A low-grade copper ore image sorting method based on deep learning

Abstract

The present invention discloses a low-grade copper ore image sorting method based on deep learning, comprising the following steps: obtaining a dual-energy X-ray image of the low-grade copper ore; constructing a time-step embedded U-Net network, inputting the dual-energy X-ray image of the low-grade copper ore into the time-step embedded U-Net network, and generating diversified dual-energy X-ray images of the copper ore; constructing an adaptive copper ore sorting model; inputting the diversified dual-energy X-ray image of the copper ore into the adaptive copper ore sorting model to sort the copper ore; the present invention optimizes the sorting process by integrating dual-energy X-ray imaging technology with data enhancement technology. Compared with traditional single-energy imaging, the present invention enhances the difference between copper minerals and waste rock in the image feature space, reduces the difficulty of the model in recognizing complex textures, reduces the amount of invalid ore entering the crushing and flotation links, and significantly reduces energy consumption and chemical agent usage.

Description

Technical Field

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

[0002] As a key mineral resource, copper plays an irreplaceable, core role in the modern industrial system. Its excellent electrical and thermal conductivity makes it a core raw material in key areas such as power transmission, new energy equipment manufacturing, and intelligent electronic devices. In particular, demand for copper is experiencing structural growth in key carbon-intensive sectors such as the three-electric systems of new energy vehicles, photovoltaic power station cable networks, and conductive components for energy storage devices.

[0003] Traditional copper ore separation technology has long relied on physical screening and chemical flotation, whose limitations permeate the entire production chain. On the one hand, ore crushing and grinding consume excessive energy, resulting in wasted resources and skyrocketing costs. On the other hand, the flotation process relies on large quantities of chemical reagents (such as collectors and depressants), which not only increases the pressure on tailings disposal but also makes residual heavy metal ions more likely to leach and contaminate soil and groundwater systems. Technically, traditional methods struggle to effectively separate fine-grained interbedded structures (such as copper minerals and gangue intergrowth) or minerals with similar physicochemical properties (such as the similar density and surface activity of chalcopyrite and pyrite), resulting in large fluctuations in concentrate grade and low recovery rates. Furthermore, the entire process, from equipment procurement to reagent supply, from waste disposal to process parameter control, requires significant capital investment and specialized manual intervention. Furthermore, the flammability of some chemical reagents and the ease with which mineral storage conditions can become uncontrollable. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a low-grade copper ore image sorting method based on deep learning, which aims to solve the problems in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for image sorting of low-grade copper ore based on deep learning, comprising the following steps:

[0006] Step S1: Acquire a dual-energy X-ray image of a low-grade copper ore, wherein the dual-energy X-ray image of the low-grade copper ore includes a dual-energy X-ray image of copper and a dual-energy X-ray image of waste rock; the dual-energy X-ray image of the low-grade copper ore is a high-energy X-ray image and a low-energy X-ray image of the low-grade copper ore;

[0007] Step S2: Embed a learnable time-step embedding module into the U-Net model, introduce cross-layer residual connections and self-attention mechanisms to construct a time-step embedding U-Net network, input the dual-energy X-ray image of the low-grade copper ore into the time-step embedding U-Net network, and generate diversified dual-energy X-ray images of the copper ore;

[0008] Step S3: constructing an adaptive copper ore separation model based on the VGG11 structure;

[0009] Step S4: inputting the diversified copper ore dual-energy X-ray images into the adaptive copper ore sorting model to sort the copper ore.

