Deep learning-based rapid ore sorting method under adaptive illumination intensity

Abstract

The invention provides a deep learning-based rapid ore sorting method under adaptive illumination intensity, and relates to the technical field of ore sorting, the deep learning-based rapid ore sorting method comprises the following steps: obtaining illumination intensity distribution information and ore three-dimensional point cloud data, and calculating ore space orientation parameters based on quaternion attitude estimation; inputting the obtained data into a layered illumination compensation network, and obtaining an illumination normalized enhanced image through adaptive brightness mapping and local light spot correction in combination with reflectivity recovery and non-local mean filtering; inputting the illumination normalized enhanced image and the ore three-dimensional point cloud data into a multi-branch feature extraction network, constructing feature high-order association, and adopting adaptive feature reconstruction and conditional random field optimization features; and inputting the reconstructed features into a classification model based on active learning, optimizing a feature manifold space by using a self-organizing mapping network, carrying out sample labeling in combination with active sampling of adaptive entropy threshold screening and uncertainty perception, and realizing ore classification and sorting through random forest integration.

Description

Technical Field

[0001] The present invention relates to the technical field of ore sorting, and in particular to a method for rapid ore sorting under adaptive light intensity based on deep learning. Background Art

[0002] With the development of intelligent manufacturing technology, intelligent identification and automated classification technologies have been widely used in the ore sorting process. Traditional ore sorting mainly relies on manual experience and judgment, which is inefficient and easily affected by subjective factors. In recent years, intelligent ore sorting systems based on machine vision and deep learning have gradually become a research hotspot. By acquiring multimodal data such as ore illumination information and three-dimensional point clouds, combined with deep learning algorithms, rapid ore classification can be achieved.

[0003] However, the existing technology has the following main problems: the uneven distribution of light intensity on the ore surface leads to inaccurate feature extraction; the traditional feature extraction method is difficult to effectively characterize the multi-scale feature information of the ore; the sample annotation cost is high, and it is difficult to obtain sufficient annotation data for model training; the single feature classification method is not robust enough and is easily disturbed by environmental factors and misjudgments. These problems seriously restrict the actual application effect of the ore intelligent sorting system.

[0004] In summary, there is an urgent need to provide a rapid ore sorting method that can adapt to complex lighting conditions, has efficient feature extraction capabilities, reduces sample annotation costs and has stable classification performance. By introducing light compensation mechanism, multimodal feature fusion strategy and active learning framework, accurate classification of ore can be achieved, and sorting efficiency and accuracy can be improved. Summary of the invention

[0005] The embodiment of the present invention provides a method for rapid sorting of ore under adaptive light intensity based on deep learning, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for rapid ore sorting under adaptive light intensity based on deep learning is provided, comprising:

[0008] Obtain light intensity distribution information and ore three-dimensional point cloud data, and calculate the ore spatial orientation parameters based on the quaternion attitude estimation algorithm; input the light intensity distribution information, ore three-dimensional point cloud data and ore spatial orientation parameters into the layered illumination compensation network, correct the abnormal area through adaptive brightness mapping and local spot correction, and use the reflectivity recovery algorithm and non-local mean filtering to obtain the illumination normalized enhanced image;

[0009] The illumination normalized enhanced image and the ore three-dimensional point cloud data are input into the multi-branch feature extraction network, and the separation features are determined by the feature decoupling module. The phase features are obtained by Hilbert transform. The phase features are constructed into feature high-order associations using the recursive neural tensor network, and the optimized reconstruction features are obtained by the adaptive feature reconstruction module and the conditional random field.

[0010] The reconstructed features are input into the classification model based on active learning. The feature manifold space is constructed and optimized using the self-organizing map network. The samples to be labeled are screened based on the adaptive entropy threshold. The uncertainty-aware active sampling strategy is used to annotate the samples to be labeled and optimize the classification model. The prediction results are fused through the random forest integration framework to obtain the ore classification results and perform ore sorting.

[0011] In an optional embodiment,

[0012] Obtain the light intensity distribution information and ore three-dimensional point cloud data, and calculate the ore spatial orientation parameters based on the quaternion attitude estimation algorithm; input the light intensity distribution information, ore three-dimensional point cloud data and ore spatial orientation parameters into the layered illumination compensation network, correct the abnormal area through adaptive brightness mapping and local spot correction, and use the reflectivity recovery algorithm and non-local mean filtering to obtain the illumination normalized enhanced image, including:

[0013] The light intensity distribution information on the surface of the ore is collected by a circumferentially distributed multispectral sensor array, and the three-dimensional point cloud data of the ore is collected by a depth camera, wherein the three-dimensional point cloud data of the ore includes spatial position information and reflection intensity information;

[0014] The three-dimensional point cloud data of the ore is projected onto the main plane and the ellipse characteristic parameters are fitted by the least square method to establish a quaternion rotation matrix, wherein the main axis direction vector of the ore is obtained according to the ellipse characteristic parameters, the rotation relationship between the main axis direction vector and the reference direction vector of the world coordinate system is converted into a quaternion representation, and the spatial orientation parameters of the ore are calculated based on the quaternion;

[0015] Inputting the illumination intensity distribution information, the three-dimensional point cloud data and the ore spatial orientation parameter into a layered illumination compensation network, wherein the layered illumination compensation network includes an adaptive brightness mapping layer and a local spot correction layer;

[0016] In the adaptive brightness mapping layer, a light attenuation model is established according to the spatial orientation parameters of the ore, the brightness compensation weights of each area on the ore surface are calculated, and the brightness of the light intensity distribution information is balanced;

[0017] In the local spot correction layer, a spot detection operator based on spatial correlation is used to locate the abnormal area, and the abnormal area is corrected by Gaussian weighting in combination with the spatial position information of the three-dimensional point cloud data of the ore to obtain a spot correction image;

[0018] The spot correction image is processed by adopting a reflectivity restoration algorithm, and the spot correction image is decomposed into the product form of a reflection component and an illumination component. The reflection component is iteratively solved by minimizing an energy function, and the reflection component is enhanced by adopting a non-local mean filter to obtain an illumination normalized enhanced image.

[0019] In an optional embodiment,

[0020] The spot detection operator based on spatial correlation locates abnormal areas including:

[0021] The multi-scale Gaussian pyramid is used to decompose the image, and the detection operator is constructed by combining the radial and angular illumination change characteristics. The candidate area of ​​the light spot is obtained through adaptive threshold segmentation, and the abnormal area is located based on the area matching criterion and gradient consistency constraint. Specifically, it includes:

[0022] Construct a multi-scale Gaussian pyramid for the input light intensity distribution information;

[0023] Based on the illumination variation characteristics in radial and angular directions of the sampling points on the ore surface, the radial detection component and the angular detection component are calculated respectively, and the spatial correlation detection operator is constructed in combination.

[0024] Performing a convolution operation on the spatial correlation detection operator and the multi-scale Gaussian pyramid to calculate a spatial correlation response value between each sampling point and a corresponding neighborhood sampling point;

[0025] Establish a local detection window with each sampling point as the center, calculate the mean and standard deviation of the light intensity distribution information in the local detection window, obtain an adaptive threshold of the sampling point, and compare the spatial correlation response value with the adaptive threshold to obtain a candidate light spot area;

[0026] Setting a region overlap rate threshold and a minimum area threshold to construct a spot region matching criterion, merging and screening the spot candidate regions, and obtaining an initial spot region;

[0027] The illumination intensity gradient and spatial position gradient of the boundary of the initial light spot area are calculated, a gradient consistency constraint is constructed, and the boundary of the initial light spot area is optimized to obtain the final abnormal light spot area positioning result.

[0028] In an optional embodiment,

[0029] The illumination normalized enhanced image and the ore three-dimensional point cloud data are input into the multi-branch feature extraction network. The separation features are determined by the feature decoupling module. The phase features are obtained by Hilbert transform. The phase features are constructed into feature high-order associations using the recursive neural tensor network. The optimized reconstructed features obtained by the adaptive feature reconstruction module and the conditional random field include:

[0030] The multi-branch feature extraction network includes an image feature branch and a point cloud feature branch. The illumination normalized enhanced image is input into the image feature branch constructed by the convolutional network based on the residual structure to obtain image features. The three-dimensional point cloud data of the ore is input into the point cloud feature branch based on the dynamic graph convolutional network to obtain point cloud features.

[0031] Performing adversarial learning based on the image features and the point cloud features, decoupling the image features and the point cloud features into shared features and private features, wherein the private features include image private features and point cloud private features;

[0032] The shared features and private features are transformed using Hilbert transform, and the phase features are obtained based on the inverse tangent operation between the transformed features and the original features;

[0033] Using a tensor weight matrix and a recursive weight matrix in a recursive neural tensor network to perform nonlinear transformation on the phase features, and constructing high-order correlation features between the phase features;

[0034] Constructing a feature weight matrix, and adaptively weighting the high-order correlation features based on the feature weight matrix to obtain reconstructed features;

[0035] The spatial dependency relationship of the reconstruction feature is constructed based on a unary potential function and a binary potential function, and the reconstruction feature is optimized by using mean field inference to obtain an optimized reconstruction feature.

