Iron-bearing mineral identification and grade estimation analysis method based on multi-scale adaptive fusion

Abstract

The application discloses a method for iron-bearing mineral recognition and grade estimation and analysis based on multi-scale adaptive fusion and belongs to the field of computer vision. The method comprises the following steps: introducing a space-depth conversion convolution module in a shallow layer of a backbone network to reserve full-pixel downsampling and losslessly extract sub-pixel level mineral characteristics; constructing an adaptive weighted feature splicing module in a neck network to realize accurate alignment and fusion of multi-scale heterogeneous semantics; connecting a three-dimensional saliency enhancement module in series at a multi-scale output end to suppress optical background noise and purify target area mineral signals; introducing an MPDIoU loss function in a segmentation decoupling head to realize accurate regression of irregularly distorted mineral boundaries; and based on a high-fidelity segmentation mask and a geological stereology principle, establishing an area-volume-mass measurement model to complete quantitative estimation of ore iron grade. The application solves the problems of large subjective error and low efficiency in the manual microscope inspection mode, and realizes high-precision segmentation and objective evaluation of iron-bearing minerals.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and geological information technology, and in particular to a method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion. Background Technology

[0002] Banded iron formations (BIFs) are one of the most important types of iron ore resources globally, widely distributed in Precambrian strata and accounting for the vast majority of global iron ore reserves. Their typical characteristic is the alternating banded distribution of iron-rich minerals (such as magnetite and hematite) and siliceous minerals. In the process of ore industrial evaluation, ore grade (i.e., iron content) is one of the core indicators for measuring the economic value and beneficiation feasibility of ore. Therefore, the accurate identification and quantitative segmentation of iron-rich minerals are of paramount importance in ore deposit genetic research, beneficiation process optimization, and grade assessment.

[0003] Currently, traditional mineral thin-section identification mainly relies on manual microscopic examination by experienced geologists. This method is not only time-consuming and labor-intensive but also prone to subjective bias, making it difficult to achieve large-scale, high-efficiency digital quantitative analysis. Ore grade determination primarily depends on chemical analysis methods or microscopic mineral quantitative analysis methods. Microscopic analysis, in particular, involves observing mineral morphology using a reflected light microscope and combining this with experience to identify different minerals and perform area statistics, thereby estimating the proportion of iron-bearing minerals. This method has long held a significant position in mineralogy research and mineral processing analysis. However, with technological advancements, these traditional methods can no longer meet the demands of current societal productivity. Therefore, transitioning from manual to intelligent methods for precise microscopic segmentation and grade estimation of iron-bearing minerals has become an urgent need in this field.

[0004] In recent years, the introduction of deep learning technology has greatly improved the automation level of rock and mineral image processing. However, when performing mineral segmentation on BIF microscopic thin section images, existing deep learning models (such as conventional YOLO series algorithms) still face the following significant technical bottlenecks: (1) Subpixel-level feature dilution and loss of fine mineral particles: In BIF samples under a microscope, some iron oxides (such as magnetite and hematite) are often distributed in fine particles or coexist with other gangue minerals in a complex manner, occupying a very small proportion of pixels. Traditional deep learning models often use strided convolution or pooling layers for downsampling in the backbone network, which will cause the pixel-level features of these small targets and their complex coexisting edge features to be severely diluted or even completely lost during feature transmission, directly affecting the segmentation accuracy. (2) Insufficient cross-layer fusion ability of multi-scale mineral embedding features: The iron-bearing minerals in BIF ores have a very large particle size range, including macroscopic continuous strip-shaped enrichment areas as well as extremely fine disseminated particles. Conventional feature pyramid structures (such as PANet) typically employ simple, equal-weighted splicing when dealing with targets with drastic scale changes. This results in low efficiency in the cross-scale transfer of low-level spatial details and high-level semantic information, failing to effectively couple global semantic information with local detail features. This can easily lead to semantic confusion in multi-scale minerals during feature fusion, resulting in false positives or false negatives. (3) Complex backgrounds and strong optical noise interference under the microscope: Due to limitations in the quality of thin-section sample preparation and the optical effects under polarizing microscopes, some minerals have similar reflectance to resin target models or associated gangue minerals, and the boundaries between different minerals are sometimes blurred. Furthermore, scratches, holes, and other optical noise are inevitably generated during sample preparation. Existing models struggle to effectively suppress these highly complex backgrounds during feature extraction, easily erroneously activating them as high-frequency response signals, leading to a decrease in the model's ability to focus on and capture target minerals in the target area. (4) Low boundary regression accuracy of irregularly bonded minerals: Mineral particles under the microscope often exhibit irregular polygons with distorted aspect ratios or are in a tightly bonded state. Traditional IoU loss functions (such as CIoU) suffer from limited geometric constraints (such as penalty degradation for fully contained states) and insufficient regression sensitivity of bounding boxes when dealing with such micro-objects. This results in severe pixel-level overflow or underfitting of the final output segmentation mask at complex phase boundaries or mineral edges.

[0005] Therefore, a deep learning model is needed that can preserve full-pixel features, enhance multi-scale fusion capabilities, effectively suppress optical background noise, and optimize boundary regression performance in order to achieve high-precision semantic segmentation of iron-bearing minerals under BIF microscope and subsequent automated grade estimation. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion. This method, targeting the embedding characteristics of minerals under a microscope, constructs a closed-loop feature processing system from non-destructive extraction of microscopic features to precise multi-scale fitting: First, information loss is prevented through full-pixel-preserving feature extraction. A spatial-depth transformation convolution (SPD-Conv) module is introduced into the shallow layers of the backbone network to replace the traditional strided convolutional layers. Utilizing the spatial-depth mapping mechanism, while achieving feature map downsampling, it ensures that sub-pixel-level fine mineral particles and microscopic symbiotic boundary features in the BIF image are fully preserved and transmitted to the deep network, providing a foundation for high-precision pixel-level segmentation. Secondly, in the neck network feature fusion stage, the conventional unidirectional or simple bidirectional paths were abandoned. An improved adaptive weighted feature concatenation (BiFPN-Concat) module was used to construct an efficient cross-scale connection network, enhancing multi-scale information fusion. This module adaptively aggregates semantic and spatial details from different levels of the backbone network through a learnable weighted feature fusion mechanism, effectively addressing the drastic scale changes in BIF samples, from micron-level fine grains to millimeter-level stripes, and eliminating identification gaps in multi-scale minerals. Subsequently, the parameterless attention mechanism SimAM module was innovatively connected in series to the output of the C3k2 module at each scale. This mechanism uses the analytical solution of the energy function to directly derive the neuron weights in the three-dimensional space and channel dimensions, adaptively purifying the texture signal of iron-bearing minerals without increasing any additional model parameters or computational burden, and effectively suppressing background interference such as microscope reflections, sample preparation scratches, and gangue minerals. Finally, accurate boundary regression optimization is achieved through the MPDIoU (Minimum Point Distance IoU) loss function; during the localization phase of model training, MPDIoU is used instead of the traditional loss function. Addressing the irregular morphology and dense adhesion of mineral particles under the microscope, this invention simplifies the similarity measurement process between the two polygons by minimizing the distance between the upper left and lower right corners of the predicted bounding box and the ground truth bounding box. This significantly enhances the model's regression dynamics and localization accuracy for deformed mineral targets, ultimately achieving pixel-level high-fidelity segmentation of mineral edge boundaries, which can be used for subsequent grade estimation analysis.

