Phosphate ore sorting method, electronic device, and medium

Abstract

The application provides a phosphate ore sorting method, an electronic device and a medium. In a preprocessing stage, a high-energy image obtained from raw ore is registered with a low-energy image as a reference, a difference image representing a gray difference is obtained from the registered high-energy image and the low-energy image, and the low-energy image, the registered high-energy image and the difference image are fused in three channels to obtain a pseudo-color image. In an intelligent identification stage, the pseudo-color image is taken as input, and a target detection model based on deep learning is used for processing, so that accurate and efficient identification of phosphate ore in the raw ore is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mineral separation, in particular to a phosphate rock separation method, an electronic device and a medium. BACKGROUND

[0002] With the increasing demand for global food security and sustainable agricultural development, phosphate rock, as an important resource for producing phosphate fertilizer and fine chemical products, has an irreplaceable strategic position in the fields of agriculture, chemical industry, food, medicine and new energy materials.

[0003] In the pre-selection and waste-throwing process of phosphate rock, a large amount of gangue minerals such as dolomite and calcite are often mixed in the raw ore, which has a wide particle size distribution, similar color and blurred boundary, and seriously restricts the separation accuracy and efficiency of phosphate rock. SUMMARY

[0004] The present application provides a phosphate rock separation method, an electronic device and a medium, which can solve one of the problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] In a first aspect, a phosphate rock separation method is provided, comprising:

[0007] In the preprocessing stage, an original high-energy image and an original low-energy image of a raw ore are obtained, the original high-energy image and the original low-energy image are obtained by dual-energy X-ray projection imaging of the raw ore; using an enhanced correlation coefficient registration algorithm, the original low-energy image is used as a reference, and the original high-energy image is registered to obtain a secondary high-energy image; from the secondary high-energy image and the original low-energy image, a difference image representing the gray difference between the secondary high-energy image and the original low-energy image is obtained; and the secondary high-energy image, the original low-energy image and the difference image are fused according to three channels to obtain a pseudo-color image; and

[0008] In the intelligent identification stage, the pseudo-color image is used as the input of a target detection model based on deep learning, and the separation result of the phosphate rock in the raw ore is obtained by processing of the target detection model.

[0009] Based on the above technical solutions, in the preprocessing stage, the obtained high-energy image of the raw ore is registered with the low-energy image as a reference, the difference image representing the gray difference is obtained from the registered high-energy image and low-energy image, and then the low-energy image, the registered high-energy image and the difference image are fused in three channels to obtain a pseudo-color image. In the intelligent identification stage, the pseudo-color image is used as the input, and the target detection model based on deep learning is used for processing, so as to realize the accurate and efficient identification of phosphate rock in the raw ore.

[0010] In a possible design of the first aspect, the target detection model adopts a YOLO model, and the YOLO model includes a feature extraction network, a feature fusion network, and a detection head.

[0011] In a possible design of the first aspect, a lightweight convolution block attention module is integrated in a path aggregation based feature pyramid network of the feature fusion network, and the lightweight convolution block attention module is configured to perform attention guidance on a channel and a spatial dimension of an intermediate feature map.

[0012] In a possible design of the first aspect, in the YOLO model, a convolution layer adopts a HardSwish activation function.

[0013]

[0014] wherein x represents an input of the HardSwish activation function, and ReLU6() represents a truncated form of a ReLU activation function in a range of 0 to 6.

[0015] In a possible design of the first aspect, the YOLO model adopts an Optuna hyperparameter automatic search framework.

[0016] In a possible design of the first aspect, the preprocessing stage further includes: performing gray scale enhancement processing and image normalization processing on the secondary high-energy image and the original low-energy image.

[0017] In a possible design of the first aspect, the gray scale enhancement processing adopts a contrast limited adaptive histogram equalization processing that divides an image into 12*12 grid regions.

[0018] In a possible design of the first aspect, the preprocessing stage further includes: performing background color replacement on the pseudo-color image to remove background noise interference.

[0019] In a second aspect, an electronic device is provided, including a processor and a memory coupled to the processor, the memory being configured to store a computer program, and the processor being configured to execute the computer program stored in the memory, so that the electronic device performs the phosphate ore sorting method according to any possible implementation manner of the first aspect.