[0036] In an optional embodiment,

[0037] The spatial dependency relationship of the reconstruction features is constructed based on the unary potential function and the binary potential function, and the reconstruction features are optimized by using the mean field inference to obtain the optimized reconstruction features, including:

[0038] By constructing unary potential functions and binary potential functions, the node attributes and spatial dependencies of the reconstructed features are characterized and combined into a conditional random field energy function. The mean field inference is used for feature optimization, and the fast Fourier transform is introduced to accelerate the iteration of the edge probability distribution on the feature grid. Specifically,

[0039] Calculate the unary attribute of the reconstructed feature to obtain a unary eigenvalue, and construct a unary potential function based on the unary eigenvalue, wherein the unary potential function is used to characterize the node attribute of the reconstructed feature; calculate the spatial distance between the reconstructed feature nodes to obtain a distance matrix, and input the distance matrix into a Gaussian kernel function to construct a binary potential function, wherein the binary potential function is used to characterize the spatial dependency of the reconstructed feature;

[0040] Based on the unary potential function and the binary potential function, a spatial dependency model of the reconstructed feature is constructed, and the spatial dependency model is expressed as a conditional random field energy function; the conditional random field energy function is input into a mean field inference model to initialize the edge probability distribution of the reconstructed feature node; based on the binary potential function, the message transmission between nodes is calculated, and the result of the message transmission is combined with the unary potential function to update the edge probability distribution;

[0041] The reconstructed feature space is divided into feature grids, the potential function value is calculated in the feature grids, and the edge probability distribution is optimized by mean field iteration using fast Fourier transform; the expected value of the reconstructed feature is calculated based on the optimized edge probability distribution, and the expected value is used as the output result of the optimized reconstructed feature.

[0042] In an optional embodiment,

[0043] Input the reconstructed features into the classification model based on active learning, use the self-organizing map network to build and optimize the feature manifold space, screen the samples to be labeled based on the adaptive entropy threshold, use the uncertainty-aware active sampling strategy to label the samples to be labeled and optimize the classification model, and obtain the ore classification results through the random forest integration framework to fuse the prediction results and perform ore sorting, including:

[0044] The reconstructed features are input into the self-organizing map network, and a topology-preserving mapping is established based on a competitive learning mechanism. The feature nonlinear dimension reduction is performed by iteratively optimizing the weight vector of the competition layer neurons to obtain a manifold feature space.

[0045] Calculating sample information entropy in the manifold feature space, constructing an adaptive entropy threshold based on the mean and standard deviation of the sample information entropy, and taking samples whose sample information entropy is greater than the adaptive entropy threshold as samples to be labeled;

[0046] Respectively obtaining the prediction uncertainty and sample representativeness of the samples to be labeled, performing a weighted combination of the prediction uncertainty and the sample representativeness to obtain a sampling score, selecting samples to be labeled based on the sampling score, and adding the labeled samples to a training set;

[0047] Randomly selecting a feature subspace in the manifold feature space, performing bootstrap sampling in the training set to obtain a data subset, and constructing multiple decision tree classifiers using the feature subspace and the data subset; inputting the samples to be classified into all the decision tree classifiers, and fusing the prediction results of all the decision tree classifiers through a voting mechanism to obtain an ore classification result;

[0048] An ore sorting operation is performed based on the ore classification result.

[0049] In an optional embodiment,

[0050] Respectively obtaining the prediction uncertainty and sample representativeness of the sample to be labeled includes:

[0051] Constructing a k-nearest neighbor structure of the sample to be labeled in the manifold feature space, calculating the geodesic distance between each sample to be labeled and the corresponding k-nearest neighbor samples; constructing a local covariance matrix based on the geodesic distance, and calculating the Mahalanobis distance of each sample to be labeled using the local covariance matrix;

[0052] Performing a convolution operation on the Mahalanobis distance and the local density function to obtain the local probability density of the sample to be labeled; calculating the manifold consistency measure of the sample to be labeled and the labeled sample set based on the local probability density, and using the manifold consistency measure as the representativeness of the sample;

[0053] The nonlinear probability prediction of the sample to be labeled is performed using the manifold Laplace eigenmap; information entropy and inter-class divergence are calculated based on the nonlinear probability prediction, and a weighted combination of the information entropy and the inter-class divergence is used as the prediction uncertainty.

[0054] In an embodiment of the present invention, a layered illumination compensation and reflectivity recovery algorithm is used to effectively eliminate the influence of uneven illumination, make the surface details of the ore clearer, and improve the accuracy of subsequent analysis; in combination with illumination information, point cloud data and ore orientation parameters, a multi-branch feature extraction network and a recursive neural tensor network are used to achieve more comprehensive ore feature modeling and improve recognition accuracy; through a self-organizing mapping network and an uncertainty-aware sampling strategy, the selection of samples to be labeled is optimized, the labeling efficiency is improved, the need for manual labeling is reduced, and the optimization of the classification model is accelerated; a random forest integrated framework is used for ore classification, and automatic operation is realized in combination with sorting equipment, so as to improve the intelligence level of ore identification and sorting and improve production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a schematic flow chart of a method for rapid ore sorting under adaptive light intensity based on deep learning according to an embodiment of the present invention;

[0056] Figure 2This is a diagram showing the effect of multi-scale detection response and processing time analysis;

[0057] Figure 3 This is a six-dimensional performance evaluation effect diagram of spatial correlation;

[0058] Figure 4 Comparative analysis of the effect before and after gradient consistency optimization;

[0059] Figure 5 The figure is a comprehensive comparison effect diagram of the performance of the method of the present invention and the traditional method. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0062] Figure 1 FIG. 1 is a flow chart of a method for rapid ore sorting under adaptive light intensity based on deep learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] Obtain light intensity distribution information and ore three-dimensional point cloud data, and calculate the ore spatial orientation parameters based on the quaternion attitude estimation algorithm; input the light intensity distribution information, ore three-dimensional point cloud data and ore spatial orientation parameters into the layered illumination compensation network, correct the abnormal area through adaptive brightness mapping and local spot correction, and use the reflectivity recovery algorithm and non-local mean filtering to obtain the illumination normalized enhanced image;

[0064] In this embodiment, the physical characteristics and geometric structure of the ore are effectively captured by comprehensively utilizing the light intensity distribution, the three-dimensional point cloud data of the ore, and the spatial orientation parameters obtained by the quaternion-based attitude estimation, providing rich information support for subsequent light compensation; the layered light compensation network is utilized to effectively correct the abnormal areas in the image through adaptive brightness mapping and local spot correction, and reduce the impact of local uneven light, thereby making the image light distribution more uniform; the reflectivity recovery algorithm and non-local mean filtering are adopted to restore the true reflection characteristics of the ore, enhance the image details and contrast, improve the overall image quality, and enhance the accuracy of subsequent target recognition and detection; through multi-data fusion and layered compensation mechanism, the method has better adaptability to complex lighting environments and local spot interference, ensuring that stable and reliable image enhancement effects can be obtained in different scenarios.

[0065] In an optional embodiment, it also includes:

[0066] The light intensity distribution information on the surface of the ore is collected by a circumferentially distributed multispectral sensor array, and the three-dimensional point cloud data of the ore is collected by a depth camera, wherein the three-dimensional point cloud data of the ore includes spatial position information and reflection intensity information;

[0067] The three-dimensional point cloud data of the ore is projected onto the main plane and the ellipse characteristic parameters are fitted by the least square method to establish a quaternion rotation matrix, wherein the main axis direction vector of the ore is obtained according to the ellipse characteristic parameters, the rotation relationship between the main axis direction vector and the reference direction vector of the world coordinate system is converted into a quaternion representation, and the spatial orientation parameters of the ore are calculated based on the quaternion;

[0068] Inputting the illumination intensity distribution information, the three-dimensional point cloud data and the ore spatial orientation parameter into a layered illumination compensation network, wherein the layered illumination compensation network includes an adaptive brightness mapping layer and a local spot correction layer;

[0069] In the adaptive brightness mapping layer, a light attenuation model is established according to the spatial orientation parameters of the ore, the brightness compensation weights of each area on the ore surface are calculated, and the brightness of the light intensity distribution information is balanced;

[0070] In the local spot correction layer, a spot detection operator based on spatial correlation is used to locate the abnormal area, and the abnormal area is corrected by Gaussian weighting in combination with the spatial position information of the three-dimensional point cloud data of the ore to obtain a spot correction image;

[0071] The spot correction image is processed by adopting a reflectivity restoration algorithm, and the spot correction image is decomposed into the product form of a reflection component and an illumination component. The reflection component is iteratively solved by minimizing an energy function, and the reflection component is enhanced by adopting a non-local mean filter to obtain an illumination normalized enhanced image.

