Fast Ore Sorting Method under Adaptive Light Intensity Based on Deep Learning

Abstract

The present invention provides a method for rapid sorting of ores under adaptive light intensity based on deep learning, which relates to the technical field of ore sorting. The method includes obtaining light intensity distribution information and three-dimensional point cloud data of ores, and calculating ore spatial orientation parameters based on quaternion attitude estimation; inputting the obtained data into a hierarchical light compensation network, and obtaining a light-normalized enhanced image through adaptive brightness mapping and local spot correction, combined with reflectance recovery and non-local mean filtering; inputting the light-normalized enhanced image and the three-dimensional point cloud data of ores into a multi-branch feature extraction network, constructing high-order feature associations, and optimizing features by using adaptive feature reconstruction and conditional random fields; inputting the reconstructed features into a classification model based on active learning, optimizing the feature manifold space by using a self-organizing mapping network, combining adaptive entropy threshold screening and uncertainty-aware active sampling for sample annotation, 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 particularly to a rapid ore sorting method based on deep learning under adaptive light intensity. Background Art

[0002] With the development of intelligent manufacturing technology, intelligent recognition and automated classification technologies in the ore sorting process have been widely applied. Traditional ore sorting mainly relies on manual experience judgment, with low efficiency and being easily affected by subjective factors. In recent years, ore intelligent sorting systems based on machine vision and deep learning have gradually become a research hotspot. By acquiring multi-modal data such as light information and three-dimensional point clouds of ores, rapid classification of ores is achieved in combination with deep learning algorithms.

[0003] However, the existing technologies mainly have the following problems: the light intensity distribution on the ore surface is uneven, resulting in inaccurate feature extraction; traditional feature extraction methods are difficult to effectively represent the multi-scale feature information of ores; the cost of sample annotation is high, and it is difficult to obtain sufficient annotated data for model training; the robustness of single-feature classification methods is insufficient, and they are easily interfered by environmental factors and cause misjudgment. These problems seriously restrict the actual application effect of ore intelligent sorting systems.

[0004] In summary, there is an urgent need to provide a rapid ore sorting method that can adapt to complex lighting conditions, has high-efficient feature extraction ability, reduces the cost of sample annotation, and has stable classification performance. By introducing a light compensation mechanism, a multi-modal feature fusion strategy, and an active learning framework, accurate classification of ores is achieved, and the sorting efficiency and accuracy are improved. Summary of the Invention

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

[0006] In the first aspect of the embodiment of the present invention,

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

[0008] Obtain the light intensity distribution information and the three-dimensional point cloud data of the ore, and calculate the ore spatial orientation parameters based on the quaternion attitude estimation algorithm; input the light intensity distribution information, the three-dimensional point cloud data of the ore, and the ore spatial orientation parameters into a hierarchical light compensation network, correct the abnormal areas through adaptive brightness mapping and local spot correction, and obtain a light-normalized enhanced image by using a reflectance recovery algorithm and non-local mean filtering;

[0009] Input the light-normalized enhanced image and the three-dimensional point cloud data of the ore into a multi-branch feature extraction network. Determine the separated features through the feature decoupling module, obtain the phase features by performing Hilbert transform on the separated features, construct the high-order feature correlations of the phase features using a recursive neural tensor network, and obtain the optimized reconstructed features through the adaptive feature reconstruction module and the conditional random field;

[0010] Input the reconstructed features into the classification model based on active learning. Use the self-organizing mapping network to construct and optimize the feature manifold space. Screen the samples to be labeled based on the adaptive entropy threshold. Adopt the uncertainty-aware active sampling strategy to label the samples to be labeled and optimize the classification model. Fusion the prediction results through the random forest ensemble framework to obtain the ore classification result and perform ore sorting.

[0011] In an alternative embodiment,

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

[0013] Collect the light intensity distribution information on the surface of the ore through a multi-spectral sensor array distributed in a circumferential direction, and collect the three-dimensional point cloud data of the ore through a depth camera, where the three-dimensional point cloud data of the ore includes spatial position information and reflection intensity information;

[0014] Project the three-dimensional point cloud data of the ore onto the main plane and fit the elliptical feature parameters by the least squares method to establish a quaternion rotation matrix. Among them, obtain the main axis direction vector of the ore according to the elliptical feature parameters, convert the rotation relationship between the main axis direction vector and the reference direction vector of the world coordinate system into a quaternion representation, and calculate the ore spatial orientation parameters based on the quaternion;

[0015] Input the light intensity distribution information, the three-dimensional point cloud data, and the ore spatial orientation parameters into the hierarchical light compensation network, where the hierarchical light compensation network includes an adaptive brightness mapping layer and a local spot correction layer;

[0016] In the adaptive brightness mapping layer, establish a light attenuation model according to the ore spatial orientation parameters, calculate the brightness compensation weights of each area on the ore surface, and perform brightness equalization on the light intensity distribution information;

[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] A reflectance recovery algorithm is used to process the spot correction image, decompose the spot correction image into the product form of a reflection component and an illumination component, iteratively solve the reflection component by minimizing the energy function, and perform enhancement processing on the reflection component by non-local mean filtering to obtain an illumination-normalized enhanced image.

[0019] In an alternative embodiment,

[0020] Locating the abnormal area by the spot detection operator based on spatial correlation includes:

[0021] The image is decomposed by multi-scale Gaussian pyramid, a detection operator is constructed by combining the illumination change characteristics in the radial and angular directions, the spot candidate area is obtained by adaptive threshold segmentation, and the abnormal area is located based on the region matching criterion and gradient consistency constraint, specifically including:

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

[0023] Based on the illumination change characteristics in the radial direction and angular direction of the ore surface sampling points, calculate the radial detection component and the angular detection component respectively, and combine them to construct a spatial correlation detection operator;

[0024] Perform a convolution operation on the spatial correlation detection operator and the multi-scale Gaussian pyramid, and calculate the spatial correlation response value of each sampling point and its corresponding neighborhood sampling point;

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

[0026] Set the region overlap rate threshold and the minimum area threshold to construct a spot region matching criterion, merge and screen the spot candidate areas, and obtain the initial spot area;

[0027] Calculate the illumination intensity gradient and spatial position gradient of the boundary of the initial spot area, construct a gradient consistency constraint, and perform boundary optimization on the initial spot area to obtain the final positioning result of the abnormal spot area.

