A method and system for detecting surface contamination of bauxite based on spectral imaging
By constructing a pollution detection model that optimizes spectral and spatial features, the problems of low sorting efficiency and misclassification of bauxite in traditional methods are solved, and high-precision detection of surface pollution in bauxite is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LULIANG UNIV
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-09
Smart Images
Figure CN122176374A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mineral detection and sorting technology, specifically relating to a method and system for detecting surface contamination in bauxite based on spectral imaging. Background Technology
[0002] Bauxite is the main raw material for producing alumina and metallic aluminum. During mining, bauxite is often found alongside impurities such as laterite and siliceous clay, leading to surface contamination. Traditional blind sorting or RGB color camera-based sorting methods cannot effectively distinguish ores with similar colors but different mineral compositions from impurities, resulting in low sorting efficiency, resource waste, and increased energy consumption.
[0003] Hyperspectral imaging technology can simultaneously acquire spatial and continuous spectral information of a target, making it possible to distinguish different substances. However, its direct application to bauxite pollution detection faces the following challenges: 1. Spectral mixing problem: The spectrum of a single pixel is often a mixture of pure bauxite, pollutants, and background signals, making direct identification difficult. 2. Different spectra for the same substance and similar spectra for different substances: The same type of mineral exhibits different spectra due to variations in water content, particle size, and surface roughness; while different minerals may exhibit similar spectral characteristics in specific bands. 3. High data dimensionality and complex processing: Hyperspectral data is large in volume and features redundant bands, putting pressure on real-time processing. 4. Limitations of traditional classification algorithms (such as SVM and random forest): These algorithms typically treat each pixel as an independent sample for classification, ignoring spatial context information between pixels, and are prone to "salt and pepper noise"-like misclassifications in areas with blurred boundaries and mixtures. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention provides a method and system for detecting surface contamination in bauxite based on spectral imaging, which features high detection accuracy, strong anti-interference ability, and suitability for industrial online applications.
[0005] To achieve the above objectives, the present invention provides the following solution: A method for detecting surface contaminants in bauxite based on spectral imaging, comprising: Three-dimensional hyperspectral data of bauxite samples were collected, and the three-dimensional hyperspectral data were corrected, filtered, and enhanced to obtain preprocessed three-dimensional hyperspectral data. Based on preprocessed 3D hyperspectral data, a pollution detection model with spectral-spatial feature co-optimization is constructed using graph convolutional networks, dynamic perception co-convolutional modules, and adaptive attention fusion modules. The hyperspectral image of the bauxite to be detected is segmented into overlapping blocks, which are then input into the pollution detection model to obtain the category probability of each pixel. Based on the category probability, a pollution distribution map of the bauxite surface is generated.
[0006] Preferably, the method for acquiring three-dimensional hyperspectral data of bauxite samples includes: An independently controllable ring LED array was used as the illumination source. Illumination of different phases was triggered sequentially. Hyperspectral data of each phase of bauxite samples on the conveyor belt were acquired synchronously using a pushbroom hyperspectral imaging system. The hyperspectral data of each phase were then fused at the pixel level to obtain the initial three-dimensional hyperspectral data. An optical encoder is installed on the side of the conveyor belt. Combined with a calibration plate of known size, the initial three-dimensional hyperspectral data is geometrically calibrated and spatially reconstructed to obtain distortion-free three-dimensional hyperspectral data. The pulse signal of the optical encoder is synchronized with the row exposure signal of the pushbroom hyperspectral imaging system, and the spatial coordinates of each acquisition row are determined by the cumulative number of pulses of the optical encoder.
[0007] Preferably, in the pollution detection model, the graph convolutional network is used to enhance the spectral features of the three-dimensional hyperspectral data by introducing a linear iterative clustering algorithm and an adaptive adjacency matrix, thereby obtaining enhanced spectral features; The dynamic perception collaborative convolution module is used to extract the local spatial-spectral texture features formed between adjacent pixels and adjacent bands in the three-dimensional hyperspectral data by using a split 3D convolution block. The adaptive attention fusion module is used to concatenate the enhanced spectral features and the local spatial-spectral texture features in the channel dimension to generate a spatial attention map, and then use the spatial attention map to perform a weighted summation of the two features to obtain the final fused features.
