A lead-zinc ore image classification and prediction model construction method based on pattern recognition

Abstract

The application discloses a kind of lead-zinc ore image classification and prediction model construction method based on pattern recognition, it is related to ore image recognition and grade prediction technical field, including the following steps: generating multi-source lead-zinc ore data frame;Lead-zinc ore image is decoupled to illumination-reflective-texture;Generate mineral candidate prompt information and carry out prompt type candidate segmentation;Generate ore domain robust features and assay label rectification weight;Generate partition sampling offset field and construct lead-zinc mineral paragenetic constraint graph;The classification prediction input package is input into lead-zinc ore classification prediction network, and the output image classification result, mineral combination probability and lead-zinc grade prediction result;According to assay feedback deviation, closed loop is updated.The application improves the lead-zinc ore image classification accuracy and grade prediction reliability, reduces the influence of reflective, humid and mineral confusing factors on identification result.

Description

Technical Field

[0001] This invention relates to the field of ore image recognition and grade prediction technology, and in particular to a method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition. Background Technology

[0002] With the increasing demand for intelligent sorting in mines, online ore grade assessment, and automation of mineral processing, intelligent classification and grade prediction technologies for lead-zinc ore images have received widespread attention. Existing technologies include methods for constructing lead-zinc ore image classification models, which improve classification accuracy for different ore types by embedding attention modules such as SE, ECA, P-CBAM, SimAM, and CA into the original image classification model. Other technologies utilize spatial attention mechanisms and multi-scale convolutional features for mineral image classification, or improve the accuracy of identifying mineral categories such as pyrite through mineral candidate region segmentation and secondary recognition. Still other technologies utilize convolutional neural networks to extract chromaticity features from mineral powder images and perform grade estimation.

[0003] However, in real-world lead-zinc ore image recognition scenarios, ore surfaces often exhibit wet reflections, overexposed reflections, mixed metallic luster, and discontinuous vein textures. Simply relying on image enhancement or attention weighting makes it difficult to distinguish between genuine metallic mineral luster and non-mineral reflections. Existing candidate region segmentation typically only outputs mineral regions, failing to further convert candidate region centers, boundaries, vein directions, and reflection confidence into constraints for subsequent network sampling locations. Current grade estimation methods often directly regress grade based on whole-image features, lacking mechanisms for correcting lead-zinc mineral coexistence relationships, easily confused pyrite relationships, and laboratory label bias. This leads to inconsistencies between image classification results, mineral combination probabilities, and lead-zinc grade predictions. Therefore, a method for constructing a lead-zinc ore image classification and prediction model that integrates reflection confidence, mineral candidate partitioning, partition sampling offset, lead-zinc mineral coexistence constraints, and laboratory feedback updates is needed.

[0004] Therefore, how to provide a method for constructing a pattern recognition-based classification and prediction model for lead-zinc ore images is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition. This invention fully utilizes image decoupling, cue-based segmentation, co-occurrence constraint graphs, and deformable convolution techniques, and describes in detail the model construction process for lead-zinc ore image classification, mineral assemblage identification, and lead-zinc grade prediction. It has the advantages of high classification accuracy, strong adaptability to reflective interference, and reliable prediction results.

[0006] A method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to an embodiment of the present invention includes the following steps: Step 1: Collect lead-zinc ore image data, sample correlation data, and laboratory correlation data. Generate multi-source lead-zinc ore data frames through consistency verification, image-laboratory binding, and confidence marking. Step 2: Decouple the illumination-reflection-texture of the lead-zinc ore images in the multi-source lead-zinc ore data frames to generate enhanced ore images, reflection confidence maps, and texture enhancement maps; Step 3: Generate mineral candidate hints based on the enhanced ore image, reflection confidence map, and texture enhancement map; generate a mineral candidate partition map through hint-based candidate segmentation. Step 4: Generate robust features of ore domains based on the mineral candidate zoning map, and generate test label correction weights based on the consistency between the test correlation data and the image cataloging results; Step 5: Generate a partitioned sampling offset field based on the mineral candidate partition map, reflectance confidence map, and laboratory label correction weights. Construct a lead-zinc mineral coexistence constraint map based on the robust features of the ore domain. The enhanced ore image, partitioned sampling offset field, and lead-zinc mineral coexistence constraint map form a classification prediction input package. Step 6: Input the classification prediction input package into the lead-zinc ore classification prediction network. The lead-zinc ore classification prediction network uses InternImage as the backbone and configures the DCNv3 deformable convolutional unit as a partitioned guided sampling unit. After partitioned confidence gating, lead-zinc mineral coexistence constraint graph constraint and grade consistency prediction, it outputs the lead-zinc ore image classification result, mineral combination probability, lead grade prediction result and zinc grade prediction result. Step 7: Update the multi-source lead-zinc ore data frame, mineral candidate zoning map, lead-zinc mineral coexistence constraint map, and lead-zinc ore classification prediction network based on the test feedback deviation.

[0007] Optionally, step one specifically includes: Collect image data, sample correlation data, and laboratory correlation data of lead-zinc ore, and extract image number, sample number, image acquisition parameters, image logging results, laboratory number, lead grade test data, and zinc grade test data respectively; A unique sample number is assigned to each lead-zinc ore sample to be processed, and this unique sample number is simultaneously configured on the sample label, image acquisition record, and submission record. Consistency checks are performed on lead-zinc ore image data, sample association data, and test association data. When the same unique sample number corresponds to multiple test numbers, the target test number is determined according to the batch of samples submitted and the test completion time. Image quality assessment is performed on lead-zinc ore image data, and confidence labels are generated based on image sharpness, saturation pixel ratio, scale bar recognition results, and background occlusion results. Image-test binding relationships are established between lead-zinc ore image data that has passed consistency verification and target test numbers. The lead-zinc ore image data, sample association data, test association data and confidence markers that have completed image-test binding are then encapsulated into a multi-source lead-zinc ore data frame.

[0008] Optionally, step two specifically includes: Perform color space conversion on lead-zinc ore images in multi-source lead-zinc ore data frames to extract luminance, saturation, and lightness channels; The illumination components are estimated based on the luminance channel, and an illumination component map is generated. The original luminance channel is divided by the illumination component map and normalized to generate a reflection component map. The original luminance channel is subtracted from the illumination component map to generate a texture residual map. Identify wet reflective areas, overexposed reflective areas, and metallic luster areas based on the reflectance component map, saturation channel, and texture residual map; A reflection confidence map is generated based on the wet reflection area, the overexposed reflection area, and the metallic luster area. The reflection confidence map includes metallic luster confidence, wet reflection confidence, overexposed reflection confidence, and texture retention confidence. The credibility of metallic luster is generated based on the intensity of the reflection component, the mean of the local gradient, and the continuity of the boundary. The credibility of wet reflection is generated based on high reflection intensity, low saturation, and low texture residual. The credibility of overexposed reflection is generated based on the proportion of saturated pixels and the proportion of low gradient region area. The credibility of texture preservation is generated based on the intensity of texture residual and the continuity of grain boundary. Edge-preserving enhancement is performed on the texture residual map to generate a texture enhancement map, and the illumination component map, reflection component map, and texture residual map are fused to generate an enhanced ore image.

[0009] Optionally, step three specifically includes: Reflection screening results are generated based on the metallic luster confidence region, wet reflection region and overexposed reflection region in the reflection confidence map, and texture boundary results are generated based on the high gradient boundary, grain boundary and striped texture in the texture enhancement map. Based on the reflection screening results, mineral candidate prompts are generated, including metallic mineral candidate prompts, galena candidate prompts, sphalerite candidate prompts, and pyrite candidate prompts; Based on the texture boundary results, vein boundary hint lines and vein boundary areas are generated, and vein background hint boxes are generated based on the low metallic luster areas and low texture areas in the enhanced ore image. The mineral candidate hint points, vein boundary hint lines, and gangue background hint boxes are combined into mineral candidate hint information. After hint-based candidate segmentation, galena candidate area, sphalerite candidate area, pyrite candidate area, gangue candidate area, oxide film candidate area, and disseminated mineralization area are generated. Post-processing is performed on each candidate region. Candidate regions that overlap with wet reflective regions or overexposed reflective regions are marked as low-confidence candidate regions, and candidate regions that overlap with metallic luster confidence regions and whose texture preservation confidence meets the preset texture preservation threshold are marked as high-confidence candidate regions. Each candidate region is used as a candidate region node, and the candidate region category, region area, region center, color mean, texture intensity, metallic luster confidence, candidate region confidence, and boundary curvature are extracted as candidate region node attributes. Adjacent edges are established based on the contact relationship of candidate area boundaries, continuous texture edges are established based on the texture direction relationship of candidate areas, and vein-related edges are established based on the crossing relationship of vein boundary hint lines. The boundary contact length, texture direction angle, and vein-related marker are used as candidate area edge attributes to generate a mineral candidate partition map.

