An intelligent sorting method for low-grade copper ore based on image multi-modal fusion

Abstract

The present invention discloses an intelligent sorting method for low-grade copper ore based on image multi-modal fusion, comprising the following steps: constructing a data set, which includes a number of dual-energy X-ray copper ore images, and the dual-energy X-ray copper ore images are high-energy X-ray copper ore images and low-energy X-ray copper ore images; constructing a two-dimensional fusion framework, which includes a feature-level extraction layer and a decision-level prediction layer; inputting the dual-energy X-ray copper ore images into the feature-level extraction layer for processing, and inputting the output of the feature-level extraction layer into the decision-level prediction layer for prediction and classification to complete the sorting of copper ore; starting from image processing and decision-level fusion, the present invention constructs a two-dimensional fusion framework of a feature-level extraction layer - decision-level prediction layer, and through feature enhancement technologies such as dual-energy subtraction, Laplace operator and wavelet transform fusion, combined with a weight voting integration mechanism, effectively improves the accuracy and robustness of ore sorting.

Description

Technical Field

[0001] The present invention relates to the technical field of copper ore sorting, and in particular to a method for intelligent sorting of low-grade copper ore based on multimodal image fusion. Background Art

[0002] As a transition metal element with excellent electrical conductivity, thermal conductivity, ductility and anti-magnetism, copper has become an indispensable basic material in the modern industrial system. It is widely used in power and telecommunications, construction engineering, transportation and electronic equipment. At the same time, its antibacterial properties also make it a key raw material for medical device manufacturing.

[0003] Traditional copper ore sorting technology mainly relies on physical and chemical methods to achieve mineral separation and enrichment, but there are many constraints in the actual application process: (1) The energy consumption of pre-treatment links such as crushing and grinding in the traditional process is high; (2) The use of chemical agents in the flotation process and improper disposal of tailings may lead to soil and water pollution; (3) When faced with ore embedded with finer particle size or minerals with similar properties, traditional technology often finds it difficult to achieve efficient separation; (4) The purchase and maintenance of large-scale equipment, the procurement of chemical agents, and the disposal of waste all constitute a significant operational burden; (5) The professional requirements of process operations increase labor costs and management difficulties; (6) The use of some chemical agents also poses potential safety hazards and increases the risks in the production process.

[0004] As an innovative solution, ore sorting technology based on advanced sensors plays an important role in improving the efficiency of mineral resource utilization, reducing environmental pollution and saving energy. This technology realizes an efficient and environmentally friendly sorting process by accurately detecting the differential characteristics inside the ore. At present, this technology system mainly uses methods (sensors) such as X-ray fluorescence spectroscopy (XRF), X-ray transmission (XRT), near-infrared rays (NIR), laser induced breakdown spectroscopy (LIBS) and dual-energy X-rays (D-XRT). The application of these methods not only improves the sorting efficiency, but also realizes a safer sorting process. More importantly, the high adaptability of the above methods enables them to adapt to the needs of different types of ore sorting, providing technical guarantees for the sustainable development of mineral resources.

[0005] At present, although ore sorting technology based on advanced sensors has made certain progress, existing research relies too much on single modality data and lacks the integration of multi-source heterogeneous information such as dual-energy X-ray images, mineral text descriptions, and spectral characteristics, which restricts the actual application effect of complex ore sorting. Summary of the invention

[0006] Aiming at the deficiencies of the existing technology, the present invention provides an intelligent sorting method for low-grade copper ore based on image multi-modal fusion, aiming to solve the problems in the background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions: An intelligent sorting method for low-grade copper ore based on image multi-modal fusion, comprising the following steps:

[0008] Step S1: Construct a data set, which includes a number of dual-energy X-ray copper ore images, and the dual-energy X-ray copper ore images are high-energy X-ray copper ore images and low-energy X-ray copper ore images;

[0009] Step S2: Construct a two-dimensional fusion framework, and the two-dimensional fusion framework includes a feature-level extraction layer and a decision-level prediction layer;

[0010] Step S3: Input the dual-energy X-ray copper ore images into the feature-level extraction layer for processing, and input the output of the feature-level extraction layer into the decision-level prediction layer for prediction and classification to complete the sorting of copper ore;

[0011] In the feature-level extraction layer, feature extraction is performed on the dual-energy X-ray copper ore images from two aspects: linear and non-linear.

