Remote sensing scene classification method and system combining human perception and Lie group characteristics

Abstract

The present invention discloses a remote sensing scene classification method and system that combines human perception with Lie group features, relating to the field of remote sensing technology. The method comprises: obtaining a remote sensing image corresponding to a target area, segmenting the remote sensing image using a superpixel segmentation method based on the inner mean of the Lie group to obtain feature maps corresponding to multiple target objects; capturing the correlation between the multiple feature maps obtained by segmentation through an external-graph attention mechanism; and aggregating and enhancing the scene features of the feature maps in combination with an internal-graph attention mechanism to obtain a scene classification result for the remote sensing image. The present invention first captures the correlation between different graphs through an external attention module, then further aggregates and enhances the features in combination with internal attention, and finally obtains a classification result. This method significantly reduces the complexity of the model while maintaining classification accuracy.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing technology, and in particular to a remote sensing scene classification method and system combining human perception with Lie group characteristics. Background Art

[0002] High-resolution remote sensing image scene classification (HRSC) is a fundamental task in remote sensing, widely used in areas closely related to human life, such as disaster management, agricultural monitoring, and urban inspection. However, due to the complex spatial layout and geometric structure of high-resolution remote sensing imagery (HRRSI), distinguishing different scenes is a key research topic in HRSC.

[0003] However, most existing models rely primarily on convolutional operations to extract high-level semantic features of a scene, but these features are often difficult for humans to intuitively understand. In reality, when humans recognize different scenes, they first focus on shallow features such as the scene's appearance, physical structure, and geometry, and then combine this with high-level semantic information to make further judgments. Therefore, existing models do not fully adhere to human cognitive logic during feature extraction and scene recognition, which to some extent limits the models' representational capabilities and comprehensibility.

[0004] Existing graph convolutional networks (GCNs) primarily rely on attention mechanisms within graphs, but fail to fully consider the inter-graph relationships and relationships within them. They also neglect the spatial location and layout combinations within a scene. In real-world scenarios, the spatial location and arrangement of different objects are crucial for distinguishing different scene categories. For example, the same building elements can form medium-density, sparse, and dense communities under different spatial distributions and layouts. Therefore, effectively capturing and constructing spatial and layout information between graphs is crucial for scene classification.

[0005] Furthermore, most existing models suffer from an imbalance between classification accuracy and model complexity. For example, models with high classification accuracy often have complex structures, high feature dimensions, and a large number of parameters. This results in poor computational performance and significant memory usage, particularly when processing HRRSI. Furthermore, relatively complex models are more prone to overfitting, particularly when training with insufficient samples, significantly reducing their generalization capabilities. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a remote sensing scene classification method and system that combines human perception and Lie group characteristics, aiming to reduce the complexity of scene classification while maintaining classification accuracy.

[0007] A first aspect of the present invention is to provide a remote sensing scene classification method combining human perception and Lie group characteristics, the method comprising:

[0008] Acquire a remote sensing image corresponding to the target area, segment the remote sensing image using a superpixel segmentation method based on the inner mean of a Lie group, and obtain feature maps corresponding to multiple target objects;

[0009] Capturing the correlation between the multiple feature maps obtained by segmentation through the out-of-graph attention mechanism;

[0010] The scene features of the feature map are aggregated and enhanced in combination with the in-graph attention mechanism to obtain the scene classification result of the remote sensing image.

[0011] According to one aspect of the above technical solution, the step of capturing the correlation between the plurality of feature maps obtained by segmentation through an out-of-graph attention mechanism includes:

[0012] Based on the out-of-graph attention mechanism, the node features of each feature graph and the edge features between any multiple feature graphs are calculated respectively;

[0013] The correlation between nodes in different feature graphs is calculated based on the node features of each feature graph and the edge features between any plurality of feature graphs.

