Method, device and medium for detecting the size of the ore entering a mill
By using a multi-scale parallel high-efficiency attention U-shaped network model and image preprocessing technology, the accuracy and real-time performance issues of ore size detection in a semi-autogenous mill were solved, achieving efficient ore size detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2024-01-18
- Publication Date
- 2026-05-26
AI Technical Summary
In the existing technology, the production status of semi-autogenous mills is easily affected by the size of the ore entering the mill, resulting in low grinding efficiency or equipment damage. Existing detection methods have low accuracy and poor real-time performance, making it difficult to achieve efficient ore size detection.
We employ a multi-scale parallel efficient attention U-shaped network model for ore image segmentation, combined with image preprocessing and enhancement techniques. We also use equivalent feature parameters to statistically determine ore block size and deep learning algorithms to improve detection accuracy.
It enables high-precision block size detection of non-uniform sized ores, reduces errors caused by segmentation difficulties, and improves the real-time performance and accuracy of detection.
Smart Images
Figure CN117911366B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of block size detection technology in ore mining, specifically relating to a method, equipment and medium for detecting the block size of ore fed into a mill. Background Technology
[0002] In mineral processing, the crushing and grinding process involves breaking down raw ore from its lumpy state into powder, ensuring the product particle size meets the requirements of subsequent flotation processes. In traditional crushing and grinding processes, crushing and grinding are completely separate, resulting in a complex process and difficult control. Semi-autogenous grinding technology integrates the multiple stages of traditional crushing and grinding into a single operation, allowing the product to be directly transferred to the flotation stage. This simplified process and convenient control have made it one of the most widely used crushing and grinding technologies today.
[0003] However, the production status of a semi-autogenous mill is easily affected by the size of the ore pieces fed into the mill. When the proportion of large ore pieces in the ore is too high or the ore hardness is too high, the difficulty of crushing the ore increases, requiring a longer grinding time. This can easily lead to the mill being overloaded, reducing grinding efficiency and even damaging the equipment. When the proportion of large ore pieces in the ore is too low or the ore hardness is too low, the semi-autogenous mill cannot effectively utilize the impact of large ore pieces on other ores, which also leads to a decrease in grinding efficiency.
[0004] It is evident that the particle size of the ore fed into the mill is a crucial factor affecting the production efficiency and stable operation of a semi-autogenous mill. On-site operators need to adjust the mill feed rate and other control parameters in real time based on the particle size of the ore to ensure the semi-autogenous mill operates at its optimal condition. Currently, most concentrators detect the particle size of the ore fed into the mill through visual inspection or manual screening. The former has low accuracy and relies heavily on operator experience, while the latter lacks real-time capability and incurs high labor and material costs.
[0005] With the widespread application of machine vision technology in the industrial field, an increasing number of ore processing plants are adopting machine vision-based ore size detection methods for mill feed size detection. This method offers good real-time performance and low cost. The machine vision-based ore size detection method mainly includes three processes: image acquisition, image segmentation, and ore size statistics. Among these, ore image segmentation is the core step, playing a decisive role in the accuracy and performance of the ore size detection algorithm. Due to the complex texture of ore images, extremely irregular ore edges, and large differences in ore size, effective image segmentation is difficult. In recent years, deep learning algorithms have achieved excellent results in image segmentation, providing possibilities for performing complex ore image segmentation. Summary of the Invention
[0006] To address the technical problem of low accuracy in particle size detection using automated vision systems due to large differences in the size of ore fed into the semi-autogenous grinding process, this invention provides a method, equipment, and medium for particle size detection of ore fed into the mill, which achieves high detection accuracy.
[0007] To achieve the above technical objectives, the present invention adopts the following technical solution:
[0008] A method for detecting the particle size of ore fed into a mill, comprising:
[0009] Step 1: Collect images of the ore being transported on the conveyor belt, manually create binarized labels, construct a sample dataset, and divide all samples into a training set and a validation set in a ratio of 8:2.
[0010] Step 2: Build a multi-scale parallel high-efficiency attention U-shaped network model. This network redesigns the encoder, skip connection layer, and decoder of the U-shaped network.
[0011] Step 3: Perform image preprocessing and image enhancement on the sample dataset images, and then use the processed images to train the model to obtain the ore image segmentation model;
[0012] Step 4: Use the ore image segmentation model to segment and predict the ore images acquired in real time from the conveyor belt to obtain ore segmentation images.
