A method and system for identifying voids in a mineral rock particle packing system
By using medical CT scans and optimizing dataset parameters with the AttentionR2U-net neural network, the accuracy and efficiency issues of void structure identification in complex mineral and rock particle accumulation systems were resolved, achieving efficient and rapid void identification and laying the foundation for subsequent analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH BEIJING
- Filing Date
- 2023-12-27
- Publication Date
- 2026-07-03
AI Technical Summary
Existing technologies are rarely and insufficiently applied in the identification of void structures in complex mineral and rock particle accumulation systems, and cannot simultaneously meet the requirements of accurate and efficient identification.
The original slice images of the mineral-rock particle accumulation system were obtained using medical CT scanning technology. Target objects were labeled using Labelme, and an AttentionR2U-net neural network was built. Using Tensorflow and Keras frameworks, the dataset parameters were optimized, including the number of datasets, image size, data augmentation methods, and image complexity. Hyperparameters were repeatedly trained and adjusted to finally achieve efficient and accurate gap recognition.
It significantly improves the model's convergence efficiency and recognition accuracy, enabling rapid and accurate identification of the void structure in ore-rock particle accumulation systems, and providing a basis for subsequent analysis.
Smart Images
Figure CN118097220B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for identifying voids in a mineral-rock particle accumulation system. Background Technology
[0002] In caving mining, the ore and rock formed through forced or natural caving are generally composed of numerous granular media of varying sizes, shapes, and complex arrangements. Under the combined influence of multiple factors, including their physical and mechanical properties and release structure, the ore particle packing structure and its spatial flow process exhibit uniqueness and complexity. The void structure affects the macroscopic mechanics and transport behavior of ore particles, thereby influencing the secondary fracturing patterns, the morphology of loosened and released bodies, the flow characteristics and emergence characteristics of small particles, plugging, and ultimately, the efficiency, safety, and economic benefits of ore extraction. As a crucial characteristic, it has attracted significant attention and research from scholars both domestically and internationally.
[0003] The measurement and identification of voids in granular systems is a prerequisite for void structure characteristic analysis. Currently, common image-based algorithms for identifying void structures in granular packing systems mainly rely on traditional image recognition methods such as threshold segmentation. These methods generally suffer from low accuracy and efficiency in void identification. Deep learning methods, on the other hand, can converge quickly and identify voids accurately, showing great potential in the intelligent identification of surface cracks, rock pores, and voids in mineral-rock granular systems. However, current research on the application of deep learning algorithms in identifying void structures in complex mineral-rock granular packing systems is limited and lacks depth, failing to simultaneously meet the requirements of both accurate and efficient identification of voids within these systems. Summary of the Invention
[0004] This invention provides a method and system for identifying voids in mineral and rock particle accumulation systems, addressing the problems of limited and insufficient research in the existing technology for identifying void structures in complex mineral and rock particle accumulation systems.
[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution: a method for identifying voids in a mineral-rock particle accumulation system, characterized by the following steps:
[0006] S1. Scan the mineral and rock particle accumulation system to obtain the original slice images of the mineral and rock particles, and preprocess the original slice images to obtain label images;
[0007] S2. Select different data parameters from the label images to obtain four types of datasets; divide the four types of datasets into training sets and test sets respectively.
[0008] S3. Build AttentionR2U-net based on Tensorflow and Keras frameworks, and train the network to obtain the model using the training set;
[0009] S4. Qualitatively and quantitatively evaluate the recognition performance of the model under the four types of datasets based on the labeled images and the original slice images, and further determine the optimal dataset parameters by comparing the binary classification confusion matrix index.
[0010] S5. Obtain the optimal dataset based on the optimal dataset parameters, and train the neural network using the optimal dataset;
[0011] S6. Train the neural network repeatedly using the optimal dataset, adjust the hyperparameters to obtain the optimal model, input the data of the mineral and rock particle accumulation system to be identified, and complete the identification of voids in the mineral and rock particle accumulation system.
