Image segmentation method for identifying mineral boundary in table ore zone

Abstract

The invention relates to the technical field of image processing and intelligent identification, and discloses an image segmentation method for identifying a mineral boundary in a table ore zone. The method comprises the steps of constructing a multi-branch single-output enhancement unit based on gating enhancement and multi-head attention combined dynamic fusion, constructing a multi-scale fusion module of a split-level fusion pyramid structure based on multi-head attention and residual fusion, constructing a trunk module based on convolutional layer and enhancement unit stacking, and constructing a multi-scale fusion module of a split-level fusion pyramid structure based on multi-head attention and residual fusion. Constructing an image recognition segmentation model by combining step-by-step down-sampling, a multi-scale fusion module and a prediction head; and obtaining an ore zone image of a shaking table operation site of the reselection workshop, inputting the ore zone image into the image recognition segmentation model to obtain a recognition result, and completing shaking table separation of the mineral particles based on the recognition result. The problem of insufficient segmentation precision caused by particle size difference, illumination interference, fuzzy boundary and real-time requirement in the existing industrial mineral separation process is solved.

Description

Technical Field

[0001] This invention relates to the field of image processing and intelligent recognition technology, and in particular to a hierarchical attention feature fusion and boundary enhancement segmentation method suitable for images of shaking table mining areas. Background Technology

[0002] In mineral processing and shaking table separation, the spatial distribution characteristics of mineral particles on the shaking table surface directly affect resource recovery efficiency and automation control accuracy. Accurate identification of mineral distribution areas, especially the precise segmentation of mineral boundaries and particle structures, is a key technological step in improving separation performance and achieving intelligent control.

[0003] Existing methods for segmenting mineral deposit images mainly rely on manual experience or traditional image processing algorithms, such as grayscale thresholding, edge detection, and morphological operations. While these methods are feasible in experimental environments, they exhibit significant limitations in complex industrial scenarios, as detailed below: First, the complex industrial environment and significant lighting interference severely impact segmentation accuracy. The surface of the shaking table often exhibits water films, reflective areas, and adherence of moist mineral particles, resulting in indistinct grayscale or color differences between the minerals and the background. Under these conditions, traditional algorithms are prone to missegmentation or missed detections, especially in edge regions and small mineral particle identification, where the missed detection rate can exceed 20%, failing to meet boundary accuracy requirements.

[0004] Secondly, the mineral particles exhibit significant size differences on the shaking table, with fine-grained minerals mixed with tailings. Traditional single-scale convolution methods or conventional convolutional neural networks (CNNs) are prone to problems such as blurred boundaries and missing targets when processing small-particle regions, leading to a decrease in overall segmentation accuracy.

[0005] Third, existing methods generally lack global semantic modeling capabilities. Traditional local convolutional or edge detection algorithms can only capture local texture features, making it difficult to establish long-range dependencies between mineral regions. In scenarios where mineral distribution is continuous or particles overlap, segmentation results are prone to fragmentation and discontinuity. Even conventional deep convolutional networks, due to their limited receptive field, struggle to effectively model cross-regional semantic relationships.

[0006] Furthermore, insufficient multi-scale feature fusion capabilities further restrict segmentation performance. Mineral belt image segmentation tasks require the simultaneous preservation of high-resolution edge information and low-resolution semantic information. Existing Feature Pyramid Networks (FPNs) or simple fusion strategies have limited performance in small-particle recognition and boundary modeling, making it difficult to balance local details and overall semantics.

[0007] Finally, industrial applications place stringent demands on real-time performance and computational efficiency. Shaking table sorting is an online processing scenario, requiring high-frame-rate image segmentation and lightweight deployment. Existing deep learning models are complex in structure and computationally intensive, making them difficult to run in real-time on embedded GPUs or limited hardware platforms, thus affecting automated sorting efficiency and system responsiveness. Summary of the Invention

[0008] This invention provides an image segmentation method for identifying mineral boundaries in a shaking table ore zone, in order to solve the problem of insufficient segmentation accuracy caused by particle size differences, illumination interference, boundary blurring, and real-time requirements in existing industrial mineral processing.