[0010] Optionally, step four specifically includes: Based on the node attributes of candidate regions in the mineral candidate zoning map, ore self-supervised characterization sample pairs are constructed, including positive sample pairs, grade-sensitive negative sample pairs, and easily confused negative sample pairs. Candidate regions with the same sample number but different perspectives or different cropping scales are combined into positive sample pairs. Candidate regions with similar image logging results but lead or zinc grade test data that differ from the preset grade difference threshold are combined into grade-sensitive negative sample pairs. Candidate regions with similar metallic luster confidence and classified as galena and pyrite candidate regions are combined into easily confused negative sample pairs. Perform masking reconstruction processing on the mineral candidate partition map, and input the masked candidate region node attributes, candidate region edge attributes and unmasked candidate region nodes into the candidate region encoder to generate candidate region representation vectors; The ore self-supervised representation sample is input into the ore self-supervised representation module, and robust features of the ore domain are generated through the image feature coding layer, the candidate region feature coding layer and the feature projection layer. Based on lead grade test data, zinc grade test data and image logging results, samples with high apparent mineralization, concealed mineralization samples, low contribution markers for galena candidate areas, low contribution markers for sphalerite candidate areas and easily confused markers for pyrite are generated. Based on samples with high apparent mineralization, samples with concealed mineralization, low contribution markers for galena candidate areas, low contribution markers for sphalerite candidate areas, and easily confused markers for pyrite, a corrective weight for the analysis label is generated.

[0011] Optionally, the step of generating a zonal sampling offset field based on the mineral candidate zoning map, the reflectance confidence map, and the test label correction weights specifically includes: Project the candidate area nodes in the mineral candidate zoning map onto the preset sampling grid, and record the center coordinates, boundary coordinates, and vein direction vector of the candidate area; The center attraction offset is generated based on the distance vector between each sampling position in the preset sampling grid and the coordinates of the center of the nearest candidate region. The boundary following offset is generated based on the candidate area boundary coordinates, candidate area boundary curvature, and vein boundary area. Generate directional constraint offsets based on the vein direction vector, texture continuity edges, and vein-related edges; Based on the reflectivity map and the test label correction weight, a sampling modulation weight is generated, and the candidate areas corresponding to the easily confused pyrite labels are marked as easily confused sampling areas. The center attraction offset, boundary following offset, direction constraint offset, and sampling modulation weight are aligned and meshed to generate a partitioned sampling offset field. Each sampling position in the partitioned sampling offset field is configured with a two-dimensional offset, sampling modulation weight, and easily confused sampling marker.

[0012] Optionally, the construction of the lead-zinc mineral co-occurrence constraint map based on the robustness characteristics of the ore domain specifically includes: A graph structure is established based on historical ore data, which includes historical logging data, mineral assemblage annotation data, and laboratory data. The graph node representation is initialized using the robust features of the ore domain. Mineral nodes, structural nodes, and grade nodes are set up. The mineral nodes include galena nodes, sphalerite nodes, pyrite nodes, and gangue nodes. The structural nodes include vein boundary nodes, disseminated mineralization nodes, and fracture nodes. The grade nodes include lead grade nodes, zinc grade nodes, and grade level nodes. Co-occurrence edges are established based on the frequency of co-occurrence of historical mineral assemblages; spatial adjacency edges are established based on the adjacency edges in the mineral candidate zoning map; lead grade contribution edges are established based on the correlation coefficient between the area proportion of galena candidate areas and lead grade test data; and zinc grade contribution edges are established based on the correlation coefficient between the area proportion of sphalerite candidate areas and zinc grade test data. Confusion edges are established based on the frequency of occurrence of easily confused markers for pyrite, and suppression edges are established based on the frequency of misjudgments of high-grade areas corresponding to gangue candidate areas; Edge weights are assigned to coexisting edges, spatially adjacent edges, lead grade contribution edges, zinc grade contribution edges, easily confused edges, and suppression edges to generate a lead-zinc mineral coexistence constraint diagram. The enhanced ore image, the partitioned sampling offset field, and the lead-zinc mineral coexistence constraint map are associated according to the sample number and image number to form a classification prediction input package.

[0013] Optionally, configuring the DCNv3 deformable convolutional unit as a partition-guided sampling unit and generating reliable mineral visual features through partition confidence gating specifically includes: An enhanced ore image is input into the image embedding layer, and initial ore visual features are generated through convolutional embedding. The lead-zinc ore classification and prediction network uses InternImage as its backbone, and configures the DCNv3 deformable convolutional units in the backbone network as partition-guided sampling units. The partition-guided sampling unit corrects the basic sampling position of the convolution sampling kernel in the DCNv3 deformable convolution unit to the partition-guided sampling position based on the two-dimensional offset in the partition sampling offset field. When the partition guidance sampling position does not fall on an integer pixel point, the corresponding sampling features are extracted by bilinear interpolation, and the sampling features are weighted and aggregated according to the sampling modulation weight in the partition sampling offset field to generate partition guidance sampling features. Retain independent feature channels for sampling points with easily confused sampling markers; The partition confidence gating layer generates partition gating coefficients based on metallic luster confidence, wet reflection confidence, overexposure reflection confidence, and texture preservation confidence. The partition gating coefficients are then multiplied position by position with the partition guided sampling features to generate reliable mineral visual features.

[0014] Optionally, the output of lead-zinc ore image classification results, mineral assemblage probabilities, lead grade prediction results, and zinc grade prediction results after being constrained by the lead-zinc mineral coexistence constraint map and grade consistency prediction specifically includes: The visual features of credible ore are mapped to mineral nodes according to the candidate area location, the corresponding features of vein boundary area and disseminated mineralization area are mapped to structure nodes, and the visual features of credible ore are mapped to lead grade node and zinc grade node through grade node mapping layer to generate graph node feature matrix. Based on the coexisting edges, spatially adjacent edges, lead grade contribution edges, zinc grade contribution edges, easily confused edges, and suppression edges in the lead-zinc mineral coexistence constraint diagram, generate an edge index matrix and an edge weight matrix. Graph message passing is performed based on the edge index matrix and edge weight matrix to obtain the updated mineral node features, structural node features, and grade node features; The updated mineral node features, structural node features, and grade node features are backmapped to the corresponding spatial locations of the credible ore visual features, and the ore features are generated by channel splicing and 1×1 convolution fusion. Input the map-constrained ore features into the classification prediction branch, mineral assemblage prediction branch, lead grade prediction branch and zinc grade prediction branch respectively to generate lead-zinc ore image classification results, mineral assemblage probability, lead grade prediction results and zinc grade prediction results; The grade consistency prediction head generates low contribution verification items for galena, low contribution verification items for sphalerite, and easily confused verification items for pyrite based on mineral assemblage probability, lead grade prediction results, and zinc grade prediction results.

