A method and system for ore classification

Abstract

The present invention discloses an ore classification method and system, which relates to the technical field of ore sorting. The method comprises: obtaining hyperspectral image data of the ore to be inspected, extracting a feature image highlighting the difference between the ore and the background from the hyperspectral image data, and obtaining an RGB image of the ore to be inspected; detecting the RGB image of the ore to be inspected, and obtaining positioning parameters; constructing a preselection box based on the positioning parameters, determining the edge contour and mask of each ore to be inspected through the preselection box, and obtaining the texture information and spectral information of the ore to be inspected; extracting features from the texture information and spectral information of the ore to be inspected, and adaptively fusing the extracted features to obtain fused features including the spectral features and texture features of the ore to be inspected; inputting the fused features into a hyperspectral image classification model for identification, and obtaining the category results of each ore to be inspected. The present invention can reduce the influence of error factors in the information collection process and improve the accuracy of ore classification.

Description

Technical Field

[0001] The present invention relates to the technical field of ore sorting, and in particular to an ore sorting method and system. Background Art

[0002] Mineral resources are the backbone of modern industrial production and have made significant contributions to science and technology, transportation and construction industries. With the continuous increase in annual mining volume, mineral reserves have shown a clear downward trend, with the characteristics of poor quality, complexity and difficulty in separation. At the same time, the sharp increase in the amount of surrounding rock during mining also leads to high energy consumption for crushing and grinding, and reduces the average grade of ore. Therefore, it is necessary to quickly identify and classify ore after mining and before washing to achieve the purpose of pretreatment, reduce costs and improve efficiency.

[0003] In the prior art, ore sorting equipment based on radiation sensors has attracted much attention. The specific process of sorting ore includes: the ore raw materials are evenly fed into the machine through a vibrating feeder or similar equipment. To ensure that the ore particles can be evenly spread in the color sorting area, multi-stage feeding is usually adopted. The first stage controls the feeding amount, and the second and third stages make the ore queue up. The ore blocks are in a single-layer stable discrete state, and the ore blocks are separated by a certain distance. When the ore particles are irradiated by a high-resolution CCD (charge-coupled device) image sensor and a matching light source, the sensor can capture the light reflected or transmitted by the surface of the ore and convert it into a digital signal. The ore is detected and distinguished by optical, acoustic, electrical and other sensors to select the target mineral. According to the output signal of the sensor, the control system drives the mechanical plate or solenoid valve to work and sort out the target mineral. The ore particles that do not meet the requirements are blown into the defective slot of the discharge hopper and flow away, while the ore that meets the requirements continues to fall into the finished product slot of the receiving hopper and flows out.

[0004] The defects of the above-mentioned prior art are that the radiation sensor is easily affected by the background and impurities in the ore during the process of collecting light reflected or transmitted from the surface of the ore, so that the collected information is not accurate enough, resulting in errors in the classification results of the ore. Summary of the invention

[0005] Based on this, it is necessary to provide an ore classification method and system to address the above technical problems.

[0006] An embodiment of the present invention provides an ore classification method, comprising:

[0007] Acquire hyperspectral image data of the ore to be inspected, extract a feature image highlighting the difference between the ore and the background from the hyperspectral image data, and obtain an RGB image of the ore to be inspected;

[0008] The target detection algorithm is used to detect the RGB image of the ore to be inspected, and the positioning parameters used to characterize the size parameters of the ore to be inspected are obtained;

[0009] A pre-selection box is constructed based on the positioning parameters to determine the position of the ore to be inspected, and the edge contour and mask of each ore to be inspected are determined through the pre-selection box to obtain the texture information and spectral information of the ore to be inspected;

[0010] Extracting features from the texture information and spectral information of the ore to be inspected, and adaptively fusing the extracted features to obtain fused features including the spectral features and texture features of the ore to be inspected;

[0011] The fused features are input into the hyperspectral image classification model for identification to obtain the category results of each ore to be inspected.

[0012] In addition, the acquisition of a hyperspectral image of the ore to be inspected specifically includes:

[0013] The ore raw materials collected from the ore-bearing layer of the mine are crushed and the ore with a particle size of 8cm×8cm is screened out;

[0014] The hyperspectral image of the screened ore to be inspected is captured by a hyperspectral camera.

[0015] In addition, the positioning parameters obtained for characterizing the size parameters of the ore to be inspected specifically include:

[0016] In the RGB image of the ore to be inspected, the characteristic images including carbonate minerals, clay minerals and diaspore are selected to construct a false color composite image;

[0017] The target detection algorithm is used to locate the false color synthetic image, and four positioning parameters x, y, w, and h are obtained, where x and y are the coordinates of the center point of the ore to be inspected, which are used as the center point coordinates of the pre-selection box; w is the maximum radial length of the ore to be inspected, which is used as the width of the pre-selection box; h is the maximum axial height of the ore to be inspected, which is used as the length of the pre-selection box.