[0072] In a specific implementation, data collection is first performed. Multiple spectral sensors are evenly arranged along the ring direction on the ore sorting conveyor belt to form a sensor array to collect light intensity distribution information on the ore surface. At the same time, a depth camera is used to collect three-dimensional point cloud data of the ore from the top to obtain the spatial position coordinates and surface reflection intensity value of the ore.

[0073] Then the ore posture is calculated, the collected three-dimensional point cloud data is projected onto the horizontal plane, and the ellipse representing the ore outline is fitted using the least squares method to obtain the major axis, minor axis and other characteristic parameters of the ellipse. The main axis direction vector of the ore is determined based on the ellipse characteristics, and the spatial rotation relationship between this vector and the preset world coordinate system reference direction vector is expressed in the form of quaternions, thereby calculating the specific spatial orientation of the ore on the conveyor belt.

[0074] Then, illumination compensation is performed, and the three types of data (illumination distribution, point cloud data, and spatial orientation) obtained above are input into a two-layer illumination compensation network. In the first layer of adaptive brightness mapping, a mathematical model of light attenuation is established according to the spatial orientation of the ore, the brightness compensation weight values ​​required for each area on the ore surface are calculated, and the original light intensity distribution is brightness-equalized. In the second layer of local spot correction, a detection operator based on spatial correlation is designed to locate abnormal spot areas on the ore surface, and then combined with the spatial position information in the three-dimensional point cloud, these abnormal areas are corrected by Gaussian weighting to obtain a corrected image.

[0075] Finally, illumination normalization is performed, and the reflectivity restoration algorithm is applied to the corrected image to decompose the image into the multiplication of the reflection component and the illumination component. The reflection component is solved by iteratively optimizing the energy function. The reflection component is then enhanced using a non-local mean filter to finally obtain an enhanced image after illumination normalization.

[0076] For example, taking an irregular iron ore as an example, eight evenly distributed spectral sensors and a depth camera are used for data collection. The sensor detects that there is obvious uneven illumination on the surface of the ore, with the center area being too bright and the edge area being darker. The point cloud data obtained by the depth camera shows that the ore is flat, about 15 cm long and about 10 cm wide.

[0077] Through ellipse fitting calculation, it is found that the ore is deflected about 30 degrees relative to the conveyor belt. Based on this posture information, the first layer of the illumination compensation network calculates that the edge area needs a 1.5-2 times brightness increase, and the center area needs a 0.7 times brightness reduction. The second layer detects two high light spot areas, with an area of ​​approximately 2cm² and 1cm² respectively, and corrects these abnormal areas based on the pixel values ​​of the surrounding normal areas.

[0078] Finally, through reflectivity restoration and non-local mean filtering, an ore surface image with uniform illumination and clear details was obtained, laying the foundation for subsequent feature extraction and classification.

[0079] In this embodiment, a spectral sensor and a depth camera are used to collect ore illumination and point cloud data to achieve information complementarity, provide an accurate basis for subsequent posture calculation and compensation, and improve the overall processing effect; the three-dimensional point cloud data is projected onto a horizontal plane, and the least squares method is used to fit an ellipse to extract the main axis direction of the ore, and its specific orientation on the conveyor belt is determined by quaternion calculation; a two-layer illumination compensation network is constructed, the first layer performs brightness mapping adjustment, and the second layer performs local spot correction, and partition processing is used to achieve overall illumination balance and abnormal area correction; a reflectivity recovery algorithm is used to separate reflection and illumination components, and non-local mean filtering is used to enhance image details and improve contrast, ultimately obtaining a normalized enhancement effect.

[0080] In an optional embodiment, locating the abnormal area using a spot detection operator based on spatial correlation includes:

[0081] The multi-scale Gaussian pyramid is used to decompose the image, and the detection operator is constructed by combining the radial and angular illumination change characteristics. The candidate area of ​​the light spot is obtained through adaptive threshold segmentation, and the abnormal area is located based on the area matching criterion and gradient consistency constraint. Specifically, it includes:

[0082] Construct a multi-scale Gaussian pyramid for the input light intensity distribution information;

[0083] Based on the illumination variation characteristics in radial and angular directions of the sampling points on the ore surface, the radial detection component and the angular detection component are calculated respectively, and the spatial correlation detection operator is constructed in combination.

[0084] Performing a convolution operation on the spatial correlation detection operator and the multi-scale Gaussian pyramid to calculate a spatial correlation response value between each sampling point and a corresponding neighborhood sampling point;

[0085] Establish a local detection window with each sampling point as the center, calculate the mean and standard deviation of the light intensity distribution information in the local detection window, obtain an adaptive threshold of the sampling point, and compare the spatial correlation response value with the adaptive threshold to obtain a candidate light spot area;

[0086] Setting a region overlap rate threshold and a minimum area threshold to construct a spot region matching criterion, merging and screening the spot candidate regions, and obtaining an initial spot region;

[0087] The illumination intensity gradient and spatial position gradient of the boundary of the initial light spot area are calculated, a gradient consistency constraint is constructed, and the boundary of the initial light spot area is optimized to obtain the final abnormal light spot area positioning result.

[0088] In a specific implementation, in the first step, the acquired light intensity distribution information is subjected to Gaussian filtering and downsampling operations to construct a multi-level image pyramid structure. The image resolution of each level is gradually reduced to form image representations at different scales for subsequent multi-scale feature analysis.

[0089] The second step is to analyze the illumination variation characteristics of the sampling points on the ore surface in the radial direction (from the center to the outside) and the angular direction (along the circumference). The radial gradient of the illumination intensity is calculated in the radial direction, and the circumferential gradient is calculated in the angular direction. The detection components of these two directions are combined to form a complete spatial correlation detection operator.

[0090] The third step is to perform convolution operations on the constructed detection operator and each layer of the multi-scale pyramid image. For each sampling point, the spatial correlation response value between it and the surrounding neighborhood sampling points is calculated. This response value reflects the degree of illumination difference between the point and the surrounding area.

[0091] The fourth step is to set a local window with each sampling point as the center, calculate the average brightness and brightness standard deviation of the pixels in the window, and determine an adaptive threshold for each sampling point based on these statistical characteristics. Compare the spatial correlation response value of the sampling point with its corresponding adaptive threshold to screen out possible light spot areas.

[0092] The fifth step is to set two threshold conditions, namely, the region overlap rate and the minimum area, to merge and screen the candidate regions obtained in the previous step. When the overlap of two candidate regions exceeds the set threshold, they are merged, and the regions with too small area are eliminated to obtain the preliminary spot region.

[0093] The sixth step is to calculate the light intensity gradient and spatial position gradient on the boundary of the initially determined spot area and construct the gradient consistency constraint. The boundary of the spot area is finely adjusted according to the gradient change characteristics, and finally the accurate position of the abnormal spot area is obtained.

[0094] Figure 2The changing trends of the detection response value and processing time of the algorithm at different image resolutions are shown. From the perspective of the detection response value, it reaches the highest value of 0.92 at the original 1024×1024 resolution, slightly decreases to 0.88 when the resolution is reduced to 512×512, and remains at a high level of 0.85 when it is further reduced to 256×256. This slow downward trend shows that the algorithm has excellent scale adaptability and can maintain high detection reliability even at low resolutions. The processing time shows a significant downward trend: it takes 825ms at the original resolution, decreases to 412ms at 512×512, and further decreases to 205ms at 256×256. The processing time decreases approximately linearly, which is consistent with the trend of decreasing the number of image pixels. It is particularly noteworthy that at the resolution of 512×512, the system achieves a good performance balance: the detection response value only decreases by 0.04, while the processing time is reduced by nearly 50%. This working point has important reference value in practical applications and provides an important basis for the resolution selection when the system is deployed.

[0095] exist Figure 3 In the figure, the spatial correlation detection response radar chart comprehensively shows the system performance from six dimensions. The radial response reaches a level of 0.85, indicating that the system has a good ability to capture the radial characteristics of the target; the angular response is 0.78, slightly lower than the radial response but still within an acceptable range; in particular, the combined response reaches a high level of 0.92, which is significantly higher than the response value of a single dimension, which confirms the synergistic enhancement effect of multi-dimensional feature fusion. The directional accuracy reaches 0.89, indicating that the system has a high accuracy in judging the target direction; the spatial consistency obtains the highest score of 0.94, indicating that the detection results maintain excellent coherence and stability in the spatial domain; the anti-noise performance reaches 0.88, proving that the system has good environmental adaptability and anti-interference ability. The balanced development of the six dimensions, without obvious shortcomings, fully demonstrates the comprehensiveness and stability of the system. The radar chart presents a relatively regular hexagon, indicating that the system has reached a similar high level in all aspects.