[0014] According to one aspect of the above technical solution, the step of calculating the correlation between nodes in different feature graphs based on the node features of each feature graph and the edge features between any plurality of feature graphs is expressed as:

[0015] ;

[0016] Where, A E is the similarity between the input graph nodes and the external units, Norm represents the normalization operation, F is a data sample with a size of n × d ,in n is the number of nodes, d For the dimension, P To learn the parameters, T represents the transpose operation, R is a real number, s is the number of external nodes;

[0017] The step also includes normalizing the similarities between the input graph nodes and the external units in terms of rows and columns, respectively, as follows:

[0018] .

[0019] According to one aspect of the above technical solution, the steps further include:

[0020] Two independent sets of external units are used for keys and values respectively, and additional external units are added for edge features. The relationship between edge features and node features is determined by shared units. The expression is ;

[0021] Where, S EN Represents a shared unit for storing connections, with dimensions of d × d , P KN 、 P VN They are all external key-value pairs of nodes. P KE 、 P VE are all external key-value pairs of edges, T Represents a transpose operation.

[0022] According to one aspect of the above technical solution, the step of aggregating and enhancing the scene features of the feature graph in combination with the in-graph attention mechanism to obtain a scene classification result of the remote sensing image includes:

[0023] Based on the out-of-graph attention mechanism, combined with the in-graph attention mechanism, global average pooling is performed on each of the feature maps along the X-axis and the Y-axis;

[0024] Merge all feature maps after global average pooling to obtain a process map;

[0025] The process graph is passed through a deep convolution layer to perform deep feature extraction, and is activated through an activation function.

[0026] According to one aspect of the above technical solution, extracting deep features from the process graph through a deep convolutional layer and performing activation processing through an activation function include:

[0027] Perform 1×1 depth convolution operations along the X-axis and Y-axis respectively, use SeLU activation function for activation processing, scale the activation value to the range of [0,1], and obtain the spatial weight W S , feature map P The size is h×w×c, where h represents height, w represents width, and c represents the number of channel layers. It is activated by the SeLU activation function and then processed by the Lie group Sigmoid kernel function to obtain the channel weight. W C , the channel weight W C With feature map P Multiply to get a new feature map P C, the spatial weight W S With feature map P Multiply to get a new feature map P S , the feature map P C With feature map P S Perform pixel-by-pixel summation to obtain the final attention map;

[0028] Among them, the expression of the attention map is:

[0029] ;

[0030] Where ⨁ represents the pixel-by-pixel summation and the channel weight W C Used to suppress some channel information, spatial weight W S Used to enhance the spatial information in some feature maps.

[0031] According to one aspect of the above technical solution, the expression for combining the out-of-graph attention and the in-graph attention is:

[0032] ;

[0033] Where, Represents the first l +1 layer of data points, Represents the first l +1 layer of edges, MPNN represents the message passing GNN network, For the l The data points of the layer, Indicates the l The edge of the layer, A is the connection matrix, Indicates the l The data points of the layer are processed by the attention mechanism to obtain the Transformer layer with self-attention. GEANet represents the external attention of the graph, Indicates the l +1 layer of edges;

[0034] The method further includes updating the node features and edge features, and the expression is:

[0035] ;

[0036] in, 、 、 、 all , l is the layer index, i 、 j is the node index, N ( i ) is the i The neighborhood of a node, f node and f edge They represent the parameters of the learnable function nodes and edge message passing respectively.

[0037] A second aspect of the present invention is to provide a remote sensing scene classification system that combines human perception and Lie group characteristics, which is applied to the method described in the above technical solution. The system includes:

[0038] The first processing module is used to obtain a remote sensing image corresponding to the target area, and segment the remote sensing image using a superpixel segmentation method based on the inner mean of the Lie group to obtain feature maps corresponding to multiple target objects;

[0039] A second processing module is configured to capture the correlation between the plurality of feature maps obtained by segmentation through an out-of-graph attention mechanism;

[0040] The third processing module is used to aggregate and enhance the scene features of the feature map in combination with the in-graph attention mechanism to obtain the scene classification result of the remote sensing image.

[0041] The third aspect of the present invention is to provide a readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the above technical solution.

[0042] The fourth aspect of the present invention is to provide an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the above technical solution when executing the program.

