Lead-zinc ore sorting method and electronic device

Abstract

The application provides a lead-zinc ore sorting method and an electronic device. In a preprocessing stage, a high-energy image of raw ore is registered with a low-energy image as a reference, a physical response difference is measured, the low-energy image and the registered high-energy image are fused to obtain a fusion image, the fusion image is colored according to a color transition rule to obtain a pseudo-color image, and the pseudo-color image is taken as input to process a target detection model based on deep learning, so that accurate and efficient identification of lead-zinc ore in the raw ore is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of mineral separation, in particular to a lead-zinc ore separation method and an electronic device. BACKGROUND

[0002] As one of important polymetallic minerals in China, lead-zinc ore is widely used in key fields such as batteries, cables, alloy smelting, corrosion-resistant coatings and photovoltaic materials, and occupies an irreplaceable position in the national defense industry and high-tech industry. China is rich in lead-zinc resources, with various types of mineralization and widely distributed deposits, but also faces many challenges such as declining ore grade, associated and symbiotic complexity, and increasing difficulty of ore dressing.

[0003] In the pre-selection and waste-throwing process of lead-zinc ore, galena is relatively easy to identify due to its large density, but there are problems such as close particle size distribution and overlapping density of sphalerite, pyrite and dolomite gangue minerals. Traditional manual selection and conventional image processing methods cannot achieve accurate and rapid ore and waste rock identification, resulting in waste of mineral resources and decline in enterprise economic benefits. SUMMARY

[0004] The application provides a lead-zinc ore separation method and an electronic device, which can solve one of the problems in the background art.

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

[0006] In a first aspect, a lead-zinc ore separation method is provided, comprising:

[0007] In the preprocessing stage, an original high-energy image and an original low-energy image of 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; using a physical response difference measure, the original low-energy image and the secondary high-energy image are fused to obtain a fusion image; and according to a color transition rule, the fusion image is colored 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 through the processing of the target detection model, the separation result of lead-zinc ore in the raw ore is obtained.

[0009] In a possible design manner of the first aspect, the physical response difference measure uses a normalized logarithmic difference :

[0010]

[0011]

[0012]

[0013] wherein, is a secondary high-energy map, is an original low-energy map, are maximum gray scale values in the secondary high-energy map and the original low-energy map, respectively, is a minimum constant set to prevent the denominator from being 0.

[0014] 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.

[0015] In a possible design of the first aspect, the YOLO model adopts a weighted intersection over union loss function L WIoU :

[0016]

[0017] wherein, denotes the Euclidean distance square of the center of the prediction box and the center of the real box , and denotes the diagonal length of the minimum circumscribed rectangle capable of enclosing the prediction box and the real box, is a dynamic weight factor for adjusting the gradient contribution of different samples.

[0018] In a possible design of the first aspect, the feature fusion network includes a C2f layer and a Concat layer, and an efficient multi-scale attention layer is arranged between the C2f layer and the Concat layer, where the efficient multi-scale attention layer is used for grouped convolution and multi-scale modeling of input features, extraction of channel attention and spatial attention information, and fusion through normalization and weight distribution mechanism.

[0019] In a possible design of the first aspect, the detection head includes a lightweight adaptive spatial feature fusion module for adaptively assigning fusion weights of different scale feature maps according to the importance of each position.

[0020] In a possible design of the first aspect, the feature extraction network adopts a ghost convolution layer.

[0021] In a possible design of the first aspect, the YOLO model adopts an Optuna hyperparameter search framework.

[0022] In a possible design of the first aspect, the preprocessing stage further includes: performing gamma correction on the secondary high-energy image and the original low-energy image.

[0023] In a second aspect, an electronic device is provided, which includes 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 lead-zinc ore sorting method according to any possible implementation manner of the first aspect.

[0024] Beneficial effects:

[0025] Based on the above technical solution, in the preprocessing stage, the obtained high-energy image of the raw ore is registered based on the low-energy image, the physical response difference is measured, the low-energy image and the registered high-energy image are fused to obtain a fusion image, the fusion image is colored according to the color transition rule to obtain a pseudo-color image, and the pseudo-color image is taken as input to process by using the target detection model based on deep learning, so that the accurate and efficient identification of the lead-zinc ore in the raw ore is realized. BRIEF DESCRIPTION OF DRAWINGS

[0026] 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 technical descriptions 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.