[0020] In a third aspect, a computer readable storage medium is provided, including a computer program or instructions, when the computer program or instructions are executed on a computer, the computer program or instructions cause the computer to perform the phosphate ore sorting method according to any possible implementation manner of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or related description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0022] Figure 1 is an X-ray diffraction analysis result schematic diagram of a phosphate rock sample provided by the present application;

[0023] Figure 2 is a collophane microstructure feature diagram provided by the present application;

[0024] Figure 3 is a phosphate rock main mineral particle size distribution diagram provided by the present application;

[0025] Figure 4 is a dual-energy XRT gray scale diagram of a sample provided by the present application, wherein, Figure 4 (a) is a high-energy phosphate rock XRT gray scale diagram, Figure 4 (b) is a low-energy phosphate rock XRT gray scale diagram;

[0026] Figure 5 is a gray scale histogram of a dual-energy XRT image of a phosphate rock provided by the present application;

[0027] Figure 6 is a high-energy phosphate rock diagram obtained after ECC registration change provided by the present application;

[0028] Figure 7 is a schematic diagram of the result of CLAHE processing of the phosphate rock gray scale diagram after ECC registration provided by the present application;

[0029] Figure 8 is an original three-channel pseudo-color diagram provided by the present application;

[0030] Figure 9 is a three-channel pseudo-color diagram processed provided by the present application;

[0031] Figure 10 is a YOLO model structure schematic diagram provided by the present application;

[0032] Figure 11 is a CBAM structure schematic diagram provided by the present application;

[0033] Figure 12 is a HardSwish and SiLU activation function schematic diagram provided by the present application;

[0034] Figure 13This is a flowchart of the phosphate rock identification algorithm implementation scheme provided in the embodiments of this application;

[0035] Figure 14 This is a schematic diagram illustrating the results of identifying XRT false-color images of phosphate rock provided in an embodiment of this application. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0037] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0039] Before introducing the embodiments of this application, a brief description of the current stage of technical research of this application will be given:

[0040] The elemental chemical composition of the phosphate rock samples collected in the study is shown in Table 1. Overall, the chemical composition of the samples is dominated by oxygen-containing compounds, exhibiting obvious multi-element symbiosis and complex compositional characteristics, providing a data foundation for subsequent process mineralogical research and optimization of intelligent mineral processing algorithms.

[0041] Table 1. Elemental chemical composition of mineral samples

[0042]

[0043] The X-ray diffraction (XRD) analysis results of the sample are shown in the figure. Figure 1 .

[0044] The phase composition of the samples was determined using an automated mineralogy image analysis system (AMICS), as shown in Table 2.

[0045] Table 2 Phase composition of phosphate rock samples

[0046]

[0047] Depend on Figure 1As shown in Table 2, the main minerals in the ore sample are dolomite and collophanite, with mass fractions of 48.84% and 30.19% respectively, and also containing a small amount of quartz, calcite, kaolinite and pyrite, with mass fractions of 8.02%, 3.98%, 1.62% and 1.09% respectively. Collophanite is the main useful mineral, and dolomite is a triclinic carbonate mineral with a crystal structure similar to calcite, which is the main gangue mineral.

[0048] The microstructure characteristics of the collophanite sample are shown in Figure 2 The figure reveals that the collophanite sample has various microstructures, which reflect the typical genetic characteristics of phosphorite in the sedimentary environment and show that there is a close cementation between the phosphorite and carbonate minerals such as dolomite. These characteristics have important reference value for judging the genesis of phosphorite, the sedimentary environment and the processing and utilization mode of the ore.

[0049] The embedded particle size characteristics are important properties of phosphorite, and the particle size distribution of the main component minerals in the phosphorite is shown in Figure 3 As shown in Figure 3

[0050] (1) The particle size distribution of the overall minerals is relatively uniform, and the proportion of the particle size greater than 150 μm and the particle size less than 75 μm is close.

[0051] (2) The particle size distribution of the main minerals dolomite and collophanite is basically the same, mainly less than 75 μm, followed by greater than 150 μm. The particle size of minerals such as calcite, quartz and pyrite is relatively fine, of which 70%~80% is less than 75 μm.

[0052] (3) The particle size of the sulfur-phosphorus-aluminum-strontium mineral is mainly 75~150 μm, accounting for nearly 50%, followed by less than 75 μm.

