91 results about "Affinity matrix" patented technology

The invention discloses a multi-view construction personnel tracking method and system based on attention perception. The method comprises the following steps: giving synchronous images from S cameras, and inputting the synchronous images into an encoder for feature extraction to obtain a multi-view feature map; transforming the multi-view feature map into a unified aerial view space by using perspective projection, and aggregating features after projection transformation of all views by using a convolutional layer; inputting the aggregated aerial view features into a decoder for decoding; a cross attention module is introduced, bird's-eye view features of a current frame and an adjacent frame are processed through 3D position coding, instance tokens are extracted as query, keys and values, an affinity matrix is generated through CNN coding similarity, features are propagated through matrix multiplication, and the bird's-eye view features of the current frame are updated. According to the method, the multi-view feature map is projected to the aerial view to realize early fusion, and a cross-frame attention mechanism is introduced, so that the problem of appearance feature distortion caused by perspective transformation is solved.

The invention provides a JSON (JavaScript Object Notation) data dynamic addition and management method based on a target keyword, which relates to the technical field of data processing, and comprises the following steps of: generating a keyword affinity matrix through grammar dependency analysis for clustering, establishing a mapping relationship between semantic identifiers and JSON nodes to form a semantic addressing table, constructing a dependency topological graph to determine a data block, and establishing a database; and accurate data addition and management based on keywords are realized. According to the method, the retrieval efficiency of the JSON data is improved, the data relevance is enhanced, the storage structure is optimized, and intelligent organization and maintenance of the data are realized.

The invention relates to an activation value re-segmentation communication method for a heterogeneous AI acceleration chip server. The activation value re-segmentation communication method comprises the following steps: scanning through chips, network interface cards and PCIe levels in the AI acceleration chip server, extracting network topology information of the server, calculating delay and bandwidth from each chip to each network interface card, and generating an affinity matrix; based on the affinity matrix, each chip is bound to a network interface card in an exclusive or shared mode through maximum matching; based on the server network topology information, the chip-network interface card affinity and the network interface card and PCIe single line bandwidth, activation value communication mode matching is carried out, and communication modes comprise a transmitting / receiving mode, a transmitting / receiving + collecting mode and a transmitting / receiving + broadcasting mode; and double-level multi-channel communication between the chip and the network interface card is realized. The method has the advantages of remarkably improving the bandwidth utilization rate, being highly adaptive to topology difference, being high in heterogeneous compatibility and the like.

The invention provides a video anomaly detection method and system based on semantic and amplitude depth collaboration, and the method comprises the steps: extracting a visual feature sequence of a video frame; enhancing the feature vector of the corresponding time step in the visual feature sequence by using the amplitude and energy of each time step to obtain an enhanced feature sequence; calculating a semantic similarity matrix of the enhanced feature sequence; calculating a sparse affinity matrix according to the amplitude of each time step in the visual feature sequence; fusion is carried out to obtain a final attention weight; and performing weighted aggregation on the value vector of the enhanced feature sequence by using the final attention weight, and inputting the value vector into a classifier to obtain a frame-level anomaly score so as to realize video anomaly detection. According to the method, through one-time parameter-free feature engineering, deep collaboration of semantics and amplitude information is forcibly realized on a model architecture level, and the problem that an existing weak supervision video anomaly detection method is insensitive to feature amplitudes is solved.

The invention discloses an SAR image weak supervision semantic segmentation method based on decision optimization pseudo label generation, and the method comprises the following steps: 1, training a multi-label classification network, and obtaining class activation mapping results of different layers of a neural network and pixel-level prediction pseudo labels; 2, fusing different levels of class activation mapping results to obtain a class activation mapping result with higher confidence; step 3, forming a decision by utilizing the relation between different layers of pseudo labels on the pixels, and extracting multi-layer decision pixel relation information; 4, constructing a relation network among pixels, and obtaining a centroid displacement vector and boundary features of each pixel in the image; 5, training a relation network between pixels by using the extracted pixel relation information to obtain a pixel semantic affinity matrix; step 6, performing pseudo tag optimization by using the pixel semantic affinity matrix; and step 7, training a semantic segmentation network by using the optimized pixel-level pseudo labels, and completing SAR image weak supervision semantic segmentation. According to the method provided by the invention, the high-quality polarized SAR image pixel-level pseudo tag can be generated, and pixel-level semantic segmentation of the SAR image under the condition of image-level annotation weak supervision is realized.