[0009] To achieve the above objectives, the present invention employs the following technical solution: This invention provides an image segmentation method for identifying mineral boundaries in a shaking table ore zone, comprising the following steps: Step 1: Construct a multi-branch single-output enhancement unit based on gating enhancement and multi-head attention combined with dynamic fusion; construct a multi-scale fusion module with a staggered fusion pyramid structure based on multi-head attention and residual fusion; construct a backbone module based on stacked convolutional layers and enhancement units; and construct an image recognition and segmentation model by combining stepwise downsampling, multi-scale fusion module and prediction head. Step 2: Obtain an image of the ore belt at the shaking table operation site in the gravity separation workshop, input the ore belt image into the image recognition and segmentation model, obtain the recognition result, and complete the shaking table separation of mineral particles based on the recognition result.

[0010] Furthermore, the image recognition and segmentation model receives an image of the ore zone, obtains multi-scale semantic features of the ore zone image at different scales through the backbone module combined with stepwise downsampling, and then performs feature fusion on the multi-scale semantic features based on multi-head attention and residual fusion through the multi-scale fusion module to output a feature fusion map. The prediction head obtains the recognition result based on the feature fusion map.

[0011] Furthermore, the enhancement unit based on gating enhancement and multi-head attention combined with dynamic fusion to construct a multi-branch single-output enhancement unit includes: constructing a local enhancement branch based on gating enhancement, constructing a global attention branch based on multi-head attention, constructing a dynamic fusion branch based on the overall distribution feature relationship of the feature image of the input features, and constructing a three-input single-output enhancement unit based on the local enhancement branch, the global attention branch, and the dynamic fusion branch.

[0012] Furthermore, the dynamic fusion branch outputs fusion weights based on the overall distribution feature relationship of the feature image combined with the input features through global average pooling, and the enhancement unit obtains fusion features based on the fusion weights combined with the output features of the local enhancement branch and the global attention branch. The overall distribution feature relationship output fusion weight based on the input features is expressed by the following formula: ; in, This represents the fusion weights of the global attention branches; This indicates a global average pooling operation; Represents a feature image; This represents the fusion weight of the local enhancement branch.

[0013] Through the above design, the self-gated local enhancement branch and the global attention branch based on multi-head attention are parallelized, and the overall distribution relationship of the feature image of the input features is used as the weight expression to adaptively balance local and global information, and better express the features at different levels.

[0014] Furthermore, the backbone module constructed based on the stacking of convolutional layers and enhancement units includes: constructing an extraction module based on convolutional layers with a kernel size of 3 and a stride of 2 and enhancement units, and constructing a backbone module based on convolutional layers with a kernel size of 3 and a stride of 1 and multiple extraction modules. In the image recognition and segmentation model, the enhancement units in each extraction module are downsampled step by step.

[0015] Furthermore, the multi-scale fusion module receives the multi-scale semantic features and outputs a fused feature map by combining residual fusion with multi-head attention in the form of guided fusion rules between different scales in the multi-scale semantic features. The guided fusion rules between different scales in the multi-scale semantic features include: sorting the multi-scale features from low to high scale according to their scale relationship, so as to... Scale semantic feature guidance Scale semantic features Scale semantic features achieve spatial alignment through multiple upsampling steps, and feature fusion is completed sequentially from low to high. The residual fusion is the lowest-scale semantic feature that is guided to be fused into a preliminary fused feature, and then the preliminary fused feature is added to the fusion of the final fused feature map.

[0016] Through the above design, combining a pyramid structure with a dynamic weighting mechanism of multi-head attention, spatial alignment of upsampling, and a fusion strategy of lateral fusion, this module can adaptively adjust the fusion ratio of features at different scales according to the image content, thereby enhancing the response of boundary regions and small-particle regions.