[0027] Figure 1 is a lead-zinc ore XRD quantitative analysis spectrum provided by the embodiments of the present application;

[0028] Figure 2 is a lead-zinc ore raw ore dual-energy XRT gray-scale image provided by the embodiments of the present application, wherein, Figure 2 (a) is a high-energy lead-zinc ore XRT gray-scale image, Figure 2 (b) is a low-energy lead-zinc ore XRT gray-scale image;

[0029] Figure 3 is a lead-zinc ore dual-energy XRT gray-scale two-dimensional joint histogram provided by the embodiments of the present application;

[0030] Figure 4 is a high-energy lead-zinc ore XRT image registration result provided by the embodiments of the present application;

[0031] Figure 5 is a high-energy lead-zinc ore image Gamma correction result provided by the embodiments of the present application;

[0032] Figure 6 is a lead-zinc mine pseudo-color map mapping result provided by an embodiment of the present application;

[0033] Figure 7 is a lead-zinc mine pseudo-color map denoising result provided by an embodiment of the present application;

[0034] Figure 8 is a L WIoU Loss function schematic diagram;

[0035] Figure 9 is an EMA module schematic diagram provided by an embodiment of the present application;

[0036] Figure 10 is a lightweight adaptive spatial feature fusion module schematic diagram provided by an embodiment of the present application;

[0037] Figure 11 is a GhostConv schematic diagram provided by an embodiment of the present application;

[0038] Figure 12 is a lead-zinc mine AI recognition network structure diagram provided by an embodiment of the present application;

[0039] Figure 13 is a lead-zinc mine recognition algorithm implementation scheme flowchart provided by an embodiment of the present application;

[0040] Figure 14 is a lead-zinc mine recognition result provided by an embodiment of the present application. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0042] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed 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 and claims and the above drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0043] 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 the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0044] Before introducing the embodiments of the present application, a brief description of the present stage of the technical research of the present application is given:

[0045] The process mineralogy research was carried out on the lead-zinc ore sample collected in this embodiment, the obtained raw ore multi-element chemical analysis results are shown in Table 1, the lead-zinc-silver chemical phase analysis results are shown in Tables 2-4, the X-ray diffraction (XRD) analysis results are shown in Figure 1 .

[0046] Table 1 Raw ore multi-element chemical analysis

[0047]

[0048] Table 2 Lead phase chemical analysis in raw ore

[0049]

[0050] Table 3 Zinc phase chemical analysis in raw ore

[0051]

[0052] Table 4 Silver phase chemical analysis in raw ore

[0053]

[0054] From the process mineralogy research, it can be seen that the content of the valuable element silver in the raw ore is 215.88 g / t; the content of lead is 2.50%; and the content of zinc is 3.73%. Among them, the silver in the form of elemental silver accounts for 19.65%, silver sulfide accounts for 63.94%, and the silver in the sulfide ore accounts for 13.49%; the lead mainly exists in the form of lead sulfide, accounting for 85.25%, and the white lead ore accounts for 5.88%; and the zinc mainly exists in the form of zinc sulfide, accounting for 86.86%, and the zinc oxide accounts for 12.65%. From the process mineralogy research, it can be seen that the main metallic minerals in the raw ore are sphalerite, galena, pyrite, siderite, etc.; and the non-metallic minerals are mainly quartz, chlorite, feldspar, mica, etc. Figure 2

[0055] It can be seen from the process mineralogy that the imaging of the lead-zinc ore under X-ray has the following characteristics: the atomic number of the high atomic number element (Pb) is 82, the density of galena is about 7.5-7.6 g / cm³, and the X-ray has a strong absorption capacity, so it presents the characteristics of low gray value and dark imaging in the XRT image, and presents a relatively strong contrast in the XRT gray value image of different energy levels; however, the atomic number of zinc (Zn) is relatively small, only 30, and the density of sphalerite is about 3.9-4.2 g / cm³, which is slightly higher than the main non-metallic gangue such as quartz and chlorite, and the image gray value of a small amount of pyrite and other interference components in the waste rock is close, and the dual-energy XRT gray value image of the lead-zinc ore is shown in Figure 2 ​​

[0056] By Figure 2 It can be seen that in the high-energy image, the imaging clarity of high-density minerals is higher, and obvious attenuation characteristics are shown, and the image shows a darker imaging feature. In the low-energy image, the penetration ability of X-rays is weak, resulting in more balanced overall imaging gray scale, but the gray scale of ore and waste rock is close, and there is a large recognition interference.