The invention discloses an equipment manufacturing energy consumption and state visual monitoring method based on an industrial internet of things, solves the problem of passive fixation of an industrial monitoring visualization technology in the prior art, and realizes autonomous and intelligent generation of an optimal visual view matched with a real-time working condition. The method comprises the following steps: collecting industrial field multi-source heterogeneous data in real time, abstracting a monitoring object as an entity node, and constructing a feature vector containing static and dynamic attributes; constructing and updating a dynamic knowledge graph based on nodes and features, wherein relation edges of the dynamic knowledge graph comprise dynamic edges generated by static connection and real-time analysis; inputting the atlas into a pre-trained multi-attention graph neural network to obtain a display priority score and a visual affinity matrix of each node; calculating coordinates of all nodes on an interface through a layout optimization algorithm in combination with the priority and the affinity matrix; and generating a visual drawing instruction according to the coordinates and the node attributes, and driving a display device to render a monitoring interface.

The invention belongs to the technical field of motion data optimization, and discloses a double-path parallel motion data optimization method fusing space-time priori and attention, and the method proposes a double-priori attention module, fuses space and time priori constraints, facilitates the improvement of the modeling capability of Transform on the dependence of a joint structure and a time sequence, and improves the accuracy of motion data optimization. And the space rationality of the skeleton structure is enhanced. An affinity matrix is constructed through a skeleton connection matrix and a learnable topological matrix, the spatial rationality and time sequence coherence are enhanced, prior bias is superimposed in attention calculation, data driving and kinematics constraint are considered, and the robustness and accuracy of the model are improved. Compared with a traditional method, the method has the advantages that the three-dimensional motion data optimized by the method is more accurate in detail processing, and particularly, the precision is obviously improved in expression of rapid gestures or tiny actions. In addition, the model structure is simplified, and the calculation efficiency is also remarkably improved.

The invention provides a remote sensing image change detection method and system based on relation perception high-order interaction, and the method comprises the steps: firstly collecting remote sensing images of the same region in different periods, constructing a dual-temporal data set, employing an independent convolution encoder to extract single-temporal features, and keeping the original temporal-spatial information; the method comprises the following steps: constructing a spatial-temporal feature pyramid through a twin network structure, introducing a time relation perception module to perform modeling on space and channel dimensions, and capturing mutual relations between pixels and between channels by using a grouping attention mechanism and an affinity matrix calculation mode to realize deep fusion between features; the high-order space-time interaction module stacks a plurality of interaction units through a recursive structure, and enhances context information modeling in a spatial dimension, so that complex high-order feature interaction expression is realized, weighted cross entropy loss and weighted IoU loss are combined, the detection capability of the model on an edge region and a difficult-to-divide region is improved, and the detection efficiency is improved. And inputting a dual-time remote sensing image into the trained model, and outputting a high-precision change detection result.

The invention provides a text travel scene adaptive illumination dynamic regulation and control method and system based on personal perception, and relates to the technical field of scene perception, and the method comprises the steps: capturing visitor multi-dimensional behavior data and scene environment state data in real time through a distributed sensor network; constructing a body perception model fusing the visitor spatial position, the moving speed, the staying duration and the physiological wake-up degree, performing space-time correlation coding on the multi-dimensional behavior data, and generating a body perception feature vector; establishing a dynamic affinity matrix based on the feature vector and scene semantic information, and mapping the dynamic affinity matrix to a spatial topological structure of lighting equipment to form a regional lighting coupling map; according to the map node weight distribution and visitor perception state clustering result, generating a partition cooperative illumination regulation and control instruction through a multi-objective optimization algorithm; and carrying out online evolution on the model by utilizing the regulated and controlled visitor behavior track and physiological index change. According to the invention, accurate matching of the illumination environment and the visitor immersion state can be realized, and the experience quality of a text travel scene is improved.

The application discloses a semi-supervised medical image segmentation method based on model self-distillation and prototype learning, and relates to the technical field of image signal processing. The semi-supervised medical image segmentation method based on model self-distillation and prototype learning comprises the following steps: S1, establishing a semi-supervised medical image segmentation dataset; S2, constructing a network architecture; S3, designing a semi-supervised medical image segmentation scheme, and building a semi-supervised medical image segmentation model according to the designed scheme; S4, training the semi-supervised medical image segmentation model by using a deep learning Pytorch framework; and S5, inputting a medical image to be segmented into the model to obtain a medical image segmentation result. The semi-supervised medical image segmentation performance is improved to a new height by using the proposed double-flow memory bank architecture, the self-distillation method based on an image block affinity matrix and the prototype synthesis method based on context matching.