[0013] Step 5: Extract ore sub-regions from the ore segmentation image, calculate the equivalent feature parameters of each ore sub-region, and use the equivalent feature parameters to calculate the proportion of ore in each block size range.
[0014] In the improved method for detecting the particle size of ore fed into the mill, the manual binarization labeling in step one is specifically achieved by using a professional image labeling tool to mark the pixel value of the ore area in the collected ore image as 255 and the pixel value of the background area as 0.
[0015] In the technical solution of the better method for detecting the size of ore fed into the mill, step two describes the construction of a multi-scale parallel high-efficiency attention U-shaped network model, which includes three parts: encoder, skip connection and decoder.
[0016] The encoder uses the Parallel Efficient Channel Attention Module (PECA) instead of the convolution operations in U-Net, and PatchMerging instead of Max Pooling in U-Net; input Grayscale images are generated using patch embedding. Feature vector , Generated through feature operations of two PECA modules The dimension contains attention information features, which are then generated through the Patch Merging layer. Feature vector Repeat the above process twice to generate Feature vector and Feature vector ;
[0017] The skip connection layer incorporates a multi-scale feature fusion interaction module (MFFI) and a bidirectional cross-attention gating unit module (BCAG), derived from the encoder feature vector. , , , The MFFI module performs information fusion and interaction to generate fused features. , , Then the encoder's multi-scale feature vector Separate and fusion features Addition to obtain new skip connection features , The BCAG module will introduce new skip connection features. and decoder Hierarchical output feature vector The weights are redistributed to obtain the decoded fusion features. Then pass the input to the decoder. Social class;
[0018] The decoder uses the PECA module to replace the convolution operation in U-net, and uses Patch Expanding to replace the upsampling operation in U-net. encoder feature vector The input to the decoder is processed by two PECA modules for feature operations, and then generated by a Patch Expanding module. eigenvectors This serves as the output of the fourth stage of the decoder; the remaining three stages of the decoder are output as follows: Decoding fusion features For input, output the first... Hierarchical feature vector The first-level output feature vector Generated by the Linear Projection module Output image.
[0019] In a superior method for detecting the particle size of ore fed into a mill, the PECA module connects efficient attention and channel attention based on cross-covariance in parallel, and then connects the Local Information Interaction (LPI) module and the Feedforward Network (FFN) module in series. The efficient attention module and the channel attention module respectively calculate the feature vector. Attention information between internal features is used to fuse the output features. Then, the local information interaction module performs local feature interaction processing based on global attention calculation. Finally, the feedforward network module outputs features of the same dimension. ;
[0020] The MFFI module is divided into two parts: fusion and interaction, providing multi-scale feature information. The dimension is extracted after a one-dimensional average pooling layer. Feature information is formed by splicing together One-dimensional multi-scale information is first processed and then fused through channel attention to generate fused multi-scale information. The interaction involves concatenating the fused multi-scale information as a header with the original feature information from each stage. Then, a two-dimensional attention calculation in both the spatial and channel domains is performed using the PECA module, enabling interaction between the feature information and the multi-scale information at each stage. Finally, the multi-scale information header is removed, and the interacting feature maps are fused to obtain the final multi-scale feature, i.e., the fused feature. ;
[0021] The BCAG module will use the jump connection features from the jump connection layer. and decoder The feature vectors output by the hierarchy Weights are redistributed using cross-attention gating units, followed by feature fusion; skip connection features are then computed. When considering the weights of cross-attention Features derived from skip connections , and From the decoder The feature vectors output by the hierarchy After performing cross-attention calculations on both, the result is output through a gating unit consisting of a convolutional layer, a normalization layer, and a sigmoid activation function layer; in the computation of the decoder... Hierarchical output feature vector When assigning weights, unlike the weights for skip connection features, From the decoder Hierarchical output feature vector , and Features derived from skip connections Finally, after a patch expanding process, features of the current layer of the decoder with the same feature dimension as the skip connections are obtained, and these features are fused to obtain the first... Decoding fusion features .