[0028] In an alternative embodiment,

[0029] Input the light-normalized enhanced image and the three-dimensional point cloud data of the ore into a multi-branch feature extraction network. Determine the separated features through the feature decoupling module, obtain the phase features by performing the Hilbert transform on the separated features, use a recursive neural tensor network to construct high-order associations of the phase features, and obtain the optimized reconstructed features through the adaptive feature reconstruction module and the conditional random field, including:

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

[0031] Perform adversarial learning based on the image features and the point cloud features, and decouple the image features and the point cloud features into common features and private features. The private features include image private features and point cloud private features;

[0032] Perform the Hilbert transform on the common features and the private features, and obtain the phase features based on the arctangent operation of the transformed features and the original features;

[0033] Perform a non-linear transformation on the phase features using the tensor weight matrix and the recursive weight matrix in the recursive neural tensor network to construct high-order association features between the phase features;

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

[0035] Construct the spatial dependence relationship of the reconstructed features based on the unary potential function and the binary potential function, and optimize the reconstructed features using mean field inference to obtain the optimized reconstructed features.

[0036] In an alternative embodiment,

[0037] Construct the spatial dependence relationship of the reconstructed features based on the unary potential function and the binary potential function, and optimize the reconstructed features using mean field inference to obtain the optimized reconstructed features, including:

[0038] By constructing the unary potential function and the binary potential function, characterize the node attributes and spatial dependence relationship of the reconstructed features, and combine them into the conditional random field energy function. Use mean field inference for feature optimization, and introduce the fast Fourier transform to accelerate the iteration of the marginal probability distribution on the feature grid, specifically including:

[0039] Calculate the unary attributes of the reconstructed features to obtain unary feature values, construct a unary potential function based on the unary feature values, and the unary potential function is used to characterize the node attributes of the reconstructed features; calculate the spatial distance between the nodes of the reconstructed features to obtain a distance matrix, and input the distance matrix into a Gaussian kernel function to construct a binary potential function, and the binary potential function is used to characterize the spatial dependence relationship of the reconstructed features;

[0040] Construct a spatial dependence model of the reconstructed features based on the unary potential function and the binary potential function, and represent the spatial dependence model as a conditional random field energy function; input the conditional random field energy function into an average field inference model, and initialize the marginal probability distribution of the reconstructed feature nodes; calculate the message passing between nodes based on the binary potential function, and combine the result of the message passing with the unary potential function to update the marginal probability distribution;

[0041] Divide the reconstructed feature space into feature grids, calculate the potential function values in the feature grids, and use the fast Fourier transform to perform average field iterative optimization on the marginal probability distribution; calculate the expected value of the reconstructed features based on the optimized marginal probability distribution, and use the expected value as the output result of the optimized reconstructed features.

[0042] In an alternative embodiment,

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

[0044] Input the reconstructed features into a self-organizing mapping network, establish a topology-preserving mapping based on a competitive learning mechanism, perform feature nonlinear dimensionality reduction by iteratively optimizing the weight vectors of the neurons in the competitive layer, and obtain a manifold feature space;

[0045] Calculate the sample information entropy in the manifold feature space, construct an adaptive entropy threshold based on the mean and standard deviation of the sample information entropy, and use the samples with sample information entropy greater than the adaptive entropy threshold as the samples to be labeled;

[0046] Obtain the prediction uncertainty and sample representativeness of the samples to be labeled respectively, perform a weighted combination of the prediction uncertainty and the sample representativeness to obtain a sampling score, select the samples to be labeled for labeling based on the sampling score, and add the labeled samples to the training set;

[0047] Randomly select a feature subspace in the manifold feature space, perform bootstrap sampling on the training set to obtain a data subset, and construct multiple decision tree classifiers using the feature subspace and the data subset; input the sample to be classified into all the decision tree classifiers, and fuse the prediction results of all the decision tree classifiers through a voting mechanism to obtain the ore classification result;

[0048] Perform ore sorting operations based on the ore classification result.

[0049] In an alternative embodiment,

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

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

[0052] Perform a convolution operation on the Mahalanobis distance and the local density function to obtain the local probability density of the sample to be labeled; calculate the manifold consistency measure between the sample to be labeled and the set of labeled samples based on the local probability density, and use the manifold consistency measure as the sample representativeness;

[0053] Perform non-linear probability prediction on the sample to be labeled using manifold Laplacian eigenmaps; calculate the information entropy and between-class divergence based on the non-linear probability prediction, and use the weighted combination of the information entropy and the between-class divergence as the prediction uncertainty.

[0054] In the embodiments of the present invention, a hierarchical illumination compensation and reflectance recovery algorithm is adopted to effectively eliminate the influence of uneven illumination, make the surface details of the ore clearer, and improve the accuracy of subsequent analysis; combining illumination information, point cloud data and ore orientation parameters, using a multi-branch feature extraction network and a recursive neural tensor network to achieve a more comprehensive ore feature modeling and improve the recognition accuracy; through a self-organizing mapping network and an uncertainty-aware sampling strategy, optimize the selection of samples to be labeled, improve the labeling efficiency, reduce the need for manual labeling, and accelerate the optimization of the classification model; adopt a random forest integration framework for ore classification, and combine sorting equipment to achieve automated operation, improve the intelligence level of ore recognition and sorting, and enhance the production efficiency. Description of the Drawings

[0055] Figure 1 It is a flowchart of the method for rapid ore sorting under adaptive illumination intensity based on deep learning in the embodiments of the present invention;