[0018] In addition, the edge contour and mask of each ore to be inspected are determined by the pre-selection frame, which specifically includes:

[0019] Convert the RGB image in each pre-selected box to the YCbCr color space and calculate the histogram of the image;

[0020] The histogram is processed by one-dimensional Gaussian filtering to obtain a smooth histogram, in which the ore and the background have different characteristic peaks respectively;

[0021] Determine the segmentation threshold according to the position between the ore characteristic peak and the background characteristic peak in the histogram, and use the segmentation threshold to eliminate the background in the RGB image to obtain the RGB image after the background is eliminated;

[0022] Perform Gaussian blur on the image and perform grayscale conversion to obtain a grayscale image;

[0023] Use contour extraction algorithm to draw edge contours of ore in each pre-selected box in the grayscale image;

[0024] Then, an edge detection algorithm is used to detect the edge of the ore in the grayscale image, find the boundary of the ore, and generate a mask of the ore in each pre-selected box according to the boundary contour.

[0025] In addition, the extracted features are adaptively fused, which specifically includes:

[0026] Different texture features are fused with spectral features to obtain channel features; the features of each channel are compressed with global information to generate channel descriptors;

[0027] Generate an attention weight vector for each channel descriptor, perform element-by-element multiplication of the weight vector with the channel descriptor of each channel, and obtain a fusion feature containing the spectral features and texture features of the ore to be inspected.

[0028] In addition, the hyperspectral image classification model includes a spectral feature learning group, a spatial feature learning group and a feature recalibration module F EFR ;

[0029] The spectral feature learning group is used to extract the spatially focused spectral features in the fusion features;

[0030] The spatial feature learning group is used to extract the spatial features of spectral focusing in the fusion features;

[0031] The feature recalibration module F EFR , which is used to dynamically adjust the weights of spectral features and spatial features.

[0032] Additionally, an ore classification system comprising:

[0033] An image acquisition module is used to obtain hyperspectral image data of the ore to be inspected, extract a feature image highlighting the difference between the ore and the background from the hyperspectral image data, and obtain an RGB image of the ore to be inspected;

[0034] A computer control module, comprising: a target detection module, a contour elimination module, a feature extraction module, and a classification module;

[0035] The target detection module is used to detect the RGB image of the ore to be inspected using a target detection algorithm to obtain positioning parameters for characterizing the size parameters of the ore to be inspected;

[0036] The contour elimination module is used to construct a pre-selection box for determining the position of the ore to be inspected based on the positioning parameters, determine the edge contour and mask of each ore to be inspected through the pre-selection box, and obtain the texture information and spectral information of the ore to be inspected;

[0037] The feature extraction module is used to extract features from the texture information and spectral information of the ore to be inspected, and adaptively fuse the extracted features to obtain fused features including the spectral features and texture features of the ore to be inspected;

[0038] The classification module is used to input the fusion features into the hyperspectral image classification model for identification, so as to obtain the classification result of each ore to be inspected.

[0039] In addition, it also includes a particle size screening module, which is used to crush the ore raw materials collected from the ore-bearing layer of the mine, and put the crushed ore into the particle size screening module to screen out the ore to be tested with a particle size of 8cm×8cm.

[0040] In addition, it also includes a material conveying module, which is used to disperse the ore to be inspected through the speed difference between conveyor belts of different levels.

[0041] In addition, it also includes an injection module, which is used to send the ore to be inspected into the corresponding secondary conveyor belt according to the classification result of each ore to be inspected.

[0042] Compared with the prior art, the above-mentioned ore classification method and system provided by the embodiment of the present invention has the following beneficial effects:

[0043] In the process of collecting light reflected or transmitted from the surface of the ore, the radiation sensor of the existing ore sorting equipment is easily affected by the background and impurities in the ore, making the collected information inaccurate, which in turn leads to errors in the classification results of the ore.