[0057] The image lead-zinc ore preprocessing scheme based on dual-energy XRT gray scale image constructed in this embodiment fuses physical modeling and visual mapping, greatly improves the visualization features of the lead-zinc ore XRT image, provides a high-quality data set input basis for the accurate recognition of the deep learning AI algorithm, and has significant engineering practical value and innovation. The gray scale two-dimensional joint histogram of the lead-zinc ore dual-energy XRT image in Figure 2 is shown in Figure 3 .

[0058] As can be seen from FIG. 4:

[0059] (1) The abscissa represents the gray value of the low-energy X-ray image, and the ordinate represents the gray value of the high-energy X-ray image. The color in the figure represents the pixel density of different gray combinations, and the brighter the color, the higher the frequency of the gray combination. This figure can intuitively reflect the transmission response relationship and material absorption difference of the raw ore under different energy rays by one-to-one corresponding statistics of the dual-energy gray pixels.

[0060] (2) As can be seen from Figure 4 , the overall pixel points are distributed in a band along the diagonal direction, indicating that there is a strong positive correlation between the high-energy and low-energy gray values, that is, the two are consistent in structure. This shows that the same ore region shows similar transmission trends under different energy rays, verifying the imaging consistency of the dual-energy XRT image acquisition system. The dense distribution area of the point cloud band represents the main part of the ore, and its gray value changes synchronously in the high-energy and low-energy channels. There are some discrete pixels below the band area, indicating that some high-density minerals have stronger absorption and lower transmission under low-energy rays, resulting in a significant decrease in gray scale and showing obvious energy spectrum difference characteristics.

[0061] (3) Figure 4 The high-density aggregation area in the upper right corner represents the background area, and the gray values in the high-energy and low-energy images are close to saturation, indicating that this area does not absorb rays. By analyzing this joint histogram, the present application can realize the energy spectrum separation of the gray response of different minerals, and provide a basis for subsequent pseudo-color mapping and algorithm optimization. This distribution rule proves that the dual-energy gray joint feature can effectively distinguish high-density metal minerals from low-density gangue, and provides a high-quality input data set for the AI recognition algorithm.

[0062] In the process of dual-energy X-ray transmission imaging of lead-zinc ore, there is often a slight angle deviation or mechanical disturbance in the imaging process of high-energy and low-energy images, resulting in sub-pixel level spatial displacement error between images. In order to realize accurate fusion and pseudo-color map mapping, the image pair needs to be accurately aligned.

[0063] The enhanced correlation coefficient image registration algorithm (ECC) is a registration strategy based on maximizing the enhanced correlation coefficient in the image gray space, which is especially suitable for the case of complex background and fuzzy boundary in lead-zinc ore images. This algorithm 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 the reference image and the image to be registered , the main goal of the ECC algorithm is to transform into by affine transformation , so that the enhanced correlation coefficient between the two images is maximized. The ECC optimization objective function is as follows:

[0064]

[0065] In the above formula, denotes the image coordinates; represents the pixel value of the reference image; represents the image to be registered after transformation by the parameter affine parameter ; and represent the mean values of the two images, respectively, represents the enhanced correlation coefficient. ECC registration takes the low-energy image as the reference and registers the high-energy image. By iteratively optimizing the affine transformation parameters, the two images are most similar in the statistical sense of gray scale. This step is the key foundation for subsequent pseudo-color map generation and mineral identification. The high-energy image obtained after ECC registration is shown in Figure 4 .

[0066] Based on the analysis of the above figure, the ECC-based image registration algorithm used can achieve accurate alignment of dual-energy XRT images at the pixel level. Compared with traditional template matching and feature point methods, ECC achieves sub-pixel level affine registration by maximizing the full image gray covariance, effectively overcoming the common material micro-shift and image jitter problems in belt conveying. The registered image has the characteristics of clear edge, consistent background, and no ghost distortion, providing high-quality image dataset input for the subsequent process.

[0067] The gray scale of the lead-zinc raw ore conveyed by the intelligent beneficiation equipment belt falls on the XRT image after the response of the XRT detector and the DA conversion, which is not strictly linearly reflected by the material difference, especially in the middle gray scale section, the problem of insufficient separability often occurs. The gamma correction can enhance the contrast between the background and the Pb-rich area by stretching and compressing the high gray value section where the zinc minerals are located, so as to enhance the performance of the subsequent differential enhancement, pseudo-color mapping and detection model. The mathematical formula of the gamma correction is as follows.