[0056] Figure 2Effect diagram for multi-scale detection response and processing time analysis;

[0057] Figure 3 Effect diagram for six-dimensional performance evaluation of spatial correlation;

[0058] Figure 4 Effect diagram for comparative analysis before and after gradient consistency optimization;

[0059] Figure 5 Effect diagram for comprehensive performance comparison between the method of the present invention and traditional methods. Detailed implementation manners

[0060] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0062] Figure 1 It is a schematic flowchart of the method for rapid ore sorting under adaptive light intensity based on deep learning in the embodiment of the present invention, as Figure 1 shown. The method includes:

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

[0064] In this embodiment, by comprehensively utilizing the light intensity distribution, the three-dimensional point cloud data of the ore, and the spatial orientation parameters obtained from the quaternion-based attitude estimation, the physical characteristics and geometric structure of the ore are effectively captured, providing rich information support for subsequent light compensation. By using a hierarchical light compensation network, through adaptive brightness mapping and local spot correction, the abnormal regions in the image are effectively corrected, reducing the impact of local uneven illumination, thereby making the illumination distribution in the image more uniform. By adopting a reflectivity recovery algorithm and non-local mean filtering, the true reflectivity characteristics of the ore can be restored, enhancing the image details and contrast, improving the overall quality of the image, and enhancing the accuracy of subsequent target recognition and detection. Through multi-data fusion and hierarchical compensation mechanisms, this method has better adaptability to complex illumination environments and local spot interference, ensuring stable and reliable image enhancement effects in different scenarios.

[0065] In an alternative embodiment, it further includes:

[0066] Collect the light intensity distribution information on the surface of the ore through a circumferentially distributed multi-spectral sensor array, and collect the three-dimensional point cloud data of the ore through a depth camera, where the three-dimensional point cloud data of the ore includes spatial position information and reflection intensity information;

[0067] Project the three-dimensional point cloud data of the ore onto the main plane and fit the elliptical feature parameters by the least squares method to establish a quaternion rotation matrix. Among them, according to the elliptical feature parameters, obtain the main axis direction vector of the ore, convert the rotation relationship between the main axis direction vector and the reference direction vector of the world coordinate system into a quaternion representation, and calculate the spatial orientation parameters of the ore based on the quaternion.

[0068] Input the light intensity distribution information, the three-dimensional point cloud data, and the spatial orientation parameters of the ore into a hierarchical light compensation network, where the hierarchical light compensation network includes an adaptive brightness mapping layer and a local spot correction layer;

[0069] In the adaptive brightness mapping layer, establish a light attenuation model according to the spatial orientation parameters of the ore, calculate the brightness compensation weights of each region on the surface of the ore, and perform brightness equalization on the light intensity distribution information;

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

[0071] The spot-corrected image is processed using a reflectance recovery algorithm, which decomposes the spot-corrected image into the product form of a reflection component and an illumination component. The reflection component is iteratively solved by minimizing the energy function, and the reflection component is enhanced using non-local means filtering to obtain an illumination-normalized enhanced image.

[0072] In a specific embodiment, data acquisition is first performed. A plurality of spectral sensors are uniformly arranged circumferentially on the ore sorting conveyor belt to form a sensor array to collect the illumination intensity distribution information on the ore surface. At the same time, a depth camera is used to collect the three-dimensional point cloud data of the ore from the top to obtain the spatial position coordinates and surface reflection intensity values of the ore.

[0073] Then, the ore attitude is calculated. The collected three-dimensional point cloud data is projected onto a horizontal plane, and the least squares method is used to fit an ellipse representing the ore contour to obtain characteristic parameters such as the major axis and minor axis of the ellipse. Based on the ellipse characteristics, the principal axis direction vector of the ore is determined, and the spatial rotation relationship between this vector and the preset reference direction vector of the world coordinate system is represented in the form of quaternions, thereby calculating the specific spatial orientation of the ore on the conveyor belt.

[0074] Next, illumination compensation is performed. The three types of data obtained above (illumination distribution, point cloud data, spatial orientation) are input into an illumination compensation network with a two-layer structure. In the first layer of adaptive brightness mapping, an illumination attenuation mathematical model is established according to the spatial orientation of the ore, and the brightness compensation weight values required for each area on the ore surface are calculated to perform brightness equalization processing on the original illumination intensity distribution. In the second layer of local spot correction, a detection operator based on spatial correlation is designed to locate the 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 in a Gaussian weighted manner to obtain a corrected image.

[0075] Finally, illumination normalization is performed. The reflectance recovery algorithm is applied to the corrected image, which decomposes the image into the multiplicative form of a reflection component and an illumination component. The reflection component is solved by iteratively optimizing the energy function. Then, the reflection component is enhanced using a non-local means filter, and finally an illumination-normalized enhanced image is obtained.

[0076] Exemplarily, taking an irregular iron ore as an example, 8 uniformly distributed spectral sensors and a depth camera are used for data acquisition. The sensors collect obvious uneven illumination on the ore surface, with the central 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 in length and about 10 cm in width.

[0077] The ore is calculated to be deflected by approximately 30 degrees relative to the conveyor belt direction through ellipse fitting. Based on this attitude information, the first layer of the illumination compensation network calculates that the edge area requires a brightness increase of 1.5 - 2 times, and the central area requires a brightness attenuation of 0.7 times. The second layer detects two high - light spot areas with areas of approximately 2 cm² and 1 cm² respectively, and corrects these abnormal areas using the pixel values of the surrounding normal areas.