[0012] Optionally, in step S1, the ore-rock particle accumulation system is scanned to obtain raw slice images of the ore-rock particles, and the raw slice images are preprocessed to obtain label images, including:
[0013] X-rays are emitted using medical CT scanning technology to perform a comprehensive spiral scan of the mineral and rock particle accumulation system in the device, and specific information about the gaps between particles inside the system is obtained through slice images.
[0014] The target region is selected by cropping in the sliced image, and the target object is labeled and preprocessed using Labelme to obtain the label image.
[0015] Optionally, in step S2, different data parameters are selected from the label images to obtain four types of datasets, including:
[0016] By selecting different data parameters, the preprocessed dataset is classified to obtain four types of datasets;
[0017] The dataset parameters include the number of datasets, image size, data augmentation method, and image complexity.
[0018] Optionally, in step S3, an AttentionR2U-net is built based on the Tensorflow and Keras frameworks, and the network is trained using a dataset to obtain the model, including:
[0019] The advantages of U-net and recurrent network and residual network are combined to build R2U-Net. Attention gates are added to the R2U-net network to build the AttentionR2U-net neural network.
[0020] The AttentionR2U-net neural network was trained using four different datasets to obtain recognition models for the corresponding datasets.
[0021] Optionally, data augmentation methods include:
[0022] By using a center-and-surround-around sliding cropping method, the slices and their corresponding label images are processed simultaneously, and the data is augmented while changing the image size. This results in a dataset of 1000 images with a size of 300-400 pixels, including both simple images with single-particle size and complex images with multiple particle sizes.
[0023] Optionally, in step S4, the recognition performance of the model under the four datasets is qualitatively and quantitatively evaluated based on the labeled image and the recognized image. The optimal dataset parameters are further determined by comparing the binary classification confusion matrix index, including:
[0024] Acquire labeled images and recognize images;
[0025] The problem of identifying gaps in sliced images is defined as a binary classification problem, and a binary classification cross-entropy loss function is constructed.
[0026] Based on the labeled image and the recognized image, and combined with the binary classification cross-entropy loss function, the optimal dataset parameters are selected by comparing the quantitative evaluation indicators of the model under the four datasets.
[0027] A void identification system for a mineral-rock particle packing system, the system being used in the aforementioned void identification method for mineral-rock particle packing systems, the system comprising:
[0028] The data processing module is used to scan the mineral and rock particle accumulation system, obtain the original slice images of the mineral and rock particles, and preprocess the original slice images to obtain label images.
[0029] The data classification module is used to select different data parameters in the label image to obtain four types of datasets; the four types of datasets are then divided into training sets and test sets.
[0030] The model building module is used to build AttentionR2U-net based on the Tensorflow and Keras frameworks, and to obtain the model by training the network using the training set.
[0031] The preliminary evaluation module is used to qualitatively and quantitatively evaluate the recognition performance of the model on four datasets based on the labeled images and the original slice images. The optimal dataset parameters are further determined by comparing the binary classification confusion matrix index.
[0032] The model training module is used to obtain the optimal dataset based on the parameters of the optimal dataset, and to train the neural network using the optimal dataset.
[0033] The void identification module is used to repeatedly train the neural network with the optimal dataset, adjust the hyperparameters to obtain the optimal model, input the data of the mineral and rock particle accumulation system to be identified, and complete the void identification of the mineral and rock particle accumulation system.
[0034] Optionally, the data processing module is used to emit X-rays using medical CT scanning technology to perform a full-range spiral scan of the mineral and rock particle accumulation system in the device, and to obtain specific information about the interparticle gaps within the system through slice images.
[0035] The target region is selected by cropping in the sliced image, and the target object is labeled by Labelme to obtain the preprocessed dataset.
[0036] Optionally, the data classification module is used to select different data parameters to classify the preprocessed dataset and obtain four types of datasets;
[0037] The dataset parameters include the number of datasets, image size, data augmentation method, and image complexity.