[0053] The particle size of the phosphorus-aluminum-calcium stone is relatively small, of which 63.30% is less than 75 μm.

[0054] According to the analysis of the process mineralogy of the sample, it is known that the sample is a phosphatic rock type ore, with a P2O5 grade of only 12.05%, belonging to a low-grade phosphorite. The main non-metallic minerals in the sample are dolomite and collophanite, with mass fractions of 48.84% and 30.19% respectively, followed by quartz and calcite, and the main metallic mineral is pyrite.

[0055] The apatite is microcrystalline, the collophanite is amorphous or microcryptocrystalline structure, the intergrowth relationship is complex, and the collophanite is mainly fine-grained irregularly embedded, with a particle size mainly less than 75 μm. At the same time, the collophanite is associated with fine-grained dolomite, most of which is oolitic structure, so it is difficult to remove dolomite by flotation.

[0056] ​Based on the mineralogical information of the phosphate rock sample, its main minerals are dolomite, collophane, and quartz, all of which have relatively similar densities. Dolomite and quartz have higher grayscale values ​​due to their strong X-ray penetrability, while collophane has a lower grayscale value due to its strong X-ray attenuation ability. Therefore, the image enhancement strategy should focus on improving grayscale contrast. In pseudo-color mapping, low grayscale areas can be mapped to blue to enhance the visual characteristics of the target minerals. The dual-energy XRT grayscale image of the sample is shown below. Figure 4 As shown.

[0057] Depend on Figure 4 It can be seen that the dual-energy XRT images of the phosphate rock samples clearly show the different responses of high and low energy X-rays to different minerals, with low energy highlighting the differences in composition and high energy ensuring penetration and morphological fidelity. Figure 4 (a) and Figure 4 The visual differences in (b) are minimal due to the similar densities of the main minerals in the sample: dolomite, phosphate rock, and quartz. Therefore, machine-based methods are needed for phosphate rock separation. Combining these two methods provides an image data foundation for subsequent differential operations, pseudo-color image construction, and deep learning-based AI recognition algorithms.

[0058] Based on this, this embodiment introduces a difference fusion strategy in the construction of pseudo-color images, and integrates dual-energy features in model training to fully explore the density response information under different energy spectra and improve the algorithm's ability to intelligently distinguish between phosphate rock areas and non-target mineral areas.

[0059] The preprocessing scheme based on the dual-energy XRT grayscale image of phosphate rock constructed in this embodiment integrates physical modeling and visual mapping, greatly enhancing the visualization features of phosphate rock XRT images. This provides a high-quality dataset input foundation for the accurate identification of deep learning AI algorithms, demonstrating significant engineering practical value and innovation. The grayscale histogram of the dual-energy XRT image of phosphate rock is shown below. Figure 5 As shown.

[0060] right Figure 5 Analysis reveals that phosphate rock and waste rock exhibit different responses at low and high energies, with their histogram distribution ranges staggered, indicating potential for enhancing contrast using dual-energy ratios or differential methods. Simultaneously, the large proportion of background pixels necessitates noise reduction processing.

[0061] In dual-energy X-ray transmission imaging of phosphate rock, slight spectral shifts or mechanical disturbances often occur between the high-energy and low-energy images during imaging, resulting in sub-pixel-level spatial displacement errors between the images. To achieve accurate fusion and pseudo-color mapping, precise alignment of the image pairs is necessary.

[0062] The enhanced correlation coefficient (ECC) registration algorithm is a registration strategy based on maximizing the enhanced correlation coefficient in the image gray space, which can effectively overcome the micro-displacement error caused by material vibration and belt movement, thereby improving the accuracy and stability of subsequent image fusion and pseudo-color map generation. Given a reference image and a to-be-registered image , the main goal of the ECC algorithm is to transform into ) through an affine transformation , so that the enhanced correlation coefficient between the two images is maximized. The ECC optimization objective function is as follows:

[0063]

[0064] In the above formula, denotes the image coordinates; denotes the pixel value of the reference image; denotes the to-be-registered image transformed by the affine transformation parameters ; and and denote the mean values of the two images, respectively, denotes the enhanced correlation coefficient. The ECC registration takes the low-energy image as the reference and registers the high-energy image, and through iterative optimization of the affine transformation parameters , the two images are made most similar in the gray statistical sense. This step is the key basis for subsequent pseudo-color map generation and mineral identification. In this embodiment, the low-energy image is selected as the reference because the low-energy image has weaker penetration ability, higher sensitivity to internal details of the ore, and clearer and more stable texture and edge information, and therefore, as a reference, it is more conducive to the registration of the high-energy image and the subsequent pseudo-color map generation. The high-energy phosphorite image obtained after ECC registration is shown in Figure 6 .