[0096] exist Figure 4In the figure, the gradient consistency optimization effect comparison chart shows the improvement effect of the optimization algorithm on four key indicators. The boundary accuracy is significantly improved, from 4.5 before optimization to 2.3 after optimization, with an improvement of nearly 50%, indicating that the boundary positioning after optimization is more accurate and reliable; the illumination gradient is improved from 0.65 to 0.88, an improvement of 35.4%, indicating that the system's adaptability to illumination changes has been greatly improved; the position gradient is improved from 0.58 to 0.85, an improvement of 46.6%, indicating that the system's sensitivity to changes in target position has been significantly enhanced; the consistency score is improved from 0.45 to 0.82, with the largest improvement of 82.2%. This significant improvement proves that the detection results after optimization are more stable and reliable. The fact that the four indicators are comprehensively improved strongly proves the effectiveness of the optimization algorithm, especially the huge improvements in the two key indicators of boundary accuracy and consistency score, which lays a solid foundation for the practical application of the system.

[0097] exist Figure 5 In the performance comprehensive comparison chart, the performance difference between this method and traditional methods is comprehensively compared through five core indicators. In terms of detection accuracy, this method reaches 94.5%, which is significantly better than the 85.2% of the traditional method, an increase of 9.3 percentage points; in terms of processing speed, this method reaches 85.2, which is 17.5% higher than the 72.5 of the traditional method, showing stronger real-time performance; in terms of noise resistance performance, this method achieves a high score of 89.3, which is 16.3% higher than the 76.8 of the traditional method, indicating that it has stronger environmental adaptability; the boundary accuracy reaches 92.1, which is 11.8% higher than the 82.4 of the traditional method, indicating that it has obvious advantages in target contour extraction; in terms of spatial consistency, this method reaches 94.0, which is better than the 85.5 of the traditional method, an increase of 9.9%, proving that the detection results are more stable and reliable. The comprehensive surpassing of the five dimensions, and the improvement rate is more than 9%, fully proves the comprehensive advantages of this method, especially the significant improvement in the two key indicators of detection accuracy and noise resistance, indicating that this method has better application prospects in practical applications.

[0098] In this embodiment, Gaussian filtering and downsampling are used to construct a multi-layer pyramid to effectively extract illumination information of different scales and provide comprehensive detail support for abnormal light spot detection. The detection operator is constructed by combining radial and circumferential illumination gradients to accurately reflect the illumination changes on the ore surface and help accurately capture the characteristics of abnormal light spots. Dynamic thresholds are set based on local statistical information to segment spatially correlated responses, and then candidate areas are merged through regional overlap and area screening to improve detection stability. The boundaries of candidate areas are finely adjusted using illumination and position gradient consistency constraints to accurately locate the abnormal light spot area and enhance the reliability of subsequent processing.

[0099] The illumination normalized enhanced image and the ore three-dimensional point cloud data are input into the multi-branch feature extraction network, and the separation features are determined by the feature decoupling module. The phase features are obtained by Hilbert transform. The phase features are constructed into feature high-order associations using the recursive neural tensor network, and the optimized reconstruction features are obtained by the adaptive feature reconstruction module and the conditional random field.

[0100] In this embodiment, the illumination normalized image and the three-dimensional point cloud data of the ore are input into a multi-branch network to realize multimodal data fusion, comprehensively extract the texture and structure information of the ore, and provide a solid data foundation for subsequent feature processing; the interference information is separated by the feature decoupling module, and the phase features are extracted by Hilbert transform, which effectively enhances the expression of key information and improves the model's ability to capture subtle features; the recursive neural tensor network is used to construct high-order correlations between phase features, deeply explore the complex relationships between various ore features, promote multi-scale feature fusion, and improve the overall expression effect; the adaptive feature reconstruction module is combined with the conditional random field to achieve accurate optimization and reconstruction of features, adjust local details, and finally output high-quality and stable reconstructed features.

[0101] In an optional embodiment, it also includes:

[0102] The multi-branch feature extraction network includes an image feature branch and a point cloud feature branch. The illumination normalized enhanced image is input into the image feature branch constructed by the convolutional network based on the residual structure to obtain image features. The three-dimensional point cloud data of the ore is input into the point cloud feature branch based on the dynamic graph convolutional network to obtain point cloud features.

[0103] Performing adversarial learning based on the image features and the point cloud features, decoupling the image features and the point cloud features into shared features and private features, wherein the private features include image private features and point cloud private features;

[0104] The shared features and private features are transformed using Hilbert transform, and the phase features are obtained based on the inverse tangent operation between the transformed features and the original features;

[0105] Using a tensor weight matrix and a recursive weight matrix in a recursive neural tensor network to perform nonlinear transformation on the phase features, and constructing high-order correlation features between the phase features;

[0106] Constructing a feature weight matrix, and adaptively weighting the high-order correlation features based on the feature weight matrix to obtain reconstructed features;

[0107] The spatial dependency relationship of the reconstruction feature is constructed based on a unary potential function and a binary potential function, and the reconstruction feature is optimized by using mean field inference to obtain an optimized reconstruction feature.

[0108] In a specific embodiment, in the multi-branch feature extraction stage, the input illumination normalized enhanced image is first preprocessed, and the image is uniformly scaled to a standard size. The image feature branch adopts a deep residual network structure, performs feature dimension reduction through the first layer convolution, and then connects multiple residual blocks in series for feature extraction. Each residual block contains a main path and a shortcut path. The convolution layer on the main path is responsible for extracting deep features, and the shortcut path ensures the effective transmission of information. Through layer-by-layer feature extraction and pooling operations, a high-dimensional image feature vector is finally obtained. At the same time, the point cloud feature branch processes the input three-dimensional point cloud data, first constructs the local connection relationship of the point cloud, and establishes a dynamic graph structure based on the neighbor information of each point. Through the dynamic graph convolutional network, the connection relationship between points can be adaptively updated to effectively capture the local structural features and global morphological information of the point cloud.

[0109] In the feature decoupling stage, a complete adversarial learning framework is constructed. The encoder network processes image features and point cloud features separately, and maps the features to the latent space through multiple layers of nonlinear transformation. In the latent space, the features are decoupled into common features that reflect common attributes and private features that express modality-specific information. The discriminator network evaluates the decoupling effect of features through adversarial training to ensure that the common features can truly reflect the public information of multimodal data, while the private features retain the unique properties of each modality. By setting appropriate loss weights, the reconstruction accuracy and feature independence are balanced in the feature decoupling process.

[0110] In the process of phase feature extraction, each characteristic component obtained by decoupling is Hilbert transformed. By constructing a Hilbert transform filter, the Fourier transform of the feature is first calculated, and then the corresponding analytical signal is obtained. Based on the real and imaginary information of the analytical signal, the phase information of the feature is obtained by inverse tangent operation. In order to improve the stability of the phase feature, phase unwrapping and jump elimination processing are also performed, and a smoothing algorithm is used to reduce the impact of noise.

[0111] In the high-order correlation construction stage, a recursive neural tensor network is used to process phase features. First, the high-dimensional feature tensor is decomposed, and the core tensor is designed to capture the complex interaction pattern between features. Through tensor contraction operations, feature information of different dimensions is effectively combined. In the recursive update process, the information flow is controlled by the gating mechanism, and the hidden state representation of the feature is continuously updated. After multiple steps of iteration, stable high-order correlation features are obtained.

[0112] An adaptive weighting strategy is used in the feature reconstruction stage. First, the discriminative ability of each feature component is evaluated, and a feature weight matrix is ​​constructed based on the discriminative score. In the feature fusion process, the dimensions of different modal features are aligned, and a unified feature representation is obtained through weighted combination. In order to improve the robustness of the features, regularization processing is also performed.

[0113] In the final spatial optimization stage, a spatial dependency model based on potential functions was constructed. By calculating the unary attributes of the nodes and the spatial distance matrix, a complete potential function system was designed in combination with the Gaussian kernel function. During the optimization process, the mean field inference method was used to iteratively update the edge probability distribution. The fast Fourier transform was used to accelerate the calculation process, and finally the optimized feature representation considering the spatial constraints was obtained.

[0114] For example, take a piece of copper ore with complex surface texture as an example:

[0115] In the feature extraction stage, the image branch uses an 18-layer residual network to extract a 256-dimensional image feature vector; the point cloud branch processes point cloud data containing 2,000 points and extracts a 128-dimensional point cloud feature vector through 3 layers of dynamic graph convolution.