[0010] Furthermore, the time step embedding U-Net network is composed of a time step embedding module, a convolution module, a residual block, a self-attention module, a transposed convolution and a convolution layer; the processing flow of the time step embedding U-Net network is as follows: first, the input, i.e., the dual-energy X-ray image of the low-grade copper ore, is encoded into an initial noise image and a time step scalar, the time step scalar is input into the time step embedding module for dimensionality increase processing, the initial noise image is input into the convolution module, the output of the convolution module is broadcasted and added to the output of the time step embedding module to obtain a first splicing feature, the first splicing feature is downsampled and broadcasted and added to the output of the time step embedding module to obtain a second splicing feature, and the second splicing feature is downsampled and broadcasted and added to the output of the time step embedding module to obtain a second splicing feature. The concatenated feature is downsampled and passed through the residual block and the self-attention module together with the output of the time step embedding module. The output of the self-attention module, the second concatenated feature and the output of the time step embedding module are concatenated to obtain a third concatenated feature. The third concatenated feature and the first concatenated feature are input into the transposed convolution together. The output of the transposed convolution is concatenated with the output of the time step embedding module to obtain a fourth concatenated feature. The fourth concatenated feature is upsampled and concatenated with the output of the time step embedding module to obtain a fifth concatenated feature. The fifth concatenated feature is mapped to the target dimension through a convolution layer to obtain the final output of the time step embedding U-Net network, i.e., the diversified copper ore dual-energy X-ray image.

[0011] The time step embedding module adopts sinusoidal position encoding.

[0012] Furthermore, the adaptive copper ore sorting model based on the VGG11 structure consists of a backbone network Stem, a first stage Stage1, a second stage Stage2, a main classifier, a feature projection module and a domain discriminator;

[0013] The processing flow of the adaptive copper ore sorting model based on the VGG11 structure is as follows: the input, i.e., the diversified copper ore dual-energy X-ray image, is sequentially passed through the backbone network, the first stage, and the second stage to obtain a second enhanced feature map, and the second enhanced feature map is respectively input into the main classifier, the feature projection module, and the domain discriminator for processing. The main classifier uses the output of the feature projection module to perform classification tasks and predict the probability of the enhanced feature map belonging to each category. The domain discriminator uses the output of the feature projection module to perform domain adaptation tasks and predict the probability of the enhanced feature map belonging to the source domain or the target domain. Finally, the output of the adaptive copper ore sorting model based on the VGG11 structure is the probability distribution of the enhanced feature map belonging to each category and the probability distribution of the enhanced feature map belonging to the source domain or the target domain.

[0014] Furthermore, the first and second stages have the same structure, both consisting of a multi-scale contextual attention module, a dynamic feature compensation unit, and an enhanced feature convolution unit. The processing flow of the first stage is as follows: the input, i.e., the output of the backbone network, passes through the multi-scale contextual attention module and the dynamic feature compensation unit in sequence, and the output of the dynamic feature compensation unit and the output of the multi-scale contextual attention module are input into the enhanced feature convolution unit for processing to obtain the first enhanced feature map, i.e., the final output of the first stage.

[0015] Among them, the enhanced feature convolution unit consists of a convolutional layer, a batch normalization layer and a ReLu activation function connected in sequence.

[0016] Furthermore, the dynamic feature compensation unit is composed of deformable convolution, dense blocks, channel attention modules, global average pooling modules, feature transformation units and Sigmoid functions; the processing flow of the dynamic feature compensation unit is as follows: the input, i.e., the output of the multi-scale context attention module, is sequentially passed through the deformable convolution and dense blocks to obtain adaptive features, the input is sequentially passed through the channel attention module, global average pooling module, feature transformation unit and Sigmoid function to obtain transformed features, the transformed features and adaptive features are spliced and then channel-fused with the input of the dynamic feature compensation unit to obtain fused features, i.e., the final output of the dynamic feature compensation unit.

[0017] Furthermore, the main classifier is composed of an adaptive average pooling layer, a flattening layer, and a fully connected layer connected in sequence;

[0018] The domain discriminator consists of a gradient reversal layer, a global average pooling layer, a first dimensionality reduction feature transformation module, a first dropout layer, a second dimensionality reduction feature transformation module, a second dropout layer and a fully connected layer connected in sequence;

[0019] The first dimensionality reduction feature transformation module and the second dimensionality reduction feature transformation module have the same structure, both consisting of a fully connected layer, a batch normalization layer and a ReLu activation function connected in sequence.