[0044] The present invention constructs a pre-selection box based on the positioning parameters of the ore to be inspected, and determines the edge contour and mask of each ore to be inspected through the pre-selection box to eliminate the background and impurities in the hyperspectral image of the ore to be inspected, thereby reducing the influence of error factors in the information collection process, obtaining accurate texture information and spectral information of the ore to be inspected, and further improving the accuracy of ore classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A method flow chart of an ore classification method provided in one embodiment;

[0046] Figure 2 A system structure diagram of an ore classification system provided in one embodiment;

[0047] Figure 3 A feature extraction and fusion process diagram of an ore classification method provided in an embodiment;

[0048] Figure 4 A classification model structure diagram of an ore classification method provided in an embodiment;

[0049] Figure 5 A residual module structure diagram of an ore classification system provided in one embodiment. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] In one embodiment, a method for classifying ore is provided, such as Figure 1 As shown, the method includes:

[0052] The hyperspectral image data of the ore to be inspected is obtained, and a feature image highlighting the difference between the ore and the background is extracted from the hyperspectral image data to obtain an RGB image of the ore to be inspected.

[0053] The target detection algorithm is used to detect the RGB image of the ore to be inspected, and the positioning parameters used to characterize the size parameters of the ore to be inspected are obtained.

[0054] A preselection box for determining the position of the ore to be inspected is constructed based on the positioning parameters, and the edge contour and mask of each ore to be inspected are determined through the preselection box to obtain the texture information and spectral information of the ore to be inspected.

[0055] Feature extraction is performed on the texture information and spectral information of the ore to be inspected, and the extracted features are adaptively fused to obtain fused features including the spectral features and texture features of the ore to be inspected.

[0056] The fused features are input into the hyperspectral image classification model for identification to obtain the category results of each ore to be inspected.

[0057] In one embodiment, a mineral classification system is provided, such as Figure 2 As shown, the system includes:

[0058] The particle size screening module is used to crush the ore raw materials collected from the ore-bearing layer of the mine, and put the crushed ore into the particle size screening module to screen out the ore to be tested with a particle size of 8cm×8cm.

[0059] The material conveying module is used to disperse the ore to be inspected by the speed difference between the conveyor belts at different levels.

[0060] The image acquisition module is used to obtain the hyperspectral image data of the ore to be inspected, extract the characteristic image highlighting the difference between the ore and the background from the hyperspectral image data, and obtain the RGB image of the ore to be inspected.

[0061] The computer control module includes: a target detection module, a contour elimination module, a feature extraction module, and a classification module.

[0062] The target detection module is used to detect the RGB image of the ore to be inspected using the target detection algorithm to obtain positioning parameters for characterizing the size parameters of the ore to be inspected;

[0063] The contour rejection module is used to construct a pre-selection box for determining the position of the ore to be inspected based on the positioning parameters, and to determine the edge contour and mask of each ore to be inspected through the pre-selection box to obtain the texture information and spectral information of the ore to be inspected.

[0064] The feature extraction module is used to extract features from the texture information and spectral information of the ore to be inspected, and adaptively fuse the extracted features to obtain fused features including the spectral features and texture features of the ore to be inspected.

[0065] The classification module is used to input the fused features into the hyperspectral image classification model for identification, and obtain the category results of each ore to be inspected.

[0066] The injection module is used to send the ore to be inspected to the corresponding secondary conveyor belt according to the classification result of each ore to be inspected.

[0067] Example 1

[0068] 1. Use a hyperspectral camera to collect hyperspectral images of ore, specifically:

[0069] The ore raw materials collected from the ore-bearing layer of the mine are crushed.

[0070] The crushed ore is put into the particle size screening module, which screens out ore with a particle size of 8cm×8cm, and uses a hyperspectral camera to capture a hyperspectral image of the screened ore to be inspected.

[0071] The ore particles output by the particle size screening module enter the material transmission system, which has three levels of conveyor belts with different speeds.

[0072] The speed of the three-stage conveyor belt of the material conveying module increases in sequence. The specific speed configuration should depend on the speed matching of the third-stage conveyor belt and the hyperspectral camera. The speed of the third-stage conveyor belt also determines the speed of ore separation. The conveying speed of the three-stage conveyor belt is configured in stages according to the scanning speed of the hyperspectral camera (2.10m / s).

[0073] The material transfer model library utilizes the speed difference between conveyor belts of different levels to disperse the ore, eliminate the phenomenon of ore overlapping and close proximity, and ensure the ore scanning quality of the image acquisition system.

[0074] The ore screened by the particle size screening module enters the material conveying system. The third-level transmission belt conveys the ore at a constant speed. The conveyor belt speed needs to match the scanning imaging speed of the hyperspectral camera.

[0075] Two groups of light sources are set up, and the power of each group of light sources is consistent, and they are perpendicular to the movement direction of the conveyor belt. The brightest light spot is located at the scanning line position of the hyperspectral camera.