[0022] In the improved method for detecting the particle size of ore fed into the mill, the preprocessing operations for the ore image in step three include: bilateral filtering, contrast-limited adaptive histogram equalization, grayscale conversion, and removal of conveyor belt and dust areas. The ore image enhancement operations include: random cropping, rotation, and flipping. These operations enhance the contrast between the ore region and its edges, reducing the impact of image noise on the segmentation results.
[0023] In a more optimized method for detecting the particle size of ore fed into a mill, the removal of the conveyor belt and dust area specifically employs a conveyor belt removal algorithm based on histogram statistics, including:
[0024] The image after grayscale processing is divided into equal columns to obtain... A slice of a cut, elongated strip of ore;
[0025] Set the grayscale range threshold for the cutting conveyor belt area and the ore area to be [value]. The grayscale frequency peak threshold is The cutting width is ;
[0026] If the range of gray-level peak values in a slice histogram falls within a set threshold... And the peak value of the gray level frequency is greater than the set peak value threshold. If so, the area to which the slice belongs is determined to be the conveyor belt area;
[0027] Based on the determined left and right boundary slices, according to the cutting width Cropping the image ensures all cropped images have the same width. This cropping area is used to crop both the original ore image and the label image.
[0028] In the improved method for detecting the particle size of ore fed into the mill, the network model is trained using the sample dataset images from step 1. The training parameters are: batch size of 8, epoch of 20, AdamW optimization algorithm, learning rate of 0.0001, and a combined loss function based on BCE Loss and Dice Loss. During training, the model is saved when the epoch is greater than 15, and the optimal model is selected from the five saved models. .
[0029] In a more optimized method for detecting the size of ore fed into a mill, step 5 involves extracting ore sub-regions from the ore segmentation image, calculating the equivalent feature parameters of each ore sub-region, and statistically analyzing the proportion of ore within each size range using these equivalent feature parameters. Specifically, this includes:
[0030] The OpenCV library functions were used to perform connected component detection and ellipse fitting on the ore segmentation image, and the feature parameters of each fitted ellipse were obtained: semi-major axis. short half shaft equivalent area Then, convert them into the corresponding three-dimensional feature parameters of the ore:
[0031]
[0032]
[0033] in, For equivalent particle size, Equivalent volume;
[0034] The ore in the image is classified into block size grades based on the equivalent particle size, and the total volume, volume percentage and mass percentage of the ore in each block size grade are calculated.
[0035]
[0036] in, , For the first Total mass and mass percentage of ore for each block size grade. , For the first Total volume and volume percentage of ore at each block size grade This represents the density of the ore.
[0037] The present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program for executing the above-described method for detecting the particle size of ore fed into a mill.
[0038] The present invention provides a computer-readable storage medium having a computer program stored thereon for storing the above-mentioned method program for detecting the particle size of ore fed into the mill and image data.
[0039] Beneficial effects
[0040] This invention enables the detection of block size in non-uniform sized ores, reducing the block size detection error caused by difficulties in segmentation and improving detection accuracy. Attached Figure Description
[0041] Figure 1 This is an overall flowchart of the method described in this invention;
[0042] Figure 2 This is a diagram of the overall deep learning network model of the method described in this invention.
[0043] Figure 3 This is a structural diagram of the parallel and efficient channel attention module of the method described in this invention;
[0044] Figure 4 This is a diagram of the multi-scale feature fusion interaction module of the method described in this invention;
[0045] Figure 5 This is a structural diagram of the bidirectional cross-attention gating unit of the method described in this invention;
[0046] Figure 6 This is a comparison curve of the actual value and the measured value under different block size grades in the non-uniform size ore test set of the method described in this invention. Detailed Implementation
[0047] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.
[0048] This embodiment provides a method for detecting the particle size of ore fed into a mill, including the following steps:
[0049] Step 1: Collect images of the ore transported on the conveyor belt, manually create binarized labels, construct a sample dataset, and divide all samples into training and validation sets according to a certain ratio. ;
[0050] Step 2: Construct a multi-scale parallel high-efficiency attention U-shaped network model. Based on the U-shaped network structure of symmetric encoder-decoder, redesign the encoder, decoder and skip connections respectively.