[0116] After feature decoupling, we obtain 64-dimensional common features (reflecting the basic morphological characteristics of the ore), as well as 192-dimensional and 64-dimensional image private features (reflecting surface texture) and point cloud private features (reflecting three-dimensional structure).

[0117] Through Hilbert transform and inverse tangent operation, each feature is converted into a phase feature to better express the periodic change law of the feature. For example, for the striped texture on the surface, the phase feature can clearly express its directionality and periodicity.

[0118] A three-layer recursive structure is set in the recursive neural tensor network, and a 32×32 tensor weight matrix is ​​used to capture feature interactions. Through five iterations, high-order features that reflect the complex associations between features are constructed.

[0119] A weight matrix is ​​constructed based on the discriminative contribution of the features, a larger weight (such as 0.8) is given to highly discriminative texture features, and a smaller weight (such as 0.2) is given to noise features to achieve adaptive fusion of features.

[0120] Finally, spatial optimization is performed on a 10×10 feature grid, and after 20 mean field iterations, optimized features with spatial consistency are obtained for subsequent classification tasks.

[0121] In this embodiment, deep residual networks and dynamic graph convolutions are used to extract features of illumination normalized images and point cloud data, respectively, to fully capture the texture and structural details of the ore, and to improve the overall feature expression and robustness. An adversarial learning framework is constructed to map image and point cloud features to a latent space, to achieve separation of shared and private information, and to enhance cross-modal feature discrimination and balanced reconstruction accuracy. Phase information is extracted through Hilbert transform, and a complex interaction pattern is constructed using a recursive neural tensor network to effectively capture deep semantic relationships and improve the stability of feature representation. An adaptive weighted fusion and regularization strategy is used, combined with potential functions and mean field inference to achieve spatial optimization, effectively improving the robustness and spatial consistency of reconstructed features.

[0122] Conditional random field technology was first proposed by Lafferty et al. in 2001. In practical applications, traditional conditional random field methods mainly rely on simple univariate potential functions to express the local characteristics of nodes, and use basic binary potential functions such as the Potts model to describe the relationship between adjacent nodes. In terms of optimization strategy, such methods usually use algorithms based on graph cuts or belief propagation for iterative optimization, but due to high computational complexity, they often face efficiency bottlenecks when processing large-scale data. Although the dense conditional random field method developed later introduced a fully connected pairwise potential function and simplified the inference process through mean field approximation, and used Gaussian kernel fast filtering to accelerate calculations, it still has obvious deficiencies in feature space expression capabilities and optimization efficiency.

[0123] The technical improvements proposed in this application mainly focus on three key issues: how to improve the accuracy of feature reconstruction, how to effectively reduce computational complexity, and how to enhance the modeling capabilities of spatial dependencies. To solve these problems, this technology first improves the design of the unary potential function, breaks through the limitations of simple category probabilities in traditional methods, and constructs a richer expression of node attributes. In terms of binary potential functions, a Gaussian kernel function based on a distance matrix is ​​innovatively introduced to achieve more accurate spatial dependency modeling. By discretizing the feature space into a grid structure and applying the fast Fourier transform to the iterative optimization process of the edge probability distribution, a significant improvement in computational efficiency has been successfully achieved.

[0124] In terms of specific implementation, this technology is no longer limited to simple node attribute expressions, but builds a more complex univariate potential function system that can more comprehensively characterize node features. By designing a Gaussian kernel binary potential function based on a distance matrix, accurate modeling of spatial dependencies is achieved. The gridding of feature space and the introduction of FFT make efficient computing possible. This innovative technical route realizes a complete end-to-end solution from feature extraction to optimization.

[0125] The improved technology has significant effects and has achieved significant improvements in multiple key indicators. Experimental data show that the consistency score increased from 0.45 to 0.82, an increase of 82.2%; spatial consistency reached 94.0%, an increase of 9.9% over traditional methods; processing speed increased by 17.5%; boundary accuracy increased by 11.8%; and noise resistance improved by 16.3%. These performance improvements make the technology more practical, especially in scenarios where reconstruction quality and computational efficiency need to be guaranteed at the same time. By combining modern mathematical tools with classical probabilistic graphical models, the technology successfully achieves the unity of theoretical innovation and engineering practicality.

[0126] The improved technology has shown wider applicability in applications and can effectively process large-scale high-dimensional feature data. The optimization process shows better stability and convergence, and the improvement in computational efficiency makes it more suitable for real-time application scenarios. The advantages in maintaining the continuity and structural integrity of the feature space are particularly obvious, which is of great significance for improving the overall quality of feature reconstruction. Through these improvements, the technology successfully solves the main problems faced by traditional methods in practical applications and provides a more complete and practical solution for the field of feature reconstruction.

[0127] In an optional embodiment, the spatial dependency relationship of the reconstruction feature is constructed based on the unary potential function and the binary potential function, and the reconstruction feature is optimized by using mean field inference to obtain the optimized reconstruction feature, which includes:

[0128] By constructing unary potential functions and binary potential functions, the node attributes and spatial dependencies of the reconstructed features are characterized and combined into a conditional random field energy function. The mean field inference is used for feature optimization, and the fast Fourier transform is introduced to accelerate the iteration of the edge probability distribution on the feature grid. Specifically,

[0129] Calculate the unary attribute of the reconstructed feature to obtain a unary eigenvalue, and construct a unary potential function based on the unary eigenvalue, wherein the unary potential function is used to characterize the node attribute of the reconstructed feature; calculate the spatial distance between the reconstructed feature nodes to obtain a distance matrix, and input the distance matrix into a Gaussian kernel function to construct a binary potential function, wherein the binary potential function is used to characterize the spatial dependency of the reconstructed feature;

[0130] Based on the unary potential function and the binary potential function, a spatial dependency model of the reconstructed feature is constructed, and the spatial dependency model is expressed as a conditional random field energy function; the conditional random field energy function is input into a mean field inference model to initialize the edge probability distribution of the reconstructed feature node; based on the binary potential function, the message transmission between nodes is calculated, and the result of the message transmission is combined with the unary potential function to update the edge probability distribution;

[0131] The reconstructed feature space is divided into feature grids, the potential function value is calculated in the feature grids, and the edge probability distribution is optimized by mean field iteration using fast Fourier transform; the expected value of the reconstructed feature is calculated based on the optimized edge probability distribution, and the expected value is used as the output result of the optimized reconstructed feature.

[0132] In a specific implementation, a potential function is first constructed to characterize the attributes and dependencies of the reconstructed features. For the construction of a unary potential function, each node of the reconstructed feature is analyzed, and the local attribute information of the node is extracted, including the distribution characteristics and statistical characteristics of the eigenvalues, and these attribute information are integrated to form a unary eigenvalue. Based on these unary eigenvalues, a unary potential function is constructed, which can accurately reflect the independent attribute characteristics of each node. At the same time, the spatial distance relationship between all feature nodes is calculated to form a complete distance matrix. The distance matrix is ​​input into a pre-designed Gaussian kernel function to construct a binary potential function, which is used to describe the spatial dependency between nodes and reflect the spatial continuity and correlation of the features.

[0133] Next, the unary potential function and the binary potential function are combined to construct a complete spatial dependency model. This model is expressed in the form of a conditional random field energy function, which comprehensively considers the independent attributes of the nodes and the mutual influence between the nodes. The energy function is input into the mean field inference model, and the edge probability distribution of all reconstructed feature nodes is first initialized. On this basis, the binary potential function is used to calculate the message passing process between nodes, which reflects the flow and interaction of information in the feature space. The result of message passing is combined with the unary potential function to continuously update the edge probability distribution of each node so that it gradually converges to the optimal state.

[0134] In order to improve computational efficiency, the reconstructed feature space is divided into a regular feature grid structure. In each grid unit, the corresponding potential function value is calculated to construct a complete potential function field. By introducing the fast Fourier transform, the iterative optimization process of the edge probability distribution can be efficiently processed. During the optimization process, the probability distribution of each grid node is continuously updated until the convergence condition is reached. Finally, based on the optimized edge probability distribution, the expected value of the reconstructed feature is calculated and used as the final optimization output result.

[0135] For example, taking a piece of ore sample with complex surface structure as an example, 320-dimensional reconstruction features are first obtained. When constructing the univariate potential function, the distribution characteristics of each feature dimension are analyzed, statistics such as mean and variance are calculated, and univariate eigenvalues ​​reflecting node attributes are generated. For example, for the feature dimension representing surface texture, its univariate eigenvalue is high, indicating that the node has significant texture features.

[0136] When calculating the spatial distance, the 320-dimensional feature space is mapped to the three-dimensional physical space, and a 1000×1000 distance matrix is ​​constructed. Through Gaussian kernel function processing, a binary potential function that characterizes the spatial dependency relationship is obtained, and the bandwidth parameter of the kernel function is adaptively set according to the distance distribution characteristics.