[0076] The hyperspectral camera used in this embodiment is a HySpex Baldur S-384N hyperspectral camera, with a working distance of 1m, a conveyor belt running speed of 2.10m / s, a pixel size of 1.82mm, a hyperspectral camera scanning width of 70cm, and a scanning bandwidth of 384 pixels. The hyperspectral camera parameters are set in advance, and the conveyor belt speed just below the hyperspectral image acquisition system is adjusted to match the hyperspectral camera scanning speed, without image distortion. A whiteboard is placed in the material transmission system, and the surface of the whiteboard is adjusted to be consistent with the surface of the ore (raised by 5cm). The whiteboard moves with the conveyor belt to calibrate the hyperspectral camera.

[0077] Based on the analysis of the spectral curves of different grades of clay-type lithium ores, the spectral sampling interval is set to the 1300-2500nm range. In this spectral range, clay minerals have diagnostic absorption characteristics that can be used to distinguish the mineral composition of the ore and thus determine the ore grade.

[0078] In this embodiment, imaging hyperspectral data of 19,840 clay-type lithium ore samples were collected. The ore grade was analyzed using expert knowledge, and a method combining chemical analysis and spectral analysis was used to mark the grade and location of all ores. All ores were divided into four categories (high grade, medium grade, low grade, and surrounding rock) to form a database for experiments and verification.

[0079] 2. Use the target detection algorithm to detect the RGB image of the ore to be inspected, specifically:

[0080] The target detection model should preferably use a single-stage lightweight network, such as SSD or YOLO series algorithms to detect and locate ores.

[0081] Three bands that can highlight the difference between ore and background are selected from the hyperspectral image to form an RGB image. The location annotations of the ore in the database are combined with the RGB image to form a target detection dataset. The target detection dataset is divided into 8:1:1, with 8 being the training set, and the validation set and test set each accounting for one tenth of the dataset.

[0082] Ore target detection actually only needs to detect and locate the ore. There are only two categories on the image, one is ore and the other is background. Referring to the application foundation of computer vision, this experiment uses the YOLO V5 model for target detection and uses the test set to evaluate the accuracy and generalization ability of the model. After many experiments, the test set accuracy of the ore target detection task is greater than 99%.

[0083] In the RGB image of the ore to be inspected, the characteristic images containing carbonate minerals, clay minerals and diaspore are selected to construct a false color synthetic image. The false color synthetic image is positioned using the target detection algorithm to obtain four positioning parameters x, y, w, and h, where x and y are the coordinates of the center point of the ore to be inspected, which are used as the center point coordinates of the pre-selection box; w is the maximum radial length of the ore to be inspected, which is used as the width of the pre-selection box; h is the maximum axial height of the ore to be inspected, which is used as the length of the pre-selection box.

[0084] 3. Determine the edge contour and mask of each ore to be inspected through the pre-selection box, and follow the steps below:

[0085] The ore target detection stage only requires three-band RGB images or single-band grayscale images, and has extremely high accuracy.

[0086] First, the RGB image in each pre-selected box is converted to the YCbCr color space, and the histogram of the image is calculated. A one-dimensional Gaussian filter is used to smooth the histogram to eliminate noise and reduce data fluctuations, making the histogram smoother and more continuous. The smoothed histogram has different characteristic peaks for the ore and background. The segmentation threshold of the ore and background is determined according to the position of the peaks on the histogram, and the background can be eliminated using the segmentation threshold.

[0087] The RGB image after background removal is further processed to obtain the location mask of the ore. This step is only performed on the pre-selected box of each ore. First, the image is Gaussian blurred to reduce the noise in the image. Next, the image is gray-converted to convert the RGB color image into a gray-scale image, and then binarized.

[0088] Use contour extraction algorithm to draw the edge contour of the ore in each pre-selected box in the grayscale image. Then use edge detection algorithm to detect the edge of the ore in the grayscale image, find the boundary of the ore, and generate the mask of the ore in each pre-selected box according to the boundary contour. This embodiment uses the simple contour approximation algorithm CHAIN_APPROX_SIMPLE. Its principle is: for a boundary contour, the adjacent pixels on its boundary are adjacent, and they are adjacent in the horizontal, vertical or diagonal direction. This continuity can be used to represent the boundary by connecting adjacent pixels. This method approximates the edge contour by removing redundant points in the contour. Specifically, it checks three adjacent points in the contour. If these three points are close to collinear, the middle point will be ignored. In this way, the same edge contour can be represented with fewer points.

[0089] The real-time intelligent hyperspectral ore sorting method needs to take into account both accuracy and the computing power of edge environment equipment, so it is designed as a two-stage ore sorting solution. In the stage of ore hyperspectral image classification, it is necessary to make full use of the texture information and spectral information of the ore. Therefore, it is not possible to directly perform principal component analysis or other simple data dimensionality reduction methods on the hyperspectral image. In order to reduce computing power consumption, the ore is classified separately according to the pre-selected box where the ore is located. This can fully retain the ore spectral information and texture information, and can also reduce the computing power requirements of edge environment equipment.

