

[0041] The pre-scanned image data is collected under preset standardized light source and camera parameters. There is no need to pursue excessively high imaging accuracy. It is only necessary to be able to stably reflect the macroscopic visual characteristics of the entire batch of rough and concentrate, so as to provide a data foundation for subsequent determination of the dominant ore category and further determination of the optimal imaging parameter scheme.

[0042] Secondly, the average gray value and standard deviation of all pixels in the entire batch of pre-scanned image data are calculated as color statistical indicators. The average gray value is the arithmetic mean of the gray values ​​of all pixels in the entire batch of pre-scanned image data, reflecting the overall brightness of the entire batch of rough and concentrate ore. A higher average gray value indicates a brighter overall ore, and vice versa. The standard deviation of gray values ​​is the square root of the sum of the squares of the differences between the gray values ​​of each pixel and the average gray value, divided by the total number of pixels. It reflects the uniformity of the brightness distribution of the entire batch of ore. A larger standard deviation indicates a more significant difference in brightness, and vice versa. Combining the average gray value and the standard deviation of gray values ​​forms a two-dimensional color statistical indicator, comprehensively characterizing the overall color attributes of the entire batch of rough and concentrate ore.

[0043] Next, the gray-level co-occurrence matrix (GLCM) of the entire batch of pre-scanned image data is calculated, and the average value of the contrast parameter is calculated based on the GLCM as a texture statistical indicator. Specifically, the GLCM is a statistical matrix that describes the correlation between the gray values ​​of pixels at different locations in an image, and can accurately reflect the texture roughness, texture direction, and other characteristics of the mineral surface. For example, when calculating the GLCM of the entire batch of pre-scanned image data, a fixed pixel distance d is first selected, such as d=1, and four directions are calculated at angles θ=0°, 45°, 90°, and 135°. Then, all valid pixels of the pre-scanned image data are traversed, and the frequency of gray value combinations (i,j) of every two pixel pairs at intervals d and at angle θ is counted. This frequency is normalized into a probability value P(i,j), where i and j are pixel gray values. Finally, all P(i,j) are arranged according to the gray value dimension, which yields the GLCM for the corresponding distance and angle.

[0044] Specifically, the contrast parameter is a statistical measure of the gray-level co-occurrence matrix, and its calculation formula is as follows: A higher contrast parameter value indicates a clearer and rougher ore texture, while a lower value indicates a smoother texture. The contrast parameters of the gray-level co-occurrence matrix in four directions (θ=0°, 45°, 90°, and 135°) are calculated respectively, and their arithmetic mean is taken as the texture statistical index of the entire batch of coarse and fine ore, thus eliminating the influence of texture direction on features.

[0045] Furthermore, the percentage of pixels with grayscale values ​​higher than a preset highlight threshold in the entire batch of pre-scanned image data is used as a gloss statistical index. For example, the preset highlight threshold can be experimentally calibrated based on the differences in gloss characteristics between magnetite and gangue. Preferably, the preset highlight threshold can be set to a grayscale value of 220. This preset highlight threshold can effectively distinguish between the highlight reflection area and the normal area on the ore surface. The number of effective pixels with grayscale values ​​greater than the preset highlight threshold in the entire batch of pre-scanned image data is counted, and the ratio of this number to the total number of pixels is calculated; this is the gloss statistical index. A higher gloss statistical index value indicates stronger reflectivity and higher gloss on the entire batch of ore surface, and vice versa. The gloss statistical index is calculated because magnetite, due to its metallic luster characteristics, typically has a higher proportion of highlight pixels than gangue minerals; therefore, the gloss statistical index can effectively distinguish the gloss attributes of the ore.

[0046] Finally, the color, texture, and gloss statistics are combined to form the macroscopic visual characteristics of the current batch of rough and concentrate ore. Specifically, the average gray value and standard deviation of gray value from the color statistics, the texture statistics, and the gloss statistics are spliced ​​together in a fixed order to form the macroscopic visual characteristics of the current batch of rough and concentrate ore. The macroscopic visual characteristics completely and quantitatively characterize the three major macroscopic visual attributes of color, texture, and gloss of the entire batch of ore, providing standardized and quantifiable feature data for subsequent determination of the dominant ore category.

[0047] Furthermore, the phrase "determining the dominant ore category of the current batch of rough and concentrate based on the macroscopic visual characteristics" includes:

[0048] Collect pre-scanned images of historical batches of rough and concentrate under preset standardized light source and camera parameters;

[0049] Macroscopic visual features are extracted from pre-scanned images from each historical batch to form multiple historical feature vectors;

[0050] Based on the actual sorting results indicators and mineral composition analysis reports corresponding to each historical batch, the historical feature vectors are divided into multiple predefined dominant ore categories;

[0051] Calculate the arithmetic mean of each dimension of all historical feature vectors corresponding to each dominant ore category to generate a category feature vector representing each dominant ore category;

[0052] Establish a mapping relationship between each dominant ore category and its corresponding category feature vector, and store it to form an ore feature database;

[0053] The similarity between the macroscopic visual features of the current batch of crude concentrate and the feature vector of each category in the ore feature database is calculated.

[0054] The dominant ore category corresponding to the category feature vector with the highest similarity calculation result is selected as the dominant ore category to which the current batch of rough and concentrate belongs.

[0055] In this embodiment, pre-scanned images of historical batches of rough and concentrate are first acquired under preset standardized light source and camera parameters. Specifically, the historical batches of rough and concentrate are continuously processed from previous refining stages, covering rough and concentrate samples from different mining areas, with different grades and mineral compositions as much as possible; the light source and camera parameters used during acquisition are kept as consistent as possible with the current batch of rough and concentrate. For example, pre-scanned images of each historical batch of rough and concentrate are acquired using a high-speed linear array camera under preset standardized light source and camera parameters, ensuring that the feature extraction conditions of historical data and current data are consistent, eliminating the interference of differences in acquisition conditions on category matching, and providing a real and reliable historical data basis for subsequent category determination.

[0056] Secondly, following the aforementioned extraction process and standards for the macroscopic visual features of the current batch of rough and concentrate, macroscopic visual features were extracted one by one from the pre-scanned images of each historical batch, forming multiple historical feature vectors. These historical feature vectors are macroscopic visual feature vectors extracted from the pre-scanned images of each historical batch of rough and concentrate, containing color, texture, and gloss statistical indicators. The extraction method is consistent with that used for the current batch of rough and concentrate. These historical feature vectors cover the distribution patterns of macroscopic visual features of different dominant ore categories and different grades of rough and concentrate, fully reflecting the batch-level visual characteristics of various ores. They provide data support for subsequent dominant ore category classification and category feature vector generation.

[0057] Furthermore, based on the actual sorting results indicators and mineral composition analysis reports corresponding to each historical batch, each historical feature vector is classified into several predefined dominant ore categories. For example, the actual sorting results indicators may include quantitative indicators such as sorting recovery rate (magnetite recovery ratio), concentrate grade (magnetite content), and tailings grade. The mineral composition analysis report refers to the ore mineral composition and content data obtained through professional testing methods such as X-ray diffraction (XRD) and X-ray fluorescence spectroscopy (XRF). The predefined multiple dominant ore categories can be divided according to production needs and ore characteristics.

[0058] For example, the dominant ore category can be divided into four types based on magnetite content: high-grade magnetite rough concentrate (magnetite content ≥ 60%), medium-grade magnetite rough concentrate (40% ≤ magnetite content < 60%), low-grade magnetite rough concentrate (20% ≤ magnetite content < 40%), and polymetallic associated rough concentrate (magnetite content < 20% and associated with multiple gangues).

[0059] For example, by combining the sorting indicators and mineral composition reports of each historical batch, the corresponding historical feature vectors can be classified into multiple dominant ore categories, thus achieving a precise correlation between feature vectors and ore categories.

[0060] Furthermore, the arithmetic mean of each dimension of all historical feature vectors corresponding to each dominant ore category is calculated to generate a category feature vector representing each dominant ore category. Specifically, for all historical feature vectors under the same dominant ore category, the arithmetic mean is calculated dimension by dimension. For example, the high-grade magnetite rough and concentrate category has four dimensions of historical feature vectors. The average value of each of the four dimensions of historical feature vectors is calculated to obtain a 4-dimensional average vector, which is the category feature vector of high-grade magnetite rough and concentrate. The category feature vector is a centralized representation of the macroscopic visual characteristics of ores of the same category, eliminating the feature fluctuations of individual batches and possessing category representativeness.

[0061] Furthermore, a mapping relationship is established between each dominant ore category and its corresponding category feature vector, forming an ore feature database. This database is constructed using a relational database, storing each dominant ore category in relation to its corresponding category feature vector, thus creating a standardized ore feature database. This database provides the foundation for rapid querying and comparison of ore categories in the current batch, serving as the data support for determining the dominant ore category.