[0007] The technical solution of this invention is as follows:

[0008] A method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion includes the following steps:

[0009] S1: Obtain reflected light microscopic images of different BIF ore samples and perform pixel-level polygon contour annotation and standardization preprocessing to construct a high-fidelity microscopic image dataset of iron-bearing minerals.

[0010] S2: Construct the YOLO-Iron-Bearing-Minerals segmentation network, which consists of a backbone network, a neck network, and a segmentation decoupling head connected in sequence.

[0011] The backbone network performs initial feature encoding on the input high-fidelity microscopic image of iron-bearing minerals, and deploys a spatial-depth transformation convolution (SPD-Conv) module in the shallow and mid-level layers to perform lossless feature downsampling and deep feature extraction. This module uses spatial segmentation, channel stitching and nonlinear fusion operations to losslessly map and reorganize information in the spatial dimension to the channel dimension, and outputs a multi-scale fidelity feature map.

[0012] The neck network receives the multi-scale fidelity feature map, uses the adaptive weighted feature concatenation (BiFPN-Concat) module to perform adaptive spatial modulation and full-channel concatenation of bidirectional cross-layer feature streams from top to bottom and bottom to top, and connects a parameter-free three-dimensional saliency enhancement module (SimAM) at the end of the multi-scale output to suppress optical background noise under the microscope and output multi-scale high signal-to-noise ratio refined features;

[0013] The segmentation decoupling head receives the multi-scale high signal-to-noise ratio refined features and performs pixel classification and instance bounding box prediction, and decodes and outputs a high-precision pixel-level segmentation mask for iron-containing minerals.

[0014] S3: Using the high-fidelity microscopic image dataset of iron-bearing minerals constructed in step S1, supervised training is performed on the YOLO-Iron-Bearing-Minerals segmentation network constructed in step S2, and the network parameters are iteratively optimized until the model converges; in the backpropagation localization loss calculation stage of model training, the minimum point distance intersection-union ratio (MPDIoU) loss function is introduced to constrain the irregular crystal phase boundary.

[0015] S4: Input the microscopic image of the BIF ore to be identified into the trained segmentation network, extract the pixel-level segmentation mask of the target iron-bearing mineral, and establish an area-volume-mass measurement mathematical model based on the Delesse principle of geological stereochemistry to complete the automated quantitative estimation of the total iron grade of the ore slice.

[0016] Furthermore, in step S1, the specific process of constructing the high-fidelity microscopic image dataset of iron-bearing minerals includes:

[0017] S1.1: Collect original digital microscopic images of different BIF ore samples under a standard reflected light microscope, and introduce publicly available BIF ore reflected light microscopic images to expand the samples and form an expanded sample image set.

[0018] S1.2: Using the Labelme polygon annotation tool, pixel-level edge delineation is performed along the true phase boundaries of iron oxides in the images within the expanded sample image set to obtain the geometric contour information of irregular minerals and generate an initial annotation file in JSON format;

[0019] S1.3: By writing a format conversion script, the initial annotation file in JSON format is uniformly converted into standard binary mask labels suitable for training segmentation network models;

[0020] S1.4: Perform resolution downsampling and pixel value normalization on the images and corresponding standard binary mask labels in the expanded sample image set, and force the 180° full-angle random rotation enhancement and disable the image mixing operation to construct a high-fidelity microscopic image dataset of iron-bearing minerals containing a training set, a validation set and a test set.

[0021] Furthermore, in step S2, the specific process of the backbone network outputting multi-scale fidelity feature maps is as follows:

[0022] The backbone network processes the input raw microscopic images. (H and W are the height and width of the original microscopic image, respectively) Initial spatial features are encoded sequentially through the first and second standard strided convolutional layers to obtain P2 scale features downsampled by a factor of 4. ; Through the first Module refinement yields refined features ; then will The transition features are obtained by inputting the first spatial-depth transformation convolution module. , After refinement by the second C3k2 module, P3-scale fidelity features are output. ;Will The transition features are obtained by inputting the second-space-depth transformation convolution module. , After refinement by the third C3k2 module, P4-scale fidelity features are output. ;

[0023] right Utilizing the third standard strided convolutional layer and the fourth The module obtains early deep features of the backbone network without pooling compression. ;Will The deepest global semantic features are extracted using the Spatial Pyramid Pooling (SPPF) and Advanced Channel Attention (C2PSA) modules. The above process can be expressed by the following formula:

[0024]

[0025] in, This represents a standard strided convolution operation that includes a two-dimensional convolutional layer, a batch normalized layer, and a non-linear activation function (SiLU). This indicates a cross-stage local fusion network; Representation space-depth transformation convolution module;

[0026] The backbone network's final output , , and They are then transmitted together to the neck network.

[0027] Furthermore, the space-depth transformation convolution module losslessly maps and reassembles information from the spatial dimension to the channel dimension through spatial segmentation, channel concatenation, and nonlinear fusion operations. The specific implementation process is as follows:

[0028] For the input feature map (in For space dimensions, (Number of channels), using a preset scaling factor The spatial partitioning operator alternately samples and demixes the data along the spatial dimensions to obtain four data points of size 1. Sub-feature map , , ,

[0029]

[0030] Subsequently, the four sub-feature maps are fully stitched together along the channel dimension (dim=1), losslessly mapping the sub-pixel-level physical boundaries and reflection signals of the spatial domain to the channel axis, generating a size of... intermediate recombination feature map :

[0031]

[0032] Finally, in order to perform nonlinear semantic representation fusion and smooth mapping on the recombined high-dimensional channel features, the following steps are taken: The input is fed into a non-stepping standard convolutional layer with a stride of 1 for smooth dimensionality reduction, and the output transition feature tensor is adjusted for different channels. ,in This indicates the number of output channels of the spatial-depth transformation convolution module.

[0033] Furthermore, the neck network performs adaptive weighted fusion and concatenates a three-dimensional self-attention mechanism to refine the feature flow through a top-down cross-layer path. The specific processing flow is as follows:

[0034] Global deepest semantic features After spatial resolution stretching by the first spatial upsampling module, the features are compared with those at the P4 scale. The first adaptive weighted feature concatenation module performs cross-layer weighted fusion, resulting in a weighted concatenated feature tensor. This tensor is then processed by the fifth C3k2 module for feature extraction, outputting P4-scale intermediate fused features. ;

[0035] P4-scale intermediate fusion features After being upsampled by the second spatial module, its spatial resolution is magnified and compared with the P3 scale features. The second adaptive weighted feature concatenation module, which is a common input, performs cross-layer concatenation and outputs a weighted concatenation feature tensor. The sixth C3k2 module then refines the features and outputs intermediate fused features at the P3 scale. ;

[0036] Will The input is cascaded into the first parametric 3D saliency enhancement module for denoising, and the output is a refined feature with high signal-to-noise ratio at the P3 scale. The above process can be expressed by the following formula:

[0037]

[0038]

[0039]

[0040] in, For spatial upsampling module, This is an adaptive weighted feature concatenation module. This is a 3D saliency enhancement module.