[0068]

[0069] In the above formula, represents the pixel value of the position in the original image, represents the gamma parameter, and the value range is usually [0.8, 2.0], represents the pixel value after gamma correction, represents the maximum gray value in the image. The value of is 1.2, and the corrected high-energy image is as shown in . Figure 5

[0070] After the gamma correction, Figure 5 the results show that the lead-zinc ore XRT registration image realizes the nonlinear redistribution of brightness on the premise of maintaining the structural continuity. This step significantly improves the subsequent differential enhancement map, pseudo-color mapping and AI recognition algorithm for weak response and small target recognition robustness.

[0071] The physical response difference measurement is one of the core steps of the dual-energy XRT image pixel-level fusion analysis, which aims to amplify the response difference of different materials under high-energy and low-energy rays, so as to generate pseudo-color map and train and optimize AI recognition algorithm. The high-energy image after ECC registration and gamma correction is denoted as , and the original low-energy image as a reference is denoted as .First, the two images are normalized to the [0, 1] interval to unify the dynamic range, and the mathematical formula is as follows:

[0072]

[0073]

[0074] In the above formula, are the maximum gray values in the high-energy and low-energy images respectively, and the logarithmic form of the normalized dual-energy difference is adopted in this embodiment, and the mathematical calculation formula is as follows:

[0075] ​

[0076] In the above formula, is the normalized logarithmic difference, is a small constant set to prevent the denominator from being zero. The subsequent gray scale image calculated by difference is mapped to a pseudo-color map based on a color lookup table to enhance the contrast between lead-zinc ore points and other waste rocks.

[0077] This embodiment constructs a pseudo-color map batch mapping process for lead-zinc ore, which is mapped into a three-channel pseudo-color map according to pixel intensity classification, and the color transition follows the following visual perception-friendly order: dark blue→light blue→green→yellow→orange→red. Low gray value is mapped to dark blue; medium gray value transitions to green and yellow; high gray value is mapped to orange-red. Both color contrast and feature continuity are considered.

[0078] To enhance the adaptability of the algorithm, this embodiment fuses segmented nonlinear mapping and custom color scale lookup table to construct a pseudo-color map mapping function, which significantly improves the visualization expression ability of different target recognition objects. At the same time, combined with batch image reading and automatic naming saving mechanism, it outputs high-quality pseudo-color map as shown in Figure 6 , effectively improving data visualization efficiency and model training availability.

[0079] As shown by Figure 6 , a large number of irregularly distributed noise points appear in the image, which not only affects the visual perception of the human eye, but also may interfere with the training and optimization of the algorithm. Therefore, this embodiment introduces a "red background denoising algorithm" based on channel rules in the pseudo-color map generation process. By setting the red channel threshold condition and combining the color difference judgment mechanism, it automatically identifies the redundant and interfering dark red background area in the image and performs pixel-level replacement processing. The lead-zinc ore pseudo-color map after denoising is shown in Figure 7 .

[0080] As shown by Figure 7 , this method effectively eliminates the image background pseudo-response, improves the overall clarity of the image, and is easier to identify, providing a more pure and higher quality input pseudo-color map image dataset basis for subsequent ore identification algorithms.

[0081] Through the above targeted preprocessing and pseudo-color map mapping algorithm, all images are divided into training set, validation set and test set in the ratio of 8:1:1, which is conducive to the training and optimization of YOLO algorithm.

[0082] The bounding box loss function, as a key part of the AI recognition algorithm loss function, can be used to measure the difference between the algorithm's predicted bounding box and the real target bounding box, and optimize the model parameters through backpropagation algorithm to improve the recognition accuracy. The CIoU (Complete Intersection over Union) loss function used by YOLOv8 is sensitive to positioning accuracy, but it may also adversely affect the overall performance of the lead-zinc ore recognition algorithm, affecting its generalization ability.