[0062] Furthermore, the macroscopic visual features of the current batch of rough and concentrate are compared with the feature vectors of each category in the ore feature database to calculate similarity. Specifically, the similarity can be obtained using Euclidean distance, calculated as follows: Where x is the macroscopic visual feature vector of the current batch, y is the category feature vector in the ore feature database, and d is the Euclidean distance. The smaller the Euclidean distance, the higher the similarity between the macroscopic visual features of the current batch of rough and concentrate and the category feature vector in the ore feature database, and the better the matching degree. According to the above Euclidean distance calculation formula, substituting the data of each dimension, the Euclidean distance between the macroscopic visual features of the current batch of rough and concentrate and each category feature vector in the ore feature database is calculated, which is used as the similarity with each dominant ore category.

[0063] Finally, the dominant ore category corresponding to the category feature vector with the highest similarity calculation result is selected as the dominant ore category of the current batch of rough and concentrate. Specifically, determining the dominant ore category of the current batch of rough and concentrate based on similarity results avoids the subjectivity of human experience judgment and ensures the accuracy and consistency of category determination. For example, if the macroscopic visual features of the current batch of rough and concentrate have the highest similarity with high-grade magnetite rough and concentrate in the ore feature database, then the dominant ore category of the current batch of rough and concentrate is determined to be high-grade magnetite rough and concentrate.

[0064] In summary, compared to existing technologies, this application extracts macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and determines the dominant ore category of the current batch of rough and concentrate based on these macroscopic visual features. This allows for a rapid and objective preliminary determination of batch ore characteristics, providing a basis for accurate matching of subsequent optimal imaging parameter schemes and adaptive adjustment of the sorting process. It effectively solves the problems of delayed ore category determination and inability to adapt to batch characteristic fluctuations in traditional mineral processing.

[0065] S20: Match the dominant ore category with the pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, call the corresponding optimal imaging parameter scheme. If the match fails, generate a new optimal imaging parameter scheme based on the macroscopic visual features.

[0066] In the beneficiation and refining stage of magnetite, traditional imaging parameters are mostly set to fixed values, which cannot adapt to the changes in visual characteristics caused by the differences in dominant ore categories in different batches. This can easily lead to insufficient visual separability between magnetite and gangue, and low accuracy in subsequent mineral identification.

[0067] For example, high-grade magnetite rough and concentrate batches, due to their high magnetite content, strong metallic luster, and dense texture, and the gangue mainly composed of quartz, require the use of red light and high illumination angle imaging parameters to enhance the differentiation of luster differences; conversely, low-grade or multi-metallic associated rough and concentrate batches, due to their low magnetite content, messy texture, and weak luster, require the use of white light and low illumination angle imaging parameters to highlight the differences in texture characteristics.

[0068] To address the aforementioned issues, this application matches the dominant ore category with a pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is invoked; if the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features.

[0069] Specifically, step S20 in the method includes:

[0070] In the pre-configured ore category-optimal imaging parameter scheme mapping library, find the record that is the same as the dominant ore category;

[0071] If the same record is found, the match is successful. The corresponding imaging parameter combination is read from the found record and directly called as the optimal imaging parameter scheme for the current batch. The imaging parameters include at least the light source color, light source intensity and illumination angle.

[0072] If no matching record is found, the match fails, and based on the macroscopic visual features, with the optimization objective of maximizing the feature separability of the target mineral and gangue in the image, an optimization search is performed within the predefined imaging parameter adjustment space.

[0073] The combination of imaging parameters that achieves the optimal goal obtained through optimization search is taken as the newly generated optimal imaging parameter scheme.

[0074] Each newly generated optimal imaging parameter scheme is mapped to the corresponding dominant ore category, and then added to the ore category-optimal imaging parameter scheme mapping library.

[0075] In this embodiment, the process first searches for records matching the dominant ore category in a pre-configured ore category-optimal imaging parameter scheme mapping library. Specifically, the pre-configured ore category-optimal imaging parameter scheme mapping library can be constructed using a key-value pair database, where the key is the name of the dominant ore category and the value is the optimal imaging parameter scheme. During the search, the currently determined dominant ore category is used as the keyword to traverse all keys in the ore category-optimal imaging parameter scheme mapping library, determining whether a completely matching dominant ore category record exists. This search process can be implemented using a database index.

[0076] Secondly, if a matching record is found, the match is successful. The corresponding imaging parameter combination is then read from the found record and directly used as the optimal imaging parameter scheme for the current batch. The imaging parameters include at least the light source color, light source intensity, and illumination angle: the light source color covers types such as white light, red light, green light, and blue light; the light source intensity is quantified as a percentage from 0-100%; and the illumination angle is the angle between the light source and the conveyor belt plane, for example, a range of 30°-90°.

[0077] For example, if the current dominant ore type is high-grade magnetite rough concentrate, after finding the corresponding record in the ore type-optimal imaging parameter scheme mapping library, its imaging parameter combination can be directly read, such as the light source color being red, the light source intensity being 75%, and the illumination angle being 60°, and used as the optimal imaging parameter scheme for the current batch.

[0078] In this way, the ore categories already existing in the ore category-optimal imaging parameter scheme mapping library can be directly retrieved and the corresponding imaging parameter combinations can be called without on-site re-adjustment, which can effectively improve the overall efficiency of the sorting process.

[0079] Secondly, if no matching record is found, the match fails, and an optimization search is performed within a predefined imaging parameter adjustment space based on macroscopic visual features, with the optimization objective of maximizing the feature separability of the target mineral and gangue in the image. Specifically, if the pre-configured ore category-optimal imaging parameter scheme mapping library does not contain a record corresponding to the current dominant ore category, the match fails, and an optimization search is required based on macroscopic visual features.

[0080] The target mineral is magnetite, while the gangue minerals are non-target minerals such as quartz and feldspar. Magnetite has a strong specular highlight due to its smooth surface, while gangue minerals have a homogeneous diffuse reflection texture due to their rough surface. Feature separability is a quantitative calculation based on the characteristics of the target mineral and gangue minerals. It characterizes the degree of difference between the target mineral and gangue minerals in visual features such as color, texture, and luster. The greater the difference, the stronger the separability and the higher the subsequent recognition accuracy.

[0081] For example, the ratio of the average gray value of the highlight pixels (typically the top 10% of pixels with the highest gray values) to the overall average gray value of the target mineral region is calculated as the highlight contrast of the target mineral region. Highlight contrast is specifically used to quantify the specular reflection intensity of magnetite; a higher highlight contrast indicates a more prominent metallic luster. Next, the absolute difference in highlight contrast between the two regions is calculated as the final feature separability quantification index, i.e.: Feature separability = |Highlight contrast of target mineral region - Highlight contrast of gangue region|.

[0082] For example, assuming that under a set of imaging parameters, the overall average gray value of the target mineral region is 100, and the average gray value of its highlight pixels (top 10%) is 250, then its highlight contrast = 250 / 100 = 2.5. In the same image, the overall average gray value of the gangue region is 90, and the average gray value of its highlight pixels is 99, then its highlight contrast = 99 / 90 = 1.1. Applying the above formula, the feature separability at this time = |2.5-1.1| = 1.4. If another set of imaging parameters reduces the highlight contrast of the target mineral region to 2.0 and increases the highlight contrast of the gangue region to 1.3, then the feature separability decreases to 0.7. Therefore, with maximizing the feature separability of the target mineral and gangue in the image as the optimization objective, it can be clearly determined that the first set of imaging parameters is better because it makes the highlight characteristics of magnetite more prominent than those of gangue.

[0083] The predefined imaging parameter adjustment space is determined by the equipment's rated parameters and experimental calibration. For example, the adjustment space for the light source color is set to {white light, red light, green light}, the adjustment space for the light source intensity is set to {30%, 40%, 50%, 60%, 70%, 80%, 90%, 100%}, and the adjustment space for the illumination angle is set to {30°, 45°, 60°, 75°, 90°}.

[0084] Among them, the optimization search can adopt an improved grid search algorithm, traverse all combinations of imaging parameters in the imaging parameter adjustment space, calculate the feature separability of the target mineral and gangue under each combination of imaging parameters, and find the optimal combination of parameters.

[0085] For example, the improved grid search algorithm can perform optimization search according to the following technical path: First, the imaging parameter adjustment space is coarsely divided, for example, each parameter is given only 3-5 representative values ​​to form the first layer of grid. These few parameter combinations are traversed, and the feature separability value corresponding to each group is calculated. Next, the results of the first round of scanning are analyzed to lock in the parameter regions with high feature separability values, and these potential regions are automatically identified, while further searching for low-performance regions is abandoned. Then, within the selected high-performance regions, a second layer of grid with a finer density is automatically generated, for example, by halving the parameter step size, and a new round of traversal and evaluation is performed within this local range, thereby more accurately locating the optimal solution.