[0015] Optionally, step seven specifically includes: The lead grade prediction deviation is generated based on the difference between the lead grade prediction results and the lead grade test data; the zinc grade prediction deviation is generated based on the difference between the zinc grade prediction results and the zinc grade test data. The laboratory feedback deviation is generated based on the lead grade prediction deviation, zinc grade prediction deviation, mineral combination probability, grade consistency verification items, and the image cataloging results after verification. The laboratory feedback deviation includes grade prediction deviation, mineral classification deviation, image quality deviation, low contribution deviation of galena, low contribution deviation of sphalerite, and pyrite confusion deviation. The credibility markers in the multi-source lead-zinc ore data frames are updated based on image quality deviations, and the candidate region credibility, candidate region node attributes, and candidate region edge attributes in the mineral candidate partitioning map are updated based on mineral classification deviations. The lead grade contribution edge weights between galena nodes and lead grade nodes are updated based on the low contribution deviation of galena; the zinc grade contribution edge weights between sphalerite nodes and zinc grade nodes are updated based on the low contribution deviation of sphalerite; and the easily confused edge weights between pyrite nodes and galena nodes are updated based on the confusion deviation of pyrite. Update the lead grade prediction branch, zinc grade prediction branch, and grade consistency prediction head in the lead-zinc ore classification prediction network based on the grade prediction deviation. Based on the types of deviations in laboratory feedback, feedback update records are established. These records are used to establish the corresponding update relationships between multi-source lead-zinc ore data frames, mineral candidate zoning maps, lead-zinc mineral coexistence constraint maps, and lead-zinc ore classification and prediction networks.

[0016] The beneficial effects of this invention are: (1) This invention decouples the illumination-reflection-texture of lead-zinc ore images to generate enhanced ore images, reflection confidence maps and texture enhancement maps. It addresses the problem of unclear mineral boundaries and difficulty in distinguishing between real mineral luster and non-mineral reflection caused by the mixture of wet reflection, overexposure reflection and metallic luster. This allows subsequent candidate segmentation to be processed based on the reliable reflection area and texture boundary, thereby improving the stability of the front-end feature expression of lead-zinc ore images. (2) The present invention generates a mineral candidate partition map based on mineral candidate prompt information, and generates a test label correction weight by combining the consistency of test association data and image cataloging results. In view of the problem that existing methods rely only on image labels and have difficulty in handling the inconsistency between apparent mineralization and actual grade, the present invention incorporates candidate area nodes, candidate area edge attributes and test correction relationships into the model construction process, thereby reducing the interference of easily confused minerals such as pyrite and galena on classification and grade prediction. (3) This invention generates a partitioned sampling offset field based on the mineral candidate partition map, the reflectivity map and the test label correction weight, and configures the DCNv3 deformable convolutional unit as the partitioned guided sampling unit. At the same time, it introduces the lead-zinc mineral coexistence constraint map and grade consistency prediction. In view of the problem that conventional classification models are difficult to constrain the sampling position, mineral coexistence relationship and lead-zinc grade output consistency at the same time, it realizes the collaborative generation of classification results, mineral combination probability and lead-zinc grade prediction results, and provides a model basis for intelligent ore sorting, grade prediction and mineral processing front-end quality control. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition, as proposed in this invention. Figure 2 This is a schematic diagram of a method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition proposed in this invention. Figure 3 This is a framework diagram of the lead-zinc ore classification and prediction network in the lead-zinc ore image classification and prediction model construction method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3 A method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition includes the following steps: Step 1: Collect lead-zinc ore image data, sample correlation data, and laboratory correlation data. Generate multi-source lead-zinc ore data frames through consistency verification, image-laboratory binding, and confidence marking. Step 2: Decouple the illumination-reflection-texture of the lead-zinc ore images in the multi-source lead-zinc ore data frames to generate enhanced ore images, reflection confidence maps, and texture enhancement maps; Step 3: Generate mineral candidate hints based on the enhanced ore image, reflection confidence map, and texture enhancement map; generate a mineral candidate partition map through hint-based candidate segmentation. Step 4: Generate robust features of ore domains based on the mineral candidate zoning map, and generate test label correction weights based on the consistency between the test correlation data and the image cataloging results; Step 5: Generate a partitioned sampling offset field based on the mineral candidate partition map, reflectance confidence map, and laboratory label correction weights. Construct a lead-zinc mineral coexistence constraint map based on the robust features of the ore domain. The enhanced ore image, partitioned sampling offset field, and lead-zinc mineral coexistence constraint map form a classification prediction input package. Step 6: Input the classification prediction input package into the lead-zinc ore classification prediction network. The lead-zinc ore classification prediction network uses InternImage as the backbone and configures the DCNv3 deformable convolutional unit as a partitioned guided sampling unit. After partitioned confidence gating, lead-zinc mineral coexistence constraint graph constraint and grade consistency prediction, it outputs the lead-zinc ore image classification result, mineral combination probability, lead grade prediction result and zinc grade prediction result. Step 7: Update the multi-source lead-zinc ore data frame, mineral candidate zoning map, lead-zinc mineral coexistence constraint map, and lead-zinc ore classification prediction network based on the test feedback deviation.