The present invention belongs to the field of data clustering technology, and provides a method and system for industrial coal-fired boiler operating condition division based on data integration clustering to obtain the operating status data of the industrial coal-fired boiler; the present invention performs integrated clustering operations on the industrial coal-fired boiler data through three steps of mixed representative nearest neighbor similarity, bipartite graph segmentation and third-order tensor integration, and realizes effective operating condition division of the boiler data; specifically, by constructing a sparse affinity submatrix through mixed representative nearest neighbor similarity, it can solve the problem that the traditional clustering method has too high computational time complexity and cannot effectively construct the affinity matrix of the coal-fired boiler data; by dividing the bipartite graph, the time for solving the characteristic problem is reduced; and by integrating the multi-base clustering results into a unified integrated clustering framework, the accuracy and robustness of the clustering are further improved while maintaining high efficiency.

The application discloses a kind of city area photovoltaic cluster dynamic grid division methods based on space-time coupling tensor, comprising: synchronously collecting the active power time series and geographic coordinates of all photovoltaic nodes in city area;Build the heterogeneous space adjacency matrix considering electrical topology constraint, calculate the static space correlation degree between any two photovoltaic nodes;Build the time-shift cross-correlation matrix that captures the moving characteristics of weather system, quantify the dynamic time correlation degree between nodes;Perform space-time feature tensor fusion, obtain global space-time affinity matrix;Photovoltaic nodes are divided using improved spectral clustering algorithm, and the optimal grid number is adaptively determined by contour coefficient, finally generate photovoltaic cluster grid division scheme at this moment;According to grid drift threshold, trigger grid reconstruction.The application can effectively consider physical topology constraint and meteorological time delay characteristics, realize the accurate aggregation of photovoltaic cluster, significantly improve the accommodation capacity and regulation flexibility of distribution network to distributed energy.

The invention provides a task scheduling method and device, equipment and a medium, and the method comprises the steps: building an inter-task communication matrix based on the task communication traffic between every two sub-tasks in a plurality of to-be-scheduled sub-tasks; constructing a network affinity matrix based on the bandwidth and the congestion index between every two task processing nodes in the plurality of task processing nodes to be scheduled; performing clustering processing on the plurality of subtasks based on the inter-task communication matrix to obtain a task cluster set, and performing clustering processing on the plurality of task processing nodes based on the network affinity matrix to obtain a node cluster set; and according to the resource demand information of each sub-task and the resource supply information of each task processing node, mapping the task cluster set and the node cluster set, and determining a mapping result between the plurality of sub-tasks and the plurality of task processing nodes. Therefore, the task cluster with dense communication can be mapped to the node cluster with high network affinity, the communication overhead is reduced, and the end-to-end time delay of the task is reduced.

The invention discloses a power distribution network dynamic partitioning method based on a composite electrical distance. The method comprises the steps that the composite electrical distance between each pair of buses in a power distribution network is calculated; constructing an affinity matrix according to the composite electrical distance, and constructing an undirected weighted graph based on the affinity matrix; calculating a Laplacian matrix for the undirected weighted graph to obtain a preliminary partition of the power distribution network; a mixed integer linear programming method is adopted to optimize the preliminary partitioning result, it is ensured that each sub-region meets set constraints, and optimized partitions are output; topological change of a power distribution network and output fluctuation of a distributed power supply are monitored in real time, and when the change exceeds a preset threshold value, a dynamic adjustment mechanism is triggered, and a partition adjustment strategy is executed. According to the method, on the basis of comprehensively describing the electrical similarity of the bus, the engineering feasibility and the running state adaptability of the partitioning result can be realized by combining constraint optimization and a dynamic adjustment mechanism.

The invention relates to the technical field of power system automation, and particularly discloses a power network fault early warning method based on a large model, and the method comprises the steps: directly mapping the data of a high-frequency power sensor into a time sequence soft prompt vector which can be understood by the large model through a lightweight encoder and a linear projection technology; deep alignment of continuous signals and discrete semantics is realized, and information loss and efficiency bottleneck caused by numerical textualization are avoided. And then, directionally retrieving an operation and maintenance knowledge base by using the soft prompt vector, introducing a cross-modal affinity matrix to construct a semantic gating mechanism, dynamically screening and enhancing retrieved key knowledge fragments according to real-time waveform characteristics, and automatically inhibiting redundant text noise irrelevant to the current working condition. And finally, deep attention interaction and reasoning are performed on a composite sequence formed by an instruction, enhanced knowledge and soft prompt through a large model, a control signal containing fault classification and disposal suggestions is output, and real-time accurate early warning considering high-frequency signal sensitivity and expert knowledge logicality is realized.