[0017] This step significantly improves the model's segmentation consistency and boundary closure in scenarios with continuous mineral distribution, overlapping particles, and blurred boundaries.

[0018] Furthermore, Scale semantic feature guidance The process of fusing scale-based semantic features includes: based on Generate query matrix based on scale semantic features. Scale semantic features are used to generate key and value matrices, which are then fused using multi-head attention.

[0019] Through the above design, by guiding the lower layer through the middle layer, and then guiding the previous fusion result through the higher layer, residual preservation is added to the attention calculation of the staggered layers to avoid semantic conflicts and spatial information loss, thus achieving more granular semantic transmission and structural integrity.

[0020] Furthermore, the prediction head obtains the recognition result based on the feature fusion map, including: the prediction head obtains a mineral category map, a segmentation mask map, and a maximum response location map as the recognition result based on the feature fusion map.

[0021] Furthermore, the image segmentation model constructs a loss function based on the main segmentation loss, boundary enhancement loss, and semantic consistency loss combined with preset weights; The main segmentation loss is constructed based on the classification accuracy of the mineral region and the background region combined with the cross-entropy function; The boundary enhancement loss is constructed based on the Dice coefficient, which combines boundary ambiguity, breakage, and missed detection phenomena. The semantic consistency loss is constructed based on semantic constraints of cosine similarity.

[0022] Furthermore, the main segmentation loss is expressed by the following formula: ; in, Indicates the primary segmentation loss; This represents the total number of pixels in the mineral belt image; This indicates the number of categories in the mineral belt image; Indicates the first The pixel in the first The function that indicates the actual label on the class; The model predicts the first The pixel in the first Probability on class; The boundary reinforcement loss is expressed by the following formula: ; in, Indicates boundary enhancement loss. Indicates the first The true boundary label of each pixel; This represents the intensity of the boundary response predicted by the model; This represents a small constant to prevent the denominator from being zero; The semantic consistency loss is expressed by the following formula: ; in, This represents the loss of semantic consistency. Indicates the number of locations for step-by-step downsampling; and They represent the first The semantic vector of each position in the fused feature map of the bottom and high layers. This represents the L2 norm of a vector.

[0023] Through the above design, the boundary enhancement loss based on Dice coefficient, the cross-scale semantic consistency loss based on cosine similarity, and the joint orchestration loss function of main segmentation cross-entropy provide better hierarchical supervision to strengthen boundary closure and cross-layer semantic alignment, thereby obtaining more accurate recognition results.

[0024] Beneficial effects: This invention provides an image segmentation method for identifying mineral boundaries in shaking table ore zones. It constructs an image recognition and segmentation model by combining a custom-designed enhancement unit, backbone module, and multi-scale fusion module with stepwise downsampling. The local-global attention enhancement design of the enhancement unit improves segmentation robustness under conditions of reflection, water film interference, and particle adhesion, enhancing the ability to distinguish mineral boundaries, tailings transition zones, and fine-grained regions. This effectively solves the problem of insufficient segmentation accuracy for small-particle minerals and tailings areas.

[0025] The multi-scale fusion module constructs a multi-scale semantic fusion path. Addressing the common problems of large differences in particle scale, blurred boundaries, and semantic breaks in industrial scenarios, it constructs a hierarchical semantic guidance path to achieve collaborative modeling of low-level spatial details and high-level semantic structures, taking into account both local details and global structure.