[0078] Finally, through reflectance recovery and non - local mean filtering, an ore surface image with uniform illumination and clear details is obtained, laying a 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, realizing information complementarity, providing an accurate basis for subsequent attitude calculation and compensation, and improving the overall processing effect; the three - dimensional point cloud data is projected onto the horizontal plane, an ellipse is fitted using the least - squares method to extract the main axis direction of the ore, and its specific orientation on the conveyor belt is determined through quaternion calculation; a two - layer illumination compensation network is constructed. The first layer performs brightness mapping adjustment, and the second layer corrects local light spots. Zonal processing realizes overall illumination balance and abnormal area correction; a reflectance recovery algorithm is used to separate the reflection and illumination components, and non - local mean filtering is used to enhance image details and improve contrast, finally obtaining a normalized enhancement effect.

[0080] In an alternative embodiment, the spot detection operator based on spatial correlation locates abnormal areas, including:

[0081] The image is decomposed using a multi - scale Gaussian pyramid, a detection operator is constructed by combining the radial and angular illumination change features, a candidate spot area is obtained through adaptive threshold segmentation, and the abnormal area is located based on the region matching criterion and gradient consistency constraint. Specifically, it includes:

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

[0083] Based on the illumination change features in the radial direction and angular direction of the ore surface sampling points, calculate the radial detection component and the angular detection component respectively, and combine them to construct a spatial correlation detection operator;

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

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

[0086] Set the threshold of the regional overlap rate and the minimum area threshold to construct the matching criterion for the spot area, merge and screen the candidate spot areas to obtain the initial spot area;

[0087] Calculate the illumination intensity gradient and the spatial position gradient of the boundary of the initial spot area, construct the gradient consistency constraint, and optimize the boundary of the initial spot area to obtain the final positioning result of the abnormal spot area.

[0088] In a specific embodiment, in the first step, the obtained illumination intensity distribution information is processed through Gaussian filtering and downsampling operations to construct a multi-level image pyramid structure. The image resolution of each level decreases gradually, forming image representations at different scales for subsequent multi-scale feature analysis.

[0089] In the second step, analyze the illumination change characteristics of the ore surface sampling points in the radial direction (from the center outwards) and the angular direction (along the circumferential direction). Calculate the radial gradient of the illumination intensity in the radial direction and the circumferential gradient in the angular direction, and combine the detection components in these two directions to form a complete spatial correlation detection operator.

[0090] In the third step, perform a convolution operation on each layer of the image in the constructed detection operator and the multi-scale pyramid. For each sampling point, calculate the spatial correlation response value between it and the surrounding neighborhood sampling points, and this response value reflects the degree of illumination difference between this point and the surrounding area.

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

[0092] In the fifth step, set two threshold conditions of the regional overlap rate and the minimum area, merge and screen the candidate areas obtained in the previous step. When the overlap degree of two candidate areas exceeds the set threshold, merge them, and at the same time, eliminate the areas with too small area to obtain the preliminary spot area.

[0093] In the sixth step, calculate the illumination intensity gradient and the spatial position gradient on the boundary of the preliminarily determined spot area, and construct the gradient consistency constraint condition. Fine-tune the boundary of the spot area according to the change characteristics of the gradient to finally obtain the accurate position of the abnormal spot area.

[0094] Figure 2Shows the changing trends of the detection response values and processing times of the algorithm at different image resolutions. From the perspective of the detection response values, the highest value of 0.92 is reached at the original 1024×1024 resolution. As the resolution decreases to 512×512, it slightly drops to 0.88, and when further decreased to 256×256, it still remains at a relatively high level of 0.85. This slow downward trend indicates that the algorithm has excellent scale adaptability and can maintain a high detection reliability even at low resolutions. In terms of the processing time, there is a significant decreasing trend: it takes 825 ms at the original resolution, reduces to 412 ms at 512×512, and further drops to 205 ms at 256×256. The processing time approximately shows a linear decrease, which is consistent with the decreasing trend of the number of image pixels. Notably, at the 512×512 resolution, the system achieves a good performance balance: the detection response value only drops by 0.04, while the processing time is reduced by nearly 50%. This operating point has important reference value in practical applications and provides an important basis for the resolution selection during system deployment.

[0095] In Figure 3 it, the radar chart of the spatial correlation detection response comprehensively demonstrates the system performance from six dimensions. The radial response reaches a level of 0.85, indicating that the system has a good capture ability for the radial features of the target; the angular response is 0.78, slightly lower than the radial response but still within an acceptable range; especially the combined response reaches a high level of 0.92, significantly higher than the response values of a single dimension, which confirms the synergistic enhancement effect of multi-dimensional feature fusion. The direction accuracy reaches 0.89, indicating that the system has a high accuracy in target direction judgment; 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 weaknesses, 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] In Figure 4Among them, the comparison chart of gradient consistency optimization effect shows the improvement effect of the optimization algorithm on four key indicators. The boundary accuracy has been significantly improved, from 4.5 before optimization to 2.3 after optimization, with an improvement rate close to 50%, indicating that the optimized boundary positioning is more accurate and reliable; the illumination gradient has increased from 0.65 to 0.88, with an increase of 35.4%, indicating that the system's adaptability to illumination changes has been greatly improved; the position gradient has improved from 0.58 to 0.85, with an increase rate of 46.6%, indicating that the system's sensitivity to target position changes has been significantly enhanced; the consistency score has increased from 0.45 to 0.82, with the largest increase rate of 82.2%. This significant improvement proves that the optimized detection results are more stable and reliable. The fact that all four indicators have been 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, laying a solid foundation for the practical application of the system.