[0051] The encoder uses the Parallel Efficient Channel Attention Module (PECA) instead of the convolution operations in U-Net, and PatchMerging instead of Max Pooling in U-Net; input Grayscale images are generated using patch embedding. Feature vector , Generated through feature operations of two PECA modules Dimensions with attentional information features Then it is generated through the Patch Merging layer. Feature vector The process is repeated twice to generate Feature vector and Feature vector ;
[0052] The skip connection layer incorporates a multi-scale feature fusion interaction module (MFFI) and a bidirectional cross-attention gating unit module (BCAG), derived from the encoder feature vector. , , , The MFFI module performs information fusion and interaction to generate fused features. , , The multi-scale feature vector of the encoder Then, separately with characteristics Addition to obtain new skip connection features The BCAG module will skip connection features. and decoder i+1 layer output The feature weights are redistributed to obtain fused features. ;
[0053] The decoder uses the PECA module to replace the convolution operation in U-net, and uses Patch Expanding to replace the upsampling operation in U-net. encoder features The input to the decoder is processed by two PECA modules for feature operations, and then generated by a Patch Expanding module. Decoder upsampling layer output After that, the three-layer decoder... Input is i=1,2,3, output is decoder features , Generated by the Linear Projection module Output image;
[0054] The PECA module connects efficient attention and channel attention based on cross-covariance in parallel, and then connects it sequentially to the Local Information Interaction (LPI) module and the Feedforward Network (FFN) module; the MFFI module is divided into two parts: fusion and interaction, involving multi-scale feature information. The dimension is extracted after a one-dimensional average pooling layer. Feature information is formed by splicing together One-dimensional multi-scale information is first processed and then fused through channel attention to generate fused multi-scale information. The interaction involves concatenating the fused multi-scale information as a header with the original feature information from each stage. Then, a two-dimensional attention calculation in both the spatial and channel domains is performed using the PECA module, enabling interaction between the feature information and the multi-scale information at each stage. Finally, the multi-scale information header is removed, and the interacting feature maps are fused to obtain the final multi-scale feature. .
[0055] The BCAG module will use the jump connection features from the jump connection layer. and decoder The feature vectors output by the hierarchy Weights are redistributed using cross-attention gating units, followed by feature fusion; skip connection features are then computed. When considering the weights of cross-attention Features derived from skip connections , and From the decoder The feature vectors output by the hierarchy After performing cross-attention calculations on both, the result is output through a gating unit consisting of a convolutional layer, a normalization layer, and a sigmoid activation function layer; in the computation of the decoder... Hierarchical output feature vector When assigning weights, unlike the weights for skip connection features, From the decoder Hierarchical output feature vector , and Features derived from skip connections Finally, after a patch expanding process, features of the current layer of the decoder with the same feature dimension as the skip connections are obtained, and these features are fused to obtain the first... Decoding fusion features .
[0056] Step 3: After preprocessing and enhancing the sample images, the model is trained to obtain the ore image segmentation model;
[0057] Step 3.1, the preprocessing operations are as follows: bilateral filtering, contrast-limited adaptive histogram equalization, grayscale conversion, and removal of conveyor belt and dust areas to enhance the contrast between the ore area and the ore edge and reduce the impact of image noise on the segmentation results.
[0058] The methods for grayscale conversion and removal of conveyor belt and dust areas are as follows:
[0059] The image after grayscale processing is divided into equal columns with a column spacing of 50. A slice of a cut, elongated strip of ore;
[0060] Analyze the range and peak value of pixel values in the histogram of each slice;
[0061] Set the gray level range threshold for the cutting conveyor belt area and the ore area to 60, the gray level frequency peak threshold to 5000, and the cutting width to 1600.
[0062] If the peak gray level frequency of the slice histogram is within the gray level range of 0-60 and the peak gray level frequency is greater than 5000, the area to which the slice belongs is determined to be the conveyor belt area.
[0063] After determining the left and right boundary slices, the left and right cropping areas are adjusted according to the cropping width to ensure that the width of all cropped images is consistent. The cropping area is then used to crop the original ore image and the label image.
[0064] Step 3.2, the ore image enhancement operation includes: random cropping, rotation, and flipping;
[0065] Step 3.3 involves training the model using sample images. The specific method is as follows:
[0066] The network model was trained using the images from the sample dataset in step 1. The training parameters were: batch size of 8, epoch of 20, AdamW optimization algorithm, learning rate of 0.0001, and a combined loss function based on BCE Loss and Dice Loss. During training, the model was saved when the epoch exceeded 15, and the optimal model was selected from the five saved models. .