[0012] In the linear aspect, the dual-energy silhouette technology is used to enhance the contrast of the mineral density difference in the dual-energy X-ray copper ore images to obtain the first enhanced feature, and the Laplace operator is used to extract the edge features of the dual-energy X-ray copper ore images to obtain the second enhanced feature;

[0013] In the non-linear processing aspect, based on the VGG16 network architecture, the multi-scale wavelet transform and the cross-modal feature interaction mechanism are fused, and the encoding-decoding structure, the skip connection mechanism and the residual learning mechanism of the U-Net network are introduced to construct the UVGG16R model, and the dual-energy X-ray copper ore images are processed by the UVGG16R model to obtain the third enhanced feature;

[0014] Among them, the output of the feature-level extraction layer is to splice and fuse the first enhanced feature, the second enhanced feature and the third enhanced feature;

[0015] In the decision-level prediction layer, the weight voting network and the multi-branch network are respectively used to predict and sort the output of the feature-level extraction layer, and the prediction and sorting results of the weight voting network and the multi-branch network are compared, and the final sorting strategy is determined through experiments.

[0016] Furthermore, the UVGG16R model includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first residual block, a second residual block, a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a fourth two-dimensional convolutional layer, a fifth two-dimensional convolutional layer, a sixth two-dimensional convolutional layer, a seventh two-dimensional convolutional layer, an eighth two-dimensional convolutional layer, and a global average pooling layer;

[0017] The processing flow of the UVGG16R model is as follows: the input is sequentially passed through the first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, and the fifth convolutional block. The output of the fifth convolutional block is sequentially passed through the first residual block and the second residual block and then upsampled and concatenated with the output of the fourth convolutional block to obtain a first concatenated feature. The first concatenated feature is sequentially passed through the first two-dimensional convolutional layer and the second two-dimensional convolutional layer and then upsampled and concatenated with the output of the third convolutional block to obtain a second concatenated feature. The second concatenated feature is sequentially passed through the third two-dimensional convolutional layer and the fourth two-dimensional convolutional layer and then upsampled and concatenated with the output of the second convolutional block to obtain a third concatenated feature. The third concatenated feature is sequentially passed through the fifth two-dimensional convolutional layer and the sixth two-dimensional convolutional layer and then upsampled and concatenated with the output of the first convolutional block to obtain a fourth concatenated feature. The fourth concatenated feature is sequentially passed through the seventh two-dimensional convolutional layer and the eighth two-dimensional convolutional layer and then input into the global average pooling layer for processing to obtain the output of the UVGG16R model.

[0018] Furthermore, based on the VGG16 network architecture, the multi-scale wavelet transform and the cross-modal feature interaction mechanism are fused, and the encoding-decoding structure, the skip connection mechanism, and the residual learning mechanism of the U-Net network are introduced. The specific process of obtaining the third enhanced feature by processing the dual-energy X-ray copper ore image through the UVGG16R model is as follows:

[0019] Perform wavelet transform processing on the dual-energy X-ray copper ore image;

[0020] Obtain the text features of the dual-energy X-ray copper ore image and encode the text features;

[0021] Input the dual-energy X-ray copper ore image after wavelet transform processing into the UVGG16R model for processing, and concatenate the output of the UVGG16R model with the encoded text features to obtain the third enhanced feature.