[0041] Furthermore, the neck network performs adaptive weighted fusion through a bottom-up cross-layer path and refines the feature stream using a three-dimensional self-attention mechanism. The specific processing flow is as follows:

[0042] Refining features at the P3 scale Using the reverse approach as the starting point for bottom-up enhancement, the data is input into the fourth standard strided convolutional layer for spatial dimensionality reduction and then fused with intermediate features at the P4 scale. and P4 scale fidelity features All three inputs are fed into the third adaptive weighted feature concatenation module to perform three-way cross-layer concatenation, outputting a weighted concatenation feature tensor; then, the seventh C3k2 module refines the features, outputting the P4-scale intermediate fused features from the bottom-up path. ;Will The data is fed into the second parametric 3D saliency enhancement module to complete denoising, and outputs refined features with high signal-to-noise ratio at the P4 scale. ;

[0043] Refining features at the P4 scale After inputting the fifth standard strided convolutional layer for downsampling spatial dimensionality reduction, it is compared with the global deepest semantic features. and early deep features of the backbone network that are not pooled All three inputs are fed into the fourth adaptive weighted feature concatenation module to perform three-way cross-layer feature fusion, outputting a weighted concatenation feature tensor; then, the eighth C3k2 module performs feature extraction, outputting P5-scale intermediate fused features. ;Will The data is fed into the third parametric 3D saliency enhancement module for denoising and purification, outputting refined features with high signal-to-noise ratio at the P5 scale. The above process can be expressed by the following formula:

[0044]

[0045]

[0046]

[0047]

[0048] in, For spatial upsampling module, This is an adaptive weighted feature concatenation module. This is a 3D saliency enhancement module.

[0049] Furthermore, the adaptive weighted feature concatenation module outputs the weighted concatenation feature tensor in the following manner:

[0050] For any cross-layer fusion node in the neck network, there is a set of n spatially aligned heterogeneous input feature maps. The system provides each input feature map Initialize a learnable scalar weight. , Use the ReLU function to obtain the non-negative mapping weights. At the same time, a minimum constant is introduced. (Preferred) Calculate the normalized weights :

[0051]

[0052] use For the corresponding input feature map Element-wise scalar multiplication based on a broadcast mechanism is performed to complete adaptive spatial modulation. Then, full concatenation is performed in the channel dimension (dim=1) to output a weighted concatenated feature tensor. :

[0053]

[0054] in, This indicates a splicing operation.

[0055] Furthermore, the 3D saliency enhancement module, based on the spatial inhibition theory in neuroscience, directly derives the 3D pixel weights by constructing an analytical solution for the energy function based on linear separability. This adaptively purifies mineral textures and suppresses background interference such as resin target molds and sample preparation scratches. The processing flow is as follows:

[0056] The 3D saliency enhancement module receives intermediate fusion features from the ends of each scale of the bidirectional fusion path in the neck network. (Where C is the number of channels, and H and W are the spatial height and width of the feature map, respectively), the 3D energy matrix E is calculated using analytical closure formulas to evaluate the saliency of each pixel deviating from the global background:

[0057]

[0058] in, The total number of pixels in a single channel; This is the activation value of the j-th pixel within the channel; , These represent the spatial mean and spatial unbiased variance of the intermediate fusion feature X in each channel, respectively. To prevent division by zero regularization coefficients, the preferred option is... ;

[0059] Finally, the Sigmoid function is used to transform the three-dimensional energy matrix. The mapping is done to normalized weights and then fused with intermediate features from the input. Element-wise multiplication is performed to complete denoising and recalibration, outputting refined features with high signal-to-noise ratio. :

[0060]

[0061] in, This indicates element-wise multiplication. This refers to high signal-to-noise ratio refining characteristics, i.e., the characteristic signals of purified minerals.

[0062] Furthermore, the segmentation decoupling head receives the multi-scale high signal-to-noise ratio refined features and performs pixel classification and instance bounding box prediction. In the bounding box regression stage of backpropagation, the minimum point distance intersection-union ratio (MPDIoU) loss function is used to overcome the geometric limitations brought about by traditional aspect ratio measures. The specific process is expressed as follows:

[0063]

[0064] in, and These are the ground truth bounding boxes and the predicted bounding boxes, respectively. and These are the physical endpoint coordinates of the top-left and bottom-right corners of the actual bounding box. and These are the coordinates of the top-left and bottom-right endpoints of the prediction box; and These are the squared Euclidean distances between the upper left and lower right corners, respectively, which directly reflect the geometric deviation of the mineral edge; and Let c be the width and height of the minimum bounding rectangle, and c be the length of the diagonal of the minimum bounding rectangle. This is the final boundary loss operator.

[0065] Furthermore, in step S4, the mathematical model for area-volume-mass calculation based on the Delas principle of geological stereochemistry is established to complete the automated quantitative estimation of the total iron grade of the ore, as follows:

[0066] After obtaining the high-precision segmentation mask, the total effective pixel area of ​​all iron-bearing minerals within the field of view is calculated using a connected component analysis algorithm. and the total pixel area of ​​effective rock facies Calculate the two-dimensional surface area ratio of iron-bearing minerals. :

[0067]

[0068] in, The effective pixel area of ​​the kth iron-bearing mineral connected region instance;

[0069] According to the Drais principle, the area ratio Equivalent mapping to 3D volume percentage Combined with the preset standard density of iron oxide and density of associated gangue Convert volume percentage to mass percentage :

[0070] Based on the mass percentage Calculate the estimated total iron grade of BIF ore :

[0071]

[0072] in, The theoretical iron content coefficient for the target mineral.

[0073] The beneficial effects of this invention are:

[0074] 1) Non-destructive retention of microscopic features. This invention introduces the SPD-Conv module into the shallow layer of the main trunk. Through spatial-to-channel dimensional mapping, it completely blocks the pixel loss of sub-pixel-level micro-hematite / magnetite particles under the microscope during the downsampling process, and significantly reduces the false detection rate of fine disseminated minerals.

[0075] 2) Accurate alignment of multi-scale heterogeneous semantics. This invention constructs a neck fusion network based on the BiFPN-Concat module and introduces a dynamic normalization weight mechanism, enabling the model to adaptively evaluate the contribution of shallow texture and deep semantics to minerals of different grain sizes, thus eliminating feature aliasing when macroscopic stripes and fine particles coexist.

[0076] 3) Effective suppression of zero-parameter optical noise. This invention uses a tandem SimAM three-dimensional attention mechanism to directly derive feature saliency through the analytical solution of the energy function. Without adding any extra parameters or computational burden, it adaptively removes microscopic reflective interferences such as resin target model and sample preparation scratches, ensuring the feature signal-to-noise ratio entering the segmentation decoding head.

[0077] 4) Limit fitting of distorted mineral boundaries. This invention introduces the MPDIoU loss function to directly constrain the physical coordinates of the endpoints, breaking the geometric blind spot of conventional methods when dealing with aspect ratio distortion and tightly bonded minerals, so that the final mask exhibits extremely high pixel-level fitting at complex grain boundaries.

[0078] 5) Non-destructive grade quantification conversion. This invention breaks through the limitations of traditional deep learning, which is limited to two-dimensional image segmentation. By combining the principles of geological stereochemistry, it realizes a rigorous mathematical mapping from pixel area to three-dimensional quality and chemical grade, providing an efficient and objective digital alternative for deposit evaluation and mineral processing optimization. Attached Figure Description

[0079] Figure 1 This is a diagram of the overall network structure framework in an embodiment of the present invention.

[0080] Figure 2 This is a network architecture diagram of the SPD-Conv module in an embodiment of the present invention.

[0081] Figure 3 This is a network architecture diagram of the BiFPN-Concat module in an embodiment of the present invention.

