Deep learning image segmentation method and system for ore and core structure analysis

Abstract

The present application relates to a deep learning image segmentation method and system for ore and core structure analysis, belonging to the technical field of material structure analysis, which solves the technical problems of insufficient multi-scale boundary recognition ability, unstable segmentation of weak boundary area, easy overlap of different mineral phases, insufficient model cross-scene generalization ability, etc. of the existing image segmentation technology. The solution is: a deep learning image segmentation method and system for ore and core structure analysis, which acquires an original image; a multi-scale feature extraction network is constructed based on the encoding end and the decoding end; a material structure prior constraint module is constructed based on the pore connectivity prior, the fracture direction consistency prior and the mineral phase consistency prior; the multi-scale feature extraction network is supervised and trained and self-supervised adapted; the image is segmented; and the image segmentation image is optimized. Compared with the prior art, the present application has the advantages of strong multi-scale boundary recognition ability, stable segmentation, strong generalization ability, wide application range, etc.

Description

Deep learning image segmentation method and system for ore and core structure analysis Technical Field

[0001] This invention belongs to the field of material structure analysis technology, specifically relating to a deep learning image segmentation method and system for ore and core structure analysis. Background Technology

[0002] Ore and core samples are key research objects in geological exploration, mineral resource evaluation, and unconventional energy development. Their internal structure contains various microscopic features such as pores, fractures, and different mineral phases. These structural parameters directly determine important physical properties of rocks, such as permeability, mechanical properties, and reservoir space. To accurately evaluate these physical properties, it is usually necessary to perform structural identification and region segmentation on sample images to extract key material characteristics such as porosity, fracture connectivity, and mineral distribution. Therefore, image segmentation, as an indispensable part of materials testing and analysis, has significant application value in geology, energy, and materials science.

[0003] Currently, ore and core images are mainly acquired using imaging equipment such as optical microscopes, scanning electron microscopes (SEM), X-rays, and industrial CT. Among these, optical microscopy images offer high resolution but are easily affected by sample staining, lighting, and focal length variations; SEM images have good texture contrast but have a small field of view and high imaging costs; industrial CT images can reflect the three-dimensional structure of materials, but issues such as noise, artifacts, and small differences in grayscale between adjacent phases make segmentation tasks challenging. Especially in cores with complex mineral compositions, the grayscale values ​​of different mineral phases often overlap, making it difficult for traditional thresholding methods to accurately distinguish them; fracture boundaries are easily broken or blurred by noise interference; and the large differences in pore structure scale result in insufficient generalization ability of algorithms based on fixed-scale filtering or region growing.

[0004] To improve segmentation accuracy, traditional methods typically employ image processing algorithms such as Otsu's adaptive thresholding, K-means clustering, region growing, edge detection, and active contouring. However, these methods generally rely on manually set parameters, making it difficult to adapt to variations in sample type, imaging conditions, and noise distribution. Furthermore, when dealing with multi-scale porous networks, weakly textured boundaries, or complex mineral mixtures, these algorithms often suffer from oversegmentation, undersegmentation, boundary blurring, and structural breakage, failing to meet the demands of high-precision material analysis.

[0005] In recent years, deep learning technology, especially convolutional neural networks, has been widely used in medical image processing, remote sensing image recognition, and industrial defect detection. Segmentation models based on U-Net, DeepLab, and SegFormer have achieved significant results on public datasets. However, directly applying existing deep learning methods in the field of materials analysis still faces the following challenges:

[0006] First, core and ore samples exhibit significant multi-scale characteristics. Different pores, fractures, and mineral phases vary considerably in morphology, size, and texture. Conventional networks, with their fixed convolutional receptive fields, struggle to simultaneously capture fine structures and large-scale mineral boundaries. Second, the significant overlap in grayscale distributions between mineral phases makes the network prone to class confusion, especially in areas with weak grayscale contrast and indistinct textures, where existing models struggle to learn stable discriminative features. Third, noise and artifacts from imaging equipment interfere with network judgment, resulting in segmentation results with discontinuous boundaries, pore adhesion, or incomplete mineral phase separation. Furthermore, deep learning models typically require a large number of labeled samples, but image annotation in materials testing is extremely costly, and differences in imaging conditions across laboratories hinder cross-scenario generalization. The lack of effective domain adaptation mechanisms often necessitates retraining or fine-tuning to adapt to different samples, leading to high costs and deployment difficulties. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies and solve the technical problems of insufficient multi-scale boundary recognition, unstable segmentation of weak boundary regions, easy overlap of different mineral phases, and insufficient generalization ability of models across scenes, this invention provides a deep learning image segmentation method and system for ore and core structure analysis.

[0008] The present invention is achieved through the following technical solutions.

[0009] This invention provides a deep learning image segmentation method for ore and core structure analysis, comprising the following steps:

[0010] S1. Obtain optical microscopic images, SEM images, and industrial CT cross-sectional images of ore and rock core as raw images, and perform normalization and noise suppression processing on the raw images;

[0011] S2. Construct a multi-scale feature extraction network based on the encoder and decoder ends;

[0012] S3. Construct a material structure prior constraint module based on priors of pore connectivity, fracture direction consistency, and mineral phase consistency.

[0013] S4. Using unlabeled target domain samples, a learning loss function is constructed through a self-supervised domain adaptation strategy. The multi-scale feature extraction network is trained and self-supervised on labeled data in the source domain based on the learning loss function and the prior constraint module of material structure.

[0014] S5. The trained multi-scale feature extraction network performs image segmentation on the image processed in step S1.

[0015] S6. Optimize the image segmentation map in step S5.

[0016] Furthermore, the step of constructing the multi-scale feature extraction network in step S2 is as follows: the encoder sets different convolution kernel sizes or different receptive fields to extract multi-scale features, and the decoder sets a feature fusion unit to progressively fuse the high-level semantic features of the decoder with the multi-scale features of the encoder. The feature fusion unit includes upsampling operation, feature alignment operation, and feature concatenation or weighted fusion operation.

[0017] Furthermore, the prior knowledge of pore connectivity in step S3 is represented by the energy of connected components. , For the energy of the connected domain, These are the weighting coefficients. The connectivity index is the i-th pore connected domain in the predicted segmentation result P of the pore category region output by the multi-scale feature extraction network.

[0018] The fracture orientation consistency prior is represented by fracture orientation consistency energy. , Energy consistent with the fracture direction. The original image spatial coordinates, To predict the direction of the fracture, The gradient direction;

[0019] The mineral phase consistency prior is represented by the mineral phase consistency prior energy. , For the a priori energy of mineral phase consistency. For local regions in the image, To predict the mineral phase category pixel set within a local region R of the image for the segmentation result P, Let I be the set of pixels within a local region R of the original image. To Texture statistical features of the corresponding region To Texture statistical features of the corresponding region;

[0020] The comprehensive prior energy in the material structure prior constraint module is: , , , These are the weighting coefficients.

[0021] Furthermore, the step of constructing the learning loss function in step S4 is as follows:

[0022] Select source domain samples Target domain samples ;

[0023] Constructing feature vectors , , , , The source domain sample feature vector, The feature vector of the target domain sample;

[0024] pass Construct the learning loss function, To learn the loss function, For feature similarity, This refers to the temperature parameter.

[0025] Furthermore, the steps in step S4 where the multi-scale feature extraction network undergoes supervised training and self-supervised adaptation on labeled data in the source domain are as follows:

[0026] Input source domain samples Supervised training is performed using cross-entropy loss, comprehensive prior energy in the material structure prior constraint module, and learning loss function.

[0027] Input target domain samples Self-supervised adaptation is achieved by learning the loss function, style normalization loss, and the comprehensive prior energy in the material and structure prior constraint module.

[0028] Furthermore, the optimization of the image segmentation map in step S5 in step S6 includes morphological post-processing strategies, pore and fracture identification strategies, mineral phase statistical analysis strategies, and structural parameter calculation strategies.