[0020] In this embodiment, step one specifically includes: Read lead-zinc ore image data, sample correlation data, and laboratory correlation data; extract image number, sample number, and image acquisition parameters from lead-zinc ore image data; extract mine area number, sampling location, and image logging results from sample correlation data; extract laboratory number, lead grade laboratory data, and zinc grade laboratory data from laboratory correlation data. A unique sample number is assigned to each lead-zinc ore sample to be processed, and this unique sample number is simultaneously configured on the sample label, image acquisition record, and submission record. Image data of lead-zinc ore is acquired at the image acquisition station, which is equipped with a diffuse reflection light source, a standard background board, and a millimeter-level scale, and the image acquisition parameters are recorded simultaneously. Perform consistency checks on lead-zinc ore image data, sample association data, and test association data; when the same unique sample number corresponds to multiple test numbers, determine the target test number according to the batch submitted for testing and the test completion time; when the test association data lacks lead grade test data or zinc grade test data, configure a test missing marker. Image quality assessment is performed on lead-zinc ore image data. Image sharpness is calculated using Laplacian variance, the proportion of saturated pixels in the brightness channel is statistically analyzed, and millimeter-level scales and standard background boards are identified. When the image sharpness is lower than a preset sharpness threshold, a blur mark is configured. When the proportion of saturated pixels exceeds a preset saturation ratio threshold, an overexposure mark is configured. When the scale is not identified or the proportion of background board occlusion exceeds a preset occlusion ratio threshold, a scale anomaly mark or a background residue mark is configured. A confidence marker is generated based on the consistency verification results and image quality evaluation results. When more than two anomalous markers exist for the same lead-zinc ore image data, the lead-zinc ore image data is marked as low-confidence image data. An image-test binding relationship is established between the lead-zinc ore image data that has passed the consistency verification and the target test number. The lead-zinc ore image data, sample association data, test association data and confidence marker that have completed the image-test binding are encapsulated into a multi-source lead-zinc ore data frame. In this embodiment, step two specifically includes: Read lead-zinc ore image data and confidence markers from multi-source lead-zinc ore data frames; when lead-zinc ore image data is marked as low confidence image data, retain the lead-zinc ore image data and reduce its enhancement weight; when lead-zinc ore image data contains overexposure markers or background residue markers, use the corresponding area as the priority analysis area for reflective area identification. Color space conversion is performed on the lead-zinc ore image data, converting the lead-zinc ore image from RGB color space to HSV color space and Lab color space, and extracting the luminance channel, saturation channel and lightness channel; median filtering is performed on the luminance channel and background grayscale correction is performed on the lightness channel. The illumination components are estimated based on the corrected luminance channels. Guided filtering is used to smooth the luminance channels to obtain the illumination component map. The original luminance channels are divided by the illumination component map and normalized to obtain the reflection component map. The difference between the original luminance channels and the illumination component map is calculated to obtain the texture residual map. Reflective areas are identified based on the reflection component map, saturation channel, and texture residual map. When the reflection component value is greater than 0.85, the saturation value is less than 0.25, and the average texture residual value is less than 0.08, the corresponding area is marked as a wet reflective area and a wet reflective label is configured. When the brightness saturation pixel ratio is greater than 5% and the average local gradient value is less than 0.05, the corresponding area is marked as an overexposed reflective area. When the reflection component value is greater than 0.70, the average local gradient value is not less than 0.05, and the area boundary is continuous with the mineral grain boundary, the corresponding area is marked as a metallic luster area. A reflection confidence map is generated based on the metallic luster region, the wet reflection region, and the overexposed reflection region. The reflection confidence map includes the confidence of metallic luster, wet reflection, overexposed reflection, and texture preservation. Edge preservation enhancement is performed on the texture residual map, confidence suppression is performed on the wet reflection region and the overexposed reflection region, and brightness abrupt changes and boundary information are preserved in the metallic luster region. The illumination component map, reflection component map, and texture residual map are fused to generate an enhanced ore image, and the texture residual map after edge preservation enhancement is normalized to generate a texture enhancement map. In this embodiment, step three specifically includes: Read the enhanced ore image, reflection confidence map, and texture enhancement map; generate reflection screening results based on the metallic luster confidence area, wet reflection area, and overexposed reflection area in the reflection confidence map; generate texture boundary results based on the high gradient boundary, particle boundary, and striped texture in the texture enhancement map. Based on the reflection screening results, mineral candidate hints are generated; when the metallic luster confidence of a certain area is not less than 0.70 and the texture retention confidence is not less than 0.60, a metallic mineral candidate hint is generated; when the average brightness of the area is not less than 180, the average saturation is not greater than 0.25, and the particle boundaries are distributed in a blocky and continuous manner, a galena candidate hint is generated; when the average lightness of the area is not greater than 120 and the local texture fluctuation amplitude is not less than 0.20, a sphalerite candidate hint is generated; when the hue of the area falls within the range of yellow to yellowish-white, the metallic luster confidence is not less than 0.70, and the boundary sharpness is not less than 0.60, a pyrite candidate hint is generated. Based on the texture boundary results, vein boundary hint lines and vein boundary regions are generated; edge detection and line segment detection are performed on the texture enhancement map to extract continuous boundary line segments; when the difference in the mean color on both sides of the boundary line segment is not less than 15 and the included angle of the texture direction is not less than 30°, the boundary line segment is marked as a vein boundary hint line, and the neighborhood area of ​​the boundary line segment is configured as a vein boundary region; when the boundary line segment is located in a wet reflective area or an overexposed reflective area, the hint weight of the vein boundary hint line is reduced. A vein background hint box is generated based on the low metallic luster and low texture regions in the enhanced ore image; a vein background hint box is generated when the metallic luster confidence of a certain connected region is less than 0.30, the texture preservation confidence is less than 0.40, and the color change amplitude is less than the preset background change threshold; when the connected region is in contact with the standard background plate region, it is excluded from the vein background hint box. The mineral candidate hint points, vein boundary hint lines, and gangue background hint boxes are combined into mineral candidate hint information and input into the hint-based candidate segmentation module. The hint-based candidate segmentation module generates a foreground seed region based on the mineral candidate hint points, constrains the candidate region boundary based on the vein boundary hint lines, generates a background suppression region based on the gangue background hint boxes, and outputs galena candidate region, sphalerite candidate region, pyrite candidate region, gangue candidate region, oxide film candidate region, and disseminated mineralization region through region growth and boundary correction. Post-processing is performed on the candidate regions; isolated candidate regions with an area of ​​less than 50 pixels are deleted, and adjacent candidate regions of the same category with a boundary distance of less than 5 pixels are merged; when the overlap area between a candidate region and a wet reflective region or an overexposed reflective region is greater than 40%, the candidate region is marked as a low-confidence candidate region; when the overlap area between a candidate region and a metallic luster confidence region is greater than 50% and the texture preservation confidence is not less than 0.60, the candidate region is marked as a high-confidence candidate region. Each candidate region is used as a candidate region node, and the candidate region category, region area, region center, color mean, texture intensity, metallic luster confidence, candidate region confidence, and boundary curvature are extracted as candidate region node attributes. Adjacent edges are established when the boundary contact length between two candidate regions is not less than 10 pixels, texture continuous edges are established when the texture direction angle between two candidate regions is less than 20°, and vein association edges are established when the vein boundary indicator line passes through two candidate regions. The candidate region edge attributes include boundary contact length, texture direction angle, and vein association marker. A mineral candidate partition map is generated based on the candidate region nodes, candidate region node attributes, adjacent edges, texture continuous edges, vein association edges, and candidate region edge attributes. In this embodiment, step four specifically includes: Read the mineral candidate zoning map and multi-source lead-zinc ore data frame; extract candidate area node attributes, candidate area edge attributes and candidate area confidence from the mineral candidate zoning map; read the image logging results, lead grade test data and zinc grade test data from the multi-source lead-zinc ore data frame; Construct self-supervised ore characterization sample pairs based on candidate region node attributes; combine candidate regions with different perspectives or different cropping scales under the same sample number into positive sample pairs; combine candidate regions with similar image logging results but lead grade or zinc grade test data that differ by more than 2 percentage points into grade-sensitive negative sample pairs; combine candidate regions with similar metallic luster confidence and categories of galena and pyrite candidate regions into easily confused negative sample pairs; Perform masking and reconstruction processing on the mineral candidate zoning map; select no more than 20% of candidate area nodes according to the candidate area ratio from high to low for masking, and input the masked candidate area node attributes, candidate area edge attributes and unmasked candidate area nodes into the candidate area encoder; the candidate area encoder outputs the candidate area representation vector, and reconstructs the category, texture intensity and metallic luster confidence of the masked candidate area based on the adjacent candidate area nodes and vein association edges; Positive sample pairs, grade-sensitive negative sample pairs, and easily confused negative sample pairs are input into the ore self-supervised representation module. The ore self-supervised representation module includes an image feature encoding layer, a candidate region feature encoding layer, and a feature projection layer. The image feature encoding layer extracts the whole image features of the enhanced ore image, the candidate region feature encoding layer extracts the node features of the mineral candidate partition map, and the feature projection layer maps the whole image features and node features to the same feature space. The feature distance of the sample pairs is adjusted within the same feature space, and the candidate region representation vector is constrained according to the occlusion reconstruction error. When the occlusion reconstruction error is less than 0.05 for three consecutive training cycles, the robust features of the ore domain are output. The grade of the sample is determined based on the lead and zinc grade test data. When the image logging result indicates strong mineralization and both the lead and zinc grade test data are below 1.0%, the corresponding sample is marked as a sample with high apparent mineralization. When the image logging result indicates weak mineralization and either the lead or zinc grade test data is not less than 3.0%, the corresponding sample is marked as a sample with concealed mineralization. When the area of ​​the galena candidate area is greater than 15% and the lead grade test data is less than 1.0%, a low contribution marker is generated for the galena candidate area. When the area of ​​the sphalerite candidate area is greater than 15% and the zinc grade test data is less than 1.0%, a low contribution marker is generated for the sphalerite candidate area. When the metallic luster confidence of the pyrite candidate area is not less than 0.75 and the lead grade test data is less than 1.0%, a pyrite easily confused marker is generated. Based on samples with high apparent mineralization, concealed mineralization, low contribution markers for galena candidate areas, low contribution markers for sphalerite candidate areas, and easily confused markers for pyrite, test label correction weights are generated. For samples with high apparent mineralization, the classification loss weight is reduced, with a weight ranging from 0.5 to 0.8. For samples with concealed mineralization, the grade prediction loss weight is increased, with a weight ranging from 1.2 to 1.8. For samples corresponding to easily confused markers for pyrite, the coexistence consistency constraint weight is increased, with a weight ranging from 1.3 to 2.0. The ore domain robustness characteristics and test label correction weights are then associated with the corresponding sample numbers. In this embodiment, step five specifically includes: Read the mineral candidate zoning map, reflectivity map, and laboratory label correction weights; project the candidate area nodes in the mineral candidate zoning map onto the preset sampling grid, and record the candidate area center coordinates, candidate area boundary coordinates, and vein direction vector; the vein direction vector is generated according to the extension direction of the vein boundary prompt line; set low weight labels for low confidence candidate areas and high weight labels for high confidence candidate areas. Calculate the distance vector between each sampling position in the preset sampling grid and the center coordinates of the nearest candidate area, and generate the center attraction offset; when the nearest candidate area is a galena candidate area, sphalerite candidate area, or disseminated mineralization area, correct the sampling position towards the center of the corresponding candidate area; when the nearest candidate area is a gangue candidate area, reduce the center attraction offset. Extract the boundary coordinates and boundary curvature of the candidate area. For the ore vein boundary area and the candidate area with a boundary curvature greater than 0.3, generate a boundary following offset along the tangent direction of the candidate area boundary. When the overlap area between the candidate area boundary and the wet reflective area or the overexposed reflective area is greater than 40%, suppress the corresponding boundary following offset. Extract vein direction vectors, texture continuous edges, and vein associated edges to generate directional constraint offsets; when a texture continuous edge connects a galena candidate area and a sphalerite candidate area, adjust the directional constraint offset upwards; when a texture continuous edge passes through a gangue candidate area, adjust the directional constraint offset downwards. Sampling modulation weights are generated based on the reflectance confidence map and the test label correction weights; the sampling modulation weights are increased for regions with a metallic luster confidence of not less than 0.70 and a texture preservation confidence of not less than 0.60; the sampling modulation weights are increased for high-confidence candidate regions; the sampling modulation weights are decreased for regions with a wet reflectance confidence of not less than 0.60 or an overexposed reflectance confidence of not less than 0.60; the sampling modulation weights are maintained for weak texture candidate regions corresponding to concealed mineralization samples; and candidate regions corresponding to easily confused pyrite markers are marked as easily confused sampling regions. The center attraction offset, boundary following offset, orientation constraint