[0020] Furthermore, we add intra-class-inter-class loss function to enhance the constraint, forcing the Euclidean distance of similar enhanced feature maps to be less than the threshold , forcing the Euclidean distance of heterogeneous enhanced feature maps to be greater than the threshold ; The intra-class-inter-class loss function is expressed as:

[0021] ;

[0022] Where, represents the intra-class-inter-class loss function; 、 are the i-th enhanced feature map and the j-th enhanced feature map respectively, 、 Belong to the same type of enhanced feature map; is the kth enhanced feature map, which belongs to the heterogeneous enhanced feature map; Represents the weight parameter.

[0023] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a low-grade copper ore image sorting method based on deep learning.

[0024] A non-volatile computer storage medium stores computer-executable instructions for executing a method for image sorting of low-grade copper ore based on deep learning.

[0025] Compared with the existing technology, the present invention has the following beneficial effects:

[0026] (1) The present invention optimizes the sorting process by integrating dual-energy X-ray imaging technology with data enhancement technology. Compared with traditional single-energy imaging, the dual-energy X-ray sensor can simultaneously capture the density and composition characteristics of the ore, enhance the difference between copper ore and waste rock in the image feature space, and thus reduce the difficulty of the model in recognizing complex textures. To address the problems of scarcity of low-grade ore samples and imbalanced distribution of positive and negative samples, the time-step embedded U-Net model is used to reconstruct the multi-scale features of the original ore image, generate synthetic samples consistent with the real mineral structure, and reduce the risk of overfitting of the model while expanding data diversity. The present invention reduces the reliance on manual experience debugging in traditional processes by constructing an automated closed loop for ore image recognition and sorting decision-making, and dynamically optimizes sorting parameters based on real-time feedback. In addition, the present invention can reduce the amount of invalid ore entering the crushing and flotation stages, significantly reduce energy consumption and chemical reagent usage, and alleviate environmental pollution pressure from the source of the process.

[0027] (2) The present invention addresses the problem of decision bias in image sorting caused by the uneven distribution of copper ore and waste rock in low-grade copper ore samples. A time-step embedded U-Net model based on the conditional diffusion process is constructed. The sample data generated by the time-step embedded U-Net model can effectively correct the decision bias of classification, reconstruct the discriminant feature space dominated by negative samples into a balanced representation space, and improve the model sorting accuracy and recall rate.

[0028] (3) To address the problem that the features in low-grade copper ore images are sparsely distributed and difficult to capture by conventional convolution models, the present invention constructs a multi-scale contextual attention module and a dynamic feature compensation unit, which improves the fixed receptive field limitation of traditional convolution and enhances the model's ability to model the association of global ore features. At the same time, the convolution sampling area is adjusted through deformable convolution to adapt to the tiny features in the ore image, and the residual connection is used to strengthen the early feature expression, effectively improving the model classification credibility.

[0029] (4) The present invention addresses the problem of model recognition errors caused by the presence of other metal impurities and sudden changes in thickness in waste rock images, which form confusing features. The gradient inversion layer and intra-class-inter-class contrast loss are used to remove the confusing features. By dynamically adjusting the adversarial coefficient, the interval between the two types of features is expanded, greatly improving the model accuracy and overall sorting ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Flow chart of the method of the present invention.

[0031] Figure 2 This is the U-Net network structure diagram for time step embedding in the present invention.

[0032] Figure 3 This is a structural diagram of the adaptive copper ore separation model based on the VGG11 structure of the present invention.

[0033] Figure 4 This is a structural diagram of the dynamic feature compensation unit of the present invention.