[0065] To further improve the gray contrast between the ore and the background in the registered image and enhance the expression of local detail information, this embodiment introduces the contrast-limited adaptive histogram equalization (CLAHE) data enhancement method to perform gray enhancement processing on the registered result image after completing the image registration based on the ECC algorithm.

[0066] Unlike traditional histogram equalization, which operates only across the entire image, CLAHE divides the image into several small grids, performing local histogram equalization on each grid and then fusing edge regions using bilinear interpolation. This effectively enhances image details while suppressing noise caused by over-enhancement. By setting the contrast enhancement threshold to its maximum value, the grayscale differences at the edges of the ore region can be effectively amplified. CLAHE divides the image into an 8×8 grid region by default; in this embodiment, to balance local image detail and overall equalization, the image is divided into a 12×12 grid region. The result of CLAHE processing on the ECC-registered phosphate rock grayscale image is shown below. Figure 7 As shown.

[0067] To fully explore the response differences of phosphate rock under high-energy and low-energy X-ray images, this embodiment designs a reasonable image preprocessing and fusion strategy. A pseudo-color difference map is constructed as input to the AI ​​recognition algorithm, thereby enhancing the density comparison between ore and waste rock and improving recognition accuracy. The specific processing flow is as follows:

[0068] (1) Image normalization processing: In order to eliminate the influence of image brightness differences on subsequent analysis, the gray values ​​of the high-energy image after the above ECC registration and CLAHE enhancement and the original low-energy image are uniformly compressed to the [0,1] range to enhance the contrast consistency between different images.

[0069] (2) Difference map generation: Based on the normalization result, the grayscale difference between the high-energy map and the low-energy map is calculated pixel by pixel to obtain a single-channel difference map. This map can effectively reflect the changes in the penetration characteristics of the ore under different energy X-rays, highlighting the difference in grayscale response between the target ore and waste rock. The specific mathematical formula is as follows:

[0070]

[0071]

[0072]

[0073] In the above formula, Indicates high-energy XRT images at the pixel level The result of normalizing the grayscale value at the maximum value. This represents the result of normalizing the grayscale values ​​of a low-energy XRT image to the maximum value. for These are high-energy and low-energy XRT images at the pixel level. grayscale at that location It is a minimal constant that prevents it from taking the value of 0. It can achieve nonlinear control. Indicates in At the pixel, based on the above normalization result and The calculated power difference feature map.

[0074] (3) Pseudo-color map construction: To further enhance the image level and feature sensitivity, the high-energy map, low-energy map and difference map are fused in a three-channel manner to construct a pseudo-color image. The specific method is: the R channel input is the power difference feature map, the G channel input is the normalized low-energy map, and the B channel input is the normalized high-energy map. The pseudo-color image generated in this way not only visually shows the contrast difference between ore and waste rock, but also provides multi-dimensional input information for subsequent deep learning algorithm training and optimization.

[0075] After the above steps, the neural network can fully utilize the complementarity of dual-energy images and capture the intensity gradient information of mineral response in the initial feature extraction stage. The final three-channel pseudo-color map is shown in Figure 8 .

[0076] In addition, in order to effectively remove background interference, a BGR color mask is constructed and the redundant red background is replaced with pure white to improve the saliency of the target ore, as shown in Figure 9 . Figure 9 Not only has good mineral differentiation in vision, but also can provide more discriminative image input for AI recognition algorithm, which verifies the effectiveness of the overall preprocessing strategy.

[0077] This embodiment collects and processes dual-energy X-ray image data set of phosphate ore, and converts the gray image into three-channel color image with high discrimination through the preprocessing strategy of fusing ECC and CLAHE and the pseudo-color mapping algorithm. Ensure the sufficiency of the training process and the generalization ability of the performance, and unify the image resolution and channel standard, which is beneficial to the subsequent input adaptation of YOLO algorithm.