[0090] This embodiment processes the ore hyperspectral image according to the preselected box of the target detection algorithm and the mask obtained by the above ore processing. Each preselected box is taken as a sample, and the ore in the preselected box is processed according to the mask, and the part outside the mask position is filled with 0. So far, the background and non-ore small particles of the ore hyperspectral image are removed, and it is ensured that there is only one ore in each sample, and there are almost no other factors that interfere with the classification of the ore hyperspectral image.

[0091] Even after screening, the particle size of the ore is not completely uniform, which requires further processing of the ore hyperspectral image. This embodiment transforms the ore hyperspectral image into a hyperspectral image with a size of 128*128, calculates the size of the hyperspectral image, and ensures that the hyperspectral image is in the center of the new image.

[0092] The ore hyperspectral image is transformed into a new image with a size of 128*128 by placing the original image at the center and then filling the surrounding area with zero padding.

[0093] Because the industrial application of ore hyperspectral image classification tasks has extremely high requirements for accuracy and robustness, this embodiment performs data enhancement on the data. Because the use of non-geometric image transformations in hyperspectral images will damage spectral information, this embodiment uses two geometric data enhancement methods, rotation and cropping, to enhance the data.

[0094] The ore hyperspectral image is combined with the annotation of ore grade category to create an ore hyperspectral image classification dataset, and the training set, test set, and validation set are divided into 8:1:1 ratios. The grade recognition ability and robustness of the model are judged based on the accuracy of the validation set.

[0095] 4. Extract the texture information and spectral information of the ore to be inspected, as follows:

[0096] The clay minerals in clay-type lithium ores have diagnostic absorption characteristics in the short-wave infrared range, and many of these diagnostic characteristics are unique to the spectrum. The lithium mineral content of clay-type lithium ores is strongly correlated with the content of clay minerals, so the spectral information of short-wave infrared is very important for analyzing the grade of clay-type lithium ores. Because clay-type lithium ores are mainly located in carbonate interlayers, there are large differences in texture characteristics (color, texture, particle size, etc.) between the ore and the carbonate surrounding rock, and rock texture information is also important for distinguishing the surrounding rock from the ore.

[0097] This embodiment uses two 3D convolution kernels of different sizes and step sizes to respectively extract the spectral information and texture features of the ore. A 3D convolution kernel with a size of 1×1×7 and a step size of (1, 1, 2) is used to extract spectral features, and a 3D convolution kernel with a size of 3×3×7 and a step size of (1, 1, 2) is used to extract texture features.

[0098] The network uses convolution kernels of different sizes to process images. Each size of the convolution kernel has its own unique receptive field size, which can capture features from local to global.

[0099] The specific extraction process is as follows: Figure 3 As shown, Figure 3 The S in it is the length and width of the image. Figure 3 The B in is the number of bands, Figure 3 The GAP in is global pooling. Figure 3 The FC in is a fully connected layer. The feature extraction module uses an attention-based adaptive texture-spectral kernel module, which can dynamically adjust the weights of different convolution kernels according to the attention mechanism, thereby achieving adaptive adjustment of the receptive field size.

[0100] Fusion and selection: After the extracted features are fused, they are selected using the attention mechanism. The attention mechanism assigns weights to the features extracted by different convolution kernels based on the channel descriptors of the features, so that the size of the receptive field can be automatically adjusted according to the different image contents.

[0101] The specific implementation is to add the texture features and the spectral features, and then input them into the pooling layer of the feature extraction module. The pooling layer compresses the global information of the channel features and generates a channel descriptor.

[0102] A fully connected layer is used to implement the channel attention mechanism, and a Softmax layer is used to compress the output of the fully connected layer into the range of (0, 1) to generate an attention weight vector for each channel.

[0103] The weight vector is multiplied element by element with the original feature map to finally obtain the ore spectrum-texture feature map. This operation strengthens the features that are helpful for ore sorting, while suppressing redundant features, helping the model adapt to the complexity of hyperspectral data, learning ore features more effectively, and maximizing computational efficiency.

[0104] This weight distribution is performed dynamically when processing different mineral hyperspectral images. For images containing different objects or features, the network will adaptively adjust the size of the receptive field to more centrally process important features. After the adaptive weight adjustment, the feature map output by the network will show different focusing effects in different areas.

[0105] This mechanism allows the network to increase the receptive field in a specific image area in order to better capture important information. For example, if the model detects that a certain area has complex texture or edge information, it may increase the receptive field by increasing the weight of the convolution kernel in that area.