[0083] To address the instability of traditional CIoU loss functions in identifying lead-zinc ore, this embodiment introduces the Weighted Intersection over Union (WIoU) loss function. This function combines the overlap between the predicted and ground truth bounding boxes, boundary offset, and target density distribution, improving the model's regression accuracy for weakly boundaryed ore targets. Its principle diagram is shown below. Figure 8 As shown. The green rectangle represents the actual bounding box, with a width and height of [missing information]. and The center coordinates are The blue rectangle represents the prediction box, with a width and height of [missing information]. and The center coordinates are The red line segment represents the Euclidean distance between the center of the predicted box and the center of the ground truth box, used to characterize the target offset; the overlapping area represents the intersection area of ​​the predicted box and the ground truth box, reflecting the degree of overlap between the two; the union region is determined by the smallest bounding rectangle formed by the two, used to calculate the intersection-union ratio.

[0084] When calculating the overlap between the blue predicted bounding box and the green target bounding box, WIoU not only considers IoU but also introduces a distance weighting factor based on the center point. This avoids the algorithm overemphasizing "easily distinguishable samples" and improves its optimization ability for "difficult-to-distinguish samples." Its mathematical formula is shown below:

[0085]

[0086] In the above formula Indicates the center of the prediction box Center of the real frame The square of the Euclidean distance This represents the diagonal length of the smallest bounding rectangle that can enclose both the predicted and ground truth bounding boxes. This is a dynamic weighting factor used to adjust the gradient contribution of different samples.

[0087] In this embodiment, to improve the AI ​​recognition algorithm's ability to model the color differences in different regions of the lead-zinc ore pseudocolor image, an efficient multi-scale attention (EMA) module is introduced into the Neck part of the YOLO network. The structural principle of the EMA module is as follows: Figure 9 As shown.

[0088] EMA aims to extract channel attention and spatial attention information by grouping and multi-scale modeling input features, and then fusing them through normalization and weight allocation mechanisms. This guides the neural network to focus more on high-response regions in lead-zinc ore images, suppressing background and noise interference. This structure effectively enhances the feature representation ability of small targets, improves overall recognition accuracy, and does not significantly increase computational cost. Specifically, the size of the input feature map is... wherein represents the number of feature map channels, represents the feature map height, represents the feature map width. First, the input feature map is divided into groups along the channel dimension, and the size of each group of features is to achieve grouped parallel computing and feature subspace separation. Subsequently, each group of features is processed in turn through the following steps for feature re-labeling:

[0089] (1) Local feature aggregation stage: two-dimensional average pooling (Avg Pool) operation is performed on each group of features to obtain global statistical features compressed along the spatial dimension, forming a channel description vector with a size of to represent the global response intensity of the group of channels.

[0090] (2) Multi-scale dependency modeling stage: the pooled features are input into two branches of different scales, respectively. The first branch generates local-scale weight coefficients after 1x1 convolution and normalization; the second branch generates global-scale weight representation using the combination of Softmax and average pooling. The features of the two scales are then interactively fused through matrix multiplication operation to obtain a multi-scale attention map containing local and global information.

[0091] (3) Feature weight generation stage: the fused attention map is weighted channel by channel, and the weight distribution range is constrained by the Sigmoid function, so that the output weight value is in the interval , thereby controlling the saliency of each channel feature.

[0092] (4) Feature reconstruction and output stage: finally, the generated attention weight is multiplied element by element with the original input feature to enhance the salient channels and suppress the non-salient channels. The weighted feature map remains the same in spatial size as the input, which is , and is finally output as the EMA module.

[0093] In the lead-zinc mine XRT pseudo-color image recognition task, the different scale target to be recognized has a very significant size difference. Although the FPN-PAN structure used in YOLOv8 can fuse multi-scale features, there are still certain limitations in feature alignment and target response consistency. To further improve the fusion expression ability of multi-scale targets, the embodiment introduces a lightweight adaptive spatial feature fusion module (ASFF) to optimize the structure of the feature fusion path of the Head detection head. The principle diagram is shown in Figure 10 .

[0094] The ASFF module can adaptively assign fusion weights of different scale feature maps according to the importance of each position, so as to realize more accurate spatial feature reconstruction. In specific implementation, the embodiment replaces the three-scale fusion output in the original YOLO Neck structure with the output channel corresponding to the ASFF module, retains the input multi-scale features, and completes spatial alignment and fusion through lightweight weight learning. The module takes input feature maps from different scale feature layers 、 、 as input, and the principle process is as follows:

[0095] (1) Multi-scale feature sampling stage: the input multi-layer feature maps 、 、 are respectively subjected to scale alignment operation, and each layer of features is adjusted to the same spatial resolution through upsampling or downsampling to obtain standardized features 、 、 . Wherein, the superscript “ ” represents the process of mapping the feature to the fusion scale, which ensures the alignment of different layer features in the spatial dimension.