[0078] For the optimization design of phosphate ore recognition algorithm, this embodiment focuses on the following three deep fusion strategies to achieve high robustness and high precision target detection effect under complex background and fuzzy boundary conditions. These three methods are closely related to the characteristics of dual-energy XRT image, combined with the structural flexibility of YOLO model, to realize pseudo-color map driven recognition optimization, as shown in Figure 10 .

[0079] In order to further improve the perception ability of YOLO to the characteristics of phosphate ore and enhance the key information expression effect in the multi-scale feature fusion process, this embodiment introduces Figure 11The illustrated lightweight convolutional block attention module (CBAM) is composed of two parts in series: channel attention and spatial attention. First, the channel description is generated by global average pooling and maximum pooling, and then the channel attention weight is calculated by the fully connected layer with shared weights, which weights the input features in the channel dimension. Then the weighted feature map is averaged and maximized in the spatial dimension, and the spatial attention map is generated by convolution operation after splicing, which weights the input features in the spatial position, thereby enhancing the network's expression ability for key regions and channel features while keeping lightweight.

[0080] Integrating CBAM into the feature fusion network (Neck) part of YOLO based on the path aggregation feature pyramid network (PAN-FPN). While keeping the overall structure of the original algorithm framework unchanged, the attention guidance of the channel and spatial dimension of the intermediate feature map effectively improves the neural network's recognition ability and region selectivity for mineral density response regions and reduces background noise interference.

[0081] Specifically, this embodiment introduces the CBAM module into the PAN-FPN feature fusion structure of YOLOv8, and performs attention enhancement on the multi-scale features from the feature extraction network (Backbone) during the fusion process after upsampling and downsampling. In this way, YOLOv8 can adaptively identify and highlight key regions when interacting with multi-level features, thereby improving the recognition ability for boundary fuzzy ores, density gradient transition regions, and noise background interference.

[0082] In summary, the introduction of CBAM not only improves the recognition accuracy of YOLOv8 in complex working conditions, but also maintains the simplicity of the network structure and the efficiency of deployment, making it suitable for application in industrial-grade online intelligent beneficiation scenarios.

[0083] In YOLO, the default activation function used is the Sigmoid linear unit (SiLU), which performs well in regular tasks. However, in the specific phosphor ore pseudo-color image recognition task, replacing the activation function is an effective strategy to further improve the algorithm's nonlinear expression ability for small targets.

[0084] This embodiment replaces the SiLU activation function with the HardSwish activation function, which is a computationally efficient and high-performance activation function widely used in lightweight deep learning models. While maintaining the model's nonlinear ability, HardSwish effectively reduces the inference delay, making it very suitable for deployment on resource-constrained beneficiation edge platforms. Its mathematical formula is as follows:

[0085]

[0086] In the above formula, x represents the input of the activation function, and ReLU6() represents the truncated form of the ReLU activation function in the range of 0 to 6.

[0087] This function has the following advantages: first, it is computationally efficient, as compared to SiLU, which involves exponential calculation, HardSwish is composed entirely of linear and ReLU, and the inference time is shorter; second, it has stronger response capability to small amplitude perturbations, improving the accuracy of ore boundary detection; third, it has good deployment compatibility, supporting fast inference and deployment on edge platforms such as Ascend Atlas, and the computation graph is more concise; finally, after combining with the pseudo-color feature channel input, HardSwish can enhance the algorithm's nonlinear response capability to weak contrast and strong noise regions. The images of HardSwish and SiLU activation functions are as shown in Figure 12 .

[0088] In the YOLOv8 structure optimization process of the present embodiment, HardSwish is preferentially selected as the activation function to construct a lightweight high-performance neural network structure more suitable for the phosphate XRT pseudo-color image recognition scenario. The replacement method specifically includes replacing all SiLU activation units in the Backbone and Neck parts one by one, while keeping the detection head (Head) detection structure unchanged to preserve the original recognition accuracy structure optimization path. From the curve shape, HardSwish and SiLU are both close to linear in the positive interval, and both show a smooth saturation trend in the negative interval, so the difference between their output curves is indeed not large. However, HardSwish is an approximate piecewise linear function, which is simpler to calculate, faster to infer, and has smaller memory overhead than SiLU function, making it very suitable for lightweight high-performance network deployment scenarios; while maintaining similar nonlinear features, HardSwish can significantly reduce inference time. This is the reason why HardSwish is chosen to replace SiLU in the YOLOv8 optimization of the present embodiment.