[0022] Furthermore, the first convolutional block and the second convolutional block have the same structure, both consisting of two sequentially connected convolutional layers and one pooling layer; the third convolutional block, the fourth convolutional block, and the fifth convolutional block have the same structure, all consisting of three sequentially connected convolutional layers and one pooling layer;

[0023] The structures of the first residual block and the second residual block are the same, including a batch normalization layer, a ReLU activation function, and a ninth two-dimensional convolutional layer connected in sequence; the processing flow of the first residual block is: the input is sequentially passed through the batch normalization layer, the ReLU activation function, and the output of the ninth two-dimensional convolutional layer is concatenated with the input of the first residual block to obtain the output of the first residual block.

[0024] Further, a weight voting network is used to predict and sort the output of the feature-level extraction layer, expressed as:

[0025] ;

[0026] In the formula, represents the final predicted output of the weight voting network belongs to the category probability distribution, that is, the output of the weight voting network; represents the number of learners in the weight voting network; represents the weight coefficient of the represents the predicted output of the belongs to the category probability distribution.

[0027] Further, the specific process of using a multi-branch network to predict and sort the output in the feature hierarchy is:

[0028] The output in the feature hierarchy is respectively input into each sub-network in the multi-branch network for processing, expressed as:

[0029] ;

[0030] In the formula, represents the output of the sub-network; represents the forward propagation function of the sub-network; represents the parameters of the represents the input, that is, the output of the feature-level extraction layer; represents the total number of sub-networks;

[0031] The multi-branch network integrates the outputs of all sub-networks into a unified total feature representation through a feature splicing layer, expressed as:

[0032] ;

[0033] In the formula, represents the total feature representation output by the multi-branch network;

[0034] The total feature representation output by the multi-branch network is further processed through a fully connected layer to generate the final prediction output of the multi-branch network. Let the weight matrix of the fully connected layer be , and the bias be , where is the number of categories in the final prediction output of the multi-branch network; the final prediction output of the multi-branch network is expressed as:

[0035] .

[0036] Furthermore, the dual-energy subtraction technique is used to enhance the contrast of the mineral density difference in the dual-energy X-ray copper ore image, and the first enhanced feature is obtained, which is expressed as:

[0037] ;

[0038] In the formula, represents the dual-energy X-ray copper ore image after enhancing the contrast of the mineral density difference, that is, the first enhanced feature; and respectively represent the input high-energy X-ray copper ore image and low-energy X-ray copper ore image; represents the pixel coordinates.

[0039] Compared with the existing technologies, the present invention has the following beneficial effects:

[0040] (1) Starting from image processing and decision-level fusion, the present invention constructs a two-dimensional fusion framework of a feature-level extraction layer - decision-level prediction layer. Through feature enhancement technologies such as dual-energy subtraction, Laplacian operator, and wavelet transform fusion, combined with a weight voting integration mechanism, the accuracy and robustness of ore sorting are effectively improved.

[0041] (2) Based on the VGG16 network architecture, the present invention can take into account both spatial and frequency characteristics by fusing multi-scale wavelet transform and cross-modal feature interaction mechanism, realize the collaborative optimization of multi-source information such as text and image, thereby improving the accuracy and robustness of ore sorting, and enhancing the model's ability to capture complex ore features.

[0042] (3) Aiming at the limitations of the traditional VGG16 network in ore sorting tasks, the present invention fuses the multi-scale wavelet transform and cross-modal feature interaction mechanism, and introduces the encoding-decoding structure and skip connection mechanism and residual learning mechanism of the U-Net network to construct the UVGG16R model. While ensuring the simplicity of the network structure, it realizes the optimal balance of classification accuracy and computational efficiency, providing reliable technical support for the efficient utilization of low-grade copper ore resources. Description of the Drawings

[0043] Figure 1 This is the flowchart of the method of the present invention.