[0029] Furthermore, the system includes:

[0030] The image acquisition and data management module is used to acquire and store optical microscopic images, SEM images, and industrial CT cross-sectional images of ore and rock cores.

[0031] An image preprocessing module is used to normalize and suppress noise in the images stored in the image acquisition and data management module.

[0032] A multi-scale feature extraction network construction module is provided, which is used to construct a multi-scale feature extraction network based on the encoder and decoder ends.

[0033] A material structure prior constraint module is used to construct a material structure prior constraint module based on pore connectivity prior, fracture direction consistency prior, and mineral phase consistency prior.

[0034] A self-supervised domain adaptation module is used for training and self-supervised adaptation through a multi-scale feature extraction network.

[0035] An image segmentation module, which is used to perform image segmentation based on a trained multi-scale feature extraction network;

[0036] An image segmentation optimization module is used to improve the rationality of image segmentation results.

[0037] Furthermore, the image segmentation optimization module includes a morphological post-processing strategy, a pore and fracture identification strategy, a mineral phase statistical analysis strategy, and a structural parameter calculation strategy. The morphological post-processing strategy is used to perform morphological operations on the initial results of the image segmentation module; the pore and fracture identification strategy is used to identify, extract, and distinguish pore structures and fracture structures in the image; the mineral phase statistical analysis strategy is used to perform statistical analysis on different mineral phases to obtain the area ratio and spatial distribution characteristics of each mineral phase; and the structural parameter calculation strategy is used to calculate and derive structural parameters.

[0038] The beneficial effects achieved by this invention are:

[0039] 1. This invention employs a multi-scale feature extraction network construction module, which significantly improves the network's ability to identify pores and fractures at different scales, especially in the segmentation of small pores and weakly textured boundaries, where it is more stable. It also employs a material structure prior constraint module to solve the category confusion problem caused by overlapping mineral phase grayscale values, improving the structural consistency and physical rationality of the segmentation structure. Furthermore, it employs a self-supervised domain adaptive module, enabling the trained multi-scale feature extraction network to adapt to different imaging devices, different geological cores, and different noise conditions, improving its cross-scene generalization ability. Finally, it employs an image segmentation optimization module to further improve the continuity and regional integrity of the segmentation boundaries, enhancing reliability.

[0040] 2. This invention also proposes a deep learning image segmentation method for ore and core structure. This method has significant advantages in structure recognition, segmentation stability and cross-scene adaptability, effectively improving the accuracy of physical structure parameter extraction in ore and core samples, and providing reliable technical support for material property analysis and resource evaluation.

[0041] Compared with existing technologies, this invention has the advantages of strong multi-scale boundary recognition capability, stable segmentation, strong generalization capability, and wide application range. Attached Figure Description

[0042] Figure 1 is a structural block diagram of the present invention;

[0043] Figure 2 is a cross-sectional image of the synthetic core constructed in an embodiment of the present invention;

[0044] Figure 3 is the original grayscale image in an embodiment of the present invention;

[0045] Figure 4 is a candidate image of the dark region obtained based on Otsu threshold segmentation in an embodiment of the present invention;

[0046] Figure 5 is a diagram of the aperture mask in an embodiment of the present invention;

[0047] Figure 6 is a diagram of the skeleton structure after the crack region has been thinned in an embodiment of the present invention.

[0048] Figure 7 is a pseudo-color structure segmentation diagram in an embodiment of the present invention;

[0049] Figure 8 is an overlay of the original grayscale image and the segmentation structure in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0051] Example 1: As shown in Figures 1 to 8, a deep learning image segmentation method for ore and core structure analysis includes the following steps:

[0052] S1. Obtain optical microscopic images, SEM images, and industrial CT cross-sectional images of ore and rock core as raw images, and perform normalization and noise suppression processing on the raw images;

[0053] Specifically, let the original image be... , Using the original image's spatial coordinates, normalization and noise suppression processing of the original image primarily improves input stability. Normalization is achieved through... deal with, The mean of the image. Image standard deviation; noise suppression processing is achieved through... For processing, Gaussian or anisotropic filtering can generally be used. Noise suppression processing is mainly used to preserve edge structure and reduce random noise.