[0086] For example, consider adjusting the light source intensity (range 0-100%) and illumination angle (range 0°-90°) in the imaging parameters: A traditional grid search algorithm with step sizes of 10% and 10° requires 10 × 10 = 100 evaluations. The improved grid search algorithm first performs a coarse scan with large step sizes of 30% and 30°, evaluating 4 × 4 = 16 times, and discovers that the optimal performance is concentrated in the intensity-60% and angle-60° region. Subsequently, a fine scan is performed only within this region with step sizes of 10% and 10°, evaluating 3 × 4 = 12 times. In total, only 28 evaluations are needed to accurately locate the optimal solution, significantly improving efficiency.

[0087] Thus, this improved grid search algorithm can systematically perform optimization within the imaging parameter adjustment space at a controllable computational cost, and finally output the imaging parameter combination that maximizes feature separability as the optimal imaging parameter scheme generated for the new ore category.

[0088] Furthermore, the optimal combination of imaging parameters obtained through optimization search, which maximizes the optimization objective, is taken as the newly generated optimal imaging parameter scheme. For example, if the imaging parameter combination obtained after optimization search is: light source color: green light, light source intensity: 80%, illumination angle: 45°, the separability of magnetite and gangue features reaches its maximum under this combination, then this imaging parameter combination is the newly generated optimal imaging parameter scheme, which can ensure the best imaging effect for the current batch of ore.

[0089] Finally, a new mapping relationship is established between each newly generated optimal imaging parameter scheme and its corresponding dominant ore category, and added to the ore category-optimal imaging parameter scheme mapping library. Specifically, the new mapping relationship still uses the dominant ore category and the optimal imaging parameter scheme as key-value pairs, inserts them into the original ore category-optimal imaging parameter scheme mapping library, and updates the index of the ore category-optimal imaging parameter scheme mapping library. When processing batches of the same dominant ore category, the new scheme can be directly called by matching the category name, realizing the dynamic updating and self-optimization of the ore category-optimal imaging parameter scheme mapping library. In continuous production applications, the success rate and efficiency of parameter matching are gradually improved, thereby reducing the computational power consumption of the sorting process.

[0090] Specifically, the construction process of the "ore category-optimal imaging parameter scheme mapping library" includes:

[0091] For each dominant ore category in the ore feature database, representative rough and concentrate samples belonging to the dominant ore category are collected;

[0092] Within the imaging parameter adjustment space, multiple sets of different imaging parameter combinations are used to acquire images of the representative coarse and concentrate samples, resulting in multiple sets of experimental images.

[0093] Calculate the visual separation index of the target mineral region and the gangue region for each group of experimental images under the corresponding combination of imaging parameters.

[0094] For each dominant ore category, the combination of imaging parameters that maximizes the separation index is selected as the optimal imaging parameter scheme for that dominant ore category.

[0095] Establish a mapping relationship between each dominant ore category and its corresponding optimal imaging parameter scheme, and store the mapping relationship to build an ore category-optimal imaging parameter scheme mapping library.

[0096] In this embodiment of the application, firstly, for each dominant ore category in the ore characteristic database, representative rough and concentrate samples belonging to the dominant ore category are collected. Specifically, representative rough and concentrate samples refer to ore samples selected from historical batches of each dominant ore category that have stable sorting indicators and typical mineral composition, ensuring that the samples can fully reflect the visual characteristics of the corresponding dominant ore category.

[0097] Secondly, within the imaging parameter adjustment space, multiple different combinations of imaging parameters are used to acquire images of representative rough and concentrate samples, resulting in multiple sets of experimental images. For example, under each combination of imaging parameters selected within the imaging parameter adjustment space, images are acquired for representative rough and concentrate samples of each dominant ore category. For instance, 10 frames of images are acquired for each representative rough and concentrate sample, resulting in multiple sets of experimental images.

[0098] Next, the separation index of visual features between the target mineral region and the gangue region was calculated for each set of experimental images under the corresponding combination of imaging parameters. The separation index quantifies the degree of difference in visual features between the target mineral and the gangue mineral. A higher separation index indicates a more significant difference in features between the two under that imaging parameter combination, which is more conducive to subsequent image recognition and sorting. It is a key quantitative standard for evaluating the quality of imaging parameters.

[0099] Furthermore, for each dominant ore category, the imaging parameter combination that yields the highest resolution index is selected as the optimal imaging parameter scheme for that dominant ore category. Specifically, for multiple sets of experimental images of the same dominant ore category, the resolution index corresponding to each set of imaging parameter combinations is calculated, and the imaging parameter combination with the highest resolution index is selected as the optimal imaging parameter scheme for that dominant ore category.

[0100] For example, when the dominant ore type is high-grade magnetite rough and concentrate, 15 separation indices are calculated under 15 different imaging parameter combinations. The imaging parameter combination with the highest separation index is selected as the optimal imaging parameter scheme for high-grade magnetite rough and concentrate. In this way, it can be ensured that the imaging parameter combination for each dominant ore type can achieve the best feature separation effect.

[0101] Finally, a mapping relationship is established between each dominant ore category and its corresponding optimal imaging parameter scheme, and stored to construct an ore category-optimal imaging parameter scheme mapping library. Specifically, each dominant ore category is associated with its corresponding optimal imaging parameter scheme, forming a set of category-parameter mapping relationships. This set can also be stored in a NoSQL database, thus constructing the ore category-optimal imaging parameter scheme mapping library. This mapping library provides the foundation for parameter matching in subsequent production processes and can be continuously improved through a dynamic addition mechanism, gradually covering more dominant ore categories and their corresponding optimal imaging parameter schemes.

[0102] Furthermore, the step of "calculating the visual separation index of the target mineral region and the gangue region in each group of experimental images under the corresponding combination of imaging parameters" includes:

[0103] In each set of experimental images, the target mineral area and gangue area are marked;

[0104] The same multidimensional visual features are extracted from the target mineral region and the gangue region respectively, wherein the visual features include color features, texture features and gloss features;

[0105] Based on the extracted multidimensional visual features, the average inter-class distance between all samples in the target mineral region and all samples in the gangue region is calculated in the feature space.

[0106] Calculate the intra-class mean distance of samples within the target mineral region and the intra-class mean distance of samples within the gangue region, respectively.

[0107] Based on the average inter-class distance, the average intra-class distance of samples within the target mineral region, and the average intra-class distance of samples within the gangue region, the separation index under the combination of imaging parameters is calculated using a preset metric formula.

[0108] In this embodiment, the target mineral region and gangue region are first marked in each set of experimental images. Specifically, professional image annotation software, such as LabelImg, is used to accurately annotate the mineral particles in each set of experimental images: the region where the target mineral is located is delineated by the differences in mineral color, luster, and texture, and the regions where gangue such as quartz and feldspar are located are also delineated, thus obtaining the target mineral region and gangue region in each set of experimental images.

[0109] Secondly, the same multidimensional visual features are extracted from both the target mineral region and the gangue region. These visual features include color features, texture features, and gloss features. For example, color features may include average gray value, standard deviation of gray value, mean hue, mean saturation, etc. Texture features may include contrast, entropy, energy, correlation, etc. of the gray-level co-occurrence matrix. Gloss features may include the proportion of highlight pixels, mean area of ​​highlight region, mean gray value of highlight region, and uniformity of highlight region distribution.

[0110] Next, based on the extracted multidimensional visual features, the average inter-class distance between all samples from the target mineral region and all samples from the gangue region is calculated in the feature space. Specifically, the feature space refers to the vector space composed of multidimensional visual features, with each target mineral region and gangue region corresponding to a point in the feature space. The average inter-class distance is the arithmetic mean of the Euclidean distances between the feature vectors of all samples from the target mineral region and the feature vectors of all samples from the gangue region; a larger average inter-class distance indicates a more significant difference in the overall features between the two types of regions.

[0111] The formula for calculating the average distance between classes is as follows: In the formula, N1 is the number of target mineral samples, N2 is the number of gangue samples, xi is the specific value of different dimensions of the multidimensional visual features of the target mineral, and yj is the specific value of different dimensions of the multidimensional visual features of the gangue. Euclidean distance

[0112] Furthermore, the intra-class average distance of samples within the target mineral region and the intra-class average distance of samples within the gangue region are calculated separately. The intra-class average distance is the arithmetic mean of the Euclidean distances between the feature vectors of all samples within the same region. The formula for calculating the intra-class average distance can be referenced from the inter-class average distance. The smaller the intra-class average distance, the more concentrated and less dispersed the features of the samples within the same region, and the stronger the feature stability.