The invention belongs to the technical field of hyperspectral image recognition, and relates to a method for recognizing spectral and spatial features of a hyperspectral image, which comprises the following steps of: dividing the hyperspectral image into N image blocks with spectral and spatial features; the method comprises the following steps: extracting and activating image blocks from a hyperspectral image, performing batch standardization processing on the image blocks to generate first-order statistical features, and performing element-by-element multiplication on the first-order statistical features and transposed outer products thereof to obtain second-order statistical features; converting the second-order statistical features into a local feature map, inputting the local feature map into a cross-orthogonal attention mechanism, calculating a horizontal attention matrix and a vertical attention matrix through an affinity matrix used for measuring correlation between feature mappings, and calculating a Chebyshev distance according to the horizontal attention matrix and the vertical attention matrix to adjust weighted features of the local feature map; and carrying out two-dimensional convolution processing on the weighted features to generate a weighted feature map, carrying out vectorization processing through a matrix, and completing identification by using a full connection layer.

A computer-implemented method of image stitching of a plurality of digital images of an infrastructure surface for defect detection is disclosed. The method includes providing a plurality of partial digital images of the infrastructure surface. The method includes extracting global positioning system metadata from data corresponding to the partial digital images. The method includes determining feature descriptions of features in one or more of the partial digital images. The method includes performing a scheduled processing sequence of the partial digital images based on the extracted global positioning system metadata, including determining an affinity matrix using the feature descriptions of adjacent partial digital images to incrementally locate each of the partial digital images such that a full view image of the infrastructure surface is produced by iteratively digitally stitching together the plurality of partial digital images.

The invention discloses a multi-view image fusion representation learning method, device and equipment based on double-layer optimization, and relates to the technical field of machine learning. The method comprises the following steps: splicing feature matrixes of a multi-view data set to obtain a multi-view data matrix; calculating an initial affinity matrix of each view, and initializing a fusion affinity matrix, a view weight set, potential feature representation and a shared mapping matrix; a bilayer optimization model is constructed to learn potential feature representations. The model comprises an objective function of an upper-layer optimization task for updating potential feature representation and sharing a mapping matrix under the constraint of a fixed fusion affinity matrix, and an objective function of a lower-layer optimization task for updating the fusion affinity matrix and a view weight set under the constraint of the fixed potential feature representation. And executing alternative optimization iteration: executing upper-layer optimization task update and first, and then executing lower-layer optimization task update and until ending. And outputting the final potential feature representation for executing the task.

The invention relates to the technical field of image processing, in particular to an unsupervised PolSAR classification method based on comparative learning, and the method comprises the steps: obtaining original PolSAR image data to form multi-view data; performing superpixel segmentation on the image to generate multi-view primitive samples in one-to-one correspondence with superpixels; obtaining potential feature representation of each view, and generating a self-representation coefficient matrix of each view through a forward self-representation module; applying cross-view relation consistency constraint by comparing loss, and optimizing a self-representation coefficient matrix of each view; fusing the self-representation coefficient matrixes of all the optimized views to generate a uniform affinity matrix, and executing spectral clustering to obtain clustering labels of element samples; according to the mapping relation between the element sample and the pixel established by the superpixel segmentation, the clustering label is mapped back to the pixel level, the image classification result is obtained, and the classification accuracy and robustness are remarkably improved.

The application provides a deep multi-view clustering method and system based on collaborative training, relates to the fields of data mining and machine learning, and specifically includes the following steps: obtaining multi-views to be clustered, and pre-training a deep self-encoding model for each view; calculating an affinity matrix of each view, setting an initial weight of each view and an initial dynamic learning factor between the views; performing iterative formal training on the deep self-encoding model of each view, updating the weight of each view and the dynamic learning factor between the views until a preset iteration stopping condition is met, and outputting final hidden layer features and clustering centers; and obtaining a clustering result based on the final hidden layer features and the clustering centers; the application designs a deep multi-view weighted graph embedding clustering algorithm based on dynamic collaborative training, adopts a dynamic collaborative training idea for multi-view data, deeply mines complementary information between the multi-views, and improves the clustering effect for multi-view data sets.