[0008] Preferably, the method for obtaining the enhanced spectral features includes: Calculate the local spectral variance map of the input 3D hyperspectral data, and dynamically adjust the superpixel granularity of the linear iterative clustering algorithm based on the local spectral variance map to construct an adaptive superpixel grid; treat each superpixel as a graph node, and use the average spectrum of all pixels within the superpixel as the initial feature of the node; Based on the cosine similarity of the initial features of the nodes and the K-nearest neighbor algorithm, an initial adjacency matrix is calculated as a topological prior. At the same time, the initial features of the nodes are input into a shared neural network, and a data-driven adjacency matrix is generated through differential calculation. The initial adjacency matrix and the data-driven adjacency matrix are weighted and summed using learnable weight parameters to obtain the final adaptive adjacency matrix; self-connections are added to the adaptive adjacency matrix and normalized to obtain the adjacency matrix used for convolution. Based on the adjacency matrix used for convolution, a graph convolutional layer with a gated mechanism is used to perform spectral context aggregation, thereby obtaining node features with enhanced spectral context and preserved boundaries. The enhanced spectral features are obtained by mapping the spectral context-enhanced and boundary-preserving node features back to the original spatial grid.
[0009] Preferably, the method for extracting the local spatial-spectral texture features includes: Based on the three-dimensional hyperspectral data, an initial feature map is obtained, the input initial feature map is processed in parallel, a spatially aware attention map and a spectral coordinated attention vector are generated, and the basic transformation features are obtained through a split 3D convolutional block. Multiply the spatial awareness attention map with the basic transformation feature to perform spatial enhancement and obtain the spatial enhancement feature; The spectral coordination attention vector is multiplied by the spatial enhancement feature to perform spectral-channel calibration and obtain the calibration feature; The calibration features are added to the linearly transformed initial feature map via residual concatenation to obtain local spatial-spectral texture features.
[0010] Preferably, the method for obtaining the final fusion features includes: A spatial attention map is generated based on the cross-covariance of the enhanced spectral features and the local spatial-spectral texture features; At multiple parallel scales, the enhanced spectral features and the local spatial-spectral texture features are weighted and fused using the spatial attention map to obtain multi-scale fused features; Based on the enhanced spectral features and local spatial-spectral texture features, a scale-gated vector is generated, and the multi-scale fusion features are dynamically weighted and summed using the scale-gated vector to obtain the final fusion features.
[0011] The present invention also provides a bauxite surface contamination detection system based on spectral imaging, for implementing the method, comprising: The data acquisition module is used to acquire three-dimensional hyperspectral data of bauxite samples, and to correct, filter and enhance the three-dimensional hyperspectral data to obtain preprocessed three-dimensional hyperspectral data. The detection model building module is used to construct a pollution detection model based on preprocessed 3D hyperspectral data, utilizing graph convolutional networks, dynamic perception collaborative convolution modules, and adaptive attention fusion modules to jointly optimize spectral and spatial features. The contamination detection module is used to segment the hyperspectral image of bauxite to be detected into overlapping blocks, input them into the contamination detection model, obtain the category probability of each pixel, and generate a contamination distribution map on the bauxite surface based on the category probability.
[0012] Preferably, the data acquisition module includes: The initial spectral acquisition unit uses an independently controllable ring LED array as an illumination source. It triggers illumination of different phases sequentially and uses a pushbroom hyperspectral imaging system to synchronously acquire hyperspectral data of each phase of bauxite samples on the conveyor belt. The hyperspectral data of each phase are then fused at the pixel level to obtain initial three-dimensional hyperspectral data. The data reconstruction unit is used to install an optical encoder on the side of the conveyor belt, and in conjunction with a calibration plate of known size, to perform geometric calibration and spatial reconstruction on the initial three-dimensional hyperspectral data to obtain distortion-free three-dimensional hyperspectral data; wherein, the pulse signal of the optical encoder is synchronized with the row exposure signal of the pushbroom hyperspectral imaging system, and the spatial coordinates of each acquisition row are determined by the cumulative number of pulses of the optical encoder.