[0082] Figure 4 This is a network architecture diagram of the SimAM module in an embodiment of the present invention. Detailed Implementation

[0083] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0084] Banded iron formations (BIFs), as the most important type of iron ore resource globally, are characterized by the alternating banded distribution of iron-rich minerals (such as magnetite and hematite) and siliceous minerals. This spatial distribution of minerals is not only a key geological clue for analyzing the genesis of ore deposits but also a core basis for determining the technological process. In ore industrial evaluation, ore grade (iron content) is a decisive indicator for measuring economic value and beneficiation feasibility. Therefore, high-precision identification and quantitative segmentation of iron-rich minerals in BIF ores under a microscope are of irreplaceable importance for geological genesis research, beneficiation process optimization, and accurate grade assessment. However, existing mineral microscopic identification and grade determination rely heavily on traditional methods, presenting significant technical bottlenecks. On the one hand, traditional chemical analysis methods are costly and cannot provide information on mineral morphology and spatial distribution; on the other hand, manual microscopic quantitative analysis heavily relies on the personal experience of geological engineers, which is not only time-consuming and labor-intensive but also easily introduces subjective bias and visual errors. This non-digital measurement method makes it difficult to achieve large-scale, efficient, and highly consistent quantitative analysis, and it can no longer meet the actual needs of modern mining production and refined assessment.

[0085] To address the aforementioned technical shortcomings, this invention proposes an intelligent and precise segmentation and grade estimation method for iron-bearing minerals under a BIF microscope. This invention effectively solves the problems of high subjectivity and low efficiency associated with traditional manual microscopic examination. It not only enables automated and high-precision boundary segmentation of complex banded iron-bearing minerals in microscopic images, but also allows for the direct extraction and statistical analysis of the area proportion of iron-rich minerals based on multi-target segmentation results, thereby objectively and rapidly inverting the ore grade.

[0086] First, the SPD-Conv module in the backbone network performs full-pixel-preserving feature downsampling on the original BIF ore microscopic image. Spatial segmentation and channel stitching operations are used to losslessly map spatial dimension information to the channel dimension, achieving resolution reduction while ensuring that sub-pixel-level fine iron oxide particles and their complex co-occurring boundary features under the microscope are not diluted, thus obtaining multi-scale fidelity feature maps. Subsequently, these multi-scale fidelity feature maps are fed into the neck network, where the BiFPN-Concat module performs adaptive spatial modulation and full channel stitching of bidirectional cross-layer feature flows from top to bottom and bottom to top. Learnable dynamic weight allocation and fast normalization mechanisms are introduced to spatially modulate features at both deep and shallow layers. Full channel stitching avoids heterogeneous semantic aliasing, adaptively aggregating cross-scale information from macroscopic stripes to microscopic disseminated minerals to construct a high-fidelity fused feature flow. Then, a parameter-free 3D saliency enhancement module (Sim) is connected in series before the multi-scale feature flow merges into the segmentation decoupling head. Based on spatial inhibition theory, the saliency of neurons is directly derived from the analytical solution of the energy function and mapped to three-dimensional attention weights. Feature adaptive recalibration is achieved through element-wise multiplication, thereby effectively suppressing optical noise interference from resin target models, sample preparation scratches, etc., and purifying the mineral signal in the target area without the burden of additional parameters. In the localization and boundary regression stages of model training, the MPDIoU loss function is introduced to completely abandon the easily degenerate aspect ratio constraint. Geometric penalties are applied directly by calculating the square Euclidean distance between the upper left and lower right corners of the predicted box and the true box. This provides a strong and dead-angle-free gradient dynamic for irregular mineral particles with aspect ratio distortion and dense adhesion, achieving pixel-level accurate fitting of complex crystal phase boundaries. Finally, the network segmentation decoupling head decodes the purified high-dimensional features and outputs a high-precision pixel-level segmentation mask that characterizes the precise distribution position and micromorphology of iron oxides, providing objective and robust digital support for the intelligent identification and quantitative analysis of BIF minerals under the microscope.

[0087] like Figure 1 As shown, in one embodiment, a method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion is provided. The method strictly follows the forward flow order of data tensors in the network and executes the following specific implementation steps:

[0088] Step S1: Dataset construction and standardization.

[0089] To train the network, this embodiment constructs an image sample set by combining images taken under a field reflective light microscope with publicly available online sample images. Using the Labelme polygon annotation tool, pixel-level edge refinement is performed along the true phase boundaries of iron oxides (magnetite, hematite) in the microscopic images, removing redundancy from the resin target model or siliceous gangue background, generating an initial JSON annotation file containing the geometric contour information of irregular mineral crystals. Then, a format conversion script is written to uniformly convert the initial JSON annotation file into standard binary mask labels suitable for segmentation network training. 152 high-quality images are divided into a training set, 44 images into a validation set, and 22 images into a test set; thus, the NEURock dataset is constructed.

[0090] Step S2: The backbone network performs micro-level full-pixel preserved feature extraction based on the SPD-Conv module.

[0091] Reference Figure 1 The network topology flow relationship shown will transform the original microscopic image. The data is fed into the front end of the backbone network for initial spatial feature encoding to obtain feature maps containing mineral basic reflectance and shallow edges; the specific mathematical flow is as follows:

[0092]

[0093] in, The original microscopic image; This is the first feature after the initial downsampling; The P2-scale features are after downsampling by a factor of 4; This is a cross-stage feature output after shallow feature refinement, containing mineral-based textures.

[0094] This embodiment introduces SPD-Conv modules at the shallow and mid-level P3 and P4 scales of the backbone network. Its network architecture diagram is as follows: Figure 2 As shown. For the input feature map (with... ∈ (For example), set the space scaling factor. The spatial partitioning operator is used to demix and shuffle it into four independent and complementary sub-feature maps:

[0095]

[0096] Where S is the spatial dimension and C is the number of channels.

[0097] Then, all sizes The sub-feature maps are concatenated along the channel dimension, and feature fusion and dimensionality reduction are performed using non-strike convolution with stride s = 1:

[0098]

[0099] in, This is the transition feature after the output channel of the SPD-Conv module is adjusted; mathematically, this process transfers redundant information from the spatial dimension to the channel dimension without loss, completely blocking pixel loss of tiny hematite or magnetite particles under the microscope during the resolution reduction process.

[0100] Finally, by continuously applying this lossless downsampling mechanism and subsequent cascading... The module performs feature extraction, and the network sequentially obtains shallow and mid-layer fidelity features:

[0101]

[0102] in, and These are the P3-scale and P4-scale fidelity features after lossless downsampling, respectively.

[0103] right Using standard strided convolution and The module obtains early deep features of the backbone network without pooling compression. ;Will The deepest global semantic features are extracted using Spatial Pyramid Pooling (SPPF) and Advanced Channel Attention (C2PSA). :

[0104]

[0105] in, These are early deep features that have not been pooled and compressed; It is the deepest global semantic feature tensor of the backbone network, which integrates Spatial Pyramid Pooling (SPPF) and Attention (C2PSA), representing the highest level of abstract global semantics of the ore in the current field of view.

[0106] Step S3: The neck network performs bidirectional multi-scale feature fusion and adaptive denoising based on the BiFPN-Concat module and the SimAM module.