[0034] Figure 5 This is the structural diagram of the domain discriminator of the present invention. DETAILED DESCRIPTION

[0035] like Figure 1 As shown, the present invention provides a technical solution: a low-grade copper ore image sorting method based on deep learning, comprising the following steps:

[0036] Step S1: Acquire a dual-energy X-ray image of a low-grade copper ore, wherein the dual-energy X-ray image of the low-grade copper ore includes a copper-containing dual-energy X-ray image and a waste rock dual-energy X-ray image; the dual-energy X-ray image of the low-grade copper ore is a high-energy X-ray image and a low-energy X-ray image of the low-grade copper ore.

[0037] Step S2: A learnable time-step embedding module is embedded in the U-Net model, and cross-layer residual connections and self-attention mechanisms are introduced to construct a time-step embedding U-Net network. The dual-energy X-ray images of low-grade copper ore are input into the time-step embedding U-Net network to generate diversified dual-energy X-ray images of copper ore (enhanced images after denoising).

[0038] Step S3: Construct an adaptive copper ore sorting model based on the VGG11 architecture. This model implements a multi-level structure evolution and dynamic parameter adaptation mechanism within a deep feature learning framework.

[0039] Step S4: inputting the diversified copper ore dual-energy X-ray images into the adaptive copper ore sorting model to sort the copper ore.

[0040] Among them, the time step is embedded in the U-Net network as follows Figure 2 As shown in the figure, the model is based on the U-Net symmetric encoding-decoding architecture and is deeply modified to address the time dependency of the diffusion model.

[0041] The time-step embedding U-Net network consists of a time-step embedding module, a convolution module, a residual block, a self-attention module, a transposed convolution and a convolution layer; the processing flow of the time-step embedding U-Net network is as follows: first, the input (dual-energy X-ray image of low-grade copper ore) is encoded into an initial noise image and a time-step scalar t, the time-step scalar t is input into the time-step embedding module for dimensionality increase, the initial noise image is input into the convolution module, the output of the convolution module is broadcasted and added to the output of the time-step embedding module to obtain a first splicing feature, the first splicing feature is downsampled and broadcasted and added to the output of the time-step embedding module to obtain a second splicing feature, the second splicing feature is downsampled and passed through the residual block and the self-attention module together with the output of the time-step embedding module, the output of the self-attention module, the second splicing feature and The output of the time-step embedding module is spliced to obtain a third spliced feature, which is input into the transposed convolution together with the first spliced feature. The output of the transposed convolution is spliced with the output of the time-step embedding module to obtain a fourth spliced feature. The fourth spliced feature is upsampled and spliced with the output of the time-step embedding module to obtain a fifth spliced feature. The fifth spliced feature is mapped to the target dimension through a 1×1 convolutional layer to obtain the final output of the time-step embedded U-Net network, namely, the diversified copper ore dual-energy X-ray image. The whole process enhances the stability of small-batch training through group normalization (GroupNorm) and SiLU activation function, and uses time-step embedding to guide the network to learn the temporal evolution of noise distribution during diffusion, forming a feature generation capability that preserves spatial details and adapts to temporal dynamics.

[0042] Among them, the time step embedding module adopts Sinusoidal Positional Encoding. Sinusoidal Positional Encoding embeds time step information into the model by using sine and cosine functions, so that the model can understand the sequential relationship of elements in the sequence.

[0043] like Figure 3 As shown in the figure, the adaptive copper ore sorting model based on the VGG11 structure consists of a backbone network (Stem), the first stage (Stage1), the second stage (Stage2), a main classifier, a feature projection module and a domain discriminator.

[0044] The processing flow of the adaptive copper ore sorting model based on the VGG11 structure is as follows: the input (diversified copper ore dual-energy X-ray image) passes through the backbone network (Stem), the first stage (Stage1) and the second stage (Stage2) in sequence to obtain the second enhanced feature map, and the second enhanced feature map is input into the main classifier, the feature projection module and the domain discriminator for processing respectively. The main classifier uses the output of the feature projection module to perform the classification task and predicts the probability of the enhanced feature map belonging to each category. The domain discriminator uses the output of the feature projection module to perform the domain adaptation task and predicts the probability of the enhanced feature map belonging to the source domain or the target domain. Finally, the output of the adaptive copper ore sorting model based on the VGG11 structure is the probability distribution of the enhanced feature map belonging to each category and the probability distribution of the enhanced feature map belonging to the source domain or the target domain.