[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: The pollution detection model, through spectral-spatial feature co-optimization, effectively overcomes the industry challenges of "different spectra for the same object" and "same spectra for different objects." The improved graph convolutional network can model complex spectral similarity relationships between pixels, while the dynamic perception co-convolutional module can accurately extract local spatial-spectral texture features. The collaborative work of both significantly improves the accuracy of identifying pollutants on bauxite surfaces, especially fundamentally improving the classification accuracy in the boundary region between pollutants and ore. Attached Figure Description
[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a flowchart of a bauxite surface contamination detection method based on spectral imaging, according to an embodiment of the present invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0018] Example 1: like Figure 1As shown, a method for detecting surface contamination in bauxite based on spectral imaging includes: S1: Acquire three-dimensional hyperspectral data of bauxite samples, and perform correction, filtering, and enhancement on the three-dimensional hyperspectral data to obtain preprocessed three-dimensional hyperspectral data. A further implementation method for acquiring three-dimensional hyperspectral data of bauxite samples includes: An independently controllable ring-shaped LED array was used as the illumination source. Illumination of different phases was triggered sequentially, and a pushbroom hyperspectral imaging system was used to simultaneously acquire hyperspectral data of each phase of the bauxite sample on the conveyor belt. The hyperspectral data of each phase were then fused at the pixel level to obtain initial three-dimensional hyperspectral data. Specifically, a pushbroom (line-scan) hyperspectral imaging system was used, with a preferred combination of 400-1000 nm (VNIR) and 1000-2500 nm (SWIR) spectral ranges to comprehensively cover the characteristic absorption bands of major minerals (such as gibbsite and boehmite) and contaminants (such as iron oxides and clay minerals) in the bauxite. Due to the irregular surface of bauxite particles, illumination from a single angle produces strong highlights (specular reflection) and shadows, severely distorting the spectrum in these areas and losing the absorption characteristics of the material itself. Therefore, an independently controllable ring-shaped LED array was used as the illumination source. While scanning the same row of data, the system rapidly triggered illumination of different phases sequentially. Phase A (low-angle light): Triggers the LEDs on one side of the ring array, highlighting textures but creating shadows.
[0019] Phase B (high-angle light): triggers the LED on the other side, producing a complementary shadow.
[0020] Phase C (vertical diffused light): Triggers all LEDs and forms uniform illumination through the diffuser plate.
[0021] Pixel-level fusion is performed on the same row of spectral data acquired from three phases (A, B, and C). The fusion algorithm employs a "maximum-median" synthesis: for each pixel's value in each band, the median value from the three phases is taken. This effectively eliminates extreme highlight values (near saturation) and shadow values (near zero), retaining spectral information closest to diffuse reflection. This greatly eliminates the influence of surface specular reflection and shadows, obtaining "diffuse reflection-dominated" spectral data that better represents the material's inherent properties.
[0022] An optical encoder is installed on the side of the conveyor belt. Combined with a calibration plate of known size, the initial three-dimensional hyperspectral data is geometrically calibrated and spatially reconstructed to obtain distortion-free three-dimensional hyperspectral data. The pulse signal of the optical encoder is synchronized with the line exposure signal of the pushbroom hyperspectral imaging system. The spatial coordinates of each acquisition line are determined by the cumulative number of pulses of the optical encoder.