[0113] Finally, based on the average inter-class distance, the average intra-class distance of samples within the target mineral region, and the average intra-class distance of samples within the gangue region, the separation index under the imaging parameter combination is calculated using a preset metric formula. The preset metric formula is: Separation Index = Average Inter-class Distance / (Average Intra-class Distance of Samples within the Target Mineral Region + Average Intra-class Distance of Samples within the Gangue Region). In this formula, the larger the average inter-class distance, and the smaller the average intra-class distances of samples within the target mineral region and the gangue region, the higher the separation index value, indicating a stronger ability to separate the visual features of the target mineral and gangue. This formula quantifies feature differences into a value between 0 and 1, facilitating comparison of the advantages and disadvantages of different imaging parameter combinations and providing a precise quantitative basis for selecting the optimal imaging parameter scheme.

[0114] In summary, compared to existing technologies, this application matches the dominant ore category with a pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is invoked; if the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features. Thus, by matching based on the ore category-optimal imaging parameter scheme mapping library, successful matching directly invokes the optimal imaging parameter scheme to improve efficiency, while failed matching dynamically generates a new optimal imaging parameter scheme to ensure adaptability. Furthermore, the ore category-optimal imaging parameter scheme mapping library can be dynamically updated, thereby ensuring that the imaging parameters accurately adapt to the characteristics of the current batch of ore and maximizing the visual separability of magnetite and gangue.

[0115] S30: At the preset imaging station, the visual imaging system is controlled to continuously acquire high-definition images of the passing coarse and concentrate ore based on the optimal imaging parameter scheme, thereby obtaining a high-definition image sequence.

[0116] In this embodiment, the preset imaging station is a fixed process station in the selection stage specifically used for high-definition image acquisition. It forms a linear spatial layout with the pre-scanning station and the sorting station to ensure the continuity of ore transportation. The visual imaging system can be composed of components such as an adjustable spectral light source, a high-resolution industrial area array camera, and an image acquisition card. It is the core equipment for acquiring high-definition images.

[0117] At the pre-set imaging station, after adjusting parameters such as light source color, intensity, and illumination angle according to the optimal imaging parameter scheme determined in the aforementioned steps, continuous, non-overlapping high-definition images are taken of the coarse and concentrate ore passing continuously and uniformly on the conveyor belt, forming a high-definition image sequence arranged in chronological order. The high-definition image sequence provides a raw, high-fidelity visual data foundation for the subsequent identification and positioning of individual ore particles.

[0118] S40: Input the high-definition image sequence into the pre-trained magnetite image recognition model array, and output the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each frame of high-definition image.

[0119] In the beneficiation and selection of magnetite, traditional manual or single-model mineral identification suffers from low efficiency, insufficient accuracy, difficulty in adapting to multiple types of ores and complex imaging conditions, and cannot accurately obtain the positional information of ore particles to support subsequent sorting operations.

[0120] To address the aforementioned issues, this application inputs the high-definition image sequence into a pre-trained magnetite image recognition model array, and outputs the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each frame of the high-definition image.

[0121] Specifically, step S40 in the method includes:

[0122] Construct and pre-train a magnetite image recognition model array, wherein the magnetite image recognition model array consists of a shared feature extraction module, M parallel sub-models and an integrated output module;

[0123] Based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-definition image sequence, the number N of sub-models to be retrieved from M parallel sub-models for integration is dynamically calculated, where N is an integer between 1 and M.

[0124] Input the current batch of high-resolution image sequences into a pre-trained magnetite image recognition model array;

[0125] The shared feature extraction module identifies and outputs the contour region image of each ore particle in a single frame of high-definition image and the corresponding image coordinate system position coordinates;

[0126] N sub-models, randomly selected from M parallel sub-models, process the contour region image of each ore particle in parallel and output a mineral composition category prediction result for each.

[0127] The integrated output module integrates the mineral composition category prediction results output by N sub-models, generates the final mineral composition category label for each ore particle, and outputs it together with the image coordinate system position coordinates.

[0128] The training process of the M parallel sub-models in the magnetite image recognition model array includes:

[0129] Based on historical production data, the training dataset consists of ore particle images and ore particle outlines in each image collected under different dominant ore categories and different optimal imaging parameter schemes.

[0130] For each ore particle profile in the training dataset, obtain the corresponding mineral composition ground truth label to form a supervision label set;

[0131] The training dataset and the supervision label set are divided into M data subset pairs using the M-fold cross-validation method.

[0132] Use the M data subsets to independently train M sub-models until each sub-model is verified to have converged.

[0133] In this embodiment, a magnetite image recognition model array is first constructed and pre-trained. This array consists of a shared feature extraction module, M parallel sub-models, and an integrated output module. The shared feature extraction module employs an improved ResNet50 convolutional neural network to uniformly extract basic visual features of the ore particles, such as edges, textures, and color distribution, eliminating feature extraction differences between different sub-models. The M parallel sub-models are M independent lightweight classification networks, each trained and run in parallel, possessing independent recognition capabilities. The integrated output module uses a voting integration strategy to integrate the prediction results of each sub-model and output them along with the image coordinates determined by the shared feature extraction module. Here, M represents the total number of parallel sub-models, and its value can be set according to production accuracy requirements. Optimally, M=10 can be chosen to balance recognition accuracy and computational efficiency.

[0134] Secondly, based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-resolution image sequence, the dynamic calculation requires retrieving the number N of sub-models participating in the integration from M parallel sub-models, where N is an integer between 1 and M. Specifically, the logic for dynamically calculating N is as follows: the more common the ore category and the smaller the difference between the optimal imaging parameter scheme and the historical average optimal imaging parameter scheme, the fewer sub-models are required, thus improving efficiency; conversely, the rarer the ore category and the greater the difference between the optimal imaging parameter scheme and the historical average optimal imaging parameter scheme, the more sub-models are required, thus ensuring accuracy. In this way, the value of N is obtained by weighting the calculation based on the commonness of the dominant ore category and the deviation of the imaging parameters, achieving adaptive adjustment of the number of sub-models and avoiding redundant calculations or insufficient accuracy.

[0135] Next, the high-resolution image sequence of the current batch is input into the pre-trained magnetite image recognition model array; the shared feature extraction module identifies and outputs the contour region image of each ore particle in a single frame of high-resolution image and its corresponding image coordinate system position coordinates. Specifically, after the high-resolution image sequence of the current batch is input into the pre-trained magnetite image recognition model array frame by frame, the shared feature extraction module first performs particle segmentation on the single frame of high-resolution image using the Canny edge detection and watershed algorithm, locates the contour boundary of each independent ore particle, and generates a particle contour mask; based on the particle contour mask, the contour region image of each particle is obtained by cropping, and the center pixel coordinates (x, y) of the particle contour are calculated as the image coordinate system position coordinates of the particle; each ore particle corresponds to a set of contour region images and image coordinate system position coordinates.

[0136] It should be noted that both Canny edge detection and the watershed algorithm are conventional algorithms in the field of image processing, which can be understood and implemented by those skilled in the art, and will not be elaborated here.

[0137] Furthermore, N sub-models randomly selected from the M parallel sub-models process the contour region image of each ore particle in parallel, and each outputs a mineral composition category prediction result. For example, if M=10 and N=6, then 6 are randomly selected from the 10 parallel sub-models. Each sub-model independently receives the contour region image of each ore particle, performs inference through its own classification network, and outputs the probability distribution of whether the ore particle belongs to magnetite or gangue. The category with the highest probability is selected as the prediction result of the sub-model. The parallel processing of the 6 selected sub-models can ensure recognition accuracy and improve recognition speed.

[0138] Finally, the integrated output module integrates the mineral composition category prediction results from the N sub-models to generate a final mineral composition category label for each ore particle, and outputs it along with its image coordinate system position coordinates. Specifically, the integrated output module adopts a hard voting strategy: it statistically analyzes the prediction results of the N sub-models and selects the category with the most votes as the final mineral composition category label for each ore particle; if there is a tie in the number of votes, the category with the highest average prediction probability from the sub-models is selected as the final mineral composition category label. Finally, the final mineral composition category label for each ore particle is combined with its image coordinate system position coordinates and output in particle order.

[0139] For example, the magnetite image recognition model array consists of a shared feature extraction module, M parallel sub-models, and an integrated output module. Its specific model structure and core parameter configuration are as follows: 1. The shared feature extraction module can use an improved ResNet50 convolutional neural network, containing 49 convolutional layers, 1 global average pooling layer, and 1 feature output layer. The fully connected classification layer of the original network is removed, retaining only the feature extraction function. The convolutional kernel size is primarily 3×3, with a stride of 1 and padding of 1. Some downsampling layers use 1×1 convolutional kernels. The activation function is ReLU. The pooling method is max pooling, with a 2×2 pooling kernel and a stride of 2.