[0097] In Figure 5 Among them, the comprehensive performance comparison chart comprehensively compares the performance differences between this method and traditional methods through five core indicators. In terms of detection accuracy, this method reaches 94.5%, significantly better than 85.2% of traditional methods, with an increase of 9.3 percentage points; in terms of processing speed, this method reaches 85.2, which is 17.5% higher than 72.5 of traditional methods, showing stronger real-time performance; in terms of anti-noise performance, this method achieves a high score of 89.3, which is 16.3% higher than 76.8 of traditional methods, indicating stronger environmental adaptability; the boundary accuracy reaches 92.1, which is 11.8% higher than 82.4 of traditional methods, indicating obvious advantages in target contour extraction; in terms of spatial consistency, this method reaches 94.0, better than 85.5 of traditional methods, with an increase of 9.9%, proving that the detection results are more stable and reliable. The comprehensive superiority in five dimensions, and the increase rate is more than 9% in all cases, fully proves the comprehensive advantages of this method, especially the significant improvements in the two key indicators of detection accuracy and anti-noise performance, 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 at different scales, providing comprehensive detailed support for abnormal light spot detection; a detection operator is constructed by combining radial and circumferential illumination gradients to accurately reflect the illumination changes on the ore surface, helping to accurately capture the characteristics of abnormal light spots; a dynamic threshold is set based on local statistical information to segment the spatially correlated responses, and then the candidate regions are merged through region overlap and area screening to improve the detection stability; the boundaries of the candidate regions are finely adjusted using the consistency constraint of illumination and position gradients to accurately locate the abnormal light spot regions and enhance the reliability of subsequent processing.

[0099] Input the light-normalized enhanced image and the three-dimensional point cloud data of the ore into the multi-branch feature extraction network. Determine the separated features through the feature decoupling module, obtain the phase features by performing Hilbert transform on the separated features, construct the high-order associations of the features using the recursive neural tensor network, and obtain the optimized reconstructed features through the adaptive feature reconstruction module and the conditional random field.

[0100] In this embodiment, input the light-normalized image and the three-dimensional point cloud data of the ore into the multi-branch network to achieve multi-modal data fusion, comprehensively extract the ore texture and structure information, and provide a solid data foundation for subsequent feature processing; separate the interference information through the feature decoupling module, and then use the Hilbert transform to extract the phase features, effectively enhancing the expression of key information and improving the model's ability to capture subtle features; adopt the recursive neural tensor network to construct the high-order associations between the phase features, deeply mine the complex relationships between various ore features, promote multi-scale feature fusion, and improve the overall expression effect; through the adaptive feature reconstruction module combined with the conditional random field, achieve precise optimization and reconstruction of the features, adjust the local details, and finally output the reconstructed features with high quality and strong stability.

[0101] In an alternative embodiment, it further includes:

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

[0103] Perform adversarial learning based on the image features and the point cloud features, and decouple the image features and the point cloud features into common features and private features. The private features include image private features and point cloud private features;

[0104] Perform Hilbert transform on the common features and the private features, and obtain the phase features based on the arctangent operation of the transformed features and the original features;

[0105] Perform non-linear transformation on the phase features using the tensor weight matrix and the recursive weight matrix in the recursive neural tensor network to construct the high-order association features between the phase features;

[0106] Construct a feature weight matrix, and perform adaptive weighting on the high-order association features based on the feature weight matrix to obtain the reconstructed features;

[0107] Construct the spatial dependence relationship of the reconstructed features based on the unary potential function and the binary potential function, and optimize the reconstructed features using mean field inference to obtain the optimized reconstructed features.

[0108] In a specific embodiment, in the multi-branch feature extraction stage, first, the input light-normalized enhanced image is preprocessed to uniformly scale the image to a standard size. The image feature branch adopts a deep residual network structure, and feature dimensionality reduction is performed through the first-layer convolution, and then multiple residual blocks are cascaded for feature extraction. Each residual block contains a main path and a shortcut path. The convolutional layers on the main path are 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, the local connection relationship of the point cloud is constructed, and a dynamic graph structure is established based on the neighbor information of each point. Through the dynamic graph convolutional network, the connection relationship between points can be adaptively updated, effectively capturing the local structural features and global morphological information of the point cloud.

[0109] In the feature decoupling link, a complete adversarial learning framework is constructed. The encoder network processes the image features and point cloud features respectively, and maps the features to the latent space through multiple non-linear transformations. In the latent space, the features are decoupled into common features reflecting common attributes and private features expressing modality-specific information. The discriminator network evaluates the decoupling effect of the features through adversarial training, ensuring that the common features can truly reflect the common information of multi-modal data, while the private features retain the unique properties of their respective modalities. By setting appropriate loss weights, the reconstruction accuracy and feature independence are balanced during the feature decoupling process.

[0110] During the phase feature extraction process, the Hilbert transform is performed on each decoupled feature component. By constructing a Hilbert transform filter, first, the Fourier transform of the feature is calculated, and then the corresponding analytic signal is obtained. Based on the real and imaginary part information of the analytic signal, the phase information of the feature is obtained through the arctangent operation. To improve the stability of the phase features, phase unwrapping and jump elimination processing are also performed, and a smoothing algorithm is used to reduce the influence of noise.

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

[0112] The feature reconstruction stage adopts an adaptive weighting strategy. First, the discriminative ability of each feature component is evaluated, and a feature weight matrix is constructed based on the discriminative scores. During the feature fusion process, ensure the dimensional alignment of different modality features, and a unified feature representation is obtained through weighted combination. To improve the robustness of the features, regularization processing is also performed.

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

[0114] Exemplarily, take a 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 2000 points and extracts a 128-dimensional point cloud feature vector through 3-layer dynamic graph convolution.

[0116] After feature decoupling, a 64-dimensional common feature (reflecting the basic morphological features of the ore) and image private features (reflecting surface texture) and point cloud private features (reflecting three-dimensional structure) of 192 dimensions and 64 dimensions respectively are obtained.

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

[0118] Three-layer recursive structures are set in the recursive neural tensor network, and a 32×32 tensor weight matrix is used to capture feature interactions. Through 5 iterations, high-order features reflecting complex associations between features are constructed.