[0038] Optionally, the model building module is used to combine the advantages of U-net with recurrent networks and residual networks to build R2U-Net, and to add attention gates to the R2U-net network to build the AttentionR2U-net neural network.
[0039] The AttentionR2U-net neural network was trained using four different datasets to obtain recognition models for the corresponding datasets.
[0040] On the one hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-mentioned method for identifying voids in a mineral-rock particle accumulation system.
[0041] On the one hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-described method for identifying voids in a mineral-rock particle accumulation system.
[0042] The above technical solution has at least the following advantages compared with the existing technology:
[0043] The above-described scheme, provided by this invention, employs medical CT scanning technology to accurately and rapidly acquire three-dimensional information on the complex voids within the mineral-rock particle accumulation system and characterize it in slice form. It then utilizes deep learning algorithms, building an AttentionR2U-net neural network based on the Tensorflow and Keras frameworks. The dataset is input into the neural network, and the optimal model is obtained through repeated training and hyperparameter adjustment. Notably, this invention starts with four dataset parameters: the number of images, image size, data augmentation methods, and image complexity. Through quantitative experiments, it explores and identifies the optimal dataset parameters that achieve excellent model training convergence speed and recognition results, thus creating the optimal dataset. The research approach and results of this invention can provide effective technical means for the efficient and accurate identification of voids and the analysis of void structure characteristics in different types of particle accumulation systems such as rock and soil, and have certain guiding significance for research in the field of image recognition related to actual collapsed mines. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0045] Figure 1 This is a schematic flowchart of a method for identifying voids in a mineral-rock particle accumulation system provided by an embodiment of the present invention;
[0046] Figure 2 This is an original slice image of the mineral and rock particle accumulation system obtained by omnidirectional scanning using an MDCT instrument, provided in an embodiment of the present invention.
[0047] Figure 3 This is a diagram showing the effect of using the optimal model to identify voids in a mineral-rock particle accumulation system, provided by an embodiment of the present invention.
[0048] Figure 4 This is a block diagram of a void identification system for a mineral and rock particle accumulation system provided in an embodiment of the present invention;
[0049] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0051] This invention addresses the shortcomings of existing technologies in identifying void structures in complex mineral and rock particle accumulation systems, where research is scarce, insufficient, and superficial. It provides a method and system for identifying voids in such systems.
[0052] like Figure 1 As shown, this embodiment of the invention provides a method for identifying voids in a mineral-rock particle accumulation system, which can be implemented by an electronic device. Figure 1 The flowchart shown is for a method to identify voids in a mineral-rock particle accumulation system. The processing flow of this method may include the following steps:
[0053] S101. Scan the mineral and rock particle accumulation system, obtain the original slice images of the mineral and rock particles, and preprocess the original slice images.
[0054] In one feasible implementation, step S101 involves scanning the ore-rock particle accumulation system to obtain raw slice images of the ore-rock particles, and preprocessing the raw slice images, including:
[0055] X-rays are emitted using medical CT scanning technology to perform a comprehensive spiral scan of the mineral and rock particle accumulation system in the device, and specific information about the gaps between particles inside the system is obtained through slice images.
[0056] The target region is selected by cropping in the sliced image, and the target object is labeled by Labelme to obtain the preprocessed dataset.
[0057] In one feasible implementation, this invention creates its own dataset for neural network training and subsequent research through a series of processes including self-built experimental equipment, CT scanning, image processing, and manual annotation. To investigate the impact of different dataset parameters on neural network training and recognition, this study uses a self-built device model (internal dimensions 25cm×25cm×30cm) to place mineral particles with diameters of 0.5-1cm, 1-1.5cm, 1.5-2.0cm, and a particle size ratio of 2.0-2.5cm to create aggregated mineral particles with different particle size ratios. The mineral particle aggregation system is then scanned omnidirectionally and spirally using Multidetector-row CT technology to obtain original slice images. Figure 2 As shown.