[0140] 2. M parallel sub-models: Each sub-model is an independent lightweight MobileNetV3-Small classification network, containing depthwise separable convolutional layers, inverse residual modules, global average pooling layers, and fully connected classification layers. The input features are vectors output by the shared feature extraction module. The inverse residual module has a scaling factor of 4 and uses the Hard-Swish activation function. The fully connected layer has an output dimension of 2, corresponding to magnetite and gangue. The loss function is cross-entropy loss, and the optimizer is Adam with a learning rate of 0.001 and a weight decay of 1e-5.

[0141] 3. Integrated Output Module: Employing a hard-voting ensemble strategy, this module receives the classification results from N selected parallel sub-models and outputs the final category based on a voting rule. The voting rule is majority voting; if the number of votes is the same, the category with the highest average predicted probability from the sub-models is selected.

[0142] The training process of the M parallel sub-models in the magnetite image recognition model array includes: collecting ore particle images and their contours from different dominant ore categories and optimal imaging parameter schemes based on historical production data, forming a training dataset; obtaining the corresponding mineral composition ground truth label for each ore particle contour in the training dataset, forming a supervision label set; using M-fold cross-validation, dividing the training dataset and the supervision label set into M data subset pairs; and independently training M sub-models using the M data subset pairs until each sub-model has been verified to converge. Specifically, the training dataset contains no fewer than 100,000 ore particle images, covering the dominant ore categories and all optimal imaging parameter schemes, with particle contours generated through annotation; the supervision label set combines XRD detection with manual annotation to ensure the accuracy of the mineral composition ground truth labels; M-fold cross-validation divides the training dataset into M subsets, using M-1 subsets for training and 1 subset for validation each time, repeating M times to obtain M independent sub-models.

[0143] For example, the training process of each sub-model can refer to the following technical path: Select one set of data subsets from the M-fold cross-validation as the validation set, and the remaining M-1 subsets as the training set. Use the images of ore particles with annotated ore particle outlines from the data subsets in the training set as input features, and the corresponding mineral composition ground truth labels as supervision labels. Use the Adam optimizer with a learning rate of 0.001 to carry out iterative training until the recognition accuracy of the validation set is ≥98% and there is no improvement for 5 consecutive rounds. The sub-model is then considered to have converged, and a trained parallel sub-model is obtained. Following the same technical path, the remaining M-1 sets of data subsets are trained independently in sequence, and finally the training of all M parallel sub-models is completed, ensuring that each sub-model has a stable and high-precision mineral composition recognition capability.

[0144] For example, after all M parallel sub-models have been trained, integrated lightweight training is performed on the entire magnetite image recognition model array. Since the M parallel sub-models have been trained independently and their parameters are stable, full retraining is unnecessary. Only targeted fine-tuning is required to achieve collaborative adaptation between modules, reducing training computation and time costs. The lightweight training process can be referenced as follows: Freeze the parameters of all parallel sub-models, and use images of ore particles under different dominant ore categories and imaging parameters in historical production data, along with the corresponding mineral composition ground truth labels, as input. Focus on fine-tuning the output layer parameters of the shared feature extraction module and the voting weights of the integrated output module. Use the Adam optimizer with a learning rate reduced to 0.0001 and the cross-entropy loss function, and iterate for 10-20 rounds until the recognition accuracy is ≥98.5% and there is no improvement for 3 consecutive rounds. This completes the lightweight training of the entire magnetite image recognition model array.

[0145] Specifically, the phrase "dynamically calculating the number N of sub-models to be retrieved from M parallel sub-models for integration based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-definition image sequence" includes:

[0146] The historical frequency of occurrence of the dominant ore categories is statistically analyzed and used as a commonness coefficient.

[0147] Calculate the absolute deviation between the optimal imaging parameter scheme and the historical average optimal imaging parameter scheme in each dimension of light source color, light source intensity and illumination angle, and calculate the average value of the absolute deviation in all dimensions as the overall deviation coefficient.

[0148] Based on the total number M of parallel sub-models, the commonness coefficient, and the overall deviation coefficient, the number N of sub-models to be retrieved from the M parallel sub-models of the magnetite image recognition model array for integration is obtained through weighted summation and rounding. N is negatively correlated with the commonness coefficient and positively correlated with the overall deviation coefficient, and N is an integer between 1 and M.

[0149] In this embodiment, the historical frequency of occurrence of the dominant ore category is first calculated as a commonness coefficient. The historical frequency of occurrence refers to the proportion of times the dominant ore category appeared in past production batches out of the total number of batches. The calculation formula is: Commonness coefficient = Number of occurrences of the dominant ore category / Total number of past batches.

[0150] For example, if a high-grade magnetite concentrate appears 600 times in 1000 previous batches, then the commonness coefficient is 0.6. The commonness coefficient reflects the frequency of ore categories; a higher coefficient indicates a more common category, more experience in identifying it among the M parallel sub-models, and fewer sub-models are needed.

[0151] Secondly, the absolute deviations between the current optimal imaging parameter scheme and the historical average optimal imaging parameter scheme are calculated in terms of light source color, light source intensity, and illumination angle. The average of these absolute deviations across all dimensions is then used as the overall deviation coefficient. Specifically, the historical average optimal imaging parameter scheme refers to the average of all historical optimal imaging parameter schemes in the ore category-optimal imaging parameter scheme mapping library across all dimensions: light source color is quantified as a numerical value (e.g., white light = 1, red light = 2, green light = 3, etc.), light source intensity is the percentage average, and illumination angle is the degree average. Then, the absolute deviations between the current optimal imaging parameter scheme and the historical average optimal imaging parameter scheme are calculated in the three dimensions of light source color, light source intensity, and illumination angle. The arithmetic mean of these three absolute deviations is then used as the overall deviation coefficient. A larger overall deviation coefficient indicates a greater difference between the currently determined optimal imaging parameter scheme and the historical average level, suggesting more unique ore visual characteristics and a greater need for sub-models.

[0152] Finally, based on the total number M of parallel sub-models, the commonness coefficient, and the overall deviation coefficient, the number N of sub-models to be retrieved from the M parallel sub-models of the magnetite image recognition model array for integration is obtained through weighted summation and rounding. N is negatively correlated with the commonness coefficient and positively correlated with the overall deviation coefficient, and N is an integer between 1 and M.

[0153] For example, the weighted average and rounding (here, rounding to the nearest integer) calculation formula is: N = round[M × (1 - commonness coefficient) + M × overall deviation coefficient], where round is the rounding function. For example, if M = 10, commonness coefficient = 0.6, and overall deviation coefficient = 0.3, then N = round[10 × (1 - 0.6) + 10 × 0.3] = round (4 + 3) = 7. After calculation, if N < 1, then take 1; if N > M, then take M, ensuring that N is within the effective range. In this way, the number of sub-models is dynamically and adaptively adjusted, ensuring both the identification efficiency of common ores and the identification accuracy of rare ores.

[0154] In summary, compared to existing technologies, this application inputs the high-definition image sequence into a pre-trained magnetite image recognition model array, outputting the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each frame of the high-definition image. Thus, by dynamically adjusting the number of models, the mineral composition category of each ore particle in each frame of the image can be accurately identified, and its position coordinates in the image coordinate system can be output simultaneously, ensuring the accuracy of mineral category recognition and the precision of location positioning, while effectively balancing recognition performance and computational efficiency.

[0155] S50: Based on the image coordinate system position coordinates, mineral composition category labels, and ore transport parameters, control the sorting execution mechanism to perform differentiated sorting actions on the corresponding ore particles.

[0156] In the beneficiation and refining process of magnetite, traditional sorting actuators mostly adopt a fixed sorting mode, which cannot combine the precise position of ore particles, mineral composition, and dynamic conveying parameters to achieve differentiated sorting actions. This easily leads to insufficient sorting accuracy between magnetite and gangue, a high rate of missorting and missed sorting, and difficulty in adapting to real-time changes in operating conditions during the conveying process.

[0157] Meanwhile, the aforementioned steps obtain the image coordinate system position coordinates and mineral composition category labels of the ore particles. This can be further combined with ore transport parameters to precisely control the sorting actuator to perform differentiated sorting actions on different ore particles, thereby improving sorting accuracy and efficiency.

[0158] Specifically, step S50 in the method includes:

[0159] The image coordinate system position coordinates of each ore particle are converted into the predicted physical coordinates of the corresponding ore particle on the sorting station plane according to the preset spatial mapping relationship between the imaging station and the sorting station.