[0043] Compared with the prior art, the remote sensing scene classification method and system combining human perception and Lie group characteristics shown in the present invention has the following beneficial effects:

[0044] This invention transforms the human perception process into a feature learning method and proposes a superpixel segmentation algorithm based on Lie group inner mean features to segment samples, thereby preserving features while reducing feature dimensionality. Secondly, the segmentation results are used to construct a graph, and a novel intra-graph and extra-graph attention mechanism is designed. This mechanism first captures the correlations between different graphs through an external attention module, then combines internal attention (Lie group space graph attention) to further aggregate and enhance features, ultimately achieving classification results. This significantly reduces model complexity while maintaining classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which:

[0046] Figure 1 Schematic diagram of the flow of a remote sensing scene classification method combining human perception and Lie group features in a first embodiment of the present invention;

[0047] Figure 2 This is a structural block diagram of a remote sensing scene classification system that combines human perception and Lie group features in the second embodiment of the present invention. DETAILED DESCRIPTION

[0048] To make the objectives, features, and advantages of the present invention more readily apparent, the following detailed description of specific embodiments of the present invention is provided in conjunction with the accompanying drawings. The accompanying drawings illustrate several embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0049] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly attached to the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0051] Example 1

[0052] See also Figure 1The first embodiment of the present invention provides a remote sensing scene classification method combining human perception and Lie group features, the method comprising steps S10 to S30:

[0053] Step S10: acquiring a remote sensing image corresponding to the target area, segmenting the remote sensing image using a superpixel segmentation method based on the inner mean of a Lie group, and obtaining feature maps corresponding to a plurality of target objects.

[0054] First of all, it should be noted that since the traditional simple linear iterative clustering (SLIC) algorithm mainly focuses on single features such as color, texture, geometry and position, it fails to fully consider the correlation between different features.

[0055] To solve this problem, this embodiment introduces Lie group machine learning into the SLIC algorithm and proposes a superpixel segmentation algorithm based on the mean within the Lie group to extract shallower features. In the feature extraction process, this embodiment transforms the human perception process into a feature learning method, as follows:

[0056] ;

[0057] Where (x, y) represents the position of the target object in the remote sensing image, that is, the pixel coordinates, (NR, NG, NB) represents the color of the scene, (Y, Cb, Cr) represents the brightness, color difference and saturation of the space, CH(x, y) represents the color histogram feature, SURF(x, y) uses the gradient information around the pixel coordinate (x, y) to describe the local area of the scene image, HOG(x, y) is the statistical calculation of the gradient direction histogram of the local area around the pixel coordinate (x, y), Gabor(x, y) represents the grayscale image of the scene, and LBP(x, y) is the binarization of the pixels around the pixel coordinate (x, y).

[0058] However, it is not enough to rely on shallow features alone. Therefore, in this embodiment, a high-level feature extraction module (HLFE) is designed to extract semantic features corresponding to remote sensing scenes in remote sensing images. The high-level feature extraction module includes multiple convolutional layers (Conv 3×3 and\or Conv 1×1 ), depth convolution layer (DepConv), batch normalization layer (BN layer), layer normalization layer (LN layer) and average pooling layer (AvgPooling layer).

[0059] First, to improve the generalization ability of the model and accelerate convergence, this embodiment uses a batch normalization layer to perform batch normalization on the input image. Next, a deep convolutional layer (DepConv) with a convolution kernel of 7×7 and a scaled exponential linear unit (SeLU) activation function is used to process the image to obtain a feature map. The obtained feature map is sent to the LGEMix module for further processing. The mathematical formula of the LGEMix module is as follows:

[0060] ;

[0061] Where, , for the i Input features to the layer encoder.

[0062] Step S20: capturing the correlation between the multiple feature maps obtained by segmentation through an out-of-graph attention mechanism.

[0063] It should be noted that in order to fully consider the different spatial layouts of the objects in the scene, in this embodiment, an undirected graph will be constructed based on the above-mentioned feature map, each superpixel in the feature map will be regarded as a node, the average features of all pixels contained in the superpixel will be used as the node features of the corresponding node, and the LGMS (Lie group manifold space distance) distance between each node will be used as the edge correspondence as the edge feature.