[0023] This embodiment provides a joint spectral filtering method based on first derivative and SG filtering to filter three-dimensional hyperspectral data. Specifically, traditional Savitzky-Golay (SG) filtering mainly smooths random noise, but bauxite spectra are characterized by broad absorption valleys. Its first derivative can amplify subtle spectral differences, but it is also extremely sensitive to noise. Therefore, this invention designs a two-step joint filtering process to obtain a spectral curve that is both smooth and feature-enhanced. The first step is SG smoothing of the original spectrum. An SG filter with a large window width (e.g., 11 points) and a low polynomial order (e.g., second order) is applied to the corrected three-dimensional hyperspectral data to obtain a smoothed spectrum. The purpose is to remove most of the high-frequency noise. The second step is SG smoothing of the derivative spectrum. The first derivative is calculated on the smoothed spectrum to obtain an initial smoothed derivative spectrum. Since differentiation amplifies noise, a second smoothing is performed on the initial derivative result using an SG filter with a smaller window width (e.g., 5 points) to obtain the final smoothed derivative spectrum. Finally, the smoothed spectrum and the final derivative smoothed spectrum are used as input features for subsequent algorithms. Smoothed spectra retain absolute reflectance information, while final derivative-smoothed spectra provide shape-sensitive features crucial for classification, and both maintain a good signal-to-noise ratio. This combined feature input of "absolute value + derivative" provides a stronger basis for the model to distinguish minerals with similar spectral shapes.
[0024] S2: Based on preprocessed 3D hyperspectral data, a pollution detection model with spectral-spatial feature co-optimization is constructed using graph convolutional networks, dynamic perception co-convolutional modules, and adaptive attention fusion modules.
[0025] A further implementation involves a graph convolutional network in the pollution detection model, which is used to enhance the spectral features of the three-dimensional hyperspectral data by introducing a linear iterative clustering algorithm and an adaptive adjacency matrix, thereby obtaining enhanced spectral features.
[0026] A further embodiment of the method for obtaining enhanced spectral features includes: The local spectral variance map of the input 3D hyperspectral data is calculated, and the superpixel granularity of the Linear Iterative Clustering (SLIC) algorithm is dynamically adjusted based on the local spectral variance map to construct an adaptive superpixel grid. Each superpixel is regarded as a graph node, and the average spectrum of all pixels within the superpixel is used as the initial feature of the node. Specifically, this embodiment deeply embeds the semantic information of spectral variance into the iterative optimization structure of the SLIC algorithm, changing its initialization, allocation, and update mechanism of cluster centers. Specifically, the standard SLIC algorithm treats each pixel as a five-dimensional feature vector in its structure and performs iterative clustering in this five-dimensional space. This embodiment first expands the five-dimensional feature vector to six dimensions, introducing local spectral variance as the sixth dimension to form a six-dimensional feature vector, thereby using spectral complexity as an intrinsic driving force for clustering at the fundamental level of the algorithm. Based on this expanded feature space, the structure of the algorithm is creatively modified as follows: Traditional SLIC initializes cluster centers uniformly in the image space. In this invention, during initialization, the image is first divided into "high variance regions" and "low variance regions" according to the local spectral variance histogram of the entire image. In high-variance regions, cluster centers are initialized with a higher density to ensure that complex boundaries and subtle contaminants are adequately resolved; in low-variance regions, they are initialized with a lower density to avoid unnecessary computational redundancy in uniform areas. This operation ensures that the distribution of superpixel "seed points" matches the content complexity of the image. The core step of SLIC is that, when assigning the nearest cluster center to each pixel, the traditional distance metric is a weighted sum of color distance and spatial distance. This invention innovates by introducing spectral variance adaptive weights. These spectral variance adaptive weights are not fixed values but a function positively correlated with local spectral variance. This means that in high-variance regions, the algorithm assigns higher weight to the "difference between the current pixel and the cluster center in the spectral variance dimension," forcing cluster boundaries to be generated more along the contours of spectral complexity; while in low-variance regions, its influence is reduced, reverting to the standard clustering pattern dominated by color and spatial proximity.