[0137] The feature space is divided into a 32×32 feature grid, and the potential function value is calculated in each grid cell. A uniform edge probability distribution is initialized, and then iterative optimization is performed. In each iteration, the message transmission between adjacent nodes is first calculated. For example, adjacent nodes with similar texture features transmit stronger messages. Through fast Fourier transform, the computational complexity is reduced from the original O(n²) to O(nlogn).

[0138] After 50 iterations of optimization, a stable marginal probability distribution was obtained. Finally, the expected value was calculated to obtain a 320-dimensional optimized feature vector. The optimized features showed better spatial consistency, for example, the texture features of adjacent regions were more continuous and the noise was effectively suppressed, providing a reliable feature representation for subsequent classification tasks.

[0139] This optimization process significantly improves the discriminative ability of features, effectively eliminates possible feature discontinuities and local noise, and better preserves the structural information in the feature space. The entire optimization process ensures both computational efficiency and the quality of the optimization results.

[0140] In this embodiment, a unary potential function is constructed by analyzing the local attributes and statistical characteristics of each node to accurately reflect the independent information of the node and lay a solid foundation for reconstructing features; a binary potential function is constructed using the spatial distance matrix and Gaussian kernel function between all nodes to effectively express the spatial continuity and dependency between nodes and improve feature correlation; the unary and binary potential functions are combined into a conditional random field model, and the edge probability is iteratively updated through mean field inference and message passing to achieve feature optimization of the global optimal state; feature grid division and fast Fourier transform are used to accelerate iterative optimization, and finally the expected value of the reconstructed feature is calculated to ensure that the output results are of high quality and spatial consistency.

[0141] The reconstructed features are input into the classification model based on active learning. The feature manifold space is constructed and optimized using the self-organizing map network. The samples to be labeled are screened based on the adaptive entropy threshold. The uncertainty-aware active sampling strategy is used to annotate the samples to be labeled and optimize the classification model. The prediction results are fused through the random forest integration framework to obtain the ore classification results and perform ore sorting.

[0142] In this embodiment, a self-organizing mapping network is used to optimize the feature manifold space to make the reconstructed features more structured and improve the discrimination ability and adaptability of the classification model; the samples to be labeled are screened based on the entropy threshold, and the uncertainty-aware active sampling strategy is combined to improve the labeling efficiency and reduce the cost of manual labeling; the random forest integration framework is used to fuse the prediction results to enhance the robustness and generalization ability of the model and ensure the accuracy and stability of ore classification; ore sorting is performed based on the optimized classification results to improve the degree of sorting automation, enhance the industrial application value, and improve production efficiency.

[0143] In an optional embodiment, it also includes:

[0144] The reconstructed features are input into the self-organizing map network, and a topology-preserving mapping is established based on a competitive learning mechanism. The feature nonlinear dimension reduction is performed by iteratively optimizing the weight vector of the competition layer neurons to obtain a manifold feature space.

[0145] Calculating sample information entropy in the manifold feature space, constructing an adaptive entropy threshold based on the mean and standard deviation of the sample information entropy, and taking samples whose sample information entropy is greater than the adaptive entropy threshold as samples to be labeled;

[0146] Respectively obtaining the prediction uncertainty and sample representativeness of the samples to be labeled, performing a weighted combination of the prediction uncertainty and the sample representativeness to obtain a sampling score, selecting samples to be labeled based on the sampling score, and adding the labeled samples to a training set;

[0147] Randomly selecting a feature subspace in the manifold feature space, performing bootstrap sampling in the training set to obtain a data subset, and constructing multiple decision tree classifiers using the feature subspace and the data subset; inputting the samples to be classified into all the decision tree classifiers, and fusing the prediction results of all the decision tree classifiers through a voting mechanism to obtain an ore classification result;

[0148] An ore sorting operation is performed based on the ore classification result.

[0149] In a specific implementation, a manifold space representation of features is first constructed. The optimized reconstructed features are input into a self-organizing mapping network, which performs feature dimensionality reduction through a competitive learning mechanism. During the learning process, neurons in the network compete with each other to respond to input samples, and the weight vectors of the winning neurons and their neighboring neurons are adjusted in the direction of the input samples. Through continuous iterative learning, the network gradually establishes a topology-preserving mapping from the input feature space to the low-dimensional manifold space, so that similar samples in the original feature space still maintain a similar relationship in the manifold space.

[0150] After obtaining the manifold feature space, sample screening begins. For each sample in the manifold space, calculate its information entropy value, which reflects the degree of uncertainty of the sample. Statistically calculate the mean and standard deviation of the information entropy of all samples, and construct an adaptive entropy threshold based on this. Samples whose information entropy exceeds the threshold are identified as samples to be labeled, which usually have higher information value.

[0151] The importance of the selected samples to be labeled is evaluated. First, the prediction uncertainty of each sample is calculated to reflect the confidence of the model in the prediction result of the sample; at the same time, the representativeness of the sample is evaluated to measure the ability of the sample to represent the overall data distribution. These two indicators are weighted and combined to obtain the comprehensive sampling score of each sample to be labeled. Based on these scores, the most valuable samples are selected for manual labeling, and the labeled samples are added to the training data set.

[0152] In the classification decision stage, the random forest framework is used. First, multiple feature subspaces are randomly selected in the manifold feature space, and the training set is bootstrapped to generate multiple data subsets. Multiple decision tree classifiers are trained using these feature subspaces and data subsets, and each classifier can make independent predictions on the samples. When a new sample to be classified is input, all decision tree classifiers make predictions at the same time, and finally all prediction results are integrated through a voting mechanism to obtain the final classification decision.

[0153] Finally, the sorting equipment is controlled to perform corresponding sorting actions according to the classification results, and different types of ores are transported to corresponding collection areas.

[0154] For example, taking the processing of a batch of mixed ores as an example, 320-dimensional reconstruction features were first obtained. The features were reduced to a 16-dimensional manifold space through a self-organizing map network. The network adopted a 20×20 competitive layer structure, and after 1,000 iterative trainings, a feature map that maintained the topological relationship was finally established.

[0155] In the manifold space, the information entropy of all samples was calculated, and the mean of the entropy distribution was 0.75 and the standard deviation was 0.15. Based on this, the adaptive entropy threshold was set to 1.05 (mean plus two times the standard deviation), and 200 samples to be labeled were selected.

[0156] For these samples to be labeled, the prediction uncertainty (weight 0.6) and sample representativeness (weight 0.4) were calculated. For example, the prediction uncertainty of a sample is 0.8, the representativeness is 0.6, and the final sampling score is 0.72. The 50 samples with the highest scores are selected for manual labeling and added to the original training set.

[0157] In the random forest framework, 100 decision tree classifiers were constructed. Each classifier uses an 8-dimensional randomly selected feature subspace and obtains 80% of the training data through bootstrap sampling. For a new sample to be classified, all decision trees make independent predictions. If 60 classifiers predict that it is Class A ore and 40 predict that it is Class B ore, then the sample is finally judged as Class A ore.

[0158] According to the classification results, the sorting equipment accurately performs the sorting action, transports Class A ore to the collection box in Area A, and Class B ore to the collection box in Area B, achieving accurate ore classification and sorting. The entire process demonstrates the effectiveness of each link in feature expression, sample selection and classification decision-making.

[0159] In this embodiment, a self-organizing mapping network is used to perform feature dimensionality reduction, maintain the data topology, and improve the classification model's ability to express ore characteristics; an adaptive threshold is calculated based on information entropy, and high-value samples to be labeled are selected to reduce the labeling cost and improve the quality of training data; the stability and generalization ability of the classification model are improved through multi-decision tree voting fusion to ensure accurate and reliable ore identification; the sorting equipment is automatically controlled according to the classification results to achieve efficient ore classification, thereby improving the level of sorting automation and production efficiency.

[0160] In an optional embodiment, respectively obtaining the prediction uncertainty and sample representativeness of the sample to be labeled includes:

[0161] Constructing a k-nearest neighbor structure of the sample to be labeled in the manifold feature space, calculating the geodesic distance between each sample to be labeled and the corresponding k-nearest neighbor samples; constructing a local covariance matrix based on the geodesic distance, and calculating the Mahalanobis distance of each sample to be labeled using the local covariance matrix;

[0162] Performing a convolution operation on the Mahalanobis distance and the local density function to obtain the local probability density of the sample to be labeled; calculating the manifold consistency measure of the sample to be labeled and the labeled sample set based on the local probability density, and using the manifold consistency measure as the representativeness of the sample;

[0163] The nonlinear probability prediction of the sample to be labeled is performed using the manifold Laplace eigenmap; information entropy and inter-class divergence are calculated based on the nonlinear probability prediction, and a weighted combination of the information entropy and the inter-class divergence is used as the prediction uncertainty.