[0054] A multi-scale feature extraction network is constructed based on the encoder and decoder ends. The steps for constructing the multi-scale feature extraction network are as follows: the encoder ends are set with different convolution kernel sizes or different receptive fields to extract multi-scale features, and the decoder ends are set with feature fusion units to progressively fuse the high-level semantic features of the decoder ends with the multi-scale features of the encoder ends. The feature fusion unit includes upsampling operation, feature alignment operation and feature concatenation or weighted fusion operation.

[0055] Specifically, the encoding end is set with different convolutional kernel sizes such as 3×3, 5×5, and 7×7. The features corresponding to the different convolutional kernel sizes are: , With different convolution kernel sizes such as 3×3, 5×5, and 7×7, the final multi-scale features are: , To address the multi-scale features at the encoding end, setting different convolution kernel sizes reduces the limitation of fixed convolution scale on structure recognition capabilities. Different kernel sizes can simultaneously capture fine pore boundaries, small-scale fracture textures, and large-scale mineral phase region contours. At the decoding end, a feature fusion unit progressively fuses the high-level semantic features from the decoding end with the multi-scale features from the encoding end. , For the decoding end Layer features, For the encoding end Layer features, For upsampling operation, To achieve convolutional fusion and thus improve the representation of mineral phase boundaries, fracture contours, and pore morphology, the multi-scale feature extraction network ultimately outputs... , For multi-scale feature extraction networks at pixel coordinates The predicted output at a given location is used to characterize the probability distribution or corresponding category label of the pixel belonging to the pore, fracture, or mineral phase category.

[0056] S3. A material structure prior constraint module is constructed based on prior knowledge of pore connectivity, fracture orientation consistency, and mineral phase consistency; among which, the pore connectivity prior is represented by the energy of the connected domain. , For the energy of the connected domain, These are the weighting coefficients. The connectivity index is the i-th pore connected domain in the predicted segmentation result P of the pore category region output by the multi-scale feature extraction network.

[0057] The fracture orientation consistency prior is represented by fracture orientation consistency energy. , Energy consistent with the fracture direction. The original image spatial coordinates, To predict the direction of the fracture, The gradient direction should be maintained; the crack direction should be kept stable, which can be achieved by utilizing the image gradient direction.

[0058] The mineral phase consistency prior is represented by the mineral phase consistency prior energy. , The mineral phase consistency prior energy primarily measures the consistency of texture within a mineral phase region. The smaller the value, the closer the predicted mineral phase region is to the actual mineral structure in terms of texture distribution. For local regions in the image, To predict the mineral phase category pixel set within a local region R of the image for the segmentation result P, Let I be the set of pixels within a local region R of the original image. To Texture statistical features of the corresponding region To Texture statistical features of the corresponding region;

[0059] The comprehensive prior energy in the material structure prior constraint module is: , , , These are the weighting coefficients.

[0060] Specifically, the material structure prior constraint module improves the structural rationality and physical interpretability of the segmentation results. Specifically, the pore connectivity prior constrains scale consistency, the fracture direction consistency prior maintains linear extension characteristics through gradient direction consistency constraints, and the mineral phase consistency prior suppresses noise interference within the region through intraphase consistency constraints based on texture statistics.

[0061] S4. Using unlabeled target domain samples, a learning loss function is constructed through a self-supervised domain adaptation strategy. The multi-scale feature extraction network is trained and self-supervised on labeled data in the source domain based on the learning loss function and the prior constraint module of material structure. This step is mainly to improve the generalization ability.

[0062] The steps for constructing the learning loss function are as follows:

[0063] Select source domain samples Target domain samples ;

[0064] Constructing feature vectors , , , , The source domain sample feature vector, The feature vector of the target domain sample;

[0065] pass Construct the learning loss function, To learn the loss function, For feature similarity, This refers to the temperature parameter.