[0058] Considering the messy nature of the raw slice data obtained from CT scans of deposited mineral particles, it is impossible to directly input it into a neural network for training. Therefore, after obtaining high-quality raw data, a series of processing steps are required to obtain a preprocessed dataset that meets the requirements for neural network training: all images are cropped to contain only gaps to filter out target regions in the images; and pixel-level manual annotation of the slice images is performed point by point using the image annotation tool Labelme.
[0059] S102. Select different data parameters from the label images to obtain four types of datasets; divide the four types of datasets into training sets and test sets respectively.
[0060] In one feasible implementation, in step S102, different data parameters are selected to obtain four types of datasets, including:
[0061] By selecting different data parameters, the preprocessed dataset is classified to obtain four types of datasets;
[0062] The dataset parameters include the number of datasets, image size, data augmentation method, and image complexity.
[0063] In one feasible implementation, the preprocessed images are further processed to create four datasets with varying numbers of images, image sizes, data augmentation methods, and image complexity. The datasets contain 200, 500, 1000, and 2000 images; the image (square) side lengths are 200, 300, 400, 500, 800, 1000, and 1200 pixels; the data augmentation methods include traditional algorithms such as rotation and flipping, and center-and-surround sliding cropping; and the image complexity is categorized into single-particle-size images and multi-particle-size images. These four datasets are used for subsequent experiments to analyze the impact of various data parameters on the neural network training and recognition process.
[0064] In one feasible implementation, the data augmentation method includes:
[0065] The center-surround sliding cropping method synchronously processes the slices and their corresponding label images. The specific operation process is as follows: starting from the upper left of the image, a small-sized cropping window is created. The size of this cropping window can be set according to requirements. It slides to the right by a certain number of pixels. After moving to the right boundary, it jumps back to the initial point and slides down by a certain number of pixels. Then, it repeats the rightward sliding, and so on, until it reaches the lower right boundary of the image. Finally, the paired images extracted from the windows in the synchronously processed slice and label image are integrated to create a dataset, thereby achieving data augmentation while changing the image size.
[0066] By using a center-and-around-the-surface cropping method, a dataset of 1000 images with a size of 300-400 pixels was obtained, which includes both simple images with single-particle size and complex images with multiple-particle size.
[0067] S103. AttentionR2U-net was built based on the Tensorflow and Keras frameworks, and the model was obtained by training the network with four different datasets.
[0068] In one feasible implementation, step S103 involves building an AttentionR2U-net based on the Tensorflow and Keras frameworks, and training the network using a dataset to obtain the model, including:
[0069] The advantages of U-net and recurrent network and residual network are combined to build R2U-Net. Attention gates are added to the R2U-net network to build the AttentionR2U-net neural network.
[0070] The AttentionR2U-net neural network was trained using four different datasets to obtain recognition models for the corresponding datasets.
[0071] In one feasible implementation, AttentionR2U-net is an architecture that combines the advantages of U-net with recurrent neural networks (RNNs) and residual networks (ResNets) to form R2U-net. Attention gates are then added to the R2U-net network to create AttentionR2U-net. During encoding and decoding, recurrent residual convolutional blocks are used to significantly increase the network depth. Simultaneously, the attention gate can automatically learn and focus on targets of different shapes and sizes, making it easy to integrate into standard CNN architectures like R2U-net. This results in higher sensitivity, prediction accuracy, and stronger generalization ability.
[0072] S104. Qualitatively and quantitatively evaluate the recognition performance of the model under the four types of datasets based on the labeled images and recognized images, and further determine the optimal dataset parameters by comparing the binary classification confusion matrix index.
[0073] In one feasible implementation, the training and recognition performance of the neural network is evaluated using a binary classification confusion matrix, primarily through label images and recognition images. The optimal dataset parameters are selected by comparing the quantitative evaluation metrics of the model across four datasets. The training set data (cropped slices and corresponding label images) is input into the constructed neural network for training in step S103. The test set is then fed into the trained model for prediction, and the prediction results are compared with the label images to analyze the model's recognition performance from both qualitative and quantitative perspectives.