[0107] The neck network receives heterogeneous multi-scale feature maps output by the backbone network. , , as well as It executes bidirectional cross-layer spatial modulation paths, both top-down and bottom-up, and introduces an improved BiFPN-Concat module (network architecture as follows). Figure 3 ) and SimAM module (network architecture such as Figure 4 Characteristic purification is performed to eliminate the characteristic overlap between macroscopic contiguous bands and microscopic scattered minerals. The specific process is as follows:

[0108] The deepest global semantic features output by the backbone network First, a spatial upsampling module is input to stretch the spatial resolution, making its resolution from... Magnify by 2 times Output upsampling transition features ; then P4-scale fidelity features at the same resolution as the backbone network output The first BiFPN-Concat module is used for cross-layer weighted fusion, and this module is the input feature map. and Initialize a learnable scalar weight. , The ReLU function is used to obtain the non-negative mapping weights. At the same time, a minimum constant is introduced. (Preferred) Calculate the normalized weights :

[0109]

[0110] use For the corresponding input feature map and Perform element-wise scalar multiplication based on a broadcast mechanism to complete adaptive spatial modulation, and then perform full concatenation in the channel dimension (dim=1):

[0111]

[0112] Output weighted concatenated feature tensor ; then The data is fed into a C3k2 module for feature extraction, and finally outputs intermediate fused features at the P4 scale. ;

[0113] Output P4-scale intermediate fused features After another spatial upsampling module (Upsample), its spatial resolution is reduced from... Magnify by 2 times Output upsampling transition features ; then, Fidelity features at the same resolution as the backbone network output at the P3 scale The input is combined with the input into the second BiFPN-Concat module to perform cross-layer concatenation, outputting a weighted concatenated feature tensor. Next, Through feature refinement using a C3k2 module, the P3-scale intermediate fused features are output. ;

[0114] To effectively suppress interference from surface scratches and matrix optical reflection under reflected light microscopy, The input is concatenated into the first parametric SimAM module for denoising. The 3D energy matrix E is calculated using analytical closure formulas, and the saliency of each pixel deviating from the global background is evaluated.

[0115]

[0116] in, The total number of pixels in a single channel; This is the activation value of the j-th pixel within the channel; , These are the intermediate fusion features. Spatial mean and spatially unbiased variance within each channel; To prevent division by zero regularization coefficients, the preferred option is... ;

[0117] Finally, the energy matrix is ​​calculated using the Sigmoid function. Mapped to normalized weights and fused with intermediate features from the input. Element-wise multiplication is performed to complete the noise reduction and recalibration:

[0118]

[0119] The final output is the first high signal-to-noise ratio refined feature. The above process can be expressed by the following formula:

[0120]

[0121] Its bottom-up path processing flow is as follows:

[0122] Refining features at the P3 scale after noise reduction and purification Using the reverse approach as the starting point for bottom-up enhancement, it first feeds the data into a standard strided convolutional layer (Conv) for spatial dimensionality reduction, increasing the resolution from... Down to Output downsampling transition features ; then, Intermediate features at the previous P4 scale And the original P4-scale fidelity features of the backbone network The three inputs are combined into the third BiFPN-Concat module to perform three-way cross-layer concatenation, outputting a weighted concatenated feature tensor. ;Will The P4-scale fusion features are output through a bottom-up path using a C3k2 module feature refinement. Finally, The signal is fed into a second parametric SimAM module for denoising, and finally outputs a second high signal-to-noise ratio refined feature. ;

[0123] The refined features of the above P4 scale Input a standard strided convolutional layer (Conv) for downsampling spatial dimensionality reduction, the resolution is reduced from... Down to Output dimensionality reduction transition features ;Will The deepest semantic features of the backbone network and early deep features of the backbone network that are not pooled All three inputs are simultaneously fed into the fourth BiFPN-Concat module to perform the final three-way cross-layer feature fusion, outputting a weighted concatenated feature tensor. ;Will Extracted from the last C3k2 module, outputting P5 scale fused features. Ultimately The signal is fed into a third parametric SimAM module for denoising and purification, outputting a third channel of high signal-to-noise ratio refined features. The above process can be expressed by the following formula:

[0124]

[0125] Step S4: Segmentation head decoding and crystal phase boundary pixel-level fitting based on MPDIoU loss constraint.

[0126] The purified characteristics , and Both are fed into the segment head. During the target localization process of network backpropagation optimization, the MPDIoU loss function is introduced to address the regression blind spot of distorted minerals. For the ground truth bounding box... With prediction box First, calculate the basic intersection-union ratio. :

[0127]

[0128]

[0129]

[0130]

[0131] in, and These are the ground truth bounding boxes and the predicted bounding boxes, respectively. and These are the physical endpoint coordinates of the top-left and bottom-right corners of the actual bounding box. and These are the coordinates of the corresponding endpoints of the predicted bounding box; and These are the squared Euclidean distances between the upper left and lower right corners, respectively, which directly reflect the geometric deviation of the mineral edge; and Let the width and height be the minimum bounding rectangle. It is the square of the length of its diagonal; As the final boundary loss operator, this mechanism abandons aspect ratio constraints and provides clear gradient regression dynamics for elongated, cohesive, irregular iron oxide particles, achieving pixel-level extreme fit between the predicted boundary and the real mineral contour.

[0132] The final output is a high-fidelity two-dimensional segmentation instance mask for iron-bearing minerals, exhibiting extremely high pixel-level fit at complex acute-angled crystal phase boundaries and sharp alternation angles. It accurately locks the target's category attributes and two-dimensional geometric boundaries, providing a spatial constraint with high spatial confidence for subsequent grade estimation.

[0133] Step S5: Automated quantitative calculation of total iron grade (TFe) of ore based on geological stereoscopic mathematical model.

[0134] After obtaining the high-precision segmentation mask, a connected component analysis algorithm is used to extract all mask instances identified as iron oxides within the field of view. The total effective pixel area is then calculated, and the total pixel area of ​​effective rock facies in the microscopic image is extracted. Subsequently, based on the principles of geological stereochemistry and prior density parameters, a multi-level mapping calculation model of "area-volume-mass-grade" is established to complete the automated and objective estimation of BIF ore iron grade. The specific mathematical derivation and process are as follows:

[0135]

[0136]

[0137]

[0138]

[0139] in, Let be the effective pixel area of ​​the k-th iron-bearing mineral connected region instance. This represents the total pixel area of ​​the effective lithophase under the current microscopic field of view; The two-dimensional surface area ratio of iron-bearing minerals; The area ratio of minerals in isotropic rock slices is statistically equivalent to their three-dimensional volume ratio, according to the Delesse principle in geological stereochemistry. This formula realizes a rigorous mathematical leap from two-dimensional image features to three-dimensional physical space. The target standard density of iron-bearing minerals is preset. The average prior density of associated gangue minerals (such as siliceous components like quartz); This represents the estimated mass percentage of iron-bearing minerals in the ore slice; This is the theoretical mass fraction coefficient of iron in the target mineral; This is the final calculated output estimate of the total iron content (TFe) of the ore at BIF (Body Intake). This mathematical model achieves efficient conversion of non-destructive image features of microscopic rock slices into industrial-grade quantitative identification indicators.