The invention discloses a pathological image segmentation method and system based on coevolution generation type difficult sample mining, and the method comprises the steps: constructing a mask synthesis engine guided by biological information, and generating a cell nucleus mask through introducing a cell affinity matrix and structure prior based on a graph theory; establishing a segmentation-oriented adversarial renderer, and aligning the generated image with a real image in a feature space by using a multi-layer feature discriminator; implementing a closed-loop co-evolution strategy, dynamically identifying vulnerability categories by using performance feedback of the segmentation model, and guiding a generator to carry out adaptive difficult sample mining; and obtaining a cell nucleus segmentation result through alternate mutual promotion of the generator and the segmentation model and regression fine tuning of real data. According to the method, the problems that in existing small sample learning, generated data lacks biological rationality and visual fidelity cannot be converted into segmentation performance are solved, and the segmentation precision and generalization ability of the model are remarkably improved under the condition of extremely few labeled data.

The embodiment of the invention provides an incremental multi-view data clustering method and system based on a cross-time consensus graph, and belongs to the technical field of artificial intelligence. The method comprises the following steps: integrating historical knowledge with a consensus affinity matrix of current view information and learning time based on kernel induction expression, and constructing a dynamic consensus graph; spectrum embedding and discrete label learning are carried out; and alternately optimizing the consensus affinity matrix, the orthogonal rotation matrix, the consensus spectrum embedding matrix and the discrete clustering label matrix of the moments in the dynamic consensus graph by adopting staged updating variables. The method can efficiently adapt to incremental environment application, and is better in clustering precision, higher in calculation efficiency, higher in time sequence stability and better in robustness.

The invention discloses a multi-view agricultural image clustering method based on enhanced multi-order similarity learning, and relates to the technical field of image clustering, and the method comprises the following steps: preprocessing image data, constructing an initial affinity matrix which is used for representing the affinity between data points, the method comprises the following steps: capturing a local structure and a neighborhood relationship of data points through first-order similarity and second-order similarity, stacking and rotating all affinity matrixes into a third-order tensor, constraining by using a weighted tensor Schatten-p norm, and optimizing all modules in a unified optimization framework. According to the method, through parallel mining of multi-order similarity, weighted tensor Schatten-p norm optimization and unified spectral clustering fusion, multi-view data complementary features are fully utilized, clustering precision and robustness are improved, and reliable technical support can be provided for precise agricultural application such as crop health monitoring and pest and disease damage detection.

The invention discloses an insulator defect feature extraction method and device based on multi-centroid transfer learning, and the method comprises the steps: collecting insulator defect image data sets of various environments based on the environment influence factors of an insulator; dividing the insulator defect image data set in combination with the defect type of the insulator to obtain a source domain defect image set and a target domain defect image set in different environments; obtaining a source domain affinity matrix and a target domain affinity matrix through similarity analysis, calculating a corresponding Laplacian matrix, and obtaining sample clustering results corresponding to the source domain and the target domain; according to the sample clustering results corresponding to the source domain and the target domain, performing spectral clustering on the samples of each insulator defect type to obtain a source domain class mass center and a target domain class mass center; and carrying out transfer learning between the source domain and the target domain through the source domain centroid and the target domain centroid to obtain a feature projection matrix so as to realize defect feature extraction which is not influenced by scene change.

The invention discloses a malignant tumor image fusion analysis method and system based on graph structure consensus synergy, and the method comprises the steps: receiving multi-modal medical image data, extracting an initial high-dimensional feature vector, and projecting the initial high-dimensional feature vector to a plurality of independent potential semantic subspaces; a feature influence value and a dimension discrimination score are calculated through a dual significance evaluation mechanism, node features are weighted, edge weights are modulated, and a sample specificity initial multi-path diagram is constructed; after node features are linearly expanded, an affinity matrix is generated and sparsified, and a robust optimization graph is output through graph structure enhancement and contrast learning; coding each plane feature by adopting a graph attention network, and generating a fusion feature through high-discrimination feature splicing and weighted summation of other features; and constructing a global aggregation graph by taking the fusion features as nodes, and aggregating global information to output a diagnosis result. The system correspondingly comprises a multi-modal feature extraction unit, a multi-modal feature analysis unit and the like, multi-modal image semantic synergy and structure optimization are achieved, and malignant tumor diagnosis accuracy is improved.