[0106] 5. Input the fused features into the hyperspectral image classification model for recognition, as follows:

[0107] The fused features are input into the hyperspectral image classification model for identification to obtain the category results of each ore to be inspected.

[0108] The structure of the hyperspectral image classification model is as follows: Figure 4 As shown in Figure 5 shown.

[0109] The hyperspectral image classification model is mainly composed of four residual modules, which receive weighted feature maps. There are three types of residual modules, namely residual module 1, residual module 2, and residual module 3. The fine-tuning of the residual block structure at different stages mainly involves the sequential changes of convolutional layers, batch normalization, and Relu activation functions.

[0110] In the model, the residual modules are divided into two groups according to their functions: spectral feature learning group and spatial feature learning group. The spectral feature learning group is residual module 1 plus residual module 2, which is used to extract spatially focused spectral features, and the spatial feature learning group is residual module 2 plus residual module 3, which is used to extract spectrally focused spatial features. Dividing the residual modules into spectral feature learning and spatial feature learning according to their functions helps the model to process different types of information more specifically, accelerate model convergence and improve performance. The spectral feature learning module focuses on extracting spectral information, while the spatial feature learning module focuses on extracting ore structure information, thereby improving the representation ability and interpretability of the model.

[0111] The hyperspectral image classification model also introduces a feature recalibration module (Feature Re-calibrationModule, FEFR ), whose purpose is to dynamically adjust the weights of spectral features and spatial features to recalibrate the importance of features. By strengthening features that are beneficial to the current task and suppressing unimportant features, the computational efficiency of the model in processing features is improved, ensuring maximum performance under limited computing resources.

[0112] The feature recalibration module first extracts features from the previous residual layer. In the hyperspectral image classification task, these features include the spatial and spectral information of the image. The module uses global average pooling (GAP) to extract global statistical information of the feature map and capture the global distribution of each feature channel. Then, a set of fully connected layers FC are used to compress and activate the extracted global information to learn the dependencies between different features. Through training, the feature recalibration module learns the weights of each feature channel, which reflect the importance of the channel in the current task. The original features are recalibrated using the learned feature weights, and the Sigmoid activation function is used to generate the weights of each feature channel. Finally, the recalibrated weights are applied to the original feature map to strengthen the features that are helpful for the current task and suppress unimportant features, helping the model adapt to the complexity of hyperspectral data and maximize computational efficiency.

[0113] Specifically, the model's processing starts with a tensor V whose shape is (B, D, H, W, C), where B represents the batch size, D represents the depth, H represents the height, W represents the width, and C represents the number of channels.

[0114] First, the tensor V enters residual block 1. In residual block 1, the tensor V first passes through a 3D convolution layer, which extracts local features through several filters. Next, the output tensor passes through a batch normalization layer to standardize the output of the convolution layer to accelerate model training and improve stability. The standardized tensor then passes through the ReLU activation function to increase nonlinearity, allowing the model to learn more complex features. After that, the tensor passes through a 3D convolution layer again to further extract features, and is once again processed by a batch normalization layer and ReLU activation function. Next, the tensor passes through the feature recalibration module (F EFR ). Feature Recalibration Module F EFR First, the global statistics of the feature map are extracted through the global average pooling (GAP) layer to capture the global distribution of each feature channel. Then, the feature map passes through the fully connected layer to compress and activate the global information to learn the dependencies between different features. Next, the module generates weights for each feature channel and recalibrates the original features. Finally, a residual connection is added (i.e., the input tensor is added to the processed tensor) and processed through the ReLU activation function to generate the output of residual block 1.

[0115] Next, the tensor enters residual block 2, and the process is similar to residual block 1. The tensor is first processed by a 3D convolution layer, a batch normalization layer, and a ReLU activation function, and then passes through a second 3D convolution layer, and then again passes through a batch normalization layer and a ReLU activation function. Subsequently, the tensor passes through the feature recalibration module F EFR , the process is the same as F in residual block 1 EFR The processing is similar. Finally, a residual connection is added and processed through the ReLU activation function to generate the output of residual block 2.

[0116] The tensor then passes through the second 3D convolutional layer for feature extraction. After this convolutional layer, the shape of the tensor becomes (B, 128, 1, H, W). Next, we resize the dimensions of the tensor so that its shape becomes (B, 128, 1, H, W). The tensor then passes through the third 3D convolutional layer and is processed by the ReLU activation function, and its shape becomes (B, C, 1, H, W).