[0140] Through the complete closed-loop processing of steps S1 to S5 described above, this invention successfully constructs an intelligent analysis architecture that ranges from non-destructive extraction of microscopic sub-pixel-level features to macroscopic ore grade estimation and analysis. This scheme not only overcomes the technical bottlenecks of existing deep learning models, such as multi-scale feature aliasing, missed detection of tiny mineral particles, and low fitting accuracy of distorted crystal phase boundaries, which are easily generated under complex microscopic lighting conditions, through the collaborative optimization of core algorithm modules such as SPD-Conv, BiFPN-Concat, and SimAM; it also transcends the theoretical limitations of simple two-dimensional image segmentation, achieving a rigorous mathematical mapping from pixel space to three-dimensional physical properties based on the principles of geological stereochemistry. Ultimately, this invention can convert BIF ore slice images acquired by a reflected light microscope into accurate engineering quantitative indicators of iron grade (TFe) in real time and with high fidelity, reducing the time-consuming, labor-intensive, and subjectively biased manual microscopic examination process. This provides objective, robust, and highly industrially valuable technical support for ore deposit genetic research, mineral processing optimization, and grade assessment.

[0141] The size and quality of the dataset are crucial factors determining the performance of deep learning algorithms. Given the limited publicly available datasets and low resolution of BIF (Browsing In-Flight) microscopic images, this invention constructs the NEURock dataset by collecting and organizing data from both online and field photography sources. The NEURock dataset includes a single category, Iron-bearing minerals (RL), and after data augmentation, contains 218 high-quality microscopic images of iron-bearing minerals.

[0142] To verify the effectiveness of the deep learning model proposed in this invention and to improve its robustness under different geological samples, this approach constructs a dedicated dataset of BIF microscopic images of iron-bearing minerals and performs rigorous data standardization and augmentation operations.

[0143] Given the highly irregular morphology and complex co-occurrence boundaries of mineral crystals (such as magnetite and hematite) under a microscope, traditional rectangular box annotations (such as Labelime) introduce a significant amount of redundant gangue background. Therefore, this solution employs Labelime software for pixel-level polygon annotation. The polygon tool is used to meticulously delineate the true phase boundaries of iron oxides, obtaining extremely accurate geometric contour information. Subsequently, a conversion script is written to uniformly transform the initial JSON annotation file into a standard mask format suitable for training the YOLOv11 segmentation model.

[0144] To further validate the effectiveness of the proposed method, this invention uses the NEURock dataset for comparison with other state-of-the-art methods. 152 images were used for model training. The test and validation sets consisted of the remaining 22 and 44 images, respectively. First, the experimental setup and model performance evaluation metrics are described in detail. Then, comparisons are made with other state-of-the-art methods on the NEURock dataset, and the results are visualized. Finally, ablation experiments are used to validate the effectiveness of each module and evaluate the synergistic effects among the modules.

[0145] This invention was implemented using the PyTorch 2.5.1 framework on Windows 11, with Python 3.9 as the programming language. The computer's CPU was an Intel Core i9 13900K, the GPU was a GeForce RTX 4090, and CUDA version 12.1 was used. The YOLO-Iron-Bearing-Minerals segmentation network was trained for 300 epochs using the AdamW optimizer, with an initial learning rate of 1×10⁻⁶. -4 And it gradually increases during training, linearly changing to 1×10 -2 The training was enhanced using methods such as Mosaic, Mixup, and HSV. To ensure the model's robustness under complex microscopic lighting and varying geometries, and to guarantee the absolute fairness and reproducibility of the ablation experiments, this embodiment specifically designed and locked the hyperparameters during the model training phase. Particularly for the physical characteristics of microscopic mineral slices—namely, the absence of absolute gravity direction and strictly defined phase boundaries—this scheme enabled 180° full-angle rotation enhancement and forcibly disabled the Mixup operation, which could cause edge feature destruction. The specific training hyperparameter configurations are shown in Table 1.

[0146] Table 1. Hyperparameter settings of the proposed method on the NEU-Rock dataset.

[0147]

[0148] To objectively and quantitatively evaluate the performance of the method of this invention and to conduct horizontal comparisons with other mainstream algorithms, this invention introduces three core evaluation metrics recognized in the field of object detection and segmentation: precision, recall, and mean average precision (mAP). Precision focuses on the model's ability to predict the target, that is, the proportion of samples that the model identifies as targets (positive examples) that are actually target samples. The higher the value of this metric, the fewer false alarms or false positives the model produces, and the stronger the reliability of the output results. Recall focuses on evaluating the model's ability to detect targets, reflecting the proportion of ground truth targets that the model successfully detects and correctly predicts. The higher the recall, the lower the false negative rate of the model. mAP, calculated by the area under the Precision-Recall curve, can simultaneously reflect the model's overall performance in terms of precision and recall. When calculating mAP, it is first necessary to determine the decision threshold of the Intersection over Union (IoU). This threshold can be set as a single value, such as 0.50 (mAP50); or it can be set as a range, such as from 0.50 to 0.95, increasing in steps of 0.05 (mAP[50,95]). In the latter case, the mAP value corresponding to each IoU threshold within the range needs to be calculated separately, and then these results are averaged. Generally, a predicted bounding box is considered a successful match only when the IoU value between it and the ground truth bounding box is greater than or equal to 0.5; any match below this threshold is considered unsuccessful.

[0149] To evaluate the effectiveness of the proposed method, it was compared with five state-of-the-art algorithms on the NEURock dataset, and the results are shown in Table 2. The algorithms used in the experiments included YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12.

[0150] Due to the limitations of standard convolutional and fusion architectures in perceiving irregular mineral edges under a microscope, the segmentation accuracy improvement of the aforementioned algorithms has reached a bottleneck. In contrast, the YOLO-Iron-Bearing-Minerals method proposed in this invention achieves significant improvements to a certain extent, with Precision reaching 69.1%, Recall reaching 67.2%, mAP50 reaching 69.8%, and mAP[50,95] reaching 40.2%. Compared with YOLOv11, while Precision decreased by 0.9%, Recall increased by 3.3%, and mAP50 and mAP[50,95] both increased by 1.7%. Experiments on the dataset show that YOLO-Iron-Bearing-Minerals has superior performance compared to the other five state-of-the-art methods.

[0151] Although the introduction of SPD-Conv in the backbone network and the BiFPN-Concat mechanism in the neck region resulted in an increase in the number of model parameters and computational cost compared to YOLOv11, this computational overhead was compensated by the superior fitting capability for microscopic fine minerals and irregular phase boundaries, significantly reducing the false negative rate of densely coexisting minerals (with a particularly noticeable improvement in recall). This fully demonstrates that the network architecture optimization of this invention has a high return on investment and application value in complex microscopic mineral scenarios.

[0152] Table 2 Performance comparison of different methods on the NEU-Rock dataset

[0153]

[0154] The segmentation results from different methods show that YOLO-Iron-Bearing-Minerals is more effective in detecting and segmenting smaller iron-bearing minerals under microscopic mineral segmentation, with a lower probability of missed detection. This superior performance is attributed to the synergistic effect of SPD-Conv, BiFPN-Concat, and SimAM, which ensures real-time performance while enhancing the response capability to small minerals and irregular, blurred phase boundaries.

[0155] To verify the independent effectiveness of each innovative component (SPD-Conv, BiFPN-Concat, SimAM) and its underlying cooperative coupling mechanism in the proposed YOLO-Iron-Bearing-Minerals segmentation network model, rigorous ablation experiments were conducted on the NEU-Rock dataset. Specific performance metrics for each network variant are shown in Table 3.