[0044] Figure 2 This is the schematic diagram of the UVGG16R model structure of the present invention. Specific implementation manners

[0045] As Figure 1 shown, the present invention provides a technical solution: an intelligent sorting method for low-grade copper ore based on image multi-modal fusion, including the following steps:

[0046] Step S1: Construct a data set, which includes a number of dual-energy X-ray copper ore images, and the dual-energy X-ray copper ore images are high-energy X-ray copper ore images and low-energy X-ray copper ore images.

[0047] Among them, in the data set of dual-energy X-ray copper ore images, the dual-energy X-ray images of concentrated ore and the dual-energy X-ray images of waste ore each account for half.

[0048] Step S2: Construct a two-dimensional fusion framework, and the two-dimensional fusion framework includes a feature-level extraction layer and a decision-level prediction layer.

[0049] Step S3: Input the dual-energy X-ray copper ore images into the feature-level extraction layer for processing, and input the output of the feature-level extraction layer into the decision-level prediction layer for prediction and classification to complete the sorting of copper ore.

[0050] Among them, in the feature-level extraction layer, the dual-energy X-ray copper ore images are feature-extracted from both linear and non-linear aspects to solve the problems such as the unclear features and blurred boundaries of low-grade copper ore images.

[0051] At the linear level, the dual-energy silhouette technique (Dual Energy Subtraction, DES) is used to enhance the contrast of the mineral density difference in the dual-energy X-ray copper ore images to obtain the first enhanced feature, and the Laplace operator is used to extract the edge features of the dual-energy X-ray copper ore images to obtain the second enhanced feature.

[0052] Among them, the dual-energy silhouette technique (Dual Energy Subtraction, DES) is used to enhance the contrast of the mineral density difference in the dual-energy X-ray copper ore images to obtain the first enhanced feature, which can be expressed as:

[0053] (1);

[0054] In the formula, represents the dual-energy X-ray copper ore image after the contrast enhancement of the mineral density difference, that is, the first enhanced feature; and respectively represent the input high-energy X-ray copper ore image and low-energy X-ray copper ore image; represent pixel coordinates.

[0055] In the non-linear processing layer, based on the VGG16 network architecture, the multi-scale wavelet transform and cross-modal feature interaction mechanism are fused, and the encoding-decoding structure, skip connection mechanism and residual learning mechanism of the U-Net network are introduced to construct the UVGG16R model. The dual-energy X-ray copper ore image is decomposed and reconstructed by the UVGG16R model to obtain the third enhanced feature.

[0056] Wavelet transform can realize the decomposition and reconstruction of multi-scale features of dual-energy X-ray copper ore images, thereby optimizing the feature expression ability of dual-energy X-ray copper ore images in both the frequency domain and the spatial domain.

[0057] Since single-modal data processing still has problems such as incomplete information, limited representation ability, and poor anti-interference ability in the feature extraction of data information, the performance of ore sorting is restricted, and it is difficult to meet the sorting requirements of complex ore features in terms of accuracy and efficiency. To solve this problem, by fusing the multi-scale wavelet transform and cross-modal feature interaction mechanism, it is possible to simultaneously take into account the spatial and frequency characteristics, realize the collaborative optimization of multi-source information, and thus improve the accuracy and robustness of ore sorting.

[0058] Specifically, perform wavelet transform processing on the dual-energy X-ray copper ore image;

[0059] Obtain the text features of the dual-energy X-ray copper ore image, perform One-hot encoding on the text features, and convert the text features into a numerical representation suitable for neural network processing.

[0060] Input the dual-energy X-ray copper ore image processed by wavelet transform into the UVGG16R model for processing, and splice the output of the UVGG16R model with the encoded text features to obtain the third enhanced feature.

[0061] As Figure 2 shown, the UVGG16R model includes the first convolutional block (Block 1), the second convolutional block (Block 2), the third convolutional block (Block 3), the fourth convolutional block (Block 4), the fifth convolutional block (Block 5), the first residual block, the second residual block, the first two-dimensional convolutional layer, the second two-dimensional convolutional layer, the third two-dimensional convolutional layer, the fourth two-dimensional convolutional layer, the fifth two-dimensional convolutional layer, the sixth two-dimensional convolutional layer, the seventh two-dimensional convolutional layer, the eighth two-dimensional convolutional layer and the global average pooling layer (GlobalAveragePooling).