[0089] To further improve the accuracy and inference efficiency of the phosphate pseudo-color image recognition AI recognition algorithm, the present embodiment introduces the Optuna hyperparameter automatic search framework based on the improved YOLO structure, and performs Bayesian optimization again for key training parameters such as learning rate, momentum coefficient, and weight decay. Compared with traditional manual parameter tuning methods, Optuna can quickly approach the optimal hyperparameter combination within a limited time, effectively accelerating the model convergence process. This optimization strategy realizes the dual improvement of accuracy and speed of the phosphate recognition algorithm under the premise of keeping the inference time basically unchanged, providing a more efficient and robust parameter configuration for the deployment of phosphate intelligent beneficiation algorithms. The best results are shown in Table 3.

[0090] Table 3 Optimal hyperparameter configuration for training of phosphate XRT pseudo-color images

[0091]

[0092] Compared with the prior art, the beneficial effects of this embodiment are as follows:

[0093] In the preprocessing stage of dual-energy XRT images, ECC image registration, CLAHE data augmentation, pseudo-color mapping, and denoising techniques were employed to enhance the feature representation of the target ore in the images. In the algorithm optimization stage, a lightweight convolutional attention module (CBAM) was further introduced, the HardSwish activation function was replaced, and the Optuna hyperparameter optimization framework was invoked to improve the accuracy and speed of the YOLO-based phosphate rock AI recognition algorithm. Through these optimizations, the model's recognition accuracy was improved to 95.8%. The final YOLO algorithm network results are shown below. Figure 10 As shown.

[0094] This embodiment first performs multi-stage optimization preprocessing on dual-energy XRT images of phosphate ore, and then optimizes the recognition algorithm to obtain a lightweight AI recognition algorithm for phosphate ore using YOLO. This algorithm has good versatility and scalability, and can be widely applied in industrial fields such as photoelectric intelligent sorting and pre-selection of waste in phosphate mines.

[0095] like Figure 13 As shown, the specific implementation details of the phosphate rock identification algorithm implementation scheme in this embodiment are as follows:

[0096] Data Acquisition and Data Set Partitioning: Obtain a dual-energy XRT image dataset of phosphate rock samples, including high-energy and low-energy channels. The acquired raw data is then rationally partitioned into training, validation, and test sets to ensure the scientific rigor and reliability of algorithm training and performance evaluation.

[0097] Dual-energy XRT image preprocessing: Registration and enhancement of high-energy and low-energy images. First, the ECC registration algorithm is used to ensure the spatial consistency of the high-energy and low-energy images, avoiding offset errors in subsequent processing. Then, the CLAHE method is used to enhance image contrast, highlighting the subtle grayscale differences between the target ore and waste rock.

[0098] Feature pseudocolor image mapping and denoising: After preprocessing, dual-energy image difference calculation is performed to obtain a single-channel feature map that reflects the density difference of the ore. Furthermore, the high-energy image, low-energy image, and difference image are combined to construct a three-channel pseudocolor image, while noise suppression processing is performed to improve the discriminative power and stability of the model's input data.

[0099] Optimization of YOLO algorithm: First, add CBAM attention mechanism in the Neck feature pyramid structure to enhance the focusing ability of the algorithm on key areas and improve the recognition performance of ore boundaries and small targets in complex backgrounds. Then replace the HardSwish activation function to improve the non-linear representation ability of the network and improve the convergence speed. Finally, combine the Optuna hyperparameter search framework to automatically optimize the learning rate, batch size, optimizer and other training hyperparameters to ensure the global optimal balance of accuracy and speed of the model.

[0100] Algorithm training and verification: input the constructed three-channel pseudo-color image into the optimized YOLO model for training and reasoning to realize automatic recognition and classification of phosphate ore and waste rock.

[0101] Result display and analysis: display and evaluate the AI algorithm recognition results through visualization software, intuitively present the rectangular recognition box, class label and confidence information as shown in Figure 14 to verify the applicability and reliability of the method in industrial scenarios.