[0074] In one feasible implementation, since the problem of gap recognition in sliced images only has two categories, gaps and background, it is treated as a binary classification problem. The binary classification evaluation metrics involved include the binary cross-entropy loss function Loss and the Dice similarity coefficient.
[0075] The formula for calculating the cross-entropy loss function (Loss) for binary classification is as follows:
[0076]
[0077] in: y is the predicted value output by the network, and y is the given sample value.
[0078] The formula for calculating the Dice similarity coefficient is as follows:
[0079]
[0080] Where: TP represents the total number of pixels that are both labeled and predicted as gaps; FP represents the total number of pixels that are labeled as background and predicted as gaps; TN represents the total number of pixels that are both labeled and predicted as backgrounds; and FN represents the total number of pixels that are labeled as gaps and predicted as backgrounds.
[0081] In one feasible implementation, to study the impact of different datasets on the neural network, this invention inputs four types of datasets into the AttentionR2U-net neural network for training, and evaluates the training and recognition performance of each dataset based on the evaluation metric of the binary classification confusion matrix. The results are as follows:
[0082] Number of images in the dataset: The training and recognition performance of neural networks tends to improve as the number of images in the dataset increases. However, there is a critical value for the number of images. Once this value is exceeded, significantly increasing the amount of data will not significantly improve the performance of the neural network.
[0083] Image size: If the image size in the dataset is too large, directly introducing it into the neural network will cause the network parameters to grow exponentially, making model training extremely difficult; if the image size is too small, the neural network will not be able to correctly learn the complete features of the target object. Therefore, the image size needs to be within a certain range to better demonstrate the performance of the neural network.
[0084] Data augmentation method: This invention innovatively proposes a "center-surround sliding cropping" data augmentation algorithm. Compared with traditional data augmentation methods, the data augmentation algorithm used in this invention has a faster convergence speed and higher recognition accuracy of the network model under the same number of training times.
[0085] Image complexity: Slice images with different particle sizes exhibit varying degrees of complexity in their data content. Single-particle-size slice images have larger and more uniform gaps, resulting in simpler data content; multi-particle-size slice images have smaller and highly non-uniform gaps, leading to more complex data content. Experiments show that selecting datasets containing both single-particle and multi-particle-size data enriches the data content, thereby improving the robustness and generalization ability of the neural network. Furthermore, when only a single dataset is available, training a model on a more complex dataset allows for acceptable gap recognition in simple CT slice images within acceptable limits.
[0086] S105. Obtain the optimal dataset based on the optimal dataset parameters, and train the neural network using the optimal dataset;
[0087] In one feasible implementation, referring to the optimal data parameters for image data, this invention proposes an optimal dataset for training a neural network, enabling the neural network to better demonstrate its performance. It creates a dataset for the "center-around sliding cropping" data augmentation algorithm, containing 1000 images, 300-400 pixels in size, and including both simple images with single-particle size and complex images with multiple particle sizes.
[0088] The AttentionR2U-net neural network was repeatedly trained on the optimal dataset, and hyperparameters were adjusted to achieve the optimal model. The optimal model was then used to predict on the test set, and the convergence and accuracy of the model were verified using both qualitative and quantitative methods. Notably, the model achieved a Dice value of 0.9859 for gap recognition, an improvement of 0.1455 compared to the traditional algorithm's Dice value of 0.8404.
[0089] S106. Repeatedly train the neural network, adjust the hyperparameters to obtain the optimal model, input the data of the mineral and rock particle accumulation system to be identified, and complete the identification of voids in the mineral and rock particle accumulation system.
[0090] In one feasible implementation, the optimal dataset parameters are determined through steps S101-S104, that is, the values or ranges of image size, total number of images, etc. are determined in the dataset. In step S105, a dataset is generated based on the optimal dataset parameters (this dataset will not be changed). The dataset is put into the neural network for repeated iterative training. Each iteration is recorded as one epoch. One epoch refers to the entire training set being traversed once in the neural network. After each epoch, the neural network automatically corrects and updates the parameters in backpropagation through the loss index. Then, the hyperparameters of the neural network are made corresponding minor improvements according to the convergence process of the model, so that the entire training process is carried out in the direction of lower loss and higher accuracy.