[0066] The steps for the multi-scale feature extraction network to undergo supervised training and self-supervised adaptation on labeled data in the source domain are as follows:

[0067] Input source domain samples Supervised training is performed using cross-entropy loss, the comprehensive prior energy in the material structure prior constraint module, and the learning loss function; that is... L represents the total training loss. For cross-entropy loss, , These are the loss weighting coefficients, used to balance the influence of prior material and structural constraints and learning loss on the total loss. These are constraints within the material structure prior constraint module. To learn the loss function;

[0068] Input target domain samples Self-supervised adaptation is achieved by learning the loss function, style normalization loss, and the comprehensive prior energy in the material and structure prior constraint module.

[0069] Specifically, by constructing a learning loss function, the multi-scale feature extraction network learns structural similarity features of images in the absence of annotations. While undergoing supervised training on labeled data in the source domain, the multi-scale feature extraction network learns structurally invariant features in target domain samples through a self-supervised domain adaptive strategy, thereby significantly improving the model's adaptability across devices and samples. The introduction of style normalization loss constrains the differences in brightness, contrast, and noise statistics between the source and target domain images, mitigating the impact of varying imaging conditions on the segmentation results and enabling the model to maintain stable output in real-world engineering environments.

[0070] S5. The trained multi-scale feature extraction network performs image segmentation on the image processed in step S1.

[0071] S6. The image segmentation map in the optimization step S5 includes morphological post-processing strategies, pore and fracture identification strategies, mineral phase statistical analysis strategies, and structural parameter calculation strategies. These ensure that the segmentation results meet the true characteristics of the core and ore samples in terms of physical structure.

[0072] The method of this invention has significant advantages in structural identification, segmentation stability, and cross-scene adaptability, effectively improving the accuracy of physical structural parameters extraction from ore and core samples, and providing reliable technical support for material property analysis and resource evaluation.

[0073] Figure 2 shows a synthetic core cross-sectional image constructed according to an embodiment of the present invention, used to simulate the spatial relationship of real mineral grains, pore structure, and fracture distribution. The background area exhibits random texture features of mineral phases, slightly brighter patches represent regions with different mineral compositions, multiple small-scale circular areas with lower grayscale represent pore structures, and striped dark lines represent fracture distribution extending horizontally. Figure 2 verifies the segmentation process of the present invention, demonstrating a stable ability to identify the internal structure of materials even without real labeled data.

[0074] Figure 4 shows the candidate dark region image obtained based on Otsu thresholding, mainly used for the initial extraction of low-grayscale regions containing pores and cracks. The black areas in the figure correspond to the set of dark pixels in the image, including pores, small cracks, and some noise points. Figure 4 illustrates the initial binarization step used in the segmentation process of this invention, which quickly distinguishes between dark and bright regions through global thresholding, providing a basic input for the subsequent structural separation of pores and cracks.

[0075] Figure 5 shows the pore mask. After the initial binarization step, the dark region contains both pore and crack structures. This invention removes isolated noise points through connected component filtering and area threshold filtering, and filters the region according to the scale characteristics of the pores to obtain a relatively pure pore distribution map. The white areas in the figure represent effective pore masks, which have been largely separated from the crack regions, further verifying the applicability of the proposed structure prior in the pore identification task. The connected component filtering mechanism includes connected component filtering and area threshold filtering, used to remove isolated noisy connected components and retain effective regions that meet the pore scale constraints.

[0076] Figure 6 shows the skeleton structure of the crack region after thinning. This embodiment uses edge detection combined with morphological dilation to extract candidate crack regions, and emphasizes the continuous direction and centerline structure of the crack through skeletonization. The linear structure in the figure accurately reflects the extension direction and connectivity path of the crack in the image, demonstrating the ability of this invention to maintain geometric consistency and directional features in crack structure recognition. Figure 6 can be further used to calculate material parameters such as crack length, orientation distribution, and connectivity.