[0160] Based on the conveyor belt speed in the ore conveying parameters and the fixed distance between the imaging station and the action point of the sorting actuator, the predicted arrival time required for each ore particle to move from the imaging moment to the action point of the sorting actuator is calculated.

[0161] Based on the mineral composition category label of each ore particle, determine the corresponding sorting action;

[0162] The sorting execution mechanism is controlled to perform corresponding sorting actions on ore particles located at the predicted physical coordinate positions at the predicted arrival time.

[0163] In this embodiment, the image coordinate system position coordinates of each ore particle are first converted into predicted physical coordinates on the sorting station plane based on a preset spatial mapping relationship between the imaging station and the sorting station. The preset spatial mapping relationship is a transformation matrix from the image coordinate system to the sorting station physical coordinate system, obtained through camera calibration and hand-eye calibration experiments, and can be generated through standard calibration. Substituting the image coordinate system position coordinates of the ore particles into the preset spatial mapping relationship, the predicted physical coordinates (X, Y) on the sorting station plane are calculated, in millimeters, accurately representing the sorting position of the ore particles in the actual physical space.

[0164] For example, the transformation relationship between the image coordinate system position coordinates (x, y, unit: pixels) and physical coordinates (X, Y, unit: millimeters) obtained after calibration is: X = 0.0625x - 0.2, Y = 0.062y + 0.1. If the image coordinate system position coordinates of a certain ore particle are (200, 150), substituting them into the calculation yields the predicted physical coordinates: X = 0.0625 × 200 - 0.2 = 12.3 mm, Y = 0.062 × 150 + 0.1 = 9.4 mm. These coordinates are the actual physical position of the ore particle when it arrives at the sorting station.

[0165] Secondly, based on the conveyor belt speed in the ore conveying parameters and the fixed distance between the imaging station and the action point of the sorting actuator, the predicted arrival time required for each ore particle to move from the imaging moment to the action point of the sorting actuator is calculated. Specifically, the formula for calculating the predicted arrival time required for each ore particle to move from the imaging moment to the action point of the sorting actuator is: Predicted arrival time = Fixed distance between the imaging station and the action point of the sorting actuator / Conveyor belt speed. The units for both must be consistent during the calculation.

[0166] Next, based on the mineral composition category label of each ore particle, the corresponding sorting action is determined. For example, the sorting action can be preset; for instance, when the mineral composition category label is magnetite, a retention conveying action is performed, meaning the sorting actuator does not operate, and the particle enters the concentrate bin with the conveyor belt. When the mineral composition category label is gangue, a high-pressure air nozzle jet removal action is performed, blowing the particle to the tailings bin. Those skilled in the art can dynamically set different mineral composition category labels and corresponding sorting actions according to actual needs, thereby achieving differentiated sorting.

[0167] Ideally, the mineral composition category label "gangue" can be further refined, such as into quartz gangue, feldspar gangue, silicate gangue, polymetallic associated minerals, etc., and further refined to adapt to different sorting actions, such as performing the "intermediate material collection" action, that is, blowing to the intermediate material silo for subsequent secondary sorting.

[0168] Finally, the control sorting actuator performs the corresponding sorting action on the ore particles located at the predicted physical coordinates when the arrival time is predicted. Specifically, the sorting actuator is the actual actuator, which can be a high-pressure air nozzle array, a robotic arm array, etc.; the control sorting actuator performs sorting or does not perform any action according to the sorting rules at the predicted arrival time based on the predicted physical coordinates (X,Y) of the ore particles, thereby achieving high-precision and high-efficiency sorting of coarse and concentrate ore.

[0169] In summary, compared to existing technologies, this application controls the sorting actuator to perform differentiated sorting actions on corresponding ore particles based on the image coordinate system position coordinates, mineral composition category labels, and ore conveying parameters. This effectively improves the sorting accuracy of magnetite and gangue, reduces mis-sorting and missed-sorting rates, and adapts to real-time conveying conditions to ensure sorting efficiency.

[0170] In summary, the embodiments of this application have at least the following technical effects:

[0171] Compared to existing technologies, this application first extracts macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and then determines the dominant ore category of the current batch of rough and concentrate based on these macroscopic visual features. This allows for a rapid and objective preliminary determination of batch ore characteristics, providing a basis for the accurate matching of subsequent optimal imaging parameter schemes and adaptive adjustment of the sorting process. It effectively solves the problems of delayed ore category determination and inability to adapt to batch characteristic fluctuations in traditional mineral processing.

[0172] Secondly, this application matches the primary ore category with a pre-configured ore category-optimal imaging parameter scheme mapping library. If a match is successful, the corresponding optimal imaging parameter scheme is invoked; if a match fails, a new optimal imaging parameter scheme is generated based on macroscopic visual features. Thus, by matching based on the ore category-optimal imaging parameter scheme mapping library, successful matching directly invokes the optimal imaging parameter scheme to improve efficiency, while failed matching dynamically generates a new optimal imaging parameter scheme to ensure adaptability. Furthermore, the ore category-optimal imaging parameter scheme mapping library can be dynamically updated, thereby ensuring that the imaging parameters accurately adapt to the characteristics of the current batch of ore and maximizing the visual separability of magnetite and gangue.

[0173] Furthermore, this application controls the visual imaging system at a preset imaging station to continuously acquire high-definition images of the passing coarse and concentrate ore based on an optimal imaging parameter scheme, resulting in a high-definition image sequence. Thus, by acquiring a high-definition image sequence based on a dynamic optimal imaging parameter scheme, high-quality data support is provided for subsequent accurate identification of mineral particles.

[0174] Furthermore, this application inputs high-definition image sequences into a pre-trained magnetite image recognition model array, outputting the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each high-definition image frame. Thus, by dynamically scheduling the number of models, the mineral composition category of each ore particle in each image frame can be accurately identified, and its position coordinates in the image coordinate system can be output simultaneously, ensuring the accuracy of mineral category recognition and the precision of location positioning, effectively balancing recognition performance and computational efficiency.

[0175] Finally, this application controls the sorting actuator to perform differentiated sorting actions on corresponding ore particles based on the image coordinate system position coordinates, mineral composition category labels, and ore conveying parameters. This effectively improves the sorting accuracy of magnetite and gangue, reduces mis-sorting and missed-sorting rates, and adapts to real-time conveying conditions to ensure sorting efficiency.

[0176] Through the above technical solutions, this application realizes the full-process collaborative optimization of the magnetite beneficiation and beneficiation stage, from the pre-judgment of ore characteristics, dynamic adaptation of optimal imaging parameters, high-definition image acquisition, to the accurate identification of mineral particles and differentiated sorting. It completely solves the pain points of poor adaptability, low identification accuracy and insufficient sorting efficiency of traditional beneficiation, greatly improves the level of intelligence, sorting accuracy and production efficiency of beneficiation, adapts to complex and ever-changing production conditions, and meets the refined and intelligent needs of modern magnetite beneficiation.

[0177] Example 2, as Figure 2 As shown, based on the same inventive concept as the magnetite beneficiation processing method integrating visual recognition provided in Embodiment 1, this embodiment of the invention also provides a magnetite beneficiation processing system integrating visual recognition, comprising:

[0178] The category determination module 11 is used to extract macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and determine the dominant ore category to which the current batch of rough and concentrate belongs based on the macroscopic visual features.

[0179] The imaging parameter configuration module 12 is used to match the dominant ore category with the pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is called. If the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features.

[0180] The image acquisition module 13 is used to control the visual imaging system to continuously acquire high-definition images of the rough and concentrate ore at a preset imaging station based on the optimal imaging parameter scheme, so as to obtain a high-definition image sequence.

[0181] The sorting and recognition module 14 is used to input the high-definition image sequence into the pre-trained magnetite image recognition model array and output the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each frame of high-definition image.

[0182] The sorting execution module 15 is used to control the sorting execution mechanism to perform differentiated sorting actions on the corresponding ore particles based on the image coordinate system position coordinates, mineral composition category labels and ore transport parameters.

[0183] Specifically, the category determination module 11 is used for:

[0184] At the pre-scanning station, using preset standardized light sources and camera parameters, a high-speed linear array camera is used to acquire pre-scan image data of the current batch of rough and concentrate.

[0185] Calculate the average gray value and standard deviation of all pixels in the entire batch of pre-scanned image data as a color statistical indicator;

[0186] Calculate the gray-level co-occurrence matrix of the entire batch of pre-scanned image data, and calculate the average value of the contrast parameter based on the gray-level co-occurrence matrix as a texture statistical index;

[0187] The percentage of pixels with gray values ​​higher than a preset highlight threshold in the entire batch of pre-scanned image data is used as a gloss statistical indicator.