[0091] One feasible implementation method is, for example Figure 3 As shown. In the analysis of model recognition results, the difficulty of recognition varies due to differences in the size of the gaps. To make the evaluation effect more reasonable, this invention redefines the Dice similarity coefficient based on the research content, dividing the gaps in the image into large gaps and small gaps according to the area occupied by each gap, and then evaluating and analyzing them separately (Dice). l and Dice s ).
[0092] In this embodiment of the invention, the method provided is based on AttentionR2U-net for void identification in mineral-rock particle accumulation systems with optimized image dataset parameters. Utilizing medical MDCT scanning technology and the AttentionR2U-net deep learning network model, it acquires three-dimensional spatial information of the relative positions of particles and voids within the mineral-rock particle accumulation system and represents this information with slice images. Based on this, a CT slice void image database is established. AttentionR2U-net is used to explore and process image data with optimal data parameters. The optimal dataset is then used to train a neural network to obtain the optimal model. Ultimately, this achieves efficient, rapid, and accurate identification of slice void images of mineral-rock particle accumulation systems. This invention significantly improves model convergence efficiency and identification accuracy, laying the foundation for subsequent extraction of relevant structural feature parameters such as porosity, void radius distribution, and throat radius distribution, as well as analysis of the macroscopic state characteristics of the particle system.
[0093] This invention addresses the shortcomings of existing research in analyzing the internal voids of ore-rock particle accumulation systems and the difficulty for neural networks in image recognition to simultaneously achieve fast convergence speed and high recognition accuracy. It delves into the impact of dataset parameters on network model training and recognition, qualitatively and quantitatively characterizing the results. An optimal dataset is created using the optimal dataset parameters, and the neural network is repeatedly trained and its parameters adjusted to obtain the optimal model. This enables efficient, rapid, and accurate intelligent identification of internal voids in particle accumulation systems.
[0094] Figure 4 This is a schematic diagram of a void identification system for a mineral-rock particle accumulation system according to the present invention. The system 300 is used in the above-mentioned void identification method for a mineral-rock particle accumulation system, and the system 300 includes:
[0095] The data processing module 310 is used to scan the mineral and rock particle accumulation system, obtain the original slice image of the mineral and rock particles, and preprocess the original slice image to obtain the label image.
[0096] The data classification module 320 is used to select different data parameters in the label image to obtain four types of datasets; the four types of datasets are then divided into training set and test set.
[0097] The model building module 330 is used to build AttentionR2U-net based on the Tensorflow and Keras frameworks, and to obtain the model by training the network using the training set.
[0098] The preliminary evaluation module 340 is used to qualitatively and quantitatively evaluate the recognition performance of the model under four types of datasets based on the label images and the original slice images, and further determine the optimal dataset parameters by comparing the binary classification confusion matrix index.
[0099] The model training module 350 is used to obtain the optimal dataset based on the optimal dataset parameters and to train the neural network using the optimal dataset.
[0100] The void recognition module 360 is used to repeatedly train the neural network with the optimal dataset, adjust the hyperparameters to obtain the optimal model, input the data of the mineral and rock particle accumulation system to be identified, and complete the void recognition of the mineral and rock particle accumulation system.
[0101] Optionally, the data processing module 310 is used to emit X-rays through medical CT scanning technology to perform an all-round spiral scan of the mineral and rock particle accumulation system in the device, and to obtain specific information about the gaps between particles inside the system through slice images.
[0102] The target region is selected by cropping in the sliced image, and the target object is labeled by Labelme to obtain the preprocessed dataset.