[0117] The tensor then enters the residual block 3 (intermediate layer). Here, the tensor is first processed by a 3D convolution layer and a batch normalization layer, and then by a second 3D convolution layer and a feature recalibration module F. EFR Processing, feature recalibration module F EFR The process is the same as above, with the addition of residual connection, and finally processed by ReLU activation function. At this time, the shape of the tensor remains (B, C, 1, H, W).

[0118] Next, the tensor enters the last residual block, and the process is similar to the previous residual block. The input tensor is first processed by a 3D convolution layer and a batch normalization layer, and then passes through a second 3D convolution layer and a feature recalibration module F. EFR Processing, and finally adding residual connection, the output result is processed by ReLU activation function.

[0119] The initial set of focused spectral feature learning modules consists of residual module 1 and module 2, which are responsible for refining spectral features with space as the focus. In this part of the process, the model will first perform a convolution operation on the spatial information to extract the spectral features of each pixel. This step can be regarded as analyzing the reflection intensity of each pixel at different wavelengths to find the similarities and differences in spectral features.

[0120] After preliminary spatial feature extraction, we resize the tensor from (B, 128, H, W, 1) to (B, 128, 1, H, W). This resize rearranges the depth dimension (D) and the channel dimension (C). This operation enables the model to refocus on the spectral features and concentrate the spectral information of each pixel.

[0121] The subsequent residual modules 2 and 3 focus on extracting spatial features with the spectrum as the focus. In this process, the model uses the spectral features extracted in the previous step through further convolution operations to conduct a deeper analysis of the spatial information in the image. This step can be seen as using spectral features to assist in identifying the spatial structure and morphology of matter.

[0122] In this way, the model can process spectral information and spatial information more specifically. Adjusting the arrangement of tensor dimensions allows the model to focus on different types of features at different stages, thereby accelerating model convergence and improving performance. The spectral feature learning module focuses on extracting spectral information, while the spatial feature learning module focuses on extracting ore structure information, thereby improving the model's representation ability and interpretability.

[0123] Finally, the tensor is processed by the global average pooling layer to compress the spatial information of each feature map channel into a value, and the shape becomes (B, C, 1, 1, 1). Then, the tensor is classified by the fully connected layer, and the shape of the output tensor is (B, classes), that is, each sample corresponds to the predicted score of a category.

[0124] This structure and processing procedure enable the model to effectively extract and process the spatial and spectral features of hyperspectral images, ensuring maximum performance with limited computing resources.

[0125] 6. Separate the ore to the corresponding secondary conveyor belt according to the location and grade of the ore, and proceed as follows:

[0126] The computer control module summarizes the ore location information output by the target detection model and the ore category information output by the hyperspectral image classification model, and generates a signal to input into the spray selection system.

[0127] The spraying system completes the sorting of each ore based on the location and category information. During the sorting process, the three different grades of ores, high, medium and low, are separated into the corresponding secondary conveyor belts in sequence.

[0128] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for ore classification, characterized in that: include: Acquire hyperspectral image data of the ore to be inspected, extract a feature image highlighting the difference between the ore and the background from the hyperspectral image data, and obtain an RGB image of the ore to be inspected; The RGB image of the ore to be inspected is detected by using the target detection algorithm to obtain positioning parameters for characterizing the size parameters of the ore to be inspected; a pre-selection box for determining the position of the ore to be inspected is constructed based on the positioning parameters, and the edge contour and mask of each ore to be inspected are determined by the pre-selection box; Processing the hyperspectral image data of the ore to be inspected according to the pre-selected frame and the edge contour and mask of each ore to be inspected to obtain new hyperspectral image data; Two 3D convolution kernels with different sizes and steps are used to extract the spectral information and texture features of the new hyperspectral image data respectively; Adaptively fuse the extracted features to obtain fused features including the spectral features and texture features of the ore to be tested; The adaptive fusion of the extracted features specifically includes: Different texture features are fused with spectral features to obtain channel features; the features of each channel are compressed with global information to generate channel descriptors; Generate an attention weight vector for each channel descriptor, perform element-by-element multiplication operation on the weight vector and the channel descriptor of each channel, and obtain a fusion feature including the spectral features and texture features of the ore to be inspected; The fused features are input into the hyperspectral image classification model for identification, and the classification results of each ore to be inspected are obtained; The hyperspectral image classification model includes a spectral feature learning group, a spatial feature learning group and a feature recalibration module F EFR The spectral feature learning group and the spatial feature learning group are both composed of 3D convolution residual modules; the spectral feature learning group is used to extract the spectral features of spatial focus in the fusion feature; the spatial feature learning group is used to extract the spatial features of spectral focus in the fusion feature; the feature recalibration module F EFR , which is used to dynamically adjust the weights of spectral features and spatial features.