[0188] The color statistics, texture statistics, and gloss statistics are combined to form the macroscopic visual characteristics of the current batch of coarse concentrate.

[0189] The category determination module 11 is further specifically used for:

[0190] Collect pre-scanned images of historical batches of rough and concentrate under preset standardized light source and camera parameters;

[0191] Macroscopic visual features are extracted from pre-scanned images from each historical batch to form multiple historical feature vectors;

[0192] Based on the actual sorting results indicators and mineral composition analysis reports corresponding to each historical batch, the historical feature vectors are divided into multiple predefined dominant ore categories;

[0193] Calculate the arithmetic mean of each dimension of all historical feature vectors corresponding to each dominant ore category to generate a category feature vector representing each dominant ore category;

[0194] Establish a mapping relationship between each dominant ore category and its corresponding category feature vector, and store it to form an ore feature database;

[0195] The similarity between the macroscopic visual features of the current batch of crude concentrate and the feature vector of each category in the ore feature database is calculated.

[0196] The dominant ore category corresponding to the category feature vector with the highest similarity calculation result is selected as the dominant ore category to which the current batch of rough and concentrate belongs.

[0197] The imaging parameter configuration module 12 is specifically used for:

[0198] In the pre-configured ore category-optimal imaging parameter scheme mapping library, find the record that is the same as the dominant ore category;

[0199] If the same record is found, the match is successful. The corresponding imaging parameter combination is read from the found record and directly called as the optimal imaging parameter scheme for the current batch. The imaging parameters include at least the light source color, light source intensity and illumination angle.

[0200] If no matching record is found, the match fails, and based on the macroscopic visual features, with the optimization objective of maximizing the feature separability of the target mineral and gangue in the image, an optimization search is performed within the predefined imaging parameter adjustment space.

[0201] The combination of imaging parameters that achieves the optimal goal obtained through optimization search is taken as the newly generated optimal imaging parameter scheme.

[0202] Each newly generated optimal imaging parameter scheme is mapped to the corresponding dominant ore category, and then added to the ore category-optimal imaging parameter scheme mapping library.

[0203] Specifically, the construction process of the "ore category-optimal imaging parameter scheme mapping library" includes:

[0204] For each dominant ore category in the ore feature database, representative rough and concentrate samples belonging to the dominant ore category are collected;

[0205] Within the imaging parameter adjustment space, multiple sets of different imaging parameter combinations are used to acquire images of the representative coarse and concentrate samples, resulting in multiple sets of experimental images.

[0206] Calculate the visual separation index of the target mineral region and the gangue region for each group of experimental images under the corresponding combination of imaging parameters.

[0207] For each dominant ore category, the combination of imaging parameters that maximizes the separation index is selected as the optimal imaging parameter scheme for that dominant ore category.

[0208] Establish a mapping relationship between each dominant ore category and its corresponding optimal imaging parameter scheme, and store the mapping relationship to build an ore category-optimal imaging parameter scheme mapping library.

[0209] Furthermore, the step of "calculating the visual separation index of the target mineral region and the gangue region in each group of experimental images under the corresponding combination of imaging parameters" includes:

[0210] In each set of experimental images, the target mineral area and gangue area are marked;

[0211] The same multidimensional visual features are extracted from the target mineral region and the gangue region respectively, wherein the visual features include color features, texture features and gloss features;

[0212] Based on the extracted multidimensional visual features, the average inter-class distance between all samples in the target mineral region and all samples in the gangue region is calculated in the feature space.

[0213] Calculate the intra-class mean distance of samples within the target mineral region and the intra-class mean distance of samples within the gangue region, respectively.

[0214] Based on the average inter-class distance, the average intra-class distance of samples within the target mineral region, and the average intra-class distance of samples within the gangue region, the separation index under the combination of imaging parameters is calculated using a preset metric formula.

[0215] The image acquisition module 13 is specifically used for:

[0216] At the preset imaging station, the control vision imaging system, based on the optimal imaging parameter scheme, continuously acquires high-definition images of the passing coarse and concentrate ore, resulting in a high-definition image sequence.

[0217] The sorting and identification module 14 is specifically used for:

[0218] Construct and pre-train a magnetite image recognition model array, wherein the magnetite image recognition model array consists of a shared feature extraction module, M parallel sub-models and an integrated output module;

[0219] Based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-definition image sequence, the number N of sub-models to be retrieved from M parallel sub-models for integration is dynamically calculated, where N is an integer between 1 and M.

[0220] Input the current batch of high-resolution image sequences into a pre-trained magnetite image recognition model array;

[0221] The shared feature extraction module identifies and outputs the contour region image of each ore particle in a single frame of high-definition image and the corresponding image coordinate system position coordinates;

[0222] N sub-models, randomly selected from M parallel sub-models, process the contour region image of each ore particle in parallel and output a mineral composition category prediction result for each.

[0223] The integrated output module integrates the mineral composition category prediction results output by N sub-models, generates the final mineral composition category label for each ore particle, and outputs it together with the image coordinate system position coordinates.

[0224] The training process of the M parallel sub-models in the magnetite image recognition model array includes:

[0225] Based on historical production data, the training dataset consists of ore particle images and ore particle outlines in each image collected under different dominant ore categories and different optimal imaging parameter schemes.

[0226] For each ore particle profile in the training dataset, obtain the corresponding mineral composition ground truth label to form a supervision label set;

[0227] The training dataset and the supervision label set are divided into M data subset pairs using the M-fold cross-validation method.

[0228] Use the M data subsets to independently train M sub-models until each sub-model is verified to have converged.

[0229] Specifically, the phrase "dynamically calculating the number N of sub-models to be retrieved from M parallel sub-models for integration based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-definition image sequence" includes:

[0230] The historical frequency of occurrence of the dominant ore categories is statistically analyzed and used as a commonness coefficient.

[0231] Calculate the absolute deviation between the optimal imaging parameter scheme and the historical average optimal imaging parameter scheme in each dimension of light source color, light source intensity and illumination angle, and calculate the average value of the absolute deviation in all dimensions as the overall deviation coefficient.

[0232] Based on the total number M of parallel sub-models, the commonness coefficient, and the overall deviation coefficient, the number N of sub-models to be retrieved from the M parallel sub-models of the magnetite image recognition model array for integration is obtained through weighted summation and rounding. N is negatively correlated with the commonness coefficient and positively correlated with the overall deviation coefficient, and N is an integer between 1 and M.

[0233] The sorting execution module 15 is specifically used for:

[0234] The image coordinate system position coordinates of each ore particle are converted into the predicted physical coordinates of the corresponding ore particle on the sorting station plane according to the preset spatial mapping relationship between the imaging station and the sorting station.

[0235] Based on the conveyor belt speed in the ore conveying parameters and the fixed distance between the imaging station and the action point of the sorting actuator, the predicted arrival time required for each ore particle to move from the imaging moment to the action point of the sorting actuator is calculated.

[0236] Based on the mineral composition category label of each ore particle, determine the corresponding sorting action;

[0237] The sorting execution mechanism is controlled to perform corresponding sorting actions on ore particles located at the predicted physical coordinate positions at the predicted arrival time.

[0238] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0239] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0240] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0241] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0242] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0243] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0244] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A magnetite beneficiation method integrating visual recognition, characterized in that, For the selection phase, the method includes: Macroscopic visual features are extracted from the pre-scanned image data of the current batch of rough and concentrate, and the dominant ore category of the current batch of rough and concentrate is determined based on the macroscopic visual features. The dominant ore category is matched with a pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is called. If the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features. At the preset imaging station, the visual imaging system is controlled to continuously acquire high-definition images of the passing coarse and concentrate ore based on the optimal imaging parameter scheme, thereby obtaining a high-definition image sequence. The high-definition image sequence is input into a pre-trained magnetite image recognition model array, and the output is the mineral composition category label of each ore particle in each frame of high-definition image and the corresponding image coordinate system position coordinates. Based on the image coordinate system position coordinates, mineral composition category labels, and ore transport parameters, the sorting execution mechanism is controlled to perform differentiated sorting actions on the corresponding ore particles.

2. The magnetite beneficiation method integrating visual recognition according to claim 1, characterized in that, Macroscopic visual features are extracted from the pre-scanned image data of the current batch of coarse and concentrate ore, including: At the pre-scanning station, using preset standardized light sources and camera parameters, a high-speed linear array camera is used to acquire pre-scan image data of the current batch of rough and concentrate. Calculate the average gray value and standard deviation of all pixels in the entire batch of pre-scanned image data as a color statistical indicator; Calculate the gray-level co-occurrence matrix of the entire batch of pre-scanned image data, and calculate the average value of the contrast parameter based on the gray-level co-occurrence matrix as a texture statistical index; The percentage of pixels with gray values ​​higher than a preset highlight threshold in the entire batch of pre-scanned image data is used as a gloss statistical indicator. The color statistics, texture statistics, and gloss statistics are combined to form the macroscopic visual characteristics of the current batch of coarse concentrate.