[0156] Table 3 Ablation experimental results on the NEU-Rock dataset

[0157]

[0158] The YOLO-Iron-Bearing-Minerals segmentation network model proposed in this invention clearly demonstrates a multi-dimensional performance leap in complex microscopic mineral scenarios. Specifically, the original model exhibits an unbalanced distribution of high precision (70.0%) and low recall (63.9%), with core metrics mAP50 and mAP[50,95] at 68.1% and 38.5%, respectively. After fully loading the SPD-Conv, BiFPN-Concat, and SimAM modules, the model achieves a significant improvement in overall evaluation metrics. Recall increases by 3.3% to 67.2%, and the comprehensive metric mAP[50,95] under stringent decision thresholds is increased to 40.2%. By constructing a rigorous multi-module collaborative feature transfer closed loop: the front end utilizes SPD-Conv to losslessly capture sub-pixel-level physical signals of micro-minerals; the middle end relies on BiFPN-Concat cross-layer weighted splicing to achieve accurate semantic alignment of multi-scale heterogeneous features; and the end end uses SimAM to perform high-precision saliency background stripping on high-fidelity aligned features. This deep coupling mechanism not only suppresses the missed detection of finely disseminated symbiotic minerals, making the segmentation mask fit the real irregular mineral phase contour perfectly, but also achieves the ultimate perception capability of complex and hidden targets with only a reasonable and slight drop in accuracy (down to 69.1%), realizing the globally optimal trade-off between precision and recall in micro-mineral feature engineering.

[0159] Further observation of the ablation experiments for other pairwise combinations in the table reveals a typical multi-module semantic mismatch phenomenon. Incomplete module combinations not only fail to surpass the original model but also generally lead to a decline in network performance, especially the combination of SPD-Conv and SimAM, where mAP50 plummets to a low of 63.2%. The root cause of this performance collapse lies in the strong coupling dependency during the deep learning feature transfer process. When the network simultaneously introduces only the front-end SPD-Conv and the back-end SimAM, although the full-pixel preservation mechanism successfully captures microscopic mineral details, it also simultaneously and losslessly amplifies high-frequency optical noise such as scratches on the resin target mold. Because the mid-level layer lacks the multi-scale semantic cross-layer alignment and structured transition provided by BiFPN-Concat, directly applying three-dimensional spatial attention to this unaligned and noise-amplified feature map causes severe attention evaluation disorder, resulting in the erroneous suppression of originally effective mineral signals.

[0160] This application integrates SPD-Conv spatial-depth mapping convolution, BiFPN-Concat multi-scale adaptive fusion network, SimAM parameterless 3D saliency refinement, and MPDIoU endpoint distance-assisted boundary regression operator to construct a closed-loop intelligent analysis system from microscopic sub-pixel phase boundary extraction to macroscopic ore total iron grade (TFe) quantification, achieving high-precision real-time segmentation and quantitative estimation of BIF iron-bearing minerals under complex reflective light microscopy conditions. This scheme addresses the issue of dilution of minute disseminated mineral features caused by traditional strided convolution during the downsampling stage by utilizing the full pixel preservation mechanism of SPD-Conv. It achieves precise alignment of multi-scale heterogeneous semantics through the dynamic weighting and channel splicing mechanism constructed by BiFPN-Concat, effectively eliminating feature aliasing when macroscopic stripes and microscopic scattered minerals coexist. Meanwhile, the SimAM module uses the analytical solution of the three-dimensional energy function to accurately suppress optical background interference such as resin target model and sample preparation scratches and purify the mineral signal in the target area under the premise of zero additional parameter load. With the help of MPDIoU to constrain the endpoint geometry of aspect ratio distortion and densely adhered targets, the pixel-level fit between the segmentation mask and the real crystal phase boundary is significantly improved. Experimental results show that the present invention achieves an mAP50 of 69.8% on the NEU-Rock dataset. Compared with existing advanced algorithms such as YOLOv12, it achieves a breakthrough improvement in segmentation accuracy and global recall. To a certain extent, it solves the technical pain points of large subjective error, easy to miss fine coexisting mineral particles, and time-consuming and labor-intensive nature of traditional manual microscopic inspection mode. It provides extremely robust technical support for mineral deposit genesis research, mineral processing optimization, and grade assessment.

[0161] Finally, it should be noted that the above embodiments are intended to illustrate the technical solutions of the present invention and do not constitute any limitation on the present invention. Those skilled in the art should fully understand that modifications to the technical solutions described in the foregoing embodiments or equivalent substitutions for any part or all of the technical features are entirely feasible. Such modifications or substitutions, as long as they do not depart from the scope of protection defined by the claims of the present invention, should be considered reasonable extensions of the present invention.

Claims

1. A method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion, characterized in that, include: S1: Obtain reflected light microscopic images of different BIF ore samples and perform pixel-level polygon contour annotation and standardization preprocessing to construct a high-fidelity microscopic image dataset of iron-bearing minerals. S2: Construct the YOLO-Iron-Bearing-Minerals segmentation network, which consists of a backbone network, a neck network, and a segmentation decoupling head connected in sequence. The backbone network performs initial feature encoding on the input high-fidelity microscopic image of iron-bearing minerals, and deploys a space-depth transformation convolution module in the middle and shallow layers to perform lossless feature downsampling and deep feature extraction. This module uses spatial segmentation, channel stitching and nonlinear fusion operations to losslessly map and reorganize the information in the spatial dimension to the channel dimension, and outputs a multi-scale fidelity feature map. The neck network receives the multi-scale fidelity feature map, uses an adaptive weighted feature stitching module to perform adaptive spatial modulation and full-channel stitching of bidirectional cross-layer feature streams from top to bottom and from bottom to top, and connects a parameter-free three-dimensional saliency enhancement module in series at the end of the multi-scale output to suppress optical background noise under the microscope and output multi-scale high signal-to-noise ratio refined features. The segmentation decoupling head receives the multi-scale high signal-to-noise ratio refined features and performs pixel classification and instance bounding box prediction, and decodes and outputs a high-precision pixel-level segmentation mask for iron-containing minerals. S3: The YOLO-Iron-Bearing-Minerals segmentation network is trained in a supervised manner using the high-fidelity microscopic image dataset of iron-bearing minerals constructed in step S1, and the network parameters are iteratively optimized until the model converges; in the backpropagation localization loss calculation stage of model training, the minimum point distance intersection-union ratio loss function is introduced to constrain the irregular crystal phase boundary. S4: Input the microscopic image of the BIF ore to be identified into the trained segmentation network, extract the pixel-level segmentation mask of the target iron-bearing mineral, and establish a mathematical model for area-volume-mass measurement based on the Deras principle of geological stereochemistry to complete the automated quantitative estimation of the total iron grade of the ore to be identified.

2. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 1, characterized in that, In step S2, the specific process of the backbone network outputting multi-scale fidelity feature maps is as follows: The backbone network processes the input raw microscopic images. Initial spatial features are encoded sequentially through a first standard strided convolutional layer and a second standard strided convolutional layer to obtain P2-scale features downsampled by 4 times. , where H and W are the height and width of the original micrograph, respectively; Through the first Module refinement yields refined features ;Will The transition features are obtained by inputting the first spatial-depth transformation convolution module. , After refinement by the second C3k2 module, P3-scale fidelity features are output. ; Will The transition features are obtained by inputting the second-space-depth transformation convolution module. , After refinement by the third C3k2 module, P4-scale fidelity features are output. ; right Utilizing the third standard strided convolutional layer and the fourth The module obtains early deep features of the backbone network without pooling compression. ;Will The deepest global semantic features are extracted using spatial pyramid pooling and advanced channel attention modules. Final output , , and And together they are transmitted to the neck network.

3. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 1 or 2, characterized in that, The spatial-depth transformation convolutional module losslessly maps and reassembles information from the spatial dimension to the channel dimension through spatial segmentation, channel concatenation, and nonlinear fusion operations. The specific implementation process is as follows: For the input feature map ,in For space dimensions, For the number of channels, use the preset scaling factor. The spatial partitioning operator alternately samples and demixes the data along the spatial dimensions to obtain four data points of size 1. Sub-feature map , , , The four sub-feature maps are fully stitched together along the channel dimension, losslessly mapping the sub-pixel-level physical boundaries and reflected signals in the spatial domain to the channel axis, generating a size of [size missing]. intermediate recombination feature map : Will The input is fed into a non-stepping standard convolutional layer with a stride of 1 for smooth dimensionality reduction, and the output transition feature tensor is adjusted for different channels. ,in This indicates the number of output channels of the spatial-depth transformation convolution module.

4. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 2, characterized in that, The neck network refines the feature stream by performing adaptive weighted fusion through a top-down cross-layer path and concatenating a 3D self-attention mechanism. The specific processing flow is as follows: Global deepest semantic features After spatial resolution stretching by the first spatial upsampling module, the features are compared with those at the P4 scale. The first adaptive weighted feature concatenation module performs cross-layer weighted fusion, resulting in a weighted concatenated feature tensor. This tensor is then refined by the fifth C3k2 module, outputting the intermediate fused features at the P4 scale. ; P4-scale intermediate fusion features After being upsampled by the second spatial module, its spatial resolution is magnified and compared with the P3 scale features. The second adaptive weighted feature concatenation module, which is a common input, performs cross-layer concatenation and outputs a weighted concatenation feature tensor. The sixth C3k2 module then refines the features and outputs intermediate fused features at the P3 scale. ; Will The input is cascaded into the first parametric 3D saliency enhancement module for denoising, and the output is a refined feature with high signal-to-noise ratio at the P3 scale. .

5. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 4, characterized in that, The neck network performs adaptive weighted fusion through a bottom-up, cross-layer path and refines the feature flow through a 3D self-attention mechanism. The specific processing flow is as follows: Refining features at the P3 scale Using the reverse approach as the starting point for bottom-up enhancement, the data is input into the fourth standard strided convolutional layer for spatial dimensionality reduction and then fused with intermediate features at the P4 scale. and P4 scale fidelity features All three inputs are fed into the third adaptive weighted feature concatenation module to perform three-way cross-layer concatenation, outputting a weighted concatenation feature tensor; then, the seventh C3k2 module refines the features, outputting the P4-scale intermediate fused features from the bottom-up path. ;Will The data is fed into the second parametric 3D saliency enhancement module to complete denoising, and outputs refined features with high signal-to-noise ratio at the P4 scale. ; Refine features at the P4 scale After inputting the fifth standard strided convolutional layer for downsampling spatial dimensionality reduction, it is compared with the global deepest semantic features. and early deep features of the backbone network that are not pooled All three inputs are fed into the fourth adaptive weighted feature concatenation module to perform three-way cross-layer feature fusion, outputting a weighted concatenation feature tensor; then, the eighth C3k2 module performs feature extraction, outputting P5-scale intermediate fused features. ;Will The data is fed into the third parameterless 3D saliency enhancement module for denoising and purification, outputting refined features with high signal-to-noise ratio at the P5 scale. .

6. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 4 or 5, characterized in that, The adaptive weighted feature concatenation module outputs the weighted concatenation feature tensor based on the following method: For any cross-layer fusion node in the neck network, there is a set of n spatially aligned heterogeneous input feature maps. The system provides each input feature map Initialize a learnable scalar weight. , Use the ReLU function to obtain the non-negative mapping weights. At the same time, a minimum constant is introduced. Calculate normalized weights : use For the corresponding input feature map Element-wise scalar multiplication based on a broadcast mechanism is performed to achieve adaptive spatial modulation, followed by full concatenation along the channel dimension to output a weighted concatenated feature tensor. : in, This indicates a splicing operation.

7. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 6, characterized in that, The 3D saliency enhancement module receives intermediate fusion features from the ends of each scale of the bidirectional fusion path in the neck network. The 3D energy matrix E is calculated using analytical closure formulas, and the saliency of each pixel deviating from the global background is evaluated. in, The total number of pixels in a single channel; This is the activation value of the j-th pixel within the channel; , These represent the spatial mean and spatial unbiased variance of the intermediate fusion feature X in each channel, respectively. To prevent division by zero regularization coefficients; Finally, the Sigmoid function is used to transform the three-dimensional energy matrix. The mapping is done to normalized weights and then fused with intermediate features from the input. Element-wise multiplication is performed to complete denoising and recalibration, outputting refined features with high signal-to-noise ratio. : in, This indicates element-wise multiplication. This refers to high signal-to-noise ratio refining characteristics, i.e., the characteristic signals of purified minerals.

8. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 1, characterized in that, The segmentation decoupling head receives multi-scale, high signal-to-noise ratio refined features and performs pixel classification and instance bounding box prediction. In the bounding box regression stage of backpropagation, it employs the minimum point distance intersection-union ratio loss function. The specific process is as follows: in, and These are the ground truth bounding boxes and the predicted bounding boxes, respectively. and These are the physical endpoint coordinates of the top-left and bottom-right corners of the actual bounding box. and These are the coordinates of the top-left and bottom-right endpoints of the prediction box; and These are the squared Euclidean distances between the top-left and bottom-right endpoints, respectively. and Let c be the width and height of the minimum bounding rectangle, and c be the length of the diagonal of the minimum bounding rectangle. This is the final boundary loss operator.

9. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 1, characterized in that, In step S4, the mathematical model for area-volume-mass calculation based on the Delacroix principle of geological stereochemistry is established to complete the automated quantitative estimation of the total iron grade of the ore, as detailed below: After obtaining the high-precision segmentation mask, the total effective pixel area of ​​all iron-bearing minerals within the field of view is calculated using a connected component analysis algorithm. and the total pixel area of ​​effective rock facies Calculate the two-dimensional surface area ratio of iron-bearing minerals. : in, The effective pixel area of ​​the kth iron-bearing mineral connected region instance; According to the Drais principle, the area ratio Equivalent mapping to 3D volume percentage Combined with the preset standard density of iron oxide and density of associated gangue Convert volume percentage to mass percentage : Based on the mass percentage Calculate the estimated total iron grade of BIF ore : in, The theoretical iron content coefficient for the target mineral.

10. The method for identifying and estimating the grade of iron-bearing minerals based on multi-scale adaptive fusion according to claim 1, characterized in that, In step S1, the specific process of constructing a high-fidelity microscopic image dataset of iron-bearing minerals includes: S1.1: Collect original digital microscopic images of different BIF ore samples under a standard reflected light microscope, and introduce publicly available BIF ore reflected light microscopic images to expand the samples and form an expanded sample image set. S1.2: Using the Labelme polygon annotation tool, pixel-level edge delineation is performed along the true phase boundaries of iron oxides in the images within the expanded sample image set to obtain the geometric contour information of irregular minerals and generate an initial annotation file in JSON format; S1.3: Convert the initial annotation files in JSON format into standard binary mask labels suitable for training segmentation network models; S1.4: Perform resolution reduction and pixel value normalization processing on the images and corresponding standard binary mask labels in the expanded sample image set, and force the 180° full-angle random rotation enhancement and disable the image aliasing operation to construct a high-fidelity microscopic image dataset of iron-bearing minerals containing a training set, a validation set and a test set.

Images