[0062] Among them, the processing flow of the UVGG16R model is as follows: The input is successively passed through the first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block, and the fifth convolutional block. The output of the fifth convolutional block is successively passed through the first residual block and the second residual block and then upsampled and concatenated with the output of the fourth convolutional block to obtain the first concatenated feature. The first concatenated feature is successively passed through the first two-dimensional convolutional layer and the second two-dimensional convolutional layer and then upsampled and concatenated with the output of the third convolutional block to obtain the second concatenated feature. The second concatenated feature is successively passed through the third two-dimensional convolutional layer and the fourth two-dimensional convolutional layer and then upsampled and concatenated with the output of the second convolutional block to obtain the third concatenated feature. The third concatenated feature is successively passed through the fifth two-dimensional convolutional layer and the sixth two-dimensional convolutional layer and then upsampled and concatenated with the output of the first convolutional block to obtain the fourth concatenated feature. The fourth concatenated feature is successively passed through the seventh two-dimensional convolutional layer and the eighth two-dimensional convolutional layer and then input into the global average pooling layer for processing to obtain the output of the UVGG16R model.

[0063] Among them, the first convolutional block and the second convolutional block have the same structure, both consisting of two successively connected convolutional layers and a pooling layer; the third convolutional block, the fourth convolutional block, and the fifth convolutional block have the same structure, all consisting of three successively connected convolutional layers and a pooling layer.

[0064] Among them, a transition layer (Transition Layer) is connected after the first residual block and the second residual block to achieve hierarchical optimization of feature extraction.

[0065] Among them, the first residual block and the second residual block have the same structure, including a batch normalization layer, a ReLU activation function, and a ninth two-dimensional convolutional layer connected in sequence; the processing flow of the first residual block is: The input is successively passed through the batch normalization layer, the ReLU activation function, and the output of the ninth two-dimensional convolutional layer and concatenated with the input of the first residual block to obtain the output of the first residual block.

[0066] Among them, the first convolutional block (Block 1), the second convolutional block (Block 2), the third convolutional block (Block 3), the fourth convolutional block (Block 4), and the fifth convolutional block (Block 5) in the UVGG16R model are the convolutional blocks in the VGG16 model. The first convolutional block (Block 1) and the second convolutional block (Block 2) have the same structure, both consisting of two successively connected convolutional layers and a pooling layer. The third convolutional block (Block 3), the fourth convolutional block (Block 4), and the fifth convolutional block (Block 5) have the same structure, all consisting of three successively connected convolutional layers and a pooling layer.

[0067] Among them, discrete wavelet transform is used for wavelet transform processing of dual-energy X-ray copper ore images.

[0068] In the decision-level prediction layer, the Weighted Voting network and the multi-branch network are respectively used to predict and sort the outputs of the feature-level extraction layer (the first enhanced feature, the second enhanced feature, and the third enhanced feature), and the prediction and sorting results of the Weighted Voting network and the multi-branch network are compared to determine the final sorting strategy through experiments.

[0069] Among them, the first enhanced feature, the second enhanced feature, and the third enhanced feature are concatenated and fused to form a unified multi-modal feature representation as the input of the decision-level prediction layer.

[0070] Among them, using the Weighted Voting network to predict and sort the output of the feature-level extraction layer is expressed as:

[0071] (2);

[0072] In the formula, represents the final predicted output of the Weighted Voting network belongs to the category probability distribution, that is, the output of the Weighted Voting network; represents the number of learners in the Weighted Voting network; represents the weight coefficient of the represents the predicted output of the belongs to the category probability distribution.