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

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

[0121] In this embodiment, a deep residual network and dynamic graph convolution are used to extract the features of the illumination-normalized image and the point cloud data respectively, fully capturing the ore texture and structural details, enhancing the overall feature expression and robustness; an adversarial learning framework is constructed to map the image and point cloud features into the latent space, realizing the separation of common and private information, enhancing the cross-modal feature discriminability and balancing the reconstruction accuracy; the phase information is extracted through Hilbert transform, and a recursive neural tensor network is used to construct complex interaction patterns, effectively capturing the deep semantic relationships and improving the stability of feature representation; an adaptive weighted fusion and regularization strategy is adopted, and the potential function and mean field inference are combined to achieve spatial optimization, effectively enhancing the robustness and spatial consistency of the reconstructed features.

[0122] The conditional random field technology was first proposed by Lafferty et al. in 2001. Traditional conditional random field methods mainly rely on simple unary potential functions to express the local features of nodes in practical applications, and use basic binary potential functions such as the Potts model to describe the relationships between adjacent nodes. In terms of optimization strategies, these methods usually use algorithms based on graph cuts or belief propagation for iterative optimization. However, due to the high computational complexity, they often face efficiency bottlenecks when dealing with large-scale data. Although the later developed dense conditional random field method introduces fully connected pairwise potential functions, simplifies the inference process through mean field approximation, and accelerates the calculation using Gaussian kernel fast filtering, there are still obvious deficiencies in the feature space expression ability 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 the computational complexity, and how to enhance the modeling ability of spatial dependence relationships. To solve these problems, this technology first improves the design of the unary potential function, breaks through the limitation of simple class probabilities in traditional methods, and constructs a richer expression method for node attributes. In terms of the binary potential function, a Gaussian kernel function based on the distance matrix is innovatively introduced to achieve more accurate modeling of spatial dependence relationships. By discretizing the feature space into a grid structure and applying the fast Fourier transform to the iterative optimization process of the marginal probability distribution, a significant improvement in computational efficiency is successfully achieved.

[0124] At the specific implementation level, this technology is no longer limited to simple node attribute expressions, but constructs a more complex unary potential function system that can more comprehensively characterize node features. By designing a Gaussian kernel binary potential function based on the distance matrix, accurate modeling of spatial dependence relationships is achieved. The grid processing of the feature space and the introduction of FFT provide the possibility for efficient calculation. This innovative technical route realizes a complete end-to-end solution from feature extraction to optimization.

[0125] The improved technical effects are remarkable, with obvious improvements achieved in multiple key indicators. Experimental data shows that the consistency score has increased from 0.45 to 0.82, with an increase rate of 82.2%; the spatial consistency reaches 94.0%, an increase of 9.9% compared with the traditional method; the processing speed has increased by 17.5%; the boundary accuracy has increased by 11.8%; the anti-noise performance has increased by 16.3%. These performance improvements make this technology have stronger practical application value, especially outstanding in scenarios that require both ensuring the reconstruction quality and computational efficiency. By combining modern mathematical tools with classical probabilistic graphical models, this technology has successfully achieved the unity of theoretical innovation and engineering practicality.

[0126] The improved technology shows a wider applicability in applications and can effectively process large-scale high-dimensional feature data. The optimization process exhibits better stability and convergence, and the improvement in computational efficiency makes it more suitable for real-time application scenarios. Its 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, this technology has successfully solved the main problems faced by traditional methods in practical applications and provided a more perfect and practical solution for the field of feature reconstruction.

[0127] In an optional embodiment, the spatial dependence relationship of the reconstructed features is constructed based on the unary potential function and the binary potential function, and the reconstructed features are optimized by mean field inference to obtain optimized reconstructed features, including:

[0128] By constructing the unary potential function and the binary potential function to characterize the node attributes and spatial dependence relationship of the reconstructed features, and combining them into a conditional random field energy function, mean field inference is used for feature optimization, and the fast Fourier transform is introduced to accelerate the iteration of the marginal probability distribution on the feature grid, specifically including:

[0129] Calculate the unary attributes of the reconstructed features to obtain unary feature values, and construct a unary potential function based on the unary feature values. The unary potential function is used to characterize the node attributes of the reconstructed features; calculate the spatial distance between the nodes of the reconstructed features to obtain a distance matrix, and input the distance matrix into the Gaussian kernel function to construct a binary potential function. The binary potential function is used to characterize the spatial dependence relationship of the reconstructed features;

[0130] Construct a spatial dependence model of the reconstructed features based on the unary potential function and the binary potential function, and represent the spatial dependence model as a conditional random field energy function; input the conditional random field energy function into the mean field inference model, and initialize the marginal probability distribution of the nodes of the reconstructed features; calculate the message passing between the nodes based on the binary potential function, and combine the result of the message passing with the unary potential function to update the marginal probability distribution;

[0131] Divide the reconstructed feature space into feature grids, calculate the potential function values in the feature grids, and use the fast Fourier transform to perform mean field iteration optimization on the marginal probability distribution; calculate the expected value of the reconstructed features based on the optimized marginal probability distribution, and use the expected value as the output result of the optimized reconstructed features.

[0132] In a specific embodiment, a potential function is first constructed to characterize the attributes and dependencies of the reconstructed features. For the construction of the unary potential function, each node of the reconstructed features is analyzed, and the local attribute information of the node is extracted, including the distribution characteristics, statistical characteristics, etc. of the feature values. These attribute information are integrated to form unary feature values. Based on these unary feature values, 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 for describing the spatial dependence relationship between nodes, reflecting 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 dependence model. This model is represented in the form of a conditional random field energy function, comprehensively considering the independent attributes of the nodes and the mutual influence between the nodes. The energy function is input into the mean field inference model. First, the marginal probability distribution of all reconstructed feature nodes is initialized. On this basis, the message passing process between nodes is calculated using the binary potential function, which reflects the flow and interaction of information in the feature space. The result of the message passing is combined with the unary potential function to continuously update the marginal probability distribution of each node, making it gradually converge to the optimal state.

[0134] To improve the calculation efficiency, the reconstructed feature space is divided into a regular feature grid structure. In each grid cell, 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 marginal 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 marginal probability distribution, the expected value of the reconstructed features is calculated, and this expected value is used as the final optimized output result.