[0027] In each iteration of updating cluster centers, instead of simply calculating the average of all pixel features within a cluster, a weighted average is used for the spectral variance dimension, with the weight being the variance value of that pixel itself. This causes the new cluster centers to "shift" towards pixels within the cluster with more dramatic spectral variations, thus more accurately anchoring them in regions with complex spectral features and guiding superpixel boundaries to better enclose potential contamination boundaries.
[0028] Based on cosine similarity and the K-nearest neighbor algorithm using initial node features, an initial adjacency matrix is calculated as a topological prior. Simultaneously, the initial node features are input into a shared neural network, and a data-driven adjacency matrix is generated through differential computation. Specifically, the initial node features are input into a parameter-shared two-layer perceptron for nonlinear transformation. Matrix multiplication is then performed on this transformed feature representation to generate an initial association strength matrix between nodes. This matrix is then fed into a differentiable sampling layer based on the Gumbel-Softmax relaxation technique, thereby achieving continuous modeling of the discrete graph's topology and ultimately generating a data-driven adjacency matrix. During this process, a temperature coefficient annealing strategy is introduced to ensure that the data-driven adjacency matrix maintains sufficient exploratory nature in the early stages of training, while gradually approximating the determined discrete graph structure in the later stages.
[0029] The initial adjacency matrix and the data-driven adjacency matrix are weighted and summed using learnable weight parameters to obtain the final adaptive adjacency matrix. Self-connections are added to the adaptive adjacency matrix and normalized to obtain the adjacency matrix used for convolution.
[0030] Based on the adjacency matrix used for convolution, a graph convolutional layer with a gating mechanism is employed to perform spectral context aggregation, obtaining node features that are enhanced in spectral context and preserve boundaries. Specifically, the features of the central node are concatenated with the aggregated features of its neighborhood, and then linearly transformed and activated by a trainable weight matrix to generate a gating vector. This gating vector, after being normalized by the sigmoid function, is used to dynamically adjust the fusion ratio between neighborhood information and the original features of the central node. When the gating value approaches 1, it strengthens the propagation of neighborhood features; when it approaches 0, it preserves the inherent features of the central node. This gated graph convolution constructs an element-wise mask along the feature dimension, enabling each node to adaptively select the degree of information retention based on its local topological environment. This effectively alleviates the problem of excessive smoothing of node features in multi-layer propagation while achieving spectral context enhancement, especially maintaining the discriminativeness of features in the boundary regions of heterogeneous nodes. This process, through end-to-end learning, enables the model to automatically enhance the feature preservation capability of the central node in the transition region between contaminants and ores, establishing an adaptive balance mechanism between spectral feature aggregation and boundary preservation.
[0031] The node features, which are enhanced with spectral context and preserved by boundaries, are mapped back to the original spatial grid to obtain enhanced spectral features.
[0032] The Dynamic Perception Collaborative Convolution Module is used to extract local spatial-spectral texture features formed between adjacent pixels and adjacent bands in three-dimensional hyperspectral data by utilizing separable 3D convolutional blocks.
[0033] A further implementation method for extracting local spatial-spectral texture features includes: Based on 3D hyperspectral data, an initial feature map is obtained. The input initial feature map is processed in parallel, and a spatially aware attention map and a spectrally coordinated attention vector are generated simultaneously. The basic transformation features are obtained through a split 3D convolutional block. The spatially aware attention map is multiplied with the basic transformation features to perform spatial enhancement and obtain spatially enhanced features. The spectrally coordinated attention vector is multiplied with the spatially enhanced features to perform spectral-channel calibration and obtain calibration features. The calibration features are added to the linearly transformed initial feature map through a residual connection to obtain local spatial-spectral texture features.