[0077] Figure 7 shows a pseudo-color structure segmentation diagram. Different colors in the diagram correspond to different material structure categories, with cyan areas representing pore structures, red areas representing fracture networks, and yellow areas representing mineral phase regions. Figure 7 visually illustrates the spatial distribution of various structures in the sample. Although this embodiment uses synthetic data and an alternative segmentation process, the discrete distribution of pores, the directional extension of fractures, and the spatial outline of mineral phases are still visible, indicating that the decoupled structure identification strategy of this invention can effectively present the multi-scale structural features inside the core material.

[0078] Figure 8 shows the overlay of the original grayscale image and the segmentation result, with a porosity of 0.440. The green outline represents the pore boundaries, and the red outline represents the crack boundaries. By spatially aligning the identified structural contours with the original image, the accuracy of the segmentation results and the boundary localization capability can be visually verified. The pore contours generated by the method of this invention closely fit the real pore area, and the crack contours are also accurately distributed along the low-brightness, slender structures, indicating that the multi-scale feature extraction network and the material structure prior constraint module proposed in this invention can effectively meet the accuracy requirements of material structure recognition.

[0079] Example 2: A deep learning image segmentation system for ore and core structure analysis includes:

[0080] The image acquisition and data management module is used to acquire and store optical microscopic images, SEM images, and industrial CT cross-sectional images of ore and rock cores.

[0081] An image preprocessing module is used to normalize and suppress noise in the images stored in the image acquisition and data management module.

[0082] A multi-scale feature extraction network construction module is provided, which is used to construct a multi-scale feature extraction network based on the encoder and decoder ends.

[0083] A material structure prior constraint module is used to construct a material structure prior constraint module based on pore connectivity prior, fracture direction consistency prior, and mineral phase consistency prior.

[0084] A self-supervised domain adaptation module is used for training and self-supervised adaptation through a multi-scale feature extraction network.

[0085] An image segmentation module, which is used to perform image segmentation based on a trained multi-scale feature extraction network;

[0086] An image segmentation optimization module is used to improve the rationality of image segmentation results.

[0087] The image segmentation optimization module includes a morphological post-processing strategy, a pore and fracture identification strategy, a mineral phase statistical analysis strategy, and a structural parameter calculation strategy. The morphological post-processing strategy is used to perform morphological operations on the initial results of the image segmentation module, eliminating isolated noise points, filling structural voids, and smoothing structural boundaries, thereby improving the structural integrity of the segmentation results. The pore and fracture identification strategy is used to identify, extract, and distinguish pore structures and fracture structures in the image. The mineral phase statistical analysis strategy is used to perform statistical analysis on different mineral phases to obtain the area ratio and spatial distribution characteristics of each mineral phase. The structural parameter calculation strategy is used to calculate and derive structural parameters, including porosity, fracture connectivity, fracture length distribution, and mineral phase ratio.

[0088] Specifically, this invention employs a multi-scale feature extraction network construction module, which significantly improves the network's ability to identify pores and fractures at different scales, especially in the segmentation of small pores and weakly textured boundaries, where it is more stable. It also employs a material structure prior constraint module, which solves the category confusion problem caused by overlapping mineral phase grayscale values, improving the structural consistency and physical rationality of the segmentation structure. Furthermore, it employs a self-supervised domain adaptive module, enabling the trained multi-scale feature extraction network to adapt to different imaging devices, different geological cores, and different noise conditions, improving its cross-scene generalization ability. Finally, it employs an image segmentation optimization module, further enhancing the continuity and regional integrity of the segmentation boundaries, thereby improving reliability.