[0045] Among them, the backbone network (Stem) consists of a two-dimensional convolutional layer (Conv2d), a batch normalization layer (BN), a ReLU activation function and a two-dimensional maximum pooling layer (MaxPool2d) connected in sequence.

[0046] Among them, the first stage (Stage1) and the second stage (Stage2) have the same structure, both consisting of a multi-scale contextual attention module (Multi-scale Contextual Attention Module), a dynamic feature compensation unit and an enhanced feature convolution unit; the processing flow of the first stage (Stage1) is: the input (the output of the backbone network) passes through the multi-scale contextual attention module and the dynamic feature compensation unit in sequence, and the output of the dynamic feature compensation unit and the output of the multi-scale contextual attention module are input into the enhanced feature convolution unit for processing to obtain the first enhanced feature map, which is the final output of the first stage (Stage1).

[0047] The second stage (Stage2) has the same structure as the first stage (Stage1), and its processing flow is not described in detail here.

[0048] Among them, the enhanced feature convolution unit consists of a convolutional layer, a batch normalization layer and a ReLu activation function connected in sequence.

[0049] like Figure 4 As shown in the figure, the dynamic feature compensation unit consists of a deformable convolution plus a residual structure, including a deformable convolution module, a dense block, a channel attention module, a global average pooling module, a feature transformation unit and a Sigmoid function; the processing flow of the dynamic feature compensation unit is as follows: the input (the output of the multi-scale context attention module) passes through the deformable convolution and the dense block in sequence to obtain the adaptive feature, the input passes through the channel attention module, the global average pooling module, the feature transformation unit and the Sigmoid function in sequence to obtain the transformed feature, the transformed feature and the adaptive feature are spliced and then channel-fused with the input of the dynamic feature compensation unit to obtain the fused feature, which is the final output of the dynamic feature compensation unit.

[0050] Among them, the dense block consists of two fully connected layers (FC) and a ReLU activation function, and the ReLU activation function is between the two fully connected layers.

[0051] The main classifier consists of an adaptive average pooling layer (AdaptiveAvgPool2d), a flattening layer (Flatten), and a fully connected layer connected in sequence; the loss function of the main classifier is:

[0052] (1);

[0053] Where, represents the loss function of the main classifier, 、 Represent the parameters of the second stage (Stage2) and the main classifier respectively; Represents the data distribution The input of the main classifier sampled from and tags The expected value of express The distribution of express distribution of represents the cross entropy loss, Indicates a given and 、 When the main classifier predicts the label probability.

[0054] like Figure 5As shown in Figure 1, the domain discriminator consists of a gradient reversal layer (GRL), a global average pooling layer, a first dimensionality reduction feature transformation module (DRFTModule), a first dropout layer (Dropout), a second dimensionality reduction feature transformation module, a second dropout layer, and a fully connected layer. The loss function of the domain discriminator is:

[0055] (2);

[0056] Where, represents the loss function of the domain discriminator, represents the parameters of the domain discriminator; Express (Real data distribution) expected value; Represents the generated data distribution middle expected value; Represents the predicted label; Represents a generator that receives and , and generate data.

[0057] Add intra-class-inter-class loss function enhancement constraints to enforce the Euclidean distance of similar enhanced feature maps to be less than the threshold , for heterogeneous enhanced feature maps, force their Euclidean distance to be greater than the threshold ; The intra-class-inter-class loss function is expressed as:

[0058] (3);

[0059] Where, represents the intra-class-inter-class loss function; 、 are the i-th enhanced feature map and the j-th enhanced feature map respectively, 、 Belong to the same type of enhanced feature map; is the kth enhanced feature map, which belongs to the heterogeneous enhanced feature map; Represents the weight parameter.