[0034] Specifically, two attention branches—spatial awareness and spectral coordination—are implemented in parallel on the input feature map. The spatial awareness branch compresses features through global pooling along the spectral and channel dimensions, then generates a spatial awareness attention map via convolutional layers. The spectral coordination branch preserves spectral depth information through 3D global pooling and constructs a spectral coordination attention vector through two consecutive 1D convolutional layers. Simultaneously, the feature transformation branch uses separate 3D convolution for basic feature extraction, separating depth-direction convolution from pointwise convolution to optimize computational efficiency. Subsequently, a dual-attention collaborative fusion mechanism is used to perform element-wise multiplication of the spatial awareness attention map and the basic transformed features to achieve spatial enhancement, and then channel-weighted summation of the spatial enhancement features and the spectral coordination attention vector completes spectral calibration. Finally, residual connections are used to add the calibrated features to the linear transformation result of the initial feature map, forming an output feature map with dual spatial and spectral awareness capabilities. This module, through the collaborative design of parallel attention mechanisms and separate convolution, achieves dynamic focusing on key spatial regions and feature bands while maintaining computational efficiency.
[0035] The adaptive attention fusion module is used to concatenate enhanced spectral features and local spatial-spectral texture features along the channel dimension to generate a spatial attention map, and then use the spatial attention map to perform a weighted summation of the two features to obtain the final fused features.
[0036] A further implementation method includes obtaining the final fusion features by: A spatial attention map is generated based on the cross-covariance between enhanced spectral features and local spatial-spectral texture features. Specifically, by calculating the cross-covariance between enhanced spectral features and local spatial-spectral texture features, the global statistical correlation between the two feature modes in the channel dimension is captured. This matrix is then mapped through a two-layer convolutional network to generate a spatial attention map with global perception capabilities.
[0037] At multiple parallel scales, spatial attention maps are used to weightedly fuse enhanced spectral features and local spatial-spectral texture features to obtain multi-scale fused features. Specifically, a fusion pathway with three parallel scales, including 1×1 convolution, 3×3 convolution, and global pooling upsampling, is constructed. In each scale pathway, a unified spatial attention map is used to adaptively weight and fuse the two types of input features, resulting in multi-scale fused features that preserve details, local context, and global semantics. Based on enhanced spectral features and local spatial-spectral texture features, a scale-gated vector is generated. This vector is then used to dynamically weight and sum the multi-scale fused features to obtain the final fused feature. Specifically, the original enhanced spectral features and local spatial-spectral texture features are concatenated and processed through a convolutional layer and a Softmax activation function to generate a spatially adaptive scale-gated vector. This vector represents the importance weights of the three scale features at each spatial location. Finally, the fused features at the three scales are dynamically weighted and summed using this gating vector, achieving adaptive aggregation of multi-scale features and obtaining a final fused feature that combines detail accuracy and contextual integrity.
[0038] S3: The hyperspectral image of the bauxite to be detected is segmented into overlapping blocks, input into the contamination detection model, and the class probability of each pixel is obtained. Based on the class probability, a contamination distribution map of the bauxite surface is generated.
[0039] Specifically, after obtaining the final fusion features, the hyperspectral image of bauxite to be detected is divided into spatially overlapping image blocks according to a preset step size, keeping the overlapping area no less than 30% of the image block size to ensure the prediction consistency of boundary pixels. These image blocks are then sequentially input into a well-trained contamination detection model, and the probability values of each pixel belonging to the three categories of pure bauxite, contaminants, and background are calculated through forward propagation. The pixel probabilities in the overlapping areas are fused using a Gaussian weighted average strategy, with higher weights in the central region and decreasing weights in the edge region. Finally, based on the fused probability matrix, the argmax function is used to determine the final category label of each pixel. By marking the contaminant category pixels with prominent colors, a bauxite surface contamination distribution map with the same spatial resolution as the original hyperspectral data is generated. This distribution map can be further improved by morphological opening and closing operations to eliminate isolated noise points, and a flood filling algorithm is used to repair internal pores in the region. Finally, a quantitative contamination detection result with clear boundaries and continuous regions is output.