[0089] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, modifications can still be made to the embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning image segmentation method for ore and core structure analysis, characterized in that: The process includes the following steps: S1, acquiring optical microscopic images, SEM images, and industrial CT cross-sectional images of the ore and rock core as raw images, and performing normalization and noise suppression processing on the raw images; S2, constructing a multi-scale feature extraction network based on the encoding and decoding ends; S3. A material structure prior constraint module is constructed based on prior knowledge of pore connectivity, fracture direction consistency, and mineral phase consistency; the pore connectivity prior is represented by the energy of the connected domain. ， For the energy of the connected domain, These are the weighting coefficients. This is the connectivity index of the i-th pore connected domain in the predicted segmentation result P of the pore category region based on the multi-scale feature extraction network output; the fracture orientation consistency prior is represented by the fracture orientation consistency energy. ， Energy consistent with the fracture direction. The original image spatial coordinates, To predict the direction of the fracture, The gradient direction is defined; the mineral phase consistency prior is represented by the mineral phase consistency prior energy. ， For the a priori energy of mineral phase consistency. For local regions in the image, To predict the mineral phase category pixel set within a local region R of the image for the segmentation result P, Let I be the set of pixels within a local region R of the original image. To Texture statistical features of the corresponding region To The corresponding region's texture statistical features; the comprehensive prior energy in the material structure prior constraint module is ， 、 、 For weighting coefficients; S4, using unlabeled target domain samples, a learning loss function is constructed through a self-supervised domain adaptation strategy. The multi-scale feature extraction network is trained and self-supervised on labeled data in the source domain based on the learning loss function and the prior constraints of material structure; the steps for constructing the learning loss function are: selecting source domain samples Target domain samples ； Constructing feature vectors 、 ， ， ， The source domain sample feature vector, The feature vector of the target domain sample; through Construct the learning loss function, To learn the loss function, For feature similarity, For temperature parameters; the steps of supervised training and self-supervised adaptation of the multi-scale feature extraction network on labeled data in the source domain are as follows: input source domain samples Supervised training is performed using cross-entropy loss, comprehensive prior energy in the material structure prior constraint module, and learning loss function. Input target domain samples S1. Self-supervised adaptation is performed by learning the loss function, style normalization loss, and the comprehensive prior energy in the material structure prior constraint module; S2. The trained multi-scale feature extraction network performs image segmentation on the image processed in step S1; S3. The image segmentation map in step S5 is optimized.

2. The deep learning image segmentation method for ore and core structure analysis according to claim 1, characterized in that: The steps in step S2 to construct a multi-scale feature extraction network are as follows: the encoder sets different convolution kernel sizes or different receptive fields to extract multi-scale features, and the decoder sets a feature fusion unit to progressively fuse the high-level semantic features of the decoder with the multi-scale features of the encoder. The feature fusion unit includes upsampling operation, feature alignment operation, and feature concatenation or weighted fusion operation.

3. The deep learning image segmentation method for ore and core structure analysis according to claim 1, characterized in that: The optimization of the image segmentation map in step S5 in step S6 includes morphological post-processing strategies, pore and fracture identification strategies, mineral phase statistical analysis strategies, and structural parameter calculation strategies.

4. A deep learning image segmentation system for ore and core structure analysis, characterized in that: This system is a deep learning image segmentation method for ore and core structure analysis as described in any one of claims 1-3. The system comprises: an image acquisition and data management module for acquiring and storing optical microscopic images, SEM images, and industrial CT cross-sectional images of the ore and core; an image preprocessing module for normalizing and noise suppression of the images stored in the image acquisition and data management module; a multi-scale feature extraction network construction module for constructing a multi-scale feature extraction network based on an encoder and decoder; a material structure prior constraint module for constructing a material structure prior constraint module based on priors for pore connectivity, fracture direction consistency, and mineral phase consistency; a self-supervised domain adaptation module for training and self-supervised adaptation using the multi-scale feature extraction network; an image segmentation module for performing image segmentation based on the trained multi-scale feature extraction network; and an image segmentation optimization module for improving the rationality of the image segmentation results.

5. The deep learning image segmentation system for ore and core structure analysis according to claim 4, characterized in that: The image segmentation optimization module includes a morphological post-processing strategy, a pore and fracture identification strategy, a mineral phase statistical analysis strategy, and a structural parameter calculation strategy. The morphological post-processing strategy is used to perform morphological operations on the initial results of the image segmentation module; the pore and fracture identification strategy is used to identify, extract, and distinguish pore structures and fracture structures in the image; the mineral phase statistical analysis strategy is used to perform statistical analysis on different mineral phases to obtain the area ratio and spatial distribution characteristics of each mineral phase. The structural parameter calculation strategy is used to calculate and derive structural parameters.

Images