[0073] The specific process of using the multi-branch network to predict and sort the output in the feature level is as follows: The output in the feature level is respectively input into each sub-network in the multi-branch network for processing, which is expressed as:

[0074] (3);

[0075] In the formula, represents the output of the sub-network; represents the forward propagation function of the sub-network; represents the parameters of the represents the input (the output of the feature-level extraction layer); represents the total number of sub-networks.

[0076] The multi-branch network integrates the outputs of all sub-networks into a unified total feature representation through the feature concatenation layer:

[0077] (4);

[0078] In the formula, represents the total feature representation output by the multi-branch network.

[0079] The total feature representation output by the multi-branch network is further processed through a fully connected layer to generate the final prediction output of the multi-branch network. Assume that the weight matrix of the fully connected layer is , and the bias is , where is the number of categories in the final prediction output of the multi-branch network; the final prediction output of the multi-branch network can be expressed as:

[0080] (5).

[0081] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent sorting method for low-grade copper ore based on image multi-modal fusion, characterized in that It includes the following steps: Step S1: Construct a data set, which includes a number of dual-energy X-ray copper ore images. The dual-energy X-ray copper ore images are high-energy X-ray copper ore images and low-energy X-ray copper ore images; Step S2: Construct a two-dimensional fusion framework, which includes a feature-level extraction layer and a decision-level prediction layer; Step S3: Input the dual-energy X-ray copper ore images into the feature-level extraction layer for processing, and input the output of the feature-level extraction layer into the decision-level prediction layer for prediction and classification to complete the sorting of copper ore; In the feature-level extraction layer, the dual-energy X-ray copper ore images are feature-extracted from both linear and nonlinear aspects; At the linear level, the dual-energy silhouette technique is used to enhance the contrast of the mineral density difference in the dual-energy X-ray copper ore images to obtain the first enhanced feature, and the Laplace operator is used to extract the edge feature of the dual-energy X-ray copper ore images to obtain the second enhanced feature; At the nonlinear processing level, based on the VGG16 network architecture, the multi-scale wavelet transform and the cross-modal feature interaction mechanism are fused, and the encoding-decoding structure, the skip connection mechanism and the residual learning mechanism of the U-Net network are introduced to construct the UVGG16R model. The dual-energy X-ray copper ore images are processed by the UVGG16R model to obtain the third enhanced feature; Among them, the output of the feature-level extraction layer is the splicing and fusion of the first enhanced feature, the second enhanced feature and the third enhanced feature; In the decision-level prediction layer, the weighted voting network and the multi-branch network are respectively used to predict and sort the output of the feature-level extraction layer, and the prediction and sorting results of the weighted voting network and the multi-branch network are compared, and the final sorting strategy is determined through experiments.

2. The intelligent sorting method for low-grade copper ore based on image multi-modal fusion according to claim 1, characterized in that: The UVGG16R model includes a first convolutional block, a second convolutional block, a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first residual block, a second residual block, a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a fourth two-dimensional convolutional layer, a fifth two-dimensional convolutional layer, a sixth two-dimensional convolutional layer, a seventh two-dimensional convolutional layer, an eighth two-dimensional convolutional layer and a global average pooling layer; The processing process of the UVGG16R model is as follows: The input is sequentially passed through the first convolutional block, the second convolutional block, the third convolutional block, the fourth convolutional block and the fifth convolutional block. The output of the fifth convolutional block is sequentially passed through the first residual block and the second residual block and then upsampled and spliced with the output of the fourth convolutional block to obtain the first spliced feature. The first spliced feature is sequentially passed through the first two-dimensional convolutional layer and the second two-dimensional convolutional layer and then upsampled and spliced with the output of the third convolutional block to obtain the second spliced feature. The second spliced feature is sequentially passed through the third two-dimensional convolutional layer and the fourth two-dimensional convolutional layer and then upsampled and spliced with the output of the second convolutional block to obtain the third spliced feature. The third spliced feature is sequentially passed through the fifth two-dimensional convolutional layer and the sixth two-dimensional convolutional layer and then upsampled and spliced with the output of the first convolutional block to obtain the fourth spliced feature. The fourth spliced feature is sequentially passed through the seventh two-dimensional convolutional layer and the eighth two-dimensional convolutional layer and then input into the global average pooling layer for processing to obtain the output of the UVGG16R model.