[0040] Example 2: The present invention also provides a bauxite surface contamination detection system based on spectral imaging, for implementing the method of Embodiment 1, comprising: The data acquisition module is used to acquire three-dimensional hyperspectral data of bauxite samples, and to correct, filter and enhance the three-dimensional hyperspectral data to obtain preprocessed three-dimensional hyperspectral data. The detection model building module is used to construct a pollution detection model based on preprocessed 3D hyperspectral data, utilizing graph convolutional networks, dynamic perception collaborative convolution modules, and adaptive attention fusion modules to jointly optimize spectral and spatial features. The contamination detection module is used to segment the hyperspectral image of bauxite to be detected into overlapping blocks, input them into the contamination detection model, obtain the class probability of each pixel, and generate a contamination distribution map on the bauxite surface based on the class probability.
[0041] A further implementation method includes a data acquisition module comprising: The initial spectral acquisition unit uses an independently controllable ring LED array as an illumination source. It triggers illumination of different phases sequentially and uses a pushbroom hyperspectral imaging system to synchronously acquire hyperspectral data of each phase of bauxite samples on the conveyor belt. The hyperspectral data of each phase are then fused at the pixel level to obtain initial three-dimensional hyperspectral data. The data reconstruction unit is used to install an optical encoder on the side of the conveyor belt and, in conjunction with a calibration plate of known size, to perform geometric calibration and spatial reconstruction on the initial three-dimensional hyperspectral data to obtain distortion-free three-dimensional hyperspectral data. The pulse signal of the optical encoder is synchronized with the line exposure signal of the pushbroom hyperspectral imaging system, and the spatial coordinates of each acquisition line are determined by the cumulative number of pulses of the optical encoder.
[0042] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for detecting surface contamination in bauxite based on spectral imaging, characterized in that, include: Three-dimensional hyperspectral data of bauxite samples were collected, and the three-dimensional hyperspectral data were corrected, filtered, and enhanced to obtain preprocessed three-dimensional hyperspectral data. Based on preprocessed 3D hyperspectral data, a pollution detection model with spectral-spatial feature co-optimization is constructed using graph convolutional networks, dynamic perception co-convolutional modules, and adaptive attention fusion modules. The hyperspectral image of the bauxite to be detected is segmented into overlapping blocks, which are then input into the pollution detection model to obtain the category probability of each pixel. Based on the category probability, a pollution distribution map of the bauxite surface is generated.
2. The method according to claim 1, characterized in that, Methods for acquiring three-dimensional hyperspectral data of bauxite samples include: An independently controllable ring LED array was used as the illumination source. Illumination of different phases was triggered sequentially. Hyperspectral data of each phase of bauxite samples on the conveyor belt were acquired synchronously using a pushbroom hyperspectral imaging system. The hyperspectral data of each phase were then fused at the pixel level to obtain the initial three-dimensional hyperspectral data. An optical encoder is installed on the side of the conveyor belt. Combined with a calibration plate of known size, the initial three-dimensional hyperspectral data is geometrically calibrated and spatially reconstructed to obtain distortion-free three-dimensional hyperspectral data. The pulse signal of the optical encoder is synchronized with the row exposure signal of the pushbroom hyperspectral imaging system, and the spatial coordinates of each acquisition row are determined by the cumulative number of pulses of the optical encoder.
3. The method according to claim 1, characterized in that, In the pollution detection model, the graph convolutional network is used to enhance the spectral features of the three-dimensional hyperspectral data by introducing a linear iterative clustering algorithm and an adaptive adjacency matrix, thereby obtaining enhanced spectral features. The dynamic perception collaborative convolution module is used to extract the local spatial-spectral texture features formed between adjacent pixels and adjacent bands in the three-dimensional hyperspectral data by using a split 3D convolution block. The adaptive attention fusion module is used to concatenate the enhanced spectral features and the local spatial-spectral texture features in the channel dimension to generate a spatial attention map, and then use the spatial attention map to perform a weighted summation of the two features to obtain the final fused features.