[0135] Exemplarily, taking the processing of an ore sample with a complex surface structure as an example, 320-dimensional reconstructed features are first obtained. When constructing the unary potential function, the distribution characteristics of each feature dimension are analyzed, and statistics such as the mean and variance are calculated to generate unary feature values reflecting the node attributes. For example, for the feature dimension representing the surface texture, its unary feature value is relatively high, indicating that this node has significant texture characteristics.

[0136] When calculating the spatial distance, a 320-dimensional feature space is mapped to a three-dimensional physical space, and a 1000×1000 distance matrix is constructed. Through the processing of the Gaussian kernel function, a binary potential function representing the spatial dependence 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 32×32 feature grids, and the potential function values are calculated in each grid cell. A uniform marginal probability distribution is initialized, and then iterative optimization is performed. In each iteration, first, the message passing between adjacent nodes is calculated. For example, stronger messages are passed between adjacent nodes with similar texture features. Through the 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 is obtained. Finally, the expected value is calculated to obtain a 320-dimensional optimized feature vector. The optimized features show better spatial consistency. For example, the texture features in adjacent regions are more continuous, and the noise is effectively suppressed, providing a reliable feature representation for subsequent classification tasks.

[0139] This optimization process significantly improves the discriminative ability of the features. The original possible feature discontinuities and local noises are effectively eliminated, and the structural information in the feature space is better preserved. The entire optimization process not only ensures the computational efficiency but also guarantees the quality of the optimization results.

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

[0141] The reconstructed features are input into a classification model based on active learning. The self-organizing mapping network is used to construct and optimize the feature manifold space. The samples to be labeled are screened based on the adaptive entropy threshold. The samples to be labeled are labeled and the classification model is optimized using an uncertainty-aware active sampling strategy. The ore classification results are obtained by fusing the prediction results through a random forest ensemble framework and the ore sorting is performed.

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

[0143] In an alternative embodiment, it further includes:

[0144] Input the reconstructed features into the self-organizing mapping network, establish a topology-preserving mapping based on the competitive learning mechanism, and perform feature non-linear dimensionality reduction by iteratively optimizing the weight vectors of the neurons in the competitive layer to obtain a manifold feature space;

[0145] Calculate the sample information entropy in the manifold feature space, construct an adaptive entropy threshold based on the mean and standard deviation of the sample information entropy, and use the samples with sample information entropy greater than the adaptive entropy threshold as the samples to be labeled;

[0146] Obtain the prediction uncertainty and sample representativeness of the samples to be labeled respectively, perform weighted combination on the prediction uncertainty and the sample representativeness to obtain a sampling score, select the samples to be labeled for labeling based on the sampling score, and add the labeled samples to the training set;

[0147] Randomly select a feature subspace in the manifold feature space, perform bootstrap sampling in the training set to obtain a data subset, and use the feature subspace and the data subset to construct multiple decision tree classifiers; input the samples to be classified into all the decision tree classifiers, and fuse the prediction results of all the decision tree classifiers through a voting mechanism to obtain the ore classification result;

[0148] Perform ore sorting operations based on the ore classification result.

[0149] In a specific implementation, first construct a manifold space representation of the features. Input the optimized reconstructed features into the self-organizing mapping network, and this network performs feature dimensionality reduction through the competitive learning mechanism. During the learning process, the neurons in the network compete with each other to respond to the input samples, and the weight vectors of the winning neurons and their neighboring neurons will be adjusted towards 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 close 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. Identify samples with information entropy exceeding this threshold as samples to be labeled, and these samples usually have high information value.

[0151] For the selected samples to be labeled, importance assessment is carried out. First, calculate the prediction uncertainty of each sample, which reflects the confidence level of the model's prediction result for this sample; at the same time, evaluate the representativeness of the sample, which measures the ability of this sample to represent the overall data distribution. Combine these two indicators through weighting to obtain the comprehensive sampling score of each sample to be labeled. Based on these scores, select the most valuable samples for manual labeling and add the labeled samples to the training dataset.

[0152] In the classification decision stage, a random forest framework is adopted. First, randomly select multiple feature subspaces in the manifold feature space, and at the same time perform bootstrap sampling on the training set to generate multiple data subsets. Use these feature subspaces and data subsets to train multiple decision tree classifiers, and each classifier can independently predict samples. When a new sample to be classified is input, all decision tree classifiers make predictions simultaneously, and finally integrate all prediction results through a voting mechanism to obtain the final classification decision.

[0153] Finally, according to the classification results, control the sorting equipment to perform corresponding sorting actions, and transport ores of different categories to the corresponding collection areas respectively.

[0154] Exemplarily, taking the processing of a batch of mixed ores as an example, first obtain 320-dimensional reconstructed features. Through a self-organizing mapping network, reduce the features to a 16-dimensional manifold space. The network adopts a 20×20 competitive layer structure and undergoes 1000 iterations of training to finally establish a feature mapping that maintains the topological relationship.

[0155] In the manifold space, calculate the information entropy of all samples, and obtain that the mean of the entropy value distribution is 0.75 and the standard deviation is 0.15. Based on this, set the adaptive entropy threshold to 1.05 (the mean plus twice the standard deviation), and screen out 200 samples to be labeled.

[0156] For these samples to be labeled, calculate the prediction uncertainty (weight 0.6) and sample representativeness (weight 0.4). For example, the prediction uncertainty of a certain sample is 0.8 and the representativeness is 0.6, and finally obtain a sampling score of 0.72. Select the 50 samples with the highest scores for manual labeling and add them to the original training set.

[0157] In the random forest framework, 100 decision tree classifiers are 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 it as ore type A and 40 predict it as ore type B, then the sample is finally determined to be ore type A.