3. The magnetite beneficiation method integrating visual recognition according to claim 1, characterized in that, The dominant ore category of the current batch of rough and concentrate is determined based on the macroscopic visual characteristics, including: Collect pre-scanned images of historical batches of rough and concentrate under preset standardized light source and camera parameters; Macroscopic visual features are extracted from pre-scanned images from each historical batch to form multiple historical feature vectors; Based on the actual sorting results indicators and mineral composition analysis reports corresponding to each historical batch, the historical feature vectors are divided into multiple predefined dominant ore categories; Calculate the arithmetic mean of each dimension of all historical feature vectors corresponding to each dominant ore category to generate a category feature vector representing each dominant ore category; Establish a mapping relationship between each dominant ore category and its corresponding category feature vector, and store it to form an ore feature database; The similarity between the macroscopic visual features of the current batch of crude concentrate and the feature vector of each category in the ore feature database is calculated. The dominant ore category corresponding to the category feature vector with the highest similarity calculation result is selected as the dominant ore category to which the current batch of rough and concentrate belongs.

4. The magnetite beneficiation method incorporating visual recognition according to claim 1, characterized in that, The dominant ore category is matched with a pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is invoked; if the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features, including: In the pre-configured ore category-optimal imaging parameter scheme mapping library, find the record that is the same as the dominant ore category; If the same record is found, the match is successful. The corresponding imaging parameter combination is read from the found record and directly called as the optimal imaging parameter scheme for the current batch. The imaging parameters include at least the light source color, light source intensity and illumination angle. If no matching record is found, the match fails, and based on the macroscopic visual features, with the optimization objective of maximizing the feature separability of the target mineral and gangue in the image, an optimization search is performed within the predefined imaging parameter adjustment space. The combination of imaging parameters that achieves the optimal goal obtained through optimization search is taken as the newly generated optimal imaging parameter scheme. Each newly generated optimal imaging parameter scheme is mapped to the corresponding dominant ore category, and then added to the ore category-optimal imaging parameter scheme mapping library.

5. The magnetite beneficiation method incorporating visual recognition according to claim 4, characterized in that, The process of constructing the ore category-optimal imaging parameter scheme mapping library includes: For each dominant ore category in the ore feature database, representative rough and concentrate samples belonging to the dominant ore category are collected; Within the imaging parameter adjustment space, multiple sets of different imaging parameter combinations are used to acquire images of the representative coarse and concentrate samples, resulting in multiple sets of experimental images. Calculate the visual separation index of the target mineral region and the gangue region for each group of experimental images under the corresponding combination of imaging parameters. For each dominant ore category, the combination of imaging parameters that maximizes the separation index is selected as the optimal imaging parameter scheme for that dominant ore category. Establish a mapping relationship between each dominant ore category and its corresponding optimal imaging parameter scheme, and store the mapping relationship to build an ore category-optimal imaging parameter scheme mapping library.

6. The magnetite beneficiation method incorporating visual recognition according to claim 4, characterized in that, Calculate the visual separation index of the target mineral region and the gangue region in each set of experimental images under the corresponding imaging parameter combination, including: In each set of experimental images, the target mineral area and gangue area are marked; The same multidimensional visual features are extracted from the target mineral region and the gangue region respectively, wherein the visual features include color features, texture features and gloss features; Based on the extracted multidimensional visual features, the average inter-class distance between all samples in the target mineral region and all samples in the gangue region is calculated in the feature space. Calculate the intra-class mean distance of samples within the target mineral region and the intra-class mean distance of samples within the gangue region, respectively. Based on the average inter-class distance, the average intra-class distance of samples within the target mineral region, and the average intra-class distance of samples within the gangue region, the separation index under the combination of imaging parameters is calculated using a preset metric formula.

7. The magnetite beneficiation method incorporating visual recognition according to claim 1, characterized in that, The high-resolution image sequence is input into a pre-trained magnetite image recognition model array, and the output includes the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each high-resolution image frame, including: Construct and pre-train a magnetite image recognition model array, wherein the magnetite image recognition model array consists of a shared feature extraction module, M parallel sub-models and an integrated output module; Based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-definition image sequence, the number N of sub-models to be retrieved from M parallel sub-models for integration is dynamically calculated, where N is an integer between 1 and M. Input the current batch of high-resolution image sequences into a pre-trained magnetite image recognition model array; The shared feature extraction module identifies and outputs the contour region image of each ore particle in a single frame of high-definition image and the corresponding image coordinate system position coordinates; N sub-models, randomly selected from M parallel sub-models, process the contour region image of each ore particle in parallel and output a mineral composition category prediction result for each. The integrated output module integrates the mineral composition category prediction results output by N sub-models, generates the final mineral composition category label for each ore particle, and outputs it together with the image coordinate system position coordinates. The training process of the M parallel sub-models in the magnetite image recognition model array includes: Based on historical production data, the training dataset consists of ore particle images and ore particle outlines in each image collected under different dominant ore categories and different optimal imaging parameter schemes. For each ore particle profile in the training dataset, obtain the corresponding mineral composition ground truth label to form a supervision label set; The training dataset and the supervision label set are divided into M data subset pairs using the M-fold cross-validation method. Use the M data subsets to independently train M sub-models until each sub-model is verified to have converged.

8. The magnetite beneficiation method integrating visual recognition according to claim 7, characterized in that, Based on the dominant ore category and optimal imaging parameter scheme corresponding to the high-resolution image sequence, the number N of sub-models N to be retrieved from M parallel sub-models for integration is dynamically calculated, including: The historical frequency of occurrence of the dominant ore categories is statistically analyzed and used as a commonness coefficient. Calculate the absolute deviation between the optimal imaging parameter scheme and the historical average optimal imaging parameter scheme in each dimension of light source color, light source intensity and illumination angle, and calculate the average value of the absolute deviation in all dimensions as the overall deviation coefficient. Based on the total number M of parallel sub-models, the commonness coefficient, and the overall deviation coefficient, the number N of sub-models to be retrieved from the M parallel sub-models of the magnetite image recognition model array for integration is obtained through weighted summation and rounding. N is negatively correlated with the commonness coefficient and positively correlated with the overall deviation coefficient, and N is an integer between 1 and M.

9. The magnetite beneficiation method incorporating visual recognition according to claim 1, characterized in that, Based on the image coordinate system position coordinates, the mineral composition category labels, and the ore transport parameters, the sorting execution mechanism is controlled to perform differentiated sorting actions on the corresponding ore particles, including: The image coordinate system position coordinates of each ore particle are converted into the predicted physical coordinates of the corresponding ore particle on the sorting station plane according to the preset spatial mapping relationship between the imaging station and the sorting station. Based on the conveyor belt speed in the ore conveying parameters and the fixed distance between the imaging station and the action point of the sorting actuator, the predicted arrival time required for each ore particle to move from the imaging moment to the action point of the sorting actuator is calculated. Based on the mineral composition category label of each ore particle, determine the corresponding sorting action; The sorting execution mechanism is controlled to perform corresponding sorting actions on ore particles located at the predicted physical coordinate positions at the predicted arrival time.

10. A magnetite beneficiation system integrating visual recognition, characterized in that, A magnetite beneficiation method for performing any one of claims 1-9, comprising: The category determination module is used to extract macroscopic visual features from the pre-scanned image data of the current batch of rough and concentrate, and determine the dominant ore category to which the current batch of rough and concentrate belongs based on the macroscopic visual features. The imaging parameter configuration module is used to match the dominant ore category with the pre-configured ore category-optimal imaging parameter scheme mapping library. If the match is successful, the corresponding optimal imaging parameter scheme is called. If the match fails, a new optimal imaging parameter scheme is generated based on the macroscopic visual features. The image acquisition module is used to control the visual imaging system to continuously acquire high-definition images of the rough and concentrate ore at a preset imaging station based on the optimal imaging parameter scheme, so as to obtain a high-definition image sequence. The sorting and recognition module is used to input the high-definition image sequence into a pre-trained magnetite image recognition model array and output the mineral composition category label and corresponding image coordinate system position coordinates of each ore particle in each frame of high-definition image. The sorting execution module is used to control the sorting execution mechanism to perform differentiated sorting actions on the corresponding ore particles based on the image coordinate system position coordinates, mineral composition category labels and ore transport parameters.

Images