4. The method according to claim 3, characterized in that, The method for obtaining the enhanced spectral features includes: Calculate the local spectral variance map of the input 3D hyperspectral data, and dynamically adjust the superpixel granularity of the linear iterative clustering algorithm based on the local spectral variance map to construct an adaptive superpixel grid; treat each superpixel as a graph node, and use the average spectrum of all pixels within the superpixel as the initial feature of the node; Based on the cosine similarity of the initial features of the nodes and the K-nearest neighbor algorithm, an initial adjacency matrix is calculated as a topological prior. At the same time, the initial features of the nodes are input into a shared neural network, and a data-driven adjacency matrix is generated through differential calculation. The initial adjacency matrix and the data-driven adjacency matrix are weighted and summed using learnable weight parameters to obtain the final adaptive adjacency matrix; self-connections are added to the adaptive adjacency matrix and normalized to obtain the adjacency matrix used for convolution. Based on the adjacency matrix used for convolution, a graph convolutional layer with a gated mechanism is used to perform spectral context aggregation, thereby obtaining node features with enhanced spectral context and preserved boundaries. The enhanced spectral features are obtained by mapping the spectral context-enhanced and boundary-preserving node features back to the original spatial grid.
5. The method according to claim 3, characterized in that, The method for extracting the local spatial-spectral texture features includes: Based on the three-dimensional hyperspectral data, an initial feature map is obtained, the input initial feature map is processed in parallel, a spatially aware attention map and a spectral coordinated attention vector are generated, and the basic transformation features are obtained through a split 3D convolutional block. Multiply the spatial awareness attention map with the basic transformation feature to perform spatial enhancement and obtain the spatial enhancement feature; The spectral coordination attention vector is multiplied by the spatial enhancement feature to perform spectral-channel calibration and obtain the calibration feature; The calibration features are added to the linearly transformed initial feature map via residual connection to obtain local spatial-spectral texture features.
6. The method according to claim 3, characterized in that, Methods for obtaining the final fusion features include: A spatial attention map is generated based on the cross-covariance of the enhanced spectral features and the local spatial-spectral texture features; At multiple parallel scales, the enhanced spectral features and the local spatial-spectral texture features are weighted and fused using the spatial attention map to obtain multi-scale fused features; Based on the enhanced spectral features and local spatial-spectral texture features, a scale-gated vector is generated, and the multi-scale fusion features are dynamically weighted and summed using the scale-gated vector to obtain the final fusion features.
7. A bauxite surface contamination detection system based on spectral imaging, used to implement the method described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire three-dimensional hyperspectral data of bauxite samples, and to correct, filter and enhance the three-dimensional hyperspectral data to obtain preprocessed three-dimensional hyperspectral data. The detection model building module is used to construct a pollution detection model based on preprocessed 3D hyperspectral data, utilizing graph convolutional networks, dynamic perception collaborative convolution modules, and adaptive attention fusion modules to jointly optimize spectral and spatial features. The contamination detection module is used to segment the hyperspectral image of bauxite to be detected into overlapping blocks, input them into the contamination detection model, obtain the category probability of each pixel, and generate a contamination distribution map on the bauxite surface based on the category probability.
8. The system according to claim 7, characterized in that, The data acquisition module includes: The initial spectral acquisition unit uses an independently controllable ring LED array as an illumination source. It triggers illumination of different phases sequentially and uses a pushbroom hyperspectral imaging system to synchronously acquire hyperspectral data of each phase of bauxite samples on the conveyor belt. The hyperspectral data of each phase are then fused at the pixel level to obtain initial three-dimensional hyperspectral data. The data reconstruction unit is used to install an optical encoder on the side of the conveyor belt, and in conjunction with a calibration plate of known size, to perform geometric calibration and spatial reconstruction on the initial three-dimensional hyperspectral data to obtain distortion-free three-dimensional hyperspectral data; wherein, the pulse signal of the optical encoder is synchronized with the row exposure signal of the pushbroom hyperspectral imaging system, and the spatial coordinates of each acquisition row are determined by the cumulative number of pulses of the optical encoder.