offset, and sampling modulation weight are aligned and meshed to generate a partitioned sampling offset field; each sampling position in the partitioned sampling offset field is configured with a two-dimensional offset, sampling modulation weight, and easily confused sampling marker. Constructing a lead-zinc mineral coexistence constraint diagram based on the robustness characteristics of ore domains involves: reading historical ore data, including historical logging data, mineral assemblage annotation data, and laboratory data; establishing a graph structure based on the historical ore data; initializing graph node representations based on the robustness characteristics of ore domains; setting mineral nodes, structural nodes, and grade nodes in the lead-zinc mineral coexistence constraint diagram; mineral nodes including galena nodes, sphalerite nodes, pyrite nodes, and gangue nodes; structural nodes including vein boundary nodes, disseminated mineralization nodes, and fracture nodes; and grade nodes including lead grade nodes, zinc grade nodes, and grade level nodes. Statistically analyze the co-occurrence frequency of historical mineral assemblages to establish co-occurrence edges; establish spatial adjacency edges based on the adjacency edges in the mineral candidate zoning map; calculate the correlation coefficient between the area proportion of galena candidate areas and lead grade test data to establish lead grade contribution edges; calculate the correlation coefficient between the area proportion of sphalerite candidate areas and zinc grade test data to establish zinc grade contribution edges; statistically analyze the occurrence frequency of easily confused markers for pyrite to establish easily confused edges; when the area proportion of gangue candidate areas in historical samples is greater than the preset gangue proportion threshold and the image logging result is low mineralization, and the lead grade test data or zinc grade test data is not less than 3.0%, it is recorded as a high-grade misjudgment, and the number of high-grade misjudgments corresponding to gangue candidate areas is statistically analyzed to establish suppression edges; Each edge in the lead-zinc mineral co-occurrence constraint map is assigned an edge weight; the normalized historical co-occurrence frequency is used as the co-occurrence edge weight; the normalized candidate area boundary contact length is used as the spatial adjacency edge weight; the correlation coefficient between the normalized candidate area area ratio and the corresponding grade test data is used as the grade contribution edge weight; the normalized frequency of occurrence of easily confused pyrite markers is used as the easily confused edge weight; the normalized frequency of misjudgment of high-grade gangue candidate areas is used as the suppression edge weight; the enhanced ore image, the partitioned sampling offset field, and the lead-zinc mineral co-occurrence constraint map are associated according to the sample number and image number to form a classification prediction input package; In this embodiment, step six specifically includes: Read the classification prediction input package, parse the enhanced ore image, partition sampling offset field and lead-zinc mineral coexistence constraint map; perform size normalization and channel normalization on the enhanced ore image, input the processed enhanced ore image into the image embedding layer, and generate the initial ore visual features through convolution embedding; A lead-zinc ore classification and prediction network was constructed, using InternImage as the backbone. The DCNv3 deformable convolutional units in the backbone network were configured as partitioned guided sampling units. The partitioned guided sampling units read the two-dimensional offset, sampling modulation weights, and easily confused sampling markers from the partitioned sampling offset field. First, the current sampling position and the base sampling position of the convolutional sampling kernel were determined. Then, the two-dimensional offset was superimposed on the base sampling position to obtain the corrected sampling position. When the corrected sampling position did not fall on an integer pixel, the feature values ​​of adjacent pixels were read using bilinear interpolation, and the sampling modulation weights were used as the weights for feature aggregation of that sampling point to generate partitioned guided sampling features. Independent feature channels were retained for sampling points with easily confused sampling markers. A partition confidence gating layer is set up to perform gating processing on the partition-guided sampling features. The partition confidence gating layer reads the confidence of metallic luster, wet reflection, overexposure reflection, and texture preservation. The confidence of the above confidence is concatenated and then input into a linear transformation layer. Then, a sigmoid activation process is used to generate a partition gating coefficient with a value range of 0 to 1. When the confidence of metallic luster is not less than 0.70 and the confidence of texture preservation is not less than 0.60, the partition gating coefficient is increased; when the confidence of wet reflection is not less than 0.60 or the confidence of overexposure reflection is not less than 0.60, the partition gating coefficient is decreased. The partition gating coefficient is multiplied with the partition-guided sampling features position by position to generate the reliable mineral visual features. Read the mineral nodes, structure nodes, grade nodes, and edge weights from the lead-zinc mineral coexistence constraint graph; map the visual features of the credible ore to the mineral nodes according to the candidate area location, and map the corresponding features of the vein boundary area and disseminated mineralization area to the structure nodes; map the visual features of the credible ore to the lead grade node and zinc grade node through the grade node mapping layer to generate grade node features; generate a graph node feature matrix based on the mineral nodes, structure nodes, and grade nodes; Based on the co-existing edges, spatial adjacency edges, lead grade contribution edges, zinc grade contribution edges, easily confused edges, and suppression edges in the lead-zinc mineral co-existence constraint diagram, generate an edge index matrix and an edge weight matrix; for each node, calculate the node association score between the node and its adjacent nodes; perform Softmax normalization on the association scores of all adjacent nodes connected to the same node to generate the weights of each adjacent node participating in graph message passing; The features of adjacent nodes are weighted and aggregated according to the graph message passing weights and then fused with the features of the current node itself. The fused node features are then subjected to ReLU nonlinear activation processing to obtain updated mineral node features, structural node features, and grade node features. For pyrite nodes and galena nodes connected by easily confused edges, the degree of difference between their node features is calculated and used as a distinguishing correction for easily confused edges in node feature updates. The updated mineral node features, structural node features, and grade node features are backmapped to the corresponding spatial locations of the credible ore visual features, and then fused using channel stitching and 1×1 convolution to generate map-constrained ore features. These map-constrained ore features are input into the classification prediction branch, and lead-zinc ore image classification results are generated through fully connected layers and Softmax normalization. The map-constrained ore features are also input into the mineral combination prediction branch, and mineral combination probabilities are generated through fully connected layers and Sigmoid activation. Finally, the map-constrained ore features are input into the lead grade prediction branch and the zinc grade prediction branch, respectively, and lead grade prediction results and zinc grade prediction results are generated through fully connected regression layers. A grade consistency prediction head is set up to perform consistency checks on the mineral assemblage probability, lead grade prediction results, and zinc grade prediction results. When the probability of a galena candidate area is higher than 0.70 and the lead grade prediction result is lower than 1.0%, a low contribution check item for galena is generated. When the probability of a sphalerite candidate area is higher than 0.70 and the zinc grade prediction result is lower than 1.0%, a low contribution check item for sphalerite is generated. When the probability of a pyrite candidate area is higher than 0.70 and the lead grade prediction result is higher than 1.0%, a pyrite easily confused check item is generated. When training the lead-zinc ore classification and prediction network, the image classification loss, mineral assemblage prediction loss, lead grade prediction loss, zinc grade prediction loss, and lead-zinc mineral co-occurrence constraint are calculated separately. Figure 1 The consistency loss and the laboratory label correction loss are weighted and summed according to the preset loss weights to obtain the joint training loss. After training, the lead-zinc ore classification and prediction network generates lead-zinc ore image classification results, mineral combination probability, lead grade prediction results, zinc grade prediction results, and grade consistency verification items. In this embodiment, step seven specifically includes: Read the lead-zinc ore image classification results, mineral combination probability, lead grade prediction results, zinc grade prediction results, and grade consistency verification items output by the lead-zinc ore classification and prediction network; read the verified image cataloging results, lead grade test data, and zinc grade test data corresponding to the same sample number; calculate the difference between the lead grade prediction results and the lead grade test data to generate the lead grade prediction deviation; calculate the difference between the zinc grade prediction results and the zinc grade test data to generate the zinc grade prediction deviation. Based on the lead grade prediction deviation, zinc grade prediction deviation, mineral combination probability, grade consistency verification items, and the reviewed image cataloging results, a laboratory feedback deviation is generated. A grade prediction deviation is generated when the lead grade prediction deviation or zinc grade prediction deviation is greater than 0.5 percentage points. A mineral classification deviation is generated when the main mineral category corresponding to the mineral combination probability is inconsistent with the reviewed image cataloging results. An image quality deviation is generated when the overlap area between the candidate area corresponding to the laboratory feedback deviation and the overexposed area, wet reflective area, or residual background area is greater than 40%. A galena low contribution deviation is generated when the galena candidate area probability is higher than 0.70 and the lead grade test data is lower than 1.0%. A sphalerite low contribution deviation is generated when the sphalerite candidate area probability is higher than 0.70 and the zinc grade test data is lower than 1.0%. A pyrite confusion deviation is generated when the pyrite candidate area probability is higher than 0.70 and the lead grade prediction result is more than 1.0 percentage point higher than the lead grade test data. Update multi-source lead-zinc ore data frames based on image quality deviations; when an image quality deviation occurs, configure the corresponding lead-zinc ore image data as low-confidence image data and mark the source of the deviation as image quality deviation; when the proportion of low-confidence image data in the same image acquisition batch is greater than 20%, configure the image acquisition batch as a verification acquisition batch; when lead grade test data or zinc grade test data is inconsistent with the image logging results continuously, configure the corresponding sample as a logging-test inconsistent sample. The mineral candidate zoning map is updated based on mineral classification deviations. When the candidate region corresponding to the mineral classification deviation is inconsistent with the image cataloging result after verification, the confidence of the corresponding candidate region is lowered. When adjacent candidate regions belong to the same mineral region after verification, the corresponding candidate region nodes are merged, and the region area, region center, color mean, and texture intensity of the merged candidate region are recalculated. When a single candidate region contains multiple mineral regions after verification, the candidate region nodes are split according to the vein boundary hint line and texture boundary results, and the candidate region node attributes and candidate region edge attributes are regenerated. The lead-zinc mineral coexistence constraint diagram is updated based on pyrite confusion bias, galena low contribution bias, and sphalerite low contribution bias. When pyrite confusion bias occurs, the weight of the easily confused edge between pyrite and galena nodes is increased. When galena low contribution bias occurs, the weight of the lead grade contribution edge between galena and lead grade nodes is decreased. When sphalerite low contribution bias occurs, the weight of the zinc grade contribution edge between sphalerite and zinc grade nodes is decreased. When the number of misjudgments of high grade corresponding to gangue candidate areas increases, the weight of the suppression edge between gangue nodes and grade level nodes is increased. The lead-zinc ore classification prediction network is updated based on grade prediction deviation. When the lead grade prediction deviation is greater than 0.5 percentage points for three consecutive times, the training weight of the lead grade prediction branch is increased. When the zinc grade prediction deviation is greater than 0.5 percentage points for three consecutive times, the training weight of the zinc grade prediction branch is increased. When mineral classification deviations are concentrated between pyrite and galena candidate areas, the training weight of samples corresponding to easily confused sampling areas is increased, while retaining the independent feature channels of pyrite and galena candidate areas. When a deviation sample corresponds to low-confidence image data, the training weight of that deviation sample in the classification prediction branch is decreased. A calibration training sample set is generated based on the deviations in laboratory feedback. The calibration training sample set includes normal samples, low-confidence image samples, catalog-test inconsistent samples, low-contribution galena samples, low-contribution sphalerite samples, and pyrite confused samples. In each round of calibration training, the proportion of deviation samples is set to 30% to 50%, and the proportion of normal samples is set to 50% to 70%. The classification loss weight is reduced for low-confidence image samples, the laboratory label correction loss weight is increased for catalog-test inconsistent samples, and the graph constraint loss weight corresponding to easily confused edges is increased for pyrite confused samples. Incremental training was performed on the lead-zinc ore classification and prediction network using a calibrated training sample set. During incremental training, the shallow parameters of the image embedding layer remained unchanged, while the partitioned guided sampling unit, partitioned confidence gating layer, lead-zinc mineral coexistence constraint map constraint layer, and grade consistency prediction head were updated. When the lead grade prediction deviation and zinc grade prediction deviation after calibrated training were both no greater than 0.5 percentage points, and the mineral classification deviation rate was less than 5%, the updated lead-zinc ore classification and prediction network was saved. A feedback update record is established based on the type of laboratory feedback deviation. The feedback update record includes sample number, image number, laboratory number, type of laboratory feedback deviation, corresponding update object, parameters before update, parameters after update, and correction training results. The feedback update record completes the corresponding update relationship between multi-source lead-zinc ore data frames, mineral candidate partition map, lead-zinc mineral coexistence constraint map, and lead-zinc ore classification prediction network. In this embodiment, each preset threshold is determined based on historical lead-zinc ore samples, manual verification results, laboratory data distribution, and model validation results. Taking the threshold for wet reflective areas as an example, lead-zinc ore image samples that have undergone manual verification are selected, and the distribution of the reflection component value, saturation value, and mean texture residual value of manually confirmed wet reflective areas is statistically analyzed. The criteria for determining wet reflective areas are a reflection component value greater than 0.85, a saturation value less than 0.25, and a mean texture residual value less than 0.08. The other thresholds are determined in the same way based on the corresponding sample distribution, manual verification results, and validation set results.