[0026] The image recognition and segmentation model designed in this invention adopts a lightweight design, which adapts to the needs of industrial online detection while ensuring accuracy, and supports multi-task joint prediction to enhance the robustness and industrial adaptability of the model. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the network structure of the image recognition and segmentation model in an embodiment of the present invention, wherein LGABlock (Local Global Attention Block) represents an enhancement unit, and HAFPAN (Hierarchical AttentionFeature Pyramid Network) represents a multi-scale fusion module; Figure 2 This is a schematic diagram of the network structure of the enhancement unit in an embodiment of the present invention; Figure 3This is a schematic diagram of the network structure of the multi-scale fusion module in an embodiment of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0030] Please see Figure 1 This application provides an image segmentation method for identifying mineral boundaries in a shaking table ore zone, comprising the following steps: Step 1: Construct a multi-branch single-output enhancement unit based on gating enhancement and multi-head attention combined with dynamic fusion; construct a multi-scale fusion module with a staggered fusion pyramid structure based on multi-head attention and residual fusion; construct a backbone module based on stacked convolutional layers and enhancement units; and construct an image recognition and segmentation model by combining stepwise downsampling, multi-scale fusion module and prediction head. Please see Figure 2 Specifically, the enhancement unit is used to optimize the representation of intermediate features in the image segmentation network. It is suitable for modeling boundary details, small-particle regions, and semantic continuity in mineral image segmentation tasks. Its specific network structure is as follows: A local enhancement branch is constructed based on gating enhancement, a global attention branch is constructed based on multi-head attention, and a dynamic fusion branch is constructed based on the overall distribution feature relationship of the input feature image. A three-input single-output enhancement unit is constructed based on the local enhancement branch, the global attention branch, and the dynamic fusion branch.

[0031] The enhancement unit first uses a gating mechanism—a local enhancement branch—to extract spatial details and edge information from the input features, enhancing local responses to obtain local features. Then, a global attention branch enhances the understanding of cross-regional structures, and multi-head attention is used to extract global features from the input features. Finally, a dynamic fusion branch adaptively adjusts the fusion ratio of local and global features based on the overall distribution of the input features to obtain fused features. The fused features output by the enhancement unit are further processed in the backbone module through dimensionality reduction via convolutional layers and feature extraction by the enhancement unit again.

[0032] The process of adaptively adjusting the fusion ratio in the dynamic fusion branch is based on the global average pooling combined with the overall distribution feature relationship of the feature image of the input features to output fusion weights. The enhancement unit obtains fusion features based on the fusion weights combined with the output features of the local enhancement branch and the global attention branch. The output fusion weights based on the overall distribution feature relationship of the feature image derived from the input features are expressed by the following formula: ; in, This represents the fusion weights of the global attention branches; This indicates a global average pooling operation; Represents a feature image; This represents the fusion weight of the local enhancement branch.

[0033] The specific network structure of the backbone module is as follows: the extraction module is constructed based on a convolutional layer with 3 kernels and a stride of 2 and enhancement units, and the backbone module is constructed based on a convolutional layer with 3 kernels and a stride of 1 and 3 extraction modules.

[0034] After constructing the main module, the enhancement units in each extraction module are downsampled step by step and then connected to the input of the multi-scale fusion module. The output of the multi-scale fusion module is then connected to the prediction head, thus completing the construction of the image recognition and segmentation model.

[0035] Please see Figure 3 The multi-scale fusion module receives multi-scale semantic features and outputs a fused feature map as the recognition result by combining residual fusion with multi-head attention in the form of guided fusion rules between different scales in the multi-scale semantic features. The multi-scale features are sorted from low to high scale according to their scale relationship, so as to... Scale semantic feature guidance The feature fusion is performed sequentially from low to high scale semantic features. After the residual fusion is guided to fuse the lowest-scale semantic features into preliminary fused features, the preliminary fused features are added to the fusion of the final fused feature map.

[0036] Scale semantic feature guidance The process of fusing scale-based semantic features includes: based on Generate query matrix based on scale semantic features. Scale semantic features are used to generate key and value matrices, which are then fused using multi-head attention.

[0037] In this embodiment, the main module stacks three extraction modules. The multi-scale semantic features finally input to the multi-scale fusion module include low-level semantic features, mid-level semantic features, and high-level semantic features. The fusion process is as follows: the mid-level semantic features guide the low-level semantic features for attention fusion, and then the high-level semantic features are further fused with the mid- and low-level fusion features after upsampling spatial alignment. The final fused feature map is output by combining the preliminary fusion features with residual fusion. In other embodiments, more extraction modules can be stacked according to the recognition accuracy requirements. Only the structure of the multi-scale fusion module needs to be adjusted. That is, the lowest-level semantic features are guided by the higher-level semantic features for fusion, and the higher-level semantic features are guided by upsampling spatial alignment for fusion. Finally, all fusion features are residually fused, which can also achieve the technical effect required by the present invention.