[0060] This constraint enhances the model's ability to discriminate boundary samples by increasing the hypersphere spacing between inter-class features and improving the compactness of intra-class features.

[0061] Among them, the first dimensionality reduction feature transformation module (DRFT Module) and the second dimensionality reduction feature transformation module have the same structure, both consisting of a fully connected layer, a batch normalization layer and a ReLu activation function connected in sequence.

[0062] During forward propagation, the gradient reversal layer keeps the features unchanged, but during backward propagation, the gradient reversal layer inverts the gradient and multiplies it by the dynamic coefficient , achieving gradient reversal; this allows the second stage (Stage2) to learn to generate feature maps that the domain discriminator cannot correctly classify, thereby achieving domain invariance.

[0063] Among them, the dynamic coefficient It changes dynamically with the iteration cycle and is expressed as:

[0064] (4);

[0065] Where, represents the exponential function; Represents an iteration cycle.

[0066] An electronic device includes a processor, a memory, and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a low-grade copper ore image sorting method based on deep learning.

[0067] A non-volatile computer storage medium stores computer-executable instructions for executing a method for image sorting of low-grade copper ore based on deep learning.

[0068] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A low-grade copper ore image sorting method based on deep learning, characterized in that: The steps include: Step S1: Acquire a dual-energy X-ray image of a low-grade copper ore, wherein the dual-energy X-ray image of the low-grade copper ore includes a dual-energy X-ray image of copper and a dual-energy X-ray image of waste rock; the dual-energy X-ray image of the low-grade copper ore is a high-energy X-ray image and a low-energy X-ray image of the low-grade copper ore; Step S2: Embed a learnable time-step embedding module into the U-Net model, introduce cross-layer residual connections and self-attention mechanisms to construct a time-step embedding U-Net network, input the dual-energy X-ray image of the low-grade copper ore into the time-step embedding U-Net network, and generate diversified dual-energy X-ray images of the copper ore; Step S3: constructing an adaptive copper ore separation model based on the VGG11 structure; Step S4: inputting the diversified copper ore dual-energy X-ray images into the adaptive copper ore sorting model to sort the copper ore; The time step embedding U-Net network is composed of a time step embedding module, a convolution module, a residual block, a self-attention module, a transposed convolution and a convolution layer. The processing flow of the time step embedding U-Net network is as follows: first, the input dual-energy X-ray image of the low-grade copper ore is encoded into an initial noise image and a time step scalar, the time step scalar is input into the time step embedding module for dimensionality increase processing, the initial noise image is input into the convolution module, the output of the convolution module and the output of the time step embedding module are broadcast added to obtain a first splicing feature, the first splicing feature is downsampled and broadcast added to the output of the time step embedding module to obtain a second splicing feature, and the second splicing feature is subtracted from the original image. The second splicing feature is down-sampled and passed through the residual block and the self-attention module together with the output of the time-step embedding module. The output of the self-attention module, the second splicing feature and the output of the time-step embedding module are spliced to obtain a third splicing feature. The third splicing feature and the first splicing feature are input into the transposed convolution together. The output of the transposed convolution is spliced with the output of the time-step embedding module to obtain a fourth splicing feature. The fourth splicing feature is up-sampled and spliced with the output of the time-step embedding module to obtain a fifth splicing feature. The fifth splicing feature is mapped to the target dimension through a convolution layer to obtain the final output of the time-step embedding U-Net network, i.e., the diversified copper ore dual-energy X-ray image.

2. The method for image separation of low-grade copper ore based on deep learning according to claim 1, characterized in that: The time step embedding module adopts sinusoidal position encoding.