[0021] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the scenario of ore image recording before crushing and grade prediction before beneficiation in a lead-zinc mine beneficiation plant.

[0022] In this scenario, ore samples were collected from underground sampling points and conveyor belt sampling points. A total of 1,800 lead-zinc ore images were collected, along with 1,800 corresponding sample-related data and 600 sets of laboratory-related data. Each set of samples included lead grade and zinc grade test data.

[0023] The image acquisition station was equipped with a diffuse light source, a standard background board, and a millimeter-scale ruler. Images were uniformly adjusted to 1024×1024 pixels, with a training set, validation set, and test set ratio of 8:1:1. The test set included samples of galena, sphalerite, pyrite, gangue, oxide film, and disseminated mineralization areas. Among the test samples, samples with mixed wet reflectivity, overexposed reflectivity, and metallic luster accounted for 18.6%.

[0024] In practice, the process described in the claims generates multi-source lead-zinc ore data frames, enhanced ore images, reflectivity maps, texture enhancement maps, mineral candidate partition maps, ore domain robust features, laboratory label correction weights, partition sampling offset fields, and lead-zinc mineral co-occurrence constraint maps. The classification prediction input package is then input into the lead-zinc ore classification prediction network.

[0025] In the comparative verification, Method A uses a lightweight CNN whole-image classification model, which generates image classification results based solely on lead-zinc ore images, and is mainly used to verify the rapid coarse screening effect; Method B uses a mineral identification model that combines candidate segmentation and secondary recognition, which first generates mineral candidate regions and then performs category recognition, and is mainly used to verify the candidate region segmentation capability; Method C uses the unmodified InternImage model, which takes enhanced ore images as input and directly outputs classification and grade prediction results, and is mainly used to verify the basic recognition capability of strong backbone networks; Method D is the method of this invention.

[0026] Table 1. Experimental Data on Image Classification and Grade Prediction of Lead-Zinc Ore

[0027] As shown in Table 1, different methods have different advantages in terms of metrics. Method A has the fastest processing speed with a single-image inference time of 24ms, making it suitable for rapid coarse screening. However, its mineral candidate partition mIoU, mineral assemblage F1, and grade prediction error are not superior, indicating that its ability to represent complex textures and multi-mineral coexisting samples is limited. Method B has the best candidate partition performance among the compared methods with a mineral candidate partition mIoU of 81.2%. Its misclassification rate for strongly reflective samples is also lower than that of Method C, indicating that it has a certain advantage in mineral region boundary delineation. However, its single-image inference time reaches 108ms, and its image classification accuracy and grade prediction error are still limited. Method C has an image classification accuracy of 91.3%, which is higher than that of Methods A and B. Its single-image inference time is 56ms, indicating that its overall image feature extraction ability and inference efficiency are relatively balanced. However, its misclassification rate for strongly reflective samples is 13.2%, and it is still prone to confusion in areas with mixed wet reflectivity and metallic luster.