[0158] According to the classification results, the sorting equipment accurately performs sorting actions, conveys ore type A to the collection bin in area A, and conveys ore type B to the collection bin in area B, achieving precise 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 for feature dimensionality reduction to maintain the data topological structure and improve the expression ability of the classification model for ore features; an adaptive threshold is calculated based on information entropy to select high-value samples to be labeled, reducing the labeling cost and improving the quality of training data; through multi-decision tree voting fusion, the stability and generalization ability of the classification model are improved to ensure accurate and reliable ore recognition; the sorting equipment is automatically controlled according to the classification results to achieve efficient ore classification, improving the sorting automation level and production efficiency.

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

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

[0162] Perform a convolution operation on the Mahalanobis distance and the local density function to obtain the local probability density of the sample to be labeled; calculate the manifold consistency measure between the sample to be labeled and the set of labeled samples based on the local probability density, and use the manifold consistency measure as the sample representativeness;

[0163] Perform non-linear probability prediction on the sample to be labeled using manifold Laplacian eigenmaps; calculate the information entropy and inter-class divergence based on the non-linear probability prediction, and use the weighted combination of the information entropy and the inter-class divergence as the prediction uncertainty.

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

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

[0166] Next, the manifold Laplacian eigenmap is used to perform probability prediction on the samples to be labeled. This non-linear mapping can preserve the local structural characteristics of the data, making the prediction results more consistent with the intrinsic 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 between-class scatter is calculated to reflect the separability between different classes. These two indicators are weighted and combined to obtain the final evaluation value of the prediction uncertainty.

[0167] Exemplarily, taking the processing of a batch of ore samples to be labeled as an example, the operation is carried out in a 16-dimensional manifold feature space. For each sample to be labeled, the 8 nearest neighbor samples are selected to construct the local structure. For example, for a certain sample A to be labeled, the geodesic distances from it to the 8 neighboring samples are calculated, and the obtained distance values 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. Using this covariance matrix, the Mahalanobis distance of sample A is calculated, and a distance metric value of 2.5 that takes into account the local distribution characteristics is obtained.

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

[0170] In the prediction stage, the manifold Laplacian eigenmap is used to predict sample A, obtaining the probability distributions belonging to different classes: 0.4 for class 1, 0.35 for class 2, and 0.25 for class 3. Based on this probability distribution, the calculated information entropy value is 0.9, and the between-class divergence value is 0.3. Setting the information entropy weight to 0.7 and the between-class divergence weight to 0.3, the final prediction uncertainty of sample A is 0.72.

[0171] These calculation results together reflect the important features of sample A in the feature space: a relatively high sample representativeness (0.75) indicates that it can well represent the data distribution characteristics, and a relatively high prediction uncertainty (0.72) indicates that it contains rich information content. These metrics help determine whether to preferentially select sample A for annotation, thereby improving the efficiency of active learning.

[0172] Through this refined sample evaluation mechanism, the most valuable samples can be intelligently selected for annotation, ensuring both the representativeness of the annotated samples and the effectiveness of improving the classification model. The entire process makes full use of the manifold structure characteristics of the data, realizing 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 manifold consistency metric is used to evaluate the importance of samples in the data distribution to optimize the selection of samples to be annotated; the manifold Laplacian eigenmap is used to preserve the local structure of the data to improve the adaptability of the classification model; the information entropy and between-class divergence are combined to evaluate the prediction uncertainty to accurately select key samples and reduce the annotation 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 entropy value of the probability distribution output by the classifier as an indicator, while density sampling relies on Euclidean distance to calculate the sample density. These methods have obvious limitations when dealing with complex non-linear data: they cannot effectively capture the intrinsic geometric structure of the data and are relatively sensitive to abnormal samples and noise. Although the improved manifold-based methods introduce local preservation properties, they often only consider simple Euclidean distance metrics and fail 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 relationships of samples in the manifold space and how to more effectively evaluate the information value of samples. The primary innovation is to introduce geodesic distance to replace the traditional Euclidean distance, which makes the distance measurement more conform to the intrinsic geometric structure of the data. Secondly, by constructing the local covariance matrix and calculating the Mahalanobis distance, the adaptive modeling of the local data distribution characteristics is realized. The design of convolving the Mahalanobis distance with the local density function effectively combines the local structural information with the global distribution characteristics.

[0176] In terms of specific implementation, this technology accurately depicts the local relationships of samples in the manifold space through the k-nearest neighbor structure and geodesic distance calculation. The introduction of the local covariance matrix enables the distance measurement to adaptively consider the local distribution characteristics of the data. Nonlinear probability prediction through manifold Laplacian eigenmaps not only preserves the manifold structure of the data but also provides reliable uncertainty estimation. Especially in the evaluation of sample representativeness, the organic combination of local density information and global structural information through manifold consistency measurement greatly improves the accuracy of sample selection.

[0177] The improved technical effects are remarkable and 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 the traditional method, the new method improves the annotation efficiency by about 30%, that is, under the same annotation budget, higher model performance can be obtained. Secondly, in terms of handling abnormal samples, the adaptive measurement 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 shows stronger adaptability, especially outstanding in dealing with high-dimensional non-linear data. By comprehensively considering the representativeness and prediction uncertainty of samples, this method can more accurately identify the samples with the most information value. Experimental results show that under the same annotation cost, the new method can reduce the annotation volume by 20%-35% compared with the traditional method while maintaining the same model performance. This significant performance improvement makes this technology have important practical value in practical applications, especially in scenarios with limited annotation resources.

[0179] Through these technical improvements, this application successfully solves the main challenges faced by traditional active learning methods in dealing with complex data. The new method not only improves the accuracy of sample selection but also enhances the robustness and efficiency of the algorithm. Especially the advantage in maintaining the essential geometric structure of the data provides a more reliable and efficient solution for dealing with high-dimensional non-linear data. These improvements make this technology have a wider range of applicability in practical applications and provide an effective way to reduce the data annotation cost.

[0180] The present invention may be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and 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