[0102] This embodiment identifies the XRT pseudo-color image of phosphate ore, and the result is shown in FIG. 14. In the figure, “ore” represents the ore target class recognized by the model, and “waste” represents the waste target class recognized by the model; the numbers above them are the confidence values output by the model, indicating the reliability of the model's judgment on the class of the detection box. The closer the value is to 1.0, the higher the model's confidence in the target belonging to the class. It can be seen that the YOLO optimization algorithm proposed in this embodiment can accurately distinguish between ore and waste rock, and the rectangular box coverage is accurate, with a confidence value of more than 0.7. At the same time, background noise is not misdetected, proving that this embodiment has good robustness and practical application value in complex working conditions. In actual industrial applications, pseudo-color images can be used as auxiliary features and combined with grayscale images for multi-modal fusion to improve the performance of existing models.

[0103] The embodiments of the present application also provide an electronic device, comprising: a processor, and a memory coupled to the processor, the memory being configured to store a computer program; and the processor being configured to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of the above embodiments.

[0104] The electronic device can be a desktop computer, a notebook computer, a palm computer, a cloud server and other computing devices. The electronic device can include, but is not limited to, a processor and a memory.

[0105] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The processor is a control center of the electronic device, and connects all parts of the device through various interfaces and lines.

[0106] The memory can be used to store the computer program, and the processor realizes various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0107] The memory can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application required by a function, etc. The data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0108] The embodiment of the present application further provides a storage medium, which is a computer readable storage medium, and the computer program is stored in the computer readable storage medium. When the computer program is executed by a processor, the steps of each method embodiment described above can be realized. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0109] The embodiment of the present application further provides a computer program product, comprising a computer program or instructions, which, when running on a computer, enable the computer to perform the method of any possible implementation manner described above.

[0110] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the technology field, without departing from the principle of the present application, can make several improvements and refinements, these improvements and refinements also be considered as the protection scope of the present application.

Claims

1. A method for separating phosphate rock, characterized in that, include: In the preprocessing stage, the original high-energy image and the original low-energy image of the raw ore are obtained. The original high-energy image and the original low-energy image are obtained by dual-energy X-ray projection imaging of the raw ore. Using the enhanced correlation coefficient registration algorithm, the original high-energy image is registered with the original low-energy image as a reference to obtain the secondary high-energy image. From the secondary high-energy image and the original low-energy image, a difference image is obtained that characterizes the grayscale difference between the secondary high-energy image and the original low-energy image; and the secondary high-energy image, the original low-energy image and the difference image are fused according to three channels to obtain a pseudo-color image; as well as, In the intelligent recognition stage, the pseudo-color image is used as the input to a deep learning-based target detection model. Through the processing of the target detection model, the sorting results of phosphate ore in the raw ore are obtained.

2. The phosphate rock beneficiation method as described in claim 1, characterized in that, The target detection model adopts the YOLO model, which includes a feature extraction network, a feature fusion network, and a detection head.

3. The phosphate rock beneficiation method as described in claim 2, characterized in that, A lightweight convolutional block attention module is integrated into the path-aggregation-based feature pyramid network of the feature fusion network. The lightweight convolutional block attention module is used to guide the attention of intermediate feature maps in terms of channel and spatial dimensions.

4. The phosphate rock beneficiation method as described in claim 2, characterized in that, In the YOLO model, the convolutional layers use the HardSwish activation function: Where x represents the input to the HardSwish activation function, and ReLU6() represents the truncated form of the ReLU activation function in the range of 0 to 6.

5. The phosphate rock beneficiation method as described in claim 2, characterized in that, The YOLO model uses the Optuna hyperparameter automatic search framework.

6. The phosphate rock beneficiation method as described in claim 1, characterized in that, The preprocessing stage further includes: performing grayscale enhancement processing and image normalization processing on the secondary high-energy image and the original low-energy image.

7. The phosphate rock beneficiation method as described in claim 6, characterized in that, The grayscale enhancement process employs contrast-limited adaptive histogram equalization, which divides the image into 12×12 grid regions.

8. The phosphate rock beneficiation method as described in claim 1, characterized in that, The preprocessing stage also includes: replacing the background color of the pseudo-color image to remove background noise interference.

9. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; and The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the phosphate rock sorting method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program or instructions that, when executed on a computer, cause the computer to perform the phosphate rock sorting method as described in any one of claims 1-8.

Images