[0096] (2) Feature compression and unified mapping stage: the standardized feature maps are respectively subjected to convolution operation to realize channel compression and semantic alignment. This process eliminates the dimensional difference between different feature layers, so that multi-scale features can be weighted fused in the same spatial scale.

[0097] (3) Feature splicing and weight generation stage: the convolution features of the three scales are spliced in the channel dimension to obtain a comprehensive feature tensor; then a convolution layer is used to further fuse the cross-layer information, and a softmax activation function is used to generate adaptive fusion weight coefficients 、 、 , and the sum of the three is 1 to balance the contribution proportion of different scale features. Adaptive weighting and fusion output stage: the generated weight coefficients are respectively multiplied with the corresponding scale features 、 、 , and then the weighted results are summed element by element to obtain the final fusion feature output.

[0098] Ablation experiments show that the introduction of the ASFF module significantly improves the model's accuracy in recognizing small targets, especially under complex conditions such as blurred edges and dense targets, where it performs more stably. Furthermore, due to the ASFF module's strong structural versatility and lightweight nature, its parameter and computational costs increase only slightly, having virtually no impact on inference speed, making it suitable for deployment on resource-constrained industrial terminal equipment. The introduction of the ASFF module effectively improves the feature fusion quality of YOLO in multi-scale lead-zinc mine XRT images, enhances the model's adaptability to complex target scenes, and is one of the important optimization strategies for achieving high-precision intelligent photoelectric mineral processing.

[0099] In standard convolution operations, all output feature maps are directly calculated from the input feature maps using convolution kernels of the same size. While this approach has good feature extraction and representation capabilities, it inevitably leads to a large number of parameters. Its high computational cost creates a bottleneck for low-cost deployment of the model.

[0100] GhostConv is a lightweight convolutional module proposed to improve the inference efficiency of convolutional neural networks. Its core idea is to reduce redundant computation by separating the feature map generation process. A diagram of GhostConv is shown below. Figure 11 As shown, in Figure 12 In Chinese, the abbreviation "Ghost" is used instead of "GhostConv".

[0101] GhostConv differs from traditional convolutional generation strategies in its design. Specifically, this module first generates a subset of main feature maps using fewer standard convolutional operations. Then, it generates more pseudo-feature maps from these main feature maps through a series of linear transformations, thus forming a complete set of output feature maps. The different transformation operators marked in the diagram represent the generation functions of these pseudo-features. Essentially, this is a low-computation feature expansion strategy that improves feature diversity and expressive power without significantly increasing model complexity. Specifically, the input feature map is denoted as... The processing flow is as follows:

[0102] (1) Main convolution feature extraction stage: First, the input feature map is processed. Perform regular convolution operations to generate basic feature maps. ,in This step only includes some of the real feature channels. In this step, only a small number of convolutional kernels are used to extract the core features, thereby reducing the main computational overhead.

[0103] (2) Feature generation and mapping stage: Due to feature redundancy in the depthwise convolution process, this module uses a linear transformation function. For feature maps The different channels are expanded to generate several auxiliary feature maps.

[0104] (3) Feature aggregation and fusion stage: the main convolution output and all generated auxiliary features are spliced or weighted superimposed in the channel dimension to form the final output feature:

[0105] (4) Identity mapping connection: in some implementations, the module can introduce an identity mapping branch (Identity) to add or splice the input feature and the output feature element by element, so as to enhance the gradient flow and feature fidelity, thereby improving the stability of neural network training and feature reuse capability.

[0106] The Optuna hyperparameter search framework is used to configure the hyperparameters before the lightweight convolution replacement operation. The results show that after replacing part of the standard convolution with GhostConv, the algorithm can significantly reduce the pure inference time while maintaining the recognition accuracy, further verifying that GhostConv can effectively improve the inference efficiency without affecting the recognition performance, and is suitable for the precise and rapid identification task of lead-zinc mines in this study which has high real-time requirements.

[0107] Compared with the prior art, the embodiment has the following beneficial effects:

[0108] In the preprocessing stage of the dual-energy XRT image, the ECC image registration method and the pseudo-color mapping technology are used to enhance the gray scale contrast and feature expression of the image.