The invention belongs to the technical field of target detection, and relates to a target detection method and device based on spatial semantic similarity knowledge distillation and a medium. Respectively inputting the image samples in the training set into a teacher model and a student model for feature extraction, and outputting a teacher feature map and a student feature map; obtaining a semantic similarity matrix and a spatial distance similarity matrix of each feature map; fusing the semantic similarity matrix and the spatial distance similarity matrix of each feature map to obtain a global affinity matrix of each feature map; generating a local neighborhood mask for the global affinity matrix of each feature map, and obtaining a spatial semantic similarity matrix of each feature map; according to the method, target detection identification is carried out based on the teacher feature map and the spatial semantic similarity matrix thereof and the student feature map and the spatial semantic similarity matrix thereof, target detection distillation loss is constructed, a student model is trained, target detection is carried out by using the trained student model, and the target detection precision is improved.

The invention discloses a city area photovoltaic cluster dynamic grid division method based on a space-time coupling tensor. The method comprises the steps that active power time sequences and geographic coordinates of all photovoltaic nodes in a city area are synchronously collected; constructing a heterogeneous space adjacency matrix considering the electrical topology constraint, and calculating the static space correlation degree between any two photovoltaic nodes; constructing a time-shifting cross-correlation matrix for capturing the moving characteristics of the weather system, and quantifying the dynamic time correlation degree between nodes; executing spatio-temporal feature tensor fusion to obtain a global spatio-temporal affinity matrix; dividing photovoltaic nodes by using an improved spectral clustering algorithm, adaptively determining an optimal grid number through a contour coefficient, and finally generating a photovoltaic cluster grid division scheme at the moment; and triggering grid reconstruction according to the grid drift threshold. According to the method, physical topology constraints and meteorological time delay characteristics can be effectively considered, accurate aggregation of photovoltaic clusters is realized, and the acceptance capability and regulation and control flexibility of a power distribution network to distributed energy sources are remarkably improved.

This invention discloses a method, apparatus, computing device, and storage medium for wire identification. The method includes: acquiring an image of a wire to be identified; preprocessing the wire image to obtain a target image; inputting the target image into a trained convolutional neural network model for feature extraction to obtain the key point coordinates and categories of the wire, the affinity matrix between key points, and the bidirectional connection vector field between key points; based on the key point coordinates and categories, using a matching algorithm to perform pairwise matching of key points to obtain a first matching result between key points; confirming the first matching result based on the affinity matrix between key points to obtain a second matching result between key points; and confirming the second matching result based on the bidirectional connection vector field between key points to obtain multiple sets of associated key points, with each set of key points identifying a wire. This scheme can improve the accuracy of wire identification in complex scenes.

The invention discloses a data exchange task scheduling method and system based on resource prediction. A resource prediction model is constructed by acquiring a resource state change vector of a computing node in a preset time period, and resource availability in a future scheduling window is predicted; obtaining a to-be-scheduled data exchange task queue, and determining task characteristics of each task; based on the task characteristics and the resource availability prediction result, calculating the dominant demand degree of the task to the resource and the implicit attraction of the resource to the task, and carrying out task and resource matching to obtain a task-resource affinity matrix; and generating a scheduling strategy according to the affinity matrix to realize data exchange task scheduling. By introducing a resource availability prediction mechanism, the resource utilization rate is improved, scheduling conflicts and resource bottlenecks are avoided, and the overall performance of a data exchange platform is enhanced.

The application discloses a network (SGM-LGAT) for constructing a multi-level graph structure guided by superpixels for HSI classification. The main steps include the following: we create different stages of the adjacency connection graph of hyperspectral data from the superpixel representation by using a hierarchical superpixel segmentation algorithm to effectively utilize the spatial topology, and obtain the adjacency matrix and the affinity matrix of different stages. Then, we also design a lite GAT module (LGAT) to prune the graph by spectral sparsification to remove some edges in the graph that are not important for the task, and use the adjacency matrix to assign a unique attention coefficient to each remaining edge. In addition, the LGAT is used to concatenate with the spectral branch to update and fuse the features. The SGM-LGAT has small calculation amount and high efficiency, can mine the features of the HSI from the perspective of multi-scale hierarchy, and proves through a large number of comparative experiments on three benchmark datasets that the performance is competitive with other deep learning-based methods.