[0038] In this embodiment, the specific feature fusion process is as follows: a query matrix is ​​extracted based on mid-level semantic features, a key matrix and a value matrix are extracted based on low-level semantic features, and then multi-head attention fusion is performed to obtain preliminary fused features. Then, after the high-level semantic features undergo two upsampling spatial alignment processes, a query matrix is ​​extracted, and the key matrix and value matrix are extracted from the preliminary fused features. Multi-head attention fusion is performed again to obtain secondary fused features. Finally, dimensionality reduction processing is used to fuse the preliminary fused features and the secondary fused features in a residual fusion manner to obtain a fused feature map.

[0039] The image segmentation model constructs a loss function based on the main segmentation loss, boundary enhancement loss, and semantic consistency loss combined with preset weights, which is expressed by the following formula: ; in, Represents the loss function; Indicates the primary segmentation loss; Indicates boundary enhancement loss; This represents the loss of semantic consistency. , as well as The values ​​represent preset weights, which are set to 1.0, 0.5, and 0.2 in this embodiment. The main segmentation loss is the basic supervision target, and is set to 1.0 to ensure its dominant position. The main segmentation loss aims to reduce boundary ambiguity, breakage, and missed detection. Considering the role of mineral boundaries in industrial sorting, it is set to 0.5. If other embodiments require higher boundary recognition accuracy, this value can be further increased. The semantic consistency loss function is set to 0.2 for micro-auxiliary constraints. If other embodiments face complex fusion or obvious semantic breakage scenarios, it can be appropriately increased.

[0040] The main segmentation loss is constructed based on the classification accuracy of the mineral region and the background region combined with the cross-entropy function, and is expressed by the following formula: ; in, This represents the total number of pixels in the mineral belt image; This indicates the number of categories in the mineral belt image; Indicates the first The pixel in the first The function that indicates the actual label on the class; The model predicts the first The pixel in the first Probability on class; The boundary enhancement loss is constructed based on the Dice coefficient, taking into account boundary ambiguity, breakage, and missed detection phenomena, and is expressed by the following formula: ; in, Indicates boundary enhancement loss. Indicates the first The true boundary label of each pixel; This represents the intensity of the boundary response predicted by the model; This represents a small constant to prevent the denominator from being zero; The semantic consistency loss is constructed based on the semantic constraints of cosine similarity, and is expressed by the following formula: ; in, Indicates the number of locations for step-by-step downsampling; and They represent the first The semantic vector of each position in the fused feature map of the bottom and high layers. This represents the L2 norm of a vector.

[0041] In this embodiment, the image recognition and segmentation model is trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 16, and 100 training epochs. An early stopping strategy is used to control the sampling on the validation set to prevent overfitting. Step 2: Obtain an image of the ore belt at the shaking table operation site in the gravity separation workshop, input the ore belt image into the image recognition and segmentation model, obtain the recognition result, and complete the shaking table separation of mineral particles based on the recognition result.

[0042] The image recognition and segmentation model receives images of the mineral zone. In the backbone module, preliminary convolution processing and spatial dimensionality reduction are performed to extract basic semantic features and compress redundant information, effectively extracting the initial edge and texture information of mineral particles, providing basic feature support for subsequent boundary modeling. The enhancement unit in the backbone module realizes the weight allocation of local and global features based on the overall distribution feature relationship of the feature image of its input features, improving the local-global feature expression, thereby enhancing the segmentation robustness under conditions of reflection, water film interference, and particle adhesion.