[0103] Optionally, the data classification module 320 is used to select different data parameters to classify the preprocessed dataset and obtain four types of datasets;
[0104] The dataset parameters include the number of datasets, image size, data augmentation method, and image complexity.
[0105] Optionally, the model building module 330 is used to combine the advantages of U-net with recurrent networks and residual networks to build R2U-Net, and to add attention gates to the R2U-net network to build the AttentionR2U-net neural network.
[0106] In this embodiment of the invention, the method provided is based on the AttentionR2U-net model with optimized parameters for image datasets to identify voids in mineral-rock particle accumulation systems. Utilizing medical MDCT scanning technology and the AttentionR2U-net deep learning network model, the three-dimensional spatial information of the relative positions of particles and voids within the mineral-rock particle accumulation system is obtained and represented by slice images. Based on this, a CT slice void image database is established. The AttentionR2U-net model is used to explore and process image data with optimal parameters. The optimal dataset is then used to train the neural network to obtain the optimal model. Ultimately, this achieves efficient, fast, and accurate identification of slice void images of mineral-rock particle accumulation systems. This invention significantly improves model convergence efficiency and recognition accuracy, laying the foundation for subsequent extraction of relevant structural feature parameters such as porosity, void radius distribution, and throat radius distribution, as well as analysis of the macroscopic state characteristics of the particle system.
[0107] This invention addresses the shortcomings of existing research in analyzing the internal voids of ore-rock particle accumulation systems and the difficulty for neural networks in image recognition to simultaneously achieve fast convergence speed and high recognition accuracy. It delves into the impact of dataset parameters on network model training and recognition, qualitatively and quantitatively characterizing the results. An optimal dataset is created using the optimal dataset parameters, and the neural network is repeatedly trained and its parameters adjusted to obtain the optimal model. This enables efficient, rapid, and accurate intelligent identification of internal voids in particle accumulation systems.
[0108] Figure 5 This is a schematic diagram of the structure of an electronic device 400 provided in an embodiment of the present invention. The electronic device 400 can vary considerably due to differences in configuration or performance. It may include one or more central processing units (CPUs) 401 and one or more memories 402. The memory 402 stores at least one instruction, which is loaded and executed by the processor 401 to implement the steps of the following method for identifying voids in a mineral-rock particle accumulation system:
[0109] S1. Scan the mineral and rock particle accumulation system to obtain the original slice images of the mineral and rock particles, and preprocess the original slice images.
[0110] S2. Select different data parameters to obtain four types of datasets;
[0111] S3. Build AttentionR2U-net based on Tensorflow and Keras frameworks, and train the network to obtain the model using four different datasets;
[0112] S4. Qualitatively and quantitatively evaluate the recognition performance of the model under the four types of datasets based on the labeled images and recognized images, and further determine the optimal dataset parameters by comparing the binary classification confusion matrix index.
[0113] S5. Obtain the optimal dataset based on the optimal dataset parameters, and train the neural network using the optimal dataset;
[0114] S6. Repeatedly train the neural network, adjust the hyperparameters to obtain the optimal model, input the data of the mineral and rock particle accumulation system to be identified, and complete the void identification of the mineral and rock particle accumulation system.
[0115] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the above-described method for identifying voids in a mineral-rock particle accumulation system. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.
[0116] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.