[0109] In the algorithm optimization stage, the WIoU loss function, the efficient multi-scale attention module (EMA), the lightweight adaptive spatial feature fusion (ASSF) module, the GhostConv lightweight convolution, and the Optuna hyperparameter optimization framework are further introduced to improve the accuracy and speed of the YOLO-based lead-zinc mine identification algorithm. Through these optimizations, the recognition accuracy of the model is improved to more than 95%.

[0110] Through the multi-stage optimization processing of the lead-zinc mine XRT image, the YOLO lightweight AI recognition algorithm for lead-zinc mines is obtained, as shown in the network result Figure 12 . The algorithm has good universality and scalability and can be widely used in the industrial field of photoelectric intelligent sorting of lead-zinc mines.

[0111] As shown in Figure 13 , the lead-zinc mine identification algorithm implementation scheme of the embodiment includes three main steps: image data preprocessing, algorithm training and optimization, and identification result display. The specific implementation details are as follows:

[0112] Image data preprocessing: For the collected lead-zinc mine dual-energy XRT images, first, the high and low energy images are spatially aligned by the ECC registration method to eliminate image offset errors caused by factors such as belt movement and collection jitter, ensuring the accuracy of subsequent analysis. Subsequently, the Gamma correction method is introduced to perform nonlinear mapping enhancement on the medium-high gray scale region to improve the contrast and feature expression ability of ore and gangue in the image. Finally, the pseudo-color mapping technology is used to convert the gray-scale image to a visual image, making the ore and waste rock regions in the image show obvious color differences, which facilitates subsequent deep learning feature extraction.

[0113] Algorithm training and optimization: In the deep learning model part, this embodiment takes YOLO network as the basic framework, further introduces Wiou loss function to improve the regression accuracy of the prediction box, and embeds EMA efficient multi-scale attention module in the Neck part to guide the network to focus on high response areas and suppress background noise interference. At the same time, combined with the GhostConv lightweight convolution structure to reduce the model calculation amount, and through the ASSF adaptive spatial feature fusion module to improve the fusion effect of multi-scale features. In order to further enhance the generalization ability of the model, this embodiment uses the Optuna hyperparameter optimization framework to automatically search and adjust the key training parameters, so as to effectively improve the precision and robustness of the model while ensuring the recognition speed.

[0114] Image recognition result display: The trained and optimized model is applied to the lead-zinc mine XRT image recognition task, and the recognition result is displayed in real time through the visualization software. The software automatically labels the ore and waste rock regions on the pseudo-color image and displays the category label and confidence value, and the recognition accuracy can be stably improved to more than 95%. This display method not only facilitates the intuitive presentation of ore automatic recognition and sorting, but also provides practical support for the engineering application of photoelectric intelligent beneficiation, and the recognition result is as shown in Figure 14

[0115] As shown in FIG. 14, based on the optimized YOLOv8 recognition algorithm of this embodiment, the lead-zinc mine dual-energy XRT pseudo-color image can be processed to realize automatic recognition and classification of ore and waste rock targets. The rectangular frame in the figure marks the detected target region, and the “ore” and “waste” labels in the frame correspond to the lead-zinc mine and waste rock categories respectively, and the numerical value is the recognition confidence. The results show that the algorithm can maintain high recognition accuracy in target regions of different sizes, shapes and partial overlaps, indicating that the algorithm has good robustness and consistency in category differentiation.

[0116] ​In addition, it can be seen from the identification result that even in the case of a relatively single background and a blurred local target boundary, the model can still accurately locate the target center region and generate a reasonable bounding box, which embodies the effectiveness of the introduced WIoU loss function and EMA attention mechanism in enhancing the target positioning accuracy. Overall, the identification result verifies the engineering application potential of the lead-zinc mine intelligent identification algorithm method proposed in the present application under complex XRT imaging conditions, which not only assists manual image labeling, but also guarantees real-time performance through operator optimization and hardware acceleration strategy to meet the operation requirements of industrial XRT sorting machines.

[0117] The embodiment of the present application further provides an electronic device, including: a processor, and a memory coupled with the processor, the memory being used for storing a computer program; the processor is used for executing the computer program stored in the memory, so that the electronic device executes the method as described in any one of the above embodiments.