[0043] Then, through progressive downsampling and channel compression convolution operations, the dimensionality is reduced to obtain multi-scale semantic features, enhancing the ability to identify small-particle minerals, tailings boundaries, and overlapping areas. The multi-scale fusion module receives multi-scale semantic features and performs dynamic weighting by combining multi-head attention. Combined with upsampling spatial alignment and lateral connectivity, it strengthens the response of boundary areas and small-particle areas to obtain a feature fusion map.

[0044] Finally, the prediction head obtains a mineral category map, a segmentation mask map, and a maximum response location map based on the feature fusion map as the recognition results. The staff can then control the mineral sorting shaker to complete the shaking sorting of mineral particles based on the above mineral category map, segmentation mask map, and maximum response location map.

[0045] The mineral category map represents the probability of mineral categories in the image, used to identify regions with different mineral types. The segmentation mask map marks boundary regions. The maximum response location map helps locate key boundary points or central regions.

[0046] The image recognition and segmentation model provided by this invention can take into account both local details and global semantic modeling, achieving accurate identification of mineral particle edges and long-range regional relationships. It possesses multi-scale feature adaptive fusion capabilities, improving the recognition accuracy of small-particle minerals and mineral boundaries. It supports a dynamic weighting mechanism, automatically adjusting the fusion ratio of local and global information based on image features. It features a lightweight structure and high frame rate processing capabilities, meeting the needs of online sorting and monitoring in industrial settings.

[0047] Finally, to verify the performance of the image recognition and segmentation model proposed in this invention, five mainstream image segmentation methods were selected as benchmarks for comparison, including U-Net, DeepLabV3+, FPN, YOLOv5-Seg, and YOLOv8-Seg. All methods were trained and tested on a uniformly constructed mineral belt image dataset, and their performance was evaluated using the same preprocessing procedures and evaluation metrics.

[0048] Specific evaluation metrics include mean Intersection over Union (mIoU), mean Dice coefficient (mDice), boundary localization accuracy (Boundary F1-score), and inference speed (Frames Per Second, FPS). Experimental results are shown in Table 1. Table 1: Performance comparison of different methods in mineral boundary identification task.

[0049]

[0050] The results in the table show that the proposed method outperforms the comparison methods in all evaluation metrics, especially in segmentation accuracy. Specifically, mIoU and mDice reach 95.2% and 95.8% respectively, representing improvements of 4.9% and 4.3% compared to YOLOv8-Seg, demonstrating higher accuracy in mineral region identification tasks. Furthermore, the Boundary F1-score reaches 91.3%, significantly better than other methods, indicating that this method has stronger closure and continuity in boundary localization, making it particularly suitable for modeling transitional regions between tailings and concentrates.

[0051] Although YOLOv8-Seg is slightly faster than this method in terms of inference speed, the image recognition and segmentation model proposed in this invention can still meet the real-time requirements of industrial sites (FPS≥20) while maintaining high accuracy, and has good deployment efficiency and engineering adaptability.

[0052] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An image segmentation method for identifying mineral boundaries in a shaking table ore zone, characterized in that, Includes the following steps: Step 1: Construct a multi-branch single-output enhancement unit based on gating enhancement and multi-head attention combined with dynamic fusion; construct a multi-scale fusion module with a staggered fusion pyramid structure based on multi-head attention and residual fusion; construct a backbone module based on stacked convolutional layers and enhancement units; and construct an image recognition and segmentation model by combining stepwise downsampling, multi-scale fusion module and prediction head. Step 2: Obtain an image of the ore belt at the shaking table operation site in the gravity separation workshop, input the ore belt image into the image recognition and segmentation model, obtain the recognition result, and complete the shaking table separation of mineral particles based on the recognition result.

2. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 1, characterized in that, The image recognition and segmentation model receives an image of the ore zone, obtains multi-scale semantic features of the ore zone image at different scales through a backbone module combined with stepwise downsampling, and then performs feature fusion on the multi-scale semantic features based on multi-head attention and residual fusion through the multi-scale fusion module to output a feature fusion map. The prediction head obtains the recognition result based on the feature fusion map.

3. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 1, characterized in that, The enhancement unit based on gating enhancement and multi-head attention combined with dynamic fusion to construct a multi-branch single-output enhancement unit includes: constructing a local enhancement branch based on gating enhancement, constructing a global attention branch based on multi-head attention, constructing a dynamic fusion branch based on the overall distribution feature relationship of the feature image of the input features, and constructing a three-input single-output enhancement unit based on the local enhancement branch, the global attention branch, and the dynamic fusion branch.

4. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 3, characterized in that, The dynamic fusion branch outputs fusion weights based on the overall distribution feature relationship of the feature image combined with the input features through global average pooling. The enhancement unit obtains fusion features based on the fusion weights and the output features of the local enhancement branch and the global attention branch. The overall distribution feature relationship output fusion weight based on the input features is expressed by the following formula: ； in, This represents the fusion weights of the global attention branches; This indicates a global average pooling operation; Represents a feature image; This represents the fusion weight of the local enhancement branch.

5. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 1, characterized in that, The backbone module constructed based on stacked convolutional layers and enhancement units includes: constructing an extraction module based on convolutional layers with 3 kernels and 2 strides and enhancement units, and constructing a backbone module based on convolutional layers with 3 kernels and 1 stride and multiple extraction modules. In the image recognition and segmentation model, the enhancement units in each extraction module are downsampled step by step.

6. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 2, characterized in that, The multi-scale fusion module receives the multi-scale semantic features and outputs a fused feature map by combining residual fusion with multi-head attention in the form of guided fusion rules between different scales in the multi-scale semantic features. The guided fusion rules between different scales in the multi-scale semantic features include: sorting the multi-scale features from low to high scale according to their scale relationship, so as to... Scale semantic feature guidance Scale semantic features Scale semantic features achieve spatial alignment through multiple upsampling steps, and feature fusion is completed sequentially from low to high. The residual fusion is the lowest-scale semantic feature that is guided to be fused into a preliminary fused feature, and then the preliminary fused feature is added to the fusion of the final fused feature map.

7. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 6, characterized in that, Scale semantic feature guidance The process of fusing scale-based semantic features includes: based on Generate query matrix based on scale semantic features. Scale semantic features are used to generate key and value matrices, which are then fused using multi-head attention.

8. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 2, characterized in that, The prediction head obtains the recognition results based on the feature fusion map, including: the prediction head obtains a mineral category map, a segmentation mask map, and a maximum response location map as recognition results based on the feature fusion map.

9. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 1, characterized in that, The image segmentation model is based on a loss function constructed by combining the main segmentation loss, boundary enhancement loss, and semantic consistency loss with preset weights. The main segmentation loss is constructed based on the classification accuracy of the mineral region and the background region combined with the cross-entropy function; The boundary enhancement loss is constructed based on the Dice coefficient, which combines boundary ambiguity, breakage, and missed detection phenomena. The semantic consistency loss is constructed based on semantic constraints of cosine similarity.

10. The image segmentation method for identifying mineral boundaries in a shaking table ore zone according to claim 9, characterized in that, The principal segmentation loss is expressed by the following formula: ； in, Indicates the primary segmentation loss; This represents the total number of pixels in the mineral belt image; This indicates the number of categories in the mineral belt image; Indicates the first The pixel in the first The function that indicates the actual label on the class; The model predicts the first The pixel in the first Probability on class; The boundary reinforcement loss is expressed by the following formula: ； in, Indicates boundary enhancement loss. Indicates the first The true boundary label of each pixel; This represents the intensity of the boundary response predicted by the model; This represents a small constant to prevent the denominator from being zero; The semantic consistency loss is expressed by the following formula: ； in, This represents the loss of semantic consistency. Indicates the number of locations for step-by-step downsampling; and They represent the first The semantic vector of each position in the fused feature map of the bottom and high layers. This represents the L2 norm of a vector.

Images