[0028] Method D corresponds to the method of this invention. Its mineral candidate partition mIoU reaches 84.6%, an improvement of 3.4 percentage points compared to Method B; the image classification accuracy reaches 93.8%, an improvement of 2.5 percentage points compared to Method C; and the mineral assemblage F1 reaches 90.4%, an improvement of 4.8 percentage points compared to Method C. The mean absolute error of lead grade decreased from 0.63 percentage points in Method C to 0.41 percentage points, and the mean absolute error of zinc grade decreased from 0.68 percentage points to 0.45 percentage points, indicating that the laboratory label correction weight, lead-zinc mineral co-occurrence constraint map, and grade consistency prediction can reduce the inconsistency between image classification results and laboratory grades. The misclassification rate of strongly reflective samples decreased from 11.6% in Method B and 13.2% in Method C to 7.1%, indicating that the reflectivity confidence map, partition sampling offset field, and partition confidence gating can reduce the interference of wet reflection and overexposure reflection in the identification of dacite and pyrite candidate areas. Although the single-sheet inference time of method D is higher than that of methods A and C, it is significantly lower than that of method B, and it achieves relatively balanced results in classification, mineral assemblage identification, grade prediction, and stability of highly reflective samples.

[0029] Based on the above implementation results, in actual lead-zinc ore image cataloging and grade prediction scenarios before selection, this invention can achieve image, sample and test data binding through multi-source lead-zinc ore data frames, guide DCNv3 deformable convolutional units to focus on candidate mineral regions through partitioned sampling offset fields, and reduce the inconsistency between classification results and test grades through lead-zinc mineral coexistence constraint maps and grade consistency prediction. Thus, while maintaining acceptable inference time, it improves the overall reliability of ore classification, mineral assemblage identification and lead-zinc grade prediction.

[0030] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition, characterized in that, Includes the following steps: Step 1: Collect lead-zinc ore image data, sample correlation data, and laboratory correlation data. Generate multi-source lead-zinc ore data frames through consistency verification, image-laboratory binding, and confidence marking. Step 2: Decouple the illumination-reflection-texture of the lead-zinc ore images in the multi-source lead-zinc ore data frames to generate enhanced ore images, reflection confidence maps, and texture enhancement maps; Step 3: Generate mineral candidate hints based on the enhanced ore image, reflection confidence map, and texture enhancement map; generate a mineral candidate partition map through hint-based candidate segmentation. Step 4: Generate robust features of ore domains based on the mineral candidate zoning map, and generate test label correction weights based on the consistency between the test correlation data and the image cataloging results; Step 5: Generate a partitioned sampling offset field based on the mineral candidate partition map, reflectance confidence map, and laboratory label correction weights. Construct a lead-zinc mineral coexistence constraint map based on the robust features of the ore domain. The enhanced ore image, partitioned sampling offset field, and lead-zinc mineral coexistence constraint map form a classification prediction input package. Step 6: Input the classification prediction input package into the lead-zinc ore classification prediction network. The lead-zinc ore classification prediction network uses InternImage as the backbone and configures the DCNv3 deformable convolutional unit as a partitioned guided sampling unit. After partitioned confidence gating, lead-zinc mineral coexistence constraint graph constraint and grade consistency prediction, it outputs the lead-zinc ore image classification result, mineral combination probability, lead grade prediction result and zinc grade prediction result. Step 7: Update the multi-source lead-zinc ore data frame, mineral candidate zoning map, lead-zinc mineral coexistence constraint map, and lead-zinc ore classification prediction network based on the test feedback deviation.

2. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, Step one specifically includes: Collect image data, sample correlation data, and laboratory correlation data of lead-zinc ore, and extract image number, sample number, image acquisition parameters, image logging results, laboratory number, lead grade test data, and zinc grade test data respectively; A unique sample number is assigned to each lead-zinc ore sample to be processed, and this unique sample number is simultaneously configured on the sample label, image acquisition record, and submission record. Consistency checks are performed on lead-zinc ore image data, sample association data, and test association data. When the same unique sample number corresponds to multiple test numbers, the target test number is determined according to the batch of samples submitted and the test completion time. Image quality assessment is performed on lead-zinc ore image data, and confidence labels are generated based on image sharpness, saturation pixel ratio, scale bar recognition results, and background occlusion results. Image-test binding relationships are established between lead-zinc ore image data that has passed consistency verification and target test numbers. The lead-zinc ore image data, sample association data, test association data and confidence markers that have completed image-test binding are then encapsulated into a multi-source lead-zinc ore data frame.

3. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, Step two specifically includes: Perform color space conversion on lead-zinc ore images in multi-source lead-zinc ore data frames to extract luminance, saturation, and lightness channels; The illumination components are estimated based on the luminance channel, and an illumination component map is generated. The original luminance channel is divided by the illumination component map and normalized to generate a reflection component map. The original luminance channel is subtracted from the illumination component map to generate a texture residual map. Identify wet reflective areas, overexposed reflective areas, and metallic luster areas based on the reflectance component map, saturation channel, and texture residual map; A reflection confidence map is generated based on the wet reflection area, the overexposed reflection area, and the metallic luster area. The reflection confidence map includes metallic luster confidence, wet reflection confidence, overexposed reflection confidence, and texture retention confidence. The credibility of metallic luster is generated based on the intensity of the reflection component, the mean of the local gradient, and the continuity of the boundary. The credibility of wet reflection is generated based on high reflection intensity, low saturation, and low texture residual. The credibility of overexposed reflection is generated based on the proportion of saturated pixels and the proportion of low gradient region area. The credibility of texture preservation is generated based on the intensity of texture residual and the continuity of grain boundary. Edge-preserving enhancement is performed on the texture residual map to generate a texture enhancement map, and the illumination component map, reflection component map, and texture residual map are fused to generate an enhanced ore image.

4. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, Step three specifically includes: Reflection screening results are generated based on the metallic luster confidence region, wet reflection region and overexposed reflection region in the reflection confidence map, and texture boundary results are generated based on the high gradient boundary, grain boundary and striped texture in the texture enhancement map. Based on the reflection screening results, mineral candidate prompts are generated, including metallic mineral candidate prompts, galena candidate prompts, sphalerite candidate prompts, and pyrite candidate prompts; Based on the texture boundary results, vein boundary hint lines and vein boundary areas are generated, and vein background hint boxes are generated based on the low metallic luster areas and low texture areas in the enhanced ore image. The mineral candidate hint points, vein boundary hint lines, and gangue background hint boxes are combined into mineral candidate hint information. After hint-based candidate segmentation, galena candidate area, sphalerite candidate area, pyrite candidate area, gangue candidate area, oxide film candidate area, and disseminated mineralization area are generated. Post-processing is performed on each candidate region. Candidate regions that overlap with wet reflective regions or overexposed reflective regions are marked as low-confidence candidate regions, and candidate regions that overlap with metallic luster confidence regions and whose texture preservation confidence meets the preset texture preservation threshold are marked as high-confidence candidate regions. Each candidate region is used as a candidate region node, and the candidate region category, region area, region center, color mean, texture intensity, metallic luster confidence, candidate region confidence, and boundary curvature are extracted as candidate region node attributes. Adjacent edges are established based on the contact relationship of candidate area boundaries, continuous texture edges are established based on the texture direction relationship of candidate areas, and vein-related edges are established based on the crossing relationship of vein boundary hint lines. The boundary contact length, texture direction angle, and vein-related marker are used as candidate area edge attributes to generate a mineral candidate partition map.

5. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, Step four specifically includes: Based on the node attributes of candidate regions in the mineral candidate zoning map, ore self-supervised characterization sample pairs are constructed, including positive sample pairs, grade-sensitive negative sample pairs, and easily confused negative sample pairs. Candidate regions with the same sample number but different perspectives or different cropping scales are combined into positive sample pairs. Candidate regions with similar image logging results but lead or zinc grade test data that differ from the preset grade difference threshold are combined into grade-sensitive negative sample pairs. Candidate regions with similar metallic luster confidence and classified as galena and pyrite candidate regions are combined into easily confused negative sample pairs. Perform masking reconstruction processing on the mineral candidate partition map, and input the masked candidate region node attributes, candidate region edge attributes and unmasked candidate region nodes into the candidate region encoder to generate candidate region representation vectors; The ore self-supervised representation sample is input into the ore self-supervised representation module, and robust features of the ore domain are generated through the image feature coding layer, the candidate region feature coding layer and the feature projection layer. Based on lead grade test data, zinc grade test data and image logging results, samples with high apparent mineralization, concealed mineralization samples, low contribution markers for galena candidate areas, low contribution markers for sphalerite candidate areas and easily confused markers for pyrite are generated. Based on samples with high apparent mineralization, samples with concealed mineralization, low contribution markers for galena candidate areas, low contribution markers for sphalerite candidate areas, and easily confused markers for pyrite, a corrective weight for the analysis label is generated.

6. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, The process of generating a zonal sampling offset field based on the mineral candidate zoning map, the reflectivity map, and the test label correction weights specifically includes: Project the candidate area nodes in the mineral candidate zoning map onto the preset sampling grid, and record the center coordinates, boundary coordinates, and vein direction vector of the candidate area; The center attraction offset is generated based on the distance vector between each sampling position in the preset sampling grid and the coordinates of the center of the nearest candidate region. The boundary following offset is generated based on the candidate area boundary coordinates, candidate area boundary curvature, and vein boundary area. Generate directional constraint offsets based on the vein direction vector, texture continuity edges, and vein-related edges; Based on the reflectivity map and the test label correction weight, the sampling modulation weight is generated, and the candidate area corresponding to the easily confused pyrite label is marked as the easily confused sampling area. The center attraction offset, boundary following offset, direction constraint offset, and sampling modulation weight are aligned and meshed to generate a partitioned sampling offset field. Each sampling position in the partitioned sampling offset field is configured with a two-dimensional offset, sampling modulation weight, and easily confused sampling marker.

7. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, The construction of the lead-zinc mineral co-occurrence constraint map based on the robustness characteristics of the ore domain specifically includes: A graph structure is established based on historical ore data, which includes historical logging data, mineral assemblage annotation data, and laboratory data. The graph node representation is initialized using the robust features of the ore domain. Mineral nodes, structural nodes, and grade nodes are set up. The mineral nodes include galena nodes, sphalerite nodes, pyrite nodes, and gangue nodes. The structural nodes include vein boundary nodes, disseminated mineralization nodes, and fracture nodes. The grade nodes include lead grade nodes, zinc grade nodes, and grade level nodes. Co-occurrence edges are established based on the frequency of co-occurrence of historical mineral assemblages; spatial adjacency edges are established based on the adjacency edges in the mineral candidate zoning map; lead grade contribution edges are established based on the correlation coefficient between the area proportion of galena candidate areas and lead grade test data; and zinc grade contribution edges are established based on the correlation coefficient between the area proportion of sphalerite candidate areas and zinc grade test data. Confusion edges are established based on the frequency of occurrence of easily confused markers for pyrite, and suppression edges are established based on the frequency of misjudgments of high-grade areas corresponding to gangue candidate areas; Edge weights are assigned to coexisting edges, spatially adjacent edges, lead grade contribution edges, zinc grade contribution edges, easily confused edges, and suppression edges to generate a lead-zinc mineral coexistence constraint diagram. The enhanced ore image, the partitioned sampling offset field, and the lead-zinc mineral coexistence constraint map are associated according to the sample number and image number to form a classification prediction input package.

8. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, The step of configuring the DCNv3 deformable convolutional unit as a partition-guided sampling unit and generating reliable mineral visual features through partition confidence gating specifically includes: An enhanced ore image is input into the image embedding layer, and initial ore visual features are generated through convolutional embedding. The lead-zinc ore classification and prediction network uses InternImage as its backbone, and configures the DCNv3 deformable convolutional units in the backbone network as partition-guided sampling units. The partition-guided sampling unit corrects the basic sampling position of the convolution sampling kernel in the DCNv3 deformable convolution unit to the partition-guided sampling position based on the two-dimensional offset in the partition sampling offset field. When the partition guidance sampling position does not fall on an integer pixel point, the corresponding sampling features are extracted by bilinear interpolation, and the sampling features are weighted and aggregated according to the sampling modulation weight in the partition sampling offset field to generate partition guidance sampling features. Retain independent feature channels for sampling points with easily confused sampling markers; The partition confidence gating layer generates partition gating coefficients based on metallic luster confidence, wet reflection confidence, overexposure reflection confidence, and texture preservation confidence. The partition gating coefficients are then multiplied position by position with the partition guided sampling features to generate reliable mineral visual features.

9. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, The output of the lead-zinc ore image classification results, mineral assemblage probabilities, lead grade prediction results, and zinc grade prediction results, constrained by the lead-zinc mineral coexistence constraint map and grade consistency prediction, specifically includes: The visual features of credible ore are mapped to mineral nodes according to the candidate area location, the corresponding features of vein boundary area and disseminated mineralization area are mapped to structure nodes, and the visual features of credible ore are mapped to lead grade node and zinc grade node through grade node mapping layer to generate graph node feature matrix. Based on the coexisting edges, spatially adjacent edges, lead grade contribution edges, zinc grade contribution edges, easily confused edges, and suppression edges in the lead-zinc mineral coexistence constraint diagram, generate an edge index matrix and an edge weight matrix. Graph message passing is performed based on the edge index matrix and edge weight matrix to obtain the updated mineral node features, structural node features, and grade node features; The updated mineral node features, structural node features, and grade node features are backmapped to the corresponding spatial locations of the credible ore visual features, and the ore features are generated by channel splicing and 1×1 convolution fusion. Input the map-constrained ore features into the classification prediction branch, mineral assemblage prediction branch, lead grade prediction branch and zinc grade prediction branch respectively to generate lead-zinc ore image classification results, mineral assemblage probability, lead grade prediction results and zinc grade prediction results; The grade consistency prediction head generates low contribution verification items for galena, low contribution verification items for sphalerite, and easily confused verification items for pyrite based on mineral assemblage probability, lead grade prediction results, and zinc grade prediction results.

10. The method for constructing a lead-zinc ore image classification and prediction model based on pattern recognition according to claim 1, characterized in that, Step seven specifically includes: The lead grade prediction deviation is generated based on the difference between the lead grade prediction results and the lead grade test data; the zinc grade prediction deviation is generated based on the difference between the zinc grade prediction results and the zinc grade test data. The laboratory feedback deviation is generated based on the lead grade prediction deviation, zinc grade prediction deviation, mineral combination probability, grade consistency verification items, and the image cataloging results after verification. The laboratory feedback deviation includes grade prediction deviation, mineral classification deviation, image quality deviation, low contribution deviation of galena, low contribution deviation of sphalerite, and pyrite confusion deviation. The credibility markers in the multi-source lead-zinc ore data frames are updated based on image quality deviations, and the candidate region credibility, candidate region node attributes, and candidate region edge attributes in the mineral candidate partitioning map are updated based on mineral classification deviations. The lead grade contribution edge weights between galena nodes and lead grade nodes are updated based on the low contribution deviation of galena; the zinc grade contribution edge weights between sphalerite nodes and zinc grade nodes are updated based on the low contribution deviation of sphalerite; and the easily confused edge weights between pyrite nodes and galena nodes are updated based on the confusion deviation of pyrite. Update the lead grade prediction branch, zinc grade prediction branch, and grade consistency prediction head in the lead-zinc ore classification prediction network based on the grade prediction deviation. Based on the types of deviations in laboratory feedback, feedback update records are established. These records are used to establish the corresponding update relationships between multi-source lead-zinc ore data frames, mineral candidate zoning maps, lead-zinc mineral coexistence constraint maps, and lead-zinc ore classification and prediction networks.

Images