[0118] The electronic device can be a desktop computer, a notebook computer, a palm computer, a cloud server and the like. The electronic device can include, but is not limited to, a processor and a memory.

[0119] 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 devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the electronic device, and is connected with various parts of the whole device through various interfaces and lines.

[0120] The memory can be used for storing 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.

[0121] The memory can mainly include a program storage area and a data storage area. The program storage area can store an operating system, application programs required by at least one function, and the like. The data storage area can store data created according to the use of the mobile phone, and the like. 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 device.

[0122] The embodiment of the present application further provides a storage medium, which is a computer readable storage medium, and a computer program is stored in the computer readable storage medium. The computer program can implement the steps of the above-mentioned various method embodiments when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, an executable file, or some intermediate form, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a U disk, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, a software distribution medium, etc.

[0123] The embodiment of the present application further provides a computer program product, which includes a computer program or instructions, and when the computer program or instructions run on a computer, the computer program or instructions make the computer execute the method of any possible implementation manner.

[0124] The above is the preferred embodiment of the present application. It should be noted that, for those skilled in the art, without departing from the principle of the present application, a number of improvements and refinements can be made, which are also considered to be within the protection scope of the present application.

Claims

1. A method of separating lead-zinc ore, characterized by, The method comprises: a preprocessing stage of obtaining an original high-energy image and an original low-energy image of an original ore, the original high-energy image and the original low-energy image being obtained by dual-energy X-ray projection imaging of the original ore; using an enhanced correlation coefficient registration algorithm, the original low-energy image is taken as a reference, and the original high-energy image is registered to obtain a secondary high-energy image; using a physical response difference measure, the original low-energy image and the secondary high-energy image are fused to obtain a fusion image; and according to a color transition rule, the fusion image is colored to obtain a false color image; and an intelligent identification stage of taking the false color image as an input of a deep learning-based target detection model, and obtaining a sorting result of lead-zinc ore in the original ore through processing of the target detection model.

2. The method of claim 1, wherein the lead-zinc ore is classified into the first and second fractions by a process of a gravity separation method. The physical response difference measure employs a normalized log difference : wherein, is a secondary high energy map, is a primary low energy map, are the maximum gray values in the secondary high energy map and the primary low energy map, respectively, is a minimum constant set to prevent the denominator from being zero.

3. The method of claim 1, wherein the lead-zinc ore is classified into the first and second fractions by a gravity separation method. The target detection model adopts a YOLO model, and the YOLO model comprises a feature extraction network, a feature fusion network and a detection head.

4. The method of claim 3, wherein the lead-zinc ore is classified into the first and second fractions by a process of a gravity separation method. The YOLO model adopts a weighted intersection over union loss function L WIoU : in, Indicates the center of the prediction box Center of the real frame The square of the Euclidean distance This represents the diagonal length of the smallest bounding rectangle that can enclose both the predicted and ground truth bounding boxes. This is a dynamic weighting factor used to adjust the gradient contribution of different samples.

5. The method of claim 3, wherein the lead-zinc ore is classified into the first and second fractions by a process of a gravity separation method. The feature fusion network comprises a C2f layer and a Concat layer, and an efficient multi-scale attention layer is arranged between the C2f layer and the Concat layer, the efficient multi-scale attention layer is used for grouping convolution and multi-scale modeling of input features, respectively extracting channel attention and spatial attention information, and fusing through a normalization and weight distribution mechanism.

6. The method of claim 3, wherein the lead-zinc ore is classified into the first and second fractions by a process of a gravity separation method. The detection head comprises a lightweight adaptive spatial feature fusion module for adaptively assigning fusion weights of different scale feature maps according to the importance of each position.

7. The method of claim 3 wherein the lead-zinc ore is classified by size. The feature extraction network adopts a ghost convolution layer. ​ 8. The method of claim 3, wherein the lead-zinc ore is classified into the first and second fractions by a process of a gravity separation method. The YOLO model adopts an Optuna hyperparameter search framework.

9. The method of claim 1 wherein the lead-zinc ore is classified by size. The preprocessing stage further comprises gamma correction of the secondary high-energy image and the original low-energy image. ​ 10. An electronic device, comprising: The electronic device comprises a processor and a memory coupled to the processor, the memory is configured to store a computer program; and the processor is configured to execute the computer program stored in the memory to enable the electronic device to perform the lead-zinc ore sorting method according to any one of claims 1-8.

Images