3. An intelligent sorting method for low-grade copper ore based on image multi-modal fusion according to claim 2, characterized in that: Based on the VGG16 network architecture, integrating multi-scale wavelet transform and cross-modal feature interaction mechanism, and introducing the encoding-decoding structure, skip connection mechanism and residual learning mechanism of the U-Net network, the specific process of processing the dual-energy X-ray copper ore image by the UVGG16R model to obtain the third enhanced feature is as follows: Perform wavelet transform processing on the dual-energy X-ray copper ore image; Obtain the text features of the dual-energy X-ray copper ore image and encode the text features; Input the dual-energy X-ray copper ore image after wavelet transform processing into the UVGG16R model for processing, and splice the output of the UVGG16R model with the encoded text features to obtain the third enhanced feature.

4. An intelligent sorting method for low-grade copper ore based on image multi-modal fusion according to claim 3, characterized in that: The structures of the first convolutional block and the second convolutional block are the same, both consisting of two convolutional layers and one pooling layer connected in sequence; the structures of the third convolutional block, the fourth convolutional block and the fifth convolutional block are the same, all consisting of three convolutional layers and one pooling layer connected in sequence; The structures of the first residual block and the second residual block are the same, including a batch normalization layer, a ReLU activation function and a ninth two-dimensional convolutional layer connected in sequence; the processing flow of the first residual block is: the input passes through the batch normalization layer, the ReLU activation function and the output of the ninth two-dimensional convolutional layer in sequence, and is spliced with the input of the first residual block to obtain the output of the first residual block.

5. An intelligent sorting method for low-grade copper ore based on image multi-modal fusion according to claim 4, characterized in that: Use the weight voting network to predict and sort the output of the feature-level extraction layer, expressed as: ； In the formula, represents the final predicted output of the weighted voting network belonging to the category , that is, the output of the weighted voting network; represents the number of learners in the weighted voting network; represents the th weight coefficient of the learner; represents the th predicted output of the learner belonging to the category probability distribution.

6. The intelligent sorting method for low-grade copper ore based on image multi-modal fusion according to claim 5, wherein: The specific process of using the multi-branch network to predict and sort the output in the feature hierarchy is: Input the output in the feature hierarchy into each sub-network in the multi-branch network for processing respectively, expressed as: ； In the formula, represents the output of the th subnetwork; represents the forward propagation function of the th subnetwork; represents the parameters of the th subnetwork; represents the input, i.e., the output of the feature-level extraction layer; represents the total number of subnets; The multi-branch network integrates the outputs of all sub-networks into a unified total feature representation through the feature splicing layer, expressed as: ； In the formula, represents the total feature representation output by the multi-branch network; The total feature representation output by the multi-branch network is further processed through a fully connected layer to generate the final prediction output of the multi-branch network. Let the weight matrix of the fully connected layer be , and the bias be , where is the number of categories in the final prediction output of the multi-branch network; the final prediction output of the multi-branch network is expressed as: 。 7. The intelligent sorting method for low-grade copper ore based on image multi-modal fusion according to claim 6, characterized in that: Use the dual-energy silhouette technology to enhance the contrast of the mineral density difference in the dual-energy X-ray copper ore image to obtain the first enhanced feature, expressed as: ； In the formula, The dual-energy X-ray copper ore image after contrast enhancement representing the mineral density difference, i.e., the first enhancement feature; and respectively represent the input high-energy X-ray copper ore image and low-energy X-ray copper ore image; represents the pixel coordinates.

Images