Claims
1. A method for identifying voids in a mineral-rock particle accumulation system, characterized in that, The method steps include: S1. Scan the mineral and rock particle accumulation system to obtain the original slice image of the mineral and rock particles, and preprocess the original slice image to obtain the label image; S2. Select different data parameters from the label images to obtain four types of datasets; divide the four types of datasets into training sets and test sets respectively. By selecting different data parameters, the preprocessed dataset is classified to obtain four types of datasets; The dataset parameters include the number of datasets, image size, data augmentation method, and image complexity. S3. Build AttentionR2U-net based on the Tensorflow and Keras frameworks, and train the network using the training set to obtain the model; The advantages of U-net and recurrent network and residual network are combined to build R2U-Net. Attention gates are added to the R2U-net network to build the AttentionR2U-net neural network. The AttentionR2U-net neural network was trained using four different datasets to obtain recognition models for the corresponding datasets. S4. Qualitatively and quantitatively evaluate the recognition performance of the model under the four types of datasets based on the labeled images and the original slice images, and further determine the optimal dataset parameters by comparing the binary classification confusion matrix index. Acquire labeled images and recognize images; The problem of identifying gaps in sliced images is defined as a binary classification problem, and a binary classification cross-entropy loss function is constructed. Based on the labeled image and the recognized image, and combined with the binary classification cross-entropy loss function, the optimal dataset parameters are selected by comparing the quantitative evaluation indicators of the model under the four datasets. S5. Based on the parameters of the optimal dataset, an optimal dataset is obtained, and a neural network is trained using the optimal dataset; S6. The neural network is repeatedly trained using the optimal dataset, and the hyperparameters are adjusted to obtain the optimal model. The data of the mineral and rock particle accumulation system to be identified is then input to complete the identification of voids in the mineral and rock particle accumulation system.
2. The method according to claim 1, characterized in that, In step S1, the mineral-rock particle accumulation system is scanned to obtain raw slice images of the mineral-rock particles, and the raw slice images are preprocessed to obtain label images, including: X-rays are emitted using medical CT scanning technology to perform a comprehensive spiral scan of the mineral and rock particle accumulation system in the device, and specific information about the gaps between particles inside the system is obtained through slice images. The target region is selected by cropping in the sliced image, and the target object is labeled and preprocessed using Labelme to obtain the label image.
3. The method according to claim 2, characterized in that, The data augmentation method includes: By using a center-and-surround-around cropping method, the slices and their corresponding label images are processed simultaneously. Data augmentation is performed while changing the image size, resulting in a dataset of 1000 images with a size of 300-400 pixels, including both simple images with single-particle size and complex images with multiple particle sizes.
4. A system for identifying voids in a mineral-rock particle accumulation system, characterized in that, The system is used for the void identification method of the mineral-rock particle accumulation system as described in any one of claims 1 to 3, the system comprising: The data processing module is used to scan the mineral and rock particle accumulation system, obtain the original slice images of the mineral and rock particles, and preprocess the original slice images to obtain label images. The data classification module is used to select different data parameters in the label image to obtain four types of datasets; the four types of datasets are then divided into training sets and test sets. The model building module is used to build AttentionR2U-net based on the Tensorflow and Keras frameworks, and to train the network using the training set to obtain the model. The preliminary evaluation module is used to qualitatively and quantitatively evaluate the recognition performance of the model on four datasets based on the labeled images and the original slice images. The optimal dataset parameters are further determined by comparing the binary classification confusion matrix index. The model training module is used to generate the optimal dataset based on the parameters of the optimal dataset, and to train the neural network using the optimal dataset. The void identification module is used to repeatedly train the neural network with the optimal dataset, adjust the hyperparameters to obtain the optimal model, input the data of the mineral and rock particle accumulation system to be identified, and complete the void identification of the mineral and rock particle accumulation system.
5. The system according to claim 4, characterized in that, The data processing module is used to emit X-rays through medical CT scanning technology to perform a full-range spiral scan of the mineral and rock particle accumulation system in the device, and to obtain specific information about the gaps between particles inside the system through slice images. The target region is selected by cropping in the sliced image, and the target object is labeled by Labelme to obtain the preprocessed dataset.
6. The system according to claim 5, characterized in that, The data classification module is used to select different data parameters to classify the preprocessed dataset and obtain four types of datasets. The dataset parameters include the number of datasets, image size, data augmentation method, and image complexity.
7. The system according to claim 6, characterized in that, The model building module is used to combine the advantages of U-net with recurrent networks and residual networks to build R2U-Net, and to add attention gates to the R2U-net network to build the AttentionR2U-net neural network. The AttentionR2U-net neural network was trained using four different datasets to obtain recognition models for the corresponding datasets.