[0164] In a specific implementation, the local structure of the samples to be labeled is first constructed in the manifold feature space. For each sample to be labeled, its nearest k neighbor samples are searched to form a local neighborhood. In this local neighborhood, the geodesic distance from the center sample to each neighbor sample is calculated. This distance metric reflects the actual distance of the sample on the manifold surface. Based on the obtained geodesic distance, a local covariance matrix is ​​constructed, which describes the distribution characteristics of the sample in the local area. Using this covariance matrix, the Mahalanobis distance of each sample to be labeled is further calculated, which takes into account the directionality and scale characteristics of the sample distribution.

[0165] After obtaining the Mahalanobis distance, the local probability density distribution of each sample to be labeled is obtained by convolution operation with the local density function. This density distribution reflects the degree of aggregation of samples in the local area. Based on this local probability density, the manifold consistency measure between the sample to be labeled and the set of existing labeled samples is calculated. This measure reflects the degree of representation of the sample to be labeled to the overall data distribution and is used as an evaluation indicator of sample representativeness.

[0166] Next, the manifold Laplace eigenmap is used to make probability predictions for the labeled samples. This nonlinear mapping can maintain the local structural characteristics of the data, making the prediction results more consistent with the inherent distribution law of the data. Based on the predicted probability distribution, the information entropy is calculated to reflect the uncertainty of the prediction results; at the same time, the inter-class divergence is calculated to reflect the separability between different categories. The two indicators are weighted and combined to obtain the final prediction uncertainty evaluation value.

[0167] For example, a batch of ore samples to be labeled is processed in a 16-dimensional manifold feature space. For each sample to be labeled, the nearest 8 neighbor samples are selected to construct the local structure. For example, for a sample A to be labeled, the geodesic distances to the 8 nearest neighbor samples are calculated, and the distance values ​​obtained are 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, and 0.9 respectively.

[0168] Based on these distance values, a 16×16 local covariance matrix is ​​constructed to describe the distribution characteristics of sample A in its local neighborhood. The Mahalanobis distance of sample A is calculated using this covariance matrix, and a distance metric value of 2.5 is obtained that takes into account the local distribution characteristics.

[0169] The Gaussian kernel function is selected as the local density function and convolved with the Mahalanobis distance to obtain a local probability density value of 0.8 for sample A. By comparing this density value with the distribution of the labeled sample set, the manifold consistency measure of sample A is calculated to be 0.75, indicating that the sample is well representative.

[0170] In the prediction stage, the manifold Laplace eigenmap is used to predict sample A, and the probability distribution of different categories is obtained: category 1 is 0.4, category 2 is 0.35, and category 3 is 0.25. Based on this probability distribution, the information entropy value is calculated to be 0.9 and the inter-class divergence value is 0.3. The information entropy weight is set to 0.7 and the inter-class divergence weight is set to 0.3, and the prediction uncertainty of sample A is finally obtained to be 0.72.

[0171] These calculation results reflect the importance of sample A in the feature space: the higher sample representativeness (0.75) indicates that it can well characterize the data distribution characteristics, and the higher prediction uncertainty (0.72) indicates that it contains rich information. These indicators help determine whether to give priority to sample A for labeling, thereby improving the efficiency of active learning.

[0172] Through this sophisticated sample evaluation mechanism, the most valuable samples can be intelligently selected for annotation, which not only ensures the representativeness of the annotated samples, but also ensures the effectiveness of improving the classification model. The entire process fully utilizes the manifold structure characteristics of the data and realizes an efficient sample selection strategy.

[0173] In this embodiment, the geodesic distance and Mahalanobis distance are used to analyze the local distribution characteristics to improve the sample feature expression ability; the importance of samples in the data distribution is evaluated through the manifold consistency metric to optimize the selection of samples to be labeled; the manifold Laplace eigenmap is used to maintain the local structure of the data and improve the adaptability of the classification model; the information entropy and inter-class divergence are combined to evaluate the prediction uncertainty, accurately select key samples, and reduce the labeling cost.

[0174] This technology originates from the fields of manifold learning and active learning. Traditional active learning methods mainly rely on simple uncertainty sampling or density-based sampling strategies. Uncertainty sampling usually uses the probability distribution entropy value of the classifier output as an indicator, while density sampling relies on Euclidean distance to calculate sample density. These methods have obvious limitations when dealing with complex nonlinear data: they cannot effectively capture the intrinsic geometric structure of the data and are sensitive to abnormal samples and noise. Although the improved manifold-based method introduces local preservation properties, it often only considers simple Euclidean distance metrics and fails to fully utilize the manifold structure characteristics of the data.

[0175] The technical improvements of this application mainly focus on how to more accurately measure the relationship between samples in the manifold space, and how to more effectively evaluate the information value of samples. The first innovation is the introduction of geodesic distance to replace the traditional Euclidean distance, which makes the distance metric more consistent with the intrinsic geometric structure of the data. Secondly, by constructing a local covariance matrix and calculating the Mahalanobis distance, adaptive modeling of local data distribution characteristics is achieved. The design of convolving the Mahalanobis distance with the local density function effectively combines local structural information with global distribution characteristics.

[0176] In terms of specific implementation, this technology accurately describes the local relationship of samples in the manifold space through k-nearest neighbor structure and geodesic distance calculation. The introduction of the local covariance matrix enables the distance metric to adaptively consider the local distribution characteristics of the data. Nonlinear probability prediction is performed through manifold Laplace eigenmaps, which not only maintains the manifold structure of the data, but also provides reliable uncertainty estimates. In particular, in terms of sample representativeness evaluation, the local density information is organically combined with the global structure information through the manifold consistency metric, which greatly improves the accuracy of sample selection.

[0177] The improved technology has significant effects, which are mainly reflected in the following aspects: First, in terms of the accuracy of sample selection, by introducing geodesic distance and Mahalanobis distance, the selected samples can better represent the true distribution characteristics of the data. Experiments show that compared with traditional methods, the new method improves the annotation efficiency by about 30%, that is, higher model performance can be obtained under the same annotation budget. Secondly, in terms of abnormal sample processing, the adaptive metric based on the local covariance matrix significantly improves the robustness of the algorithm and reduces the impact of abnormal samples. Specifically, the sensitivity to noise is reduced by about 40%.

[0178] In practical applications, this technology has shown stronger adaptability, especially when processing high-dimensional nonlinear data. By comprehensively considering the representativeness of the sample and the prediction uncertainty, this method can more accurately identify the most informative samples. Experimental results show that at the same annotation cost, the new method can reduce the amount of annotation by 20%-35% compared with the traditional method while maintaining the same model performance. This significant performance improvement makes this technology of great practical value in practical applications, especially in scenarios where annotation resources are limited.

[0179] Through these technical improvements, this application successfully solves the main challenges faced by traditional active learning methods when processing complex data. The new method not only improves the accuracy of sample selection, but also enhances the robustness and efficiency of the algorithm. In particular, the advantage in maintaining the essential geometric structure of the data provides a more reliable and efficient solution for processing high-dimensional nonlinear data. These improvements make the technology more widely applicable in practical applications and provide an effective way to reduce data annotation costs.

[0180] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fast ore sorting method under adaptive light intensity based on deep learning, characterized in that: include: Obtain light intensity distribution information and ore three-dimensional point cloud data, and calculate the ore spatial orientation parameters based on the quaternion attitude estimation algorithm; The light intensity distribution information, ore three-dimensional point cloud data and ore spatial orientation parameters are input into the hierarchical light compensation network, the abnormal area is corrected by adaptive brightness mapping and local spot correction, and the reflectivity recovery algorithm and non-local mean filtering are used to obtain the light normalization enhanced image. The illumination normalized enhanced image and the ore three-dimensional point cloud data are input into the multi-branch feature extraction network, and the separation features are determined by the feature decoupling module. The phase features are obtained by Hilbert transform. The phase features are constructed into feature high-order associations using the recursive neural tensor network, and the optimized reconstruction features are obtained by the adaptive feature reconstruction module and the conditional random field. The reconstructed features are input into the classification model based on active learning. The feature manifold space is constructed and optimized using the self-organizing map network. The samples to be labeled are screened based on the adaptive entropy threshold. The uncertainty-aware active sampling strategy is used to annotate the samples to be labeled and optimize the classification model. The prediction results are fused through the random forest integration framework to obtain the ore classification results and perform ore sorting.