2. A method for ore classification according to claim 1, characterized in that: The step of obtaining a hyperspectral image of the ore to be inspected specifically includes: The ore raw materials collected from the ore-bearing layer of the mine are crushed and the ore with a particle size of 8cm×8cm is screened out; The hyperspectral image of the screened ore to be inspected is captured by a hyperspectral camera.

3. A method for ore classification as claimed in claim 1, characterized in that: The positioning parameters obtained for characterizing the size parameters of the ore to be inspected specifically include: In the RGB image of the ore to be inspected, the characteristic images including carbonate minerals, clay minerals and diaspore are selected to construct a false color composite image; The target detection algorithm is used to locate the false color synthetic image, and four positioning parameters x, y, w, and h are obtained, where x and y are the coordinates of the center point of the ore to be inspected, which are used as the center point coordinates of the pre-selection box; w is the maximum radial length of the ore to be inspected, which is used as the width of the pre-selection box; h is the maximum axial height of the ore to be inspected, which is used as the length of the pre-selection box.

4. A method for ore classification as claimed in claim 3, characterized in that: The step of determining the edge contour and mask of each ore to be inspected by means of a pre-selected frame specifically includes: Convert the RGB image in each pre-selected box to the YCbCr color space and calculate the histogram of the image; The histogram is processed by one-dimensional Gaussian filtering to obtain a smooth histogram, in which the ore and the background have different characteristic peaks respectively; Determine the segmentation threshold according to the position between the ore characteristic peak and the background characteristic peak in the histogram, and use the segmentation threshold to eliminate the background in the RGB image to obtain the RGB image after the background is eliminated; Perform Gaussian blur on the image and perform grayscale conversion to obtain a grayscale image; Use contour extraction algorithm to draw edge contours of ore in each pre-selected box in the grayscale image; Then, an edge detection algorithm is used to detect the edge of the ore in the grayscale image, find the boundary of the ore, and generate a mask of the ore in each pre-selected box according to the boundary contour.

5. A system for classifying ore, characterized in that: include: An image acquisition module is used to obtain hyperspectral image data of the ore to be inspected, extract a feature image highlighting the difference between the ore and the background from the hyperspectral image data, and obtain an RGB image of the ore to be inspected; A computer control module, comprising: a target detection module, a contour elimination module, a feature extraction module, and a classification module; The target detection module is used to detect the RGB image of the ore to be inspected using the target detection algorithm to obtain positioning parameters for characterizing the size parameters of the ore to be inspected; based on the positioning parameters, a pre-selection box is constructed to determine the position of the ore to be inspected, and the edge contour and mask of each ore to be inspected are determined by the pre-selection box; The contour rejection module is used to process the hyperspectral image data of the ore to be inspected according to the pre-selected box and the edge contour and mask of each ore to be inspected to obtain new hyperspectral image data; The feature extraction module is used to respectively extract the spectral information and texture features of the new hyperspectral image data using two 3D convolution kernels of different sizes and step sizes; adaptively fuse the extracted features to obtain fused features including the spectral features and texture features of the ore to be inspected; Among them, the extracted features are adaptively fused, specifically including: The different texture features are fused with the spectral features to obtain the channel features; the features of each channel are compressed with global information to generate a channel descriptor; an attention weight vector of each channel descriptor is generated, and the weight vector is element-wise multiplied with the channel descriptor of each channel to obtain a fusion feature containing the spectral features and texture features of the ore to be inspected; The classification module is used to input the fusion features into the hyperspectral image classification model for identification, and obtain the classification result of each ore to be inspected; Among them, the hyperspectral image classification model includes spectral feature learning group, spatial feature learning group and feature recalibration module F EFR , the spectral feature learning group and the spatial feature learning group are both composed of 3D convolution residual modules; the spectral feature learning group is used to extract the spectral features of spatial focus in the fusion feature; the spatial feature learning group is used to extract the spatial features of spectral focus in the fusion feature; the feature recalibration module F EFR , which is used to dynamically adjust the weights of spectral features and spatial features.

6. A mineral sorting system as claimed in claim 5, characterized in that: It also includes a particle size screening module, which is used to crush the ore raw materials collected from the ore-bearing layer of the mine, and put the crushed ore into the particle size screening module to screen out the ore to be tested with a particle size of 8cm×8cm.

7. A mineral classification system as claimed in claim 5, characterized in that: It also includes a material conveying module, which is used to disperse the ore to be inspected through the speed difference between conveyor belts at different levels.

8. An ore classification system as claimed in claim 5, characterized in that: It also includes an injection module, which is used to send the ore to be inspected into the corresponding secondary conveyor belt according to the classification result of each ore to be inspected.

Images