[0067] Step 4: Segment the real-time acquired ore image to obtain a segmented image with a resolution of 512*512, and use the model trained in Step 3. The segmented image is segmented and predicted, and then the segmented predicted image is reassembled to the original image size to obtain the ore segmentation image.
[0068] Step 5: Extract ore sub-regions from the ore segmentation image obtained in Step 4, and obtain equivalent feature parameters for each ore sub-region. Calculate the proportion of ore within the corresponding block size range using these equivalent feature parameters. The method is as follows:
[0069] The OpenCV functions `findContours` and `fitEllipse` are used to perform connected component detection and ellipse fitting on the segmentation results, and the semi-major axis of the fitted ellipse is obtained. short half shaft and equivalent area Image feature parameters are obtained, and then the statistical two-dimensional shape feature parameters are transformed into three-dimensional feature parameters of the ore through a certain transformation relationship. The transformation formula is as follows:
[0070]
[0071]
[0072] in, For equivalent particle size, The equivalent volume is thus determined, and the equivalent volume corresponding to the ore sub-region is determined accordingly.
[0073] Let the first The proportion of ore mass for each block size grade is as follows: The total mass of ore is The volume percentage of the ore is The total volume of the ore is , Given the density of the ore, according to the following formula, the mass percentage of different ore size grades is equal to their volume percentage;
[0074]
[0075] Based on the screening data of non-uniform ore, the ore was divided into six block size grades: greater than 100mm, 80-100mm, 60-80mm, 40-60mm, 20-40mm, and 0-20mm. This method was used to statistically analyze the proportion of ore within different ranges, and the comparison curves between the actual and measured values for each block size grade are shown below. Figure 6 As shown in (a) to (f).
[0076] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.
Claims
1. A method for detecting the particle size of ore fed into a mill, characterized in that, include: Step 1: Collect images of the ore being transported on the conveyor belt, manually create binarized labels, construct a sample dataset, and divide all samples into training and validation sets according to the proportions. Step 2: Construct a multi-scale parallel high-efficiency attention U-shaped network model; Step 2 describes the construction of a multi-scale parallel efficient attention U-shaped network model, which includes three parts: encoder, skip connections, and decoder. The encoder uses the Parallel Efficient Channel Attention Module (PECA) to replace the convolution operation in U-net, and uses PatchMerging to replace Max Pooling in U-net. The skip connection layer incorporates the Multi-Scale Feature Fusion Interaction (MFFI) module and the Bidirectional Cross-Attention Gating Unit (BCAG) module. The decoder uses the PECA module to replace the convolution operation in U-net, and uses Patch Expanding to replace the upsampling operation in U-net. Step 3: Perform image preprocessing and image enhancement on the sample dataset images, and then use the processed images to train the model to obtain the ore image segmentation model; Step 4: Use the ore image segmentation model to segment and predict the ore images acquired in real time from the conveyor belt to obtain ore segmentation images. Step 5: Extract the ore sub-regions from the ore segmentation image, calculate the equivalent feature parameters of each ore sub-region, and use the equivalent feature parameters to calculate the proportion of ore in each block size range.
2. The method according to claim 1, characterized in that, The manual binarization label creation described in step 1 involves using a professional image label creation tool to mark the pixel value of the ore area in the collected ore image as 255 and the pixel value of the background area as 0.
3. The method according to claim 1, characterized in that, In the encoder: Input Grayscale images are generated using patch embedding. Feature vector , Generated through feature operations of two PECA modules The dimension contains attention information features, which are then generated through the Patch Merging layer. Feature vector Repeat the above process twice to generate Feature vector and Feature vector ; In the skip connection layer: features from the encoder feature vector , , , The MFFI module performs information fusion and interaction to generate fused features. , , Then the encoder's multi-scale feature vector Separate and fusion features Addition to obtain new skip connection features , The BCAG module will introduce new skip connection features. and decoder Hierarchical output feature vector The weights are redistributed to obtain the decoded fusion features. Then pass the input to the decoder. Social class; In the decoder: encoder feature vector The input to the decoder is processed by two PECA modules for feature operations, and then generated by a Patch Expanding module. eigenvectors This serves as the output of the fourth stage of the decoder. The remaining three layers of the decoder are as follows: Decoding fusion features For input, output the first... Hierarchical feature vector The first-level output feature vector Generated by the Linear Projection module Output image.