3. The method for image separation of low-grade copper ore based on deep learning according to claim 2, characterized in that: The adaptive copper ore sorting model based on VGG11 structure consists of a backbone network Stem, the first stage Stage1, the second stage Stage2, a main classifier, a feature projection module and a domain discriminator; The processing flow of the adaptive copper ore sorting model based on the VGG11 structure is as follows: the input, i.e., the diversified copper ore dual-energy X-ray image, is sequentially passed through the backbone network, the first stage, and the second stage to obtain a second enhanced feature map, and the second enhanced feature map is respectively input into the main classifier, the feature projection module, and the domain discriminator for processing. The main classifier uses the output of the feature projection module to perform classification tasks and predict the probability of the enhanced feature map belonging to each category. The domain discriminator uses the output of the feature projection module to perform domain adaptation tasks and predict the probability of the enhanced feature map belonging to the source domain or the target domain. Finally, the output of the adaptive copper ore sorting model based on the VGG11 structure is the probability distribution of the enhanced feature map belonging to each category and the probability distribution of the enhanced feature map belonging to the source domain or the target domain.

4. The method for image sorting of low-grade copper ore based on deep learning according to claim 3, characterized in that: The first and second stages have the same structure, both consisting of a multi-scale contextual attention module, a dynamic feature compensation unit, and an enhanced feature convolution unit. The processing flow of the first stage is as follows: the input, i.e., the output of the backbone network, passes through the multi-scale contextual attention module and the dynamic feature compensation unit in sequence. The output of the dynamic feature compensation unit and the output of the multi-scale contextual attention module are input into the enhanced feature convolution unit for processing, thereby obtaining the first enhanced feature map, i.e., the final output of the first stage. Among them, the enhanced feature convolution unit consists of a convolutional layer, a batch normalization layer and a ReLu activation function connected in sequence.

5. The method for image separation of low-grade copper ore based on deep learning according to claim 4, characterized in that: The dynamic feature compensation unit consists of a deformable convolution, a dense block, a channel attention module, a global average pooling module, a feature transformation unit and a Sigmoid function. The processing flow of the dynamic feature compensation unit is as follows: the input, i.e., the output of the multi-scale context attention module, is sequentially passed through the deformable convolution and dense block to obtain adaptive features, the input is sequentially passed through the channel attention module, the global average pooling module, the feature transformation unit and the Sigmoid function to obtain transformed features, the transformed features and the adaptive features are spliced and then channel-fused with the input of the dynamic feature compensation unit to obtain fused features, i.e., the final output of the dynamic feature compensation unit.

6. The method for image separation of low-grade copper ore based on deep learning according to claim 5, characterized in that: The main classifier consists of an adaptive average pooling layer, a flattening layer and a fully connected layer connected in sequence; The domain discriminator consists of a gradient reversal layer, a global average pooling layer, a first dimensionality reduction feature transformation module, a first dropout layer, a second dimensionality reduction feature transformation module, a second dropout layer and a fully connected layer connected in sequence; The first dimensionality reduction feature transformation module and the second dimensionality reduction feature transformation module have the same structure, both consisting of a fully connected layer, a batch normalization layer and a ReLu activation function connected in sequence.

7. The method for image sorting of low-grade copper ore based on deep learning according to claim 6, characterized in that: Add intra-class-inter-class loss function enhancement constraints to force the Euclidean distance of similar enhanced feature maps to be less than the threshold , forcing the Euclidean distance of heterogeneous enhanced feature maps to be greater than the threshold ; The intra-class-inter-class loss function is expressed as: ； Where, represents the intra-class-inter-class loss function; 、 are the i-th enhanced feature map and the j-th enhanced feature map respectively, 、 Belong to the same type of enhanced feature map; is the kth enhanced feature map, which belongs to the heterogeneous enhanced feature map; Represents the weight parameter.

8. An electronic device, characterized in that: The invention comprises a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a low-grade copper ore image sorting method based on deep learning as described in any one of claims 1 to 7.

9. A non-volatile computer storage medium storing computer-executable instructions, characterized in that: The computer-executable instructions execute the low-grade copper ore image sorting method based on deep learning as described in any one of claims 1-7.

Images