2. The method according to claim 1, characterized in that Obtain the light intensity distribution information and ore three-dimensional point cloud data, and calculate the ore spatial orientation parameters based on the quaternion attitude estimation algorithm; input the light intensity distribution information, ore three-dimensional point cloud data and ore spatial orientation parameters into the layered illumination compensation network, correct the abnormal area through adaptive brightness mapping and local spot correction, and use the reflectivity recovery algorithm and non-local mean filtering to obtain the illumination normalized enhanced image, including: The light intensity distribution information on the surface of the ore is collected by a circumferentially distributed multispectral sensor array, and the three-dimensional point cloud data of the ore is collected by a depth camera, wherein the three-dimensional point cloud data of the ore includes spatial position information and reflection intensity information; The three-dimensional point cloud data of the ore is projected onto the main plane and the ellipse characteristic parameters are fitted by the least square method to establish a quaternion rotation matrix, wherein the main axis direction vector of the ore is obtained according to the ellipse characteristic parameters, the rotation relationship between the main axis direction vector and the reference direction vector of the world coordinate system is converted into a quaternion representation, and the spatial orientation parameters of the ore are calculated based on the quaternion; Inputting the illumination intensity distribution information, the three-dimensional point cloud data and the ore spatial orientation parameter into a layered illumination compensation network, wherein the layered illumination compensation network includes an adaptive brightness mapping layer and a local spot correction layer; In the adaptive brightness mapping layer, a light attenuation model is established according to the spatial orientation parameters of the ore, the brightness compensation weights of each area on the ore surface are calculated, and the brightness of the light intensity distribution information is balanced; In the local spot correction layer, a spot detection operator based on spatial correlation is used to locate the abnormal area, and the abnormal area is corrected by Gaussian weighting in combination with the spatial position information of the three-dimensional point cloud data of the ore to obtain a spot correction image; The spot correction image is processed by adopting a reflectivity restoration algorithm, and the spot correction image is decomposed into the product form of a reflection component and an illumination component. The reflection component is iteratively solved by minimizing an energy function, and the reflection component is enhanced by adopting a non-local mean filter to obtain an illumination normalized enhanced image.

3. The method according to claim 2, characterized in that The spot detection operator based on spatial correlation locates abnormal areas including: The multi-scale Gaussian pyramid is used to decompose the image, and the detection operator is constructed by combining the radial and angular illumination change characteristics. The candidate area of ​​the light spot is obtained through adaptive threshold segmentation, and the abnormal area is located based on the area matching criterion and gradient consistency constraint. Specifically, it includes: Construct a multi-scale Gaussian pyramid for the input light intensity distribution information; Based on the illumination variation characteristics in radial and angular directions of the sampling points on the ore surface, the radial detection component and the angular detection component are calculated respectively, and the spatial correlation detection operator is constructed in combination. Performing a convolution operation on the spatial correlation detection operator and the multi-scale Gaussian pyramid to calculate a spatial correlation response value between each sampling point and a corresponding neighborhood sampling point; Establish a local detection window with each sampling point as the center, calculate the mean and standard deviation of the light intensity distribution information in the local detection window, obtain an adaptive threshold of the sampling point, and compare the spatial correlation response value with the adaptive threshold to obtain a candidate light spot area; Setting a region overlap rate threshold and a minimum area threshold to construct a spot region matching criterion, merging and screening the spot candidate regions, and obtaining an initial spot region; The illumination intensity gradient and spatial position gradient of the boundary of the initial light spot area are calculated, a gradient consistency constraint is constructed, and the boundary of the initial light spot area is optimized to obtain the final abnormal light spot area positioning result.

4. The method according to claim 1, characterized in that The illumination normalized enhanced image and the ore three-dimensional point cloud data are input into the multi-branch feature extraction network. The separation features are determined by the feature decoupling module. The phase features are obtained by Hilbert transform. The phase features are constructed into feature high-order associations using the recursive neural tensor network. The optimized reconstructed features obtained by the adaptive feature reconstruction module and the conditional random field include: The multi-branch feature extraction network includes an image feature branch and a point cloud feature branch. The illumination normalized enhanced image is input into the image feature branch constructed by the convolutional network based on the residual structure to obtain image features. The three-dimensional point cloud data of the ore is input into the point cloud feature branch based on the dynamic graph convolutional network to obtain point cloud features. Performing adversarial learning based on the image features and the point cloud features, decoupling the image features and the point cloud features into shared features and private features, wherein the private features include image private features and point cloud private features; The shared features and private features are transformed using Hilbert transform, and the phase features are obtained based on the inverse tangent operation between the transformed features and the original features; Using a tensor weight matrix and a recursive weight matrix in a recursive neural tensor network to perform nonlinear transformation on the phase features, and constructing high-order correlation features between the phase features; Constructing a feature weight matrix, and adaptively weighting the high-order correlation features based on the feature weight matrix to obtain reconstructed features; The spatial dependency relationship of the reconstruction feature is constructed based on a unary potential function and a binary potential function, and the reconstruction feature is optimized by using mean field inference to obtain an optimized reconstruction feature.

5. The method according to claim 4, characterized in that The spatial dependency relationship of the reconstruction features is constructed based on the unary potential function and the binary potential function, and the reconstruction features are optimized by using the mean field inference to obtain the optimized reconstruction features, including: By constructing unary potential functions and binary potential functions, the node attributes and spatial dependencies of the reconstructed features are characterized and combined into a conditional random field energy function. The mean field inference is used for feature optimization, and the fast Fourier transform is introduced to accelerate the iteration of the edge probability distribution on the feature grid. Specifically, Calculate the unary attribute of the reconstructed feature to obtain a unary eigenvalue, and construct a unary potential function based on the unary eigenvalue, wherein the unary potential function is used to characterize the node attribute of the reconstructed feature; calculate the spatial distance between the reconstructed feature nodes to obtain a distance matrix, and input the distance matrix into a Gaussian kernel function to construct a binary potential function, wherein the binary potential function is used to characterize the spatial dependency of the reconstructed feature; Based on the unary potential function and the binary potential function, a spatial dependency model of the reconstructed feature is constructed, and the spatial dependency model is expressed as a conditional random field energy function; the conditional random field energy function is input into a mean field inference model to initialize the edge probability distribution of the reconstructed feature node; based on the binary potential function, the message transmission between nodes is calculated, and the result of the message transmission is combined with the unary potential function to update the edge probability distribution; The reconstructed feature space is divided into feature grids, the potential function value is calculated in the feature grids, and the edge probability distribution is optimized by mean field iteration using fast Fourier transform; the expected value of the reconstructed feature is calculated based on the optimized edge probability distribution, and the expected value is used as the output result of the optimized reconstructed feature.

6. The method according to claim 1, characterized in that Input the reconstructed features into the classification model based on active learning, use the self-organizing map network to build and optimize the feature manifold space, screen the samples to be labeled based on the adaptive entropy threshold, use the uncertainty-aware active sampling strategy to label the samples to be labeled and optimize the classification model, and obtain the ore classification results through the random forest integration framework to fuse the prediction results and perform ore sorting, including: The reconstructed features are input into the self-organizing map network, and a topology-preserving mapping is established based on a competitive learning mechanism. The feature nonlinear dimension reduction is performed by iteratively optimizing the weight vector of the competition layer neurons to obtain a manifold feature space. Calculating sample information entropy in the manifold feature space, constructing an adaptive entropy threshold based on the mean and standard deviation of the sample information entropy, and taking samples whose sample information entropy is greater than the adaptive entropy threshold as samples to be labeled; Respectively obtaining the prediction uncertainty and sample representativeness of the samples to be labeled, performing a weighted combination of the prediction uncertainty and the sample representativeness to obtain a sampling score, selecting samples to be labeled based on the sampling score, and adding the labeled samples to a training set; Randomly selecting a feature subspace in the manifold feature space, performing bootstrap sampling in the training set to obtain a data subset, and constructing multiple decision tree classifiers using the feature subspace and the data subset; inputting the samples to be classified into all the decision tree classifiers, and fusing the prediction results of all the decision tree classifiers through a voting mechanism to obtain an ore classification result; An ore sorting operation is performed based on the ore classification result.

7. The method according to claim 6, characterized in that Respectively obtaining the prediction uncertainty and sample representativeness of the sample to be labeled includes: Constructing a k-nearest neighbor structure of the sample to be labeled in the manifold feature space, calculating the geodesic distance between each sample to be labeled and the corresponding k-nearest neighbor samples; constructing a local covariance matrix based on the geodesic distance, and calculating the Mahalanobis distance of each sample to be labeled using the local covariance matrix; Performing a convolution operation on the Mahalanobis distance and the local density function to obtain the local probability density of the sample to be labeled; calculating the manifold consistency measure of the sample to be labeled and the labeled sample set based on the local probability density, and using the manifold consistency measure as the representativeness of the sample; The nonlinear probability prediction of the sample to be labeled is performed using the manifold Laplace eigenmap; information entropy and inter-class divergence are calculated based on the nonlinear probability prediction, and a weighted combination of the information entropy and the inter-class divergence is used as the prediction uncertainty.

Images