4. The method according to claim 3, characterized in that, The use of the PECA module, MFFI module and BCAG module; The PECA module connects efficient attention and channel attention based on cross-covariance in parallel, and then connects the Local Information Interaction (LPI) module and the Feedforward Network (FFN) module in series. The efficient attention module and the channel attention module calculate the feature vectors respectively. Attention information between internal features is used to fuse the output features. Then, the local information interaction module performs local feature interaction processing based on global attention calculation. Finally, the feedforward network module outputs features of the same dimension. ; The MFFI module is divided into two parts: fusion and interaction, providing multi-scale feature information. The dimension is extracted after a one-dimensional average pooling layer. Feature information is formed by splicing together One-dimensional multi-scale information is then fused through channel attention to generate fused multi-scale information. The interaction involves concatenating multi-scale information as a header with the original feature information from each stage, then performing two-dimensional attention calculations in the spatial and channel domains using the PECA module. This allows the feature information from each stage to interact with the multi-scale information. Finally, the multi-scale information header is removed, and the interacted feature maps are fused to obtain the final multi-scale feature, i.e., the fused feature. ; The BCAG module will use the jump connection features from the jump connection layer. and decoder The feature vectors output by the hierarchy We use cross-attention gating units to redistribute weights before feature fusion. In calculating skip connection features When considering the weights of cross-attention Features derived from skip connections , and From the decoder The feature vectors output by the hierarchy After performing cross-attention calculation on the two, the calculation result is output through a gated unit consisting of a convolutional layer, a normalization layer, and a sigmoid activation function layer. In the computation decoder Hierarchical output feature vector When assigning weights, unlike the weights for skip connection features, From the decoder Hierarchical output feature vector , and Features derived from skip connections Finally, after a patch expanding process, features of the current layer of the decoder with the same feature dimension as the skip connections are obtained, and these features are fused to obtain the first... Decoding fusion features .
5. The method according to claim 1, characterized in that, The preprocessing operations for the ore image in step 3 include: bilateral filtering, contrast-limited adaptive histogram equalization, grayscale conversion, and removal of conveyor belt and dust areas. The ore image enhancement operations include: random cropping, rotation, and flipping.
6. The method according to claim 5, characterized in that, The removal of the conveyor belt and dust area specifically uses a conveyor belt removal algorithm based on histogram statistics, including: The image after grayscale processing is divided into equal columns to obtain... A slice of a cut, elongated strip of ore; Set the grayscale range threshold for the cutting conveyor belt area and the ore area to be [value]. The grayscale frequency peak threshold is The cutting width is ; If the range of gray-level peak values in a slice histogram falls within a set threshold... And the peak value of the gray level frequency is greater than the set peak value threshold. If so, the area to which the slice belongs is determined to be the conveyor belt area; Based on the determined left and right boundary slices, according to the cutting width Cropping the images ensures that all cropped images have the same width.
7. The method according to claim 1, characterized in that, Step 3 uses the data-augmented training set for model training, with the training sample resolution... The batch size is 8, the epoch is 20, the AdamW optimization algorithm is used, the learning rate is 0.0001, and the loss function is a combination loss function based on BCE Loss and Dice Loss, defined as follows: ； in The standard BCE Loss function, This is the standard Dice Loss function.
8. The method according to claim 1, characterized in that, Step 5 involves extracting ore sub-regions from the ore segmentation image, calculating the equivalent feature parameters for each ore sub-region, and statistically analyzing the proportion of ore in each size range using these equivalent feature parameters. Specifically, this includes: The OpenCV library functions were used to perform connected component detection and ellipse fitting on the ore segmentation image, and the feature parameters of each fitted ellipse were obtained: semi-major axis. short half shaft equivalent area Then, convert them into the corresponding three-dimensional feature parameters of the ore: ； ； in, For equivalent particle size, Equivalent volume; The ore in the image is classified into block size grades based on the equivalent particle size, and the total volume, volume percentage and mass percentage of the ore in each block size grade are calculated. ； in, , For the first Total mass and mass percentage of ore for each block size grade. , For the first Total volume and volume percentage of ore at each block size grade This represents the density of the ore.
9. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the processor, the processor causes the processor to implement the method as described in any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.