[0064] Specifically, use G=(V, E) represents an undirected graph, V is a node set, E is the edge set. The total number of nodes in the graph is recorded as N , the total number of edges is recorded as M . i The characteristics of ni , the node i With node j The characteristics of the edges between eij ,in, i 、 j ∈ V The characteristics of each node are expressed as a matrix F ∈ Rn × d To store, the features of each edge are used E ∈ Rm × d to store.

[0065] In this embodiment, the off-graph attention mechanism, also known as the external-graph attention mechanism, is used as a supplement to the internal-graph attention mechanism.

[0066] The step of capturing the correlation between the plurality of feature maps obtained by segmentation through an out-of-graph attention mechanism includes:

[0067] Based on the out-of-graph attention mechanism, the node features of each feature graph and the edge features between any multiple feature graphs are calculated respectively;

[0068] The correlation between nodes in different feature graphs is calculated based on the node features of each feature graph and the edge features between any plurality of feature graphs.

[0069] Specifically, in this embodiment, the self-attention of the graph can be regarded as a linear combination of the features of each node to replace the original node features. However, it only considers the association between each node in a single graph and ignores the hidden connections between nodes in different graphs. Therefore, in order to effectively focus on the features between different graphs, this embodiment obtains the correlation between nodes in different graphs by calculating the attention between each node feature of the graph and the external unit, which makes up for the problem of insufficient capacity and adaptability of the attention mechanism within the graph. Its mathematical expression is:

[0070] ;

[0071] Where, A E is the similarity between the input graph nodes and the external units, Norm represents the normalization operation, F is a data sample with a size of n × d ,in n is the number of nodes, d For the dimension, P To learn the parameters, T represents the transpose operation, R is a real number, s is the number of external nodes;

[0072] The step also includes normalizing the similarities between the input graph nodes and the external units in terms of rows and columns, respectively, as follows:

[0073] .

[0074] To improve network performance, this embodiment uses two independent sets of external units for key and value, respectively. Furthermore, to fully utilize the edge information of the input graph, this embodiment also adds additional external units specifically for edge features. Furthermore, a shared unit is used to record the relationship between edges and nodes, expressed as:

[0075] ;

[0076] Where, S ENRepresents a shared unit for storing connections, with dimensions of d × d , P KN 、 P VN is the external key-value pair of the node, P KE 、 P VE is the external key-value pair of the edge, T Represents a transpose operation.

[0077] In step S30 , the scene features of the feature graph are aggregated and enhanced in combination with the in-graph attention mechanism to obtain a scene classification result of the remote sensing image.

[0078] In this embodiment, the steps of aggregating and enhancing the scene features of the feature graph in combination with the intra-graph attention mechanism to obtain the scene classification result of the remote sensing image include:

[0079] Based on the out-of-graph attention mechanism, combined with the in-graph attention mechanism, global average pooling is performed on each of the feature maps along the X-axis and the Y-axis;

[0080] Merge all feature maps after global average pooling to obtain a process map;

[0081] The process graph is passed through a deep convolution layer to perform deep feature extraction, and is activated through an activation function.

[0082] This embodiment mainly highlights the important spatial positions on the target object feature map and suppresses channels with less information. It uses the previous Lie group spatial attention mechanism for optimization and then proposes the Lie group spatial graph attention mechanism as the intra-graph attention mechanism shown in this embodiment.

[0083] Specifically, extracting deep features from the process graph through a deep convolutional layer and performing activation processing through an activation function include:

[0084] Among them, a 1×1 depth convolution operation is performed along the X-axis and Y-axis respectively, and the SeLU activation function is used for activation processing to scale the activation value to the range of [0,1] to obtain the spatial weight W S , feature map P The size is h×w×c, where h represents height, w represents width, and c represents the number of channel layers. It is activated by the SeLU activation function and then processed by the Lie group Sigmoid kernel function to obtain the channel weight. W C , the channel weight W C With feature map PMultiply to get a new feature map P C , the spatial weight W S With feature map P Multiply to get a new feature map P S , the feature map P C With feature map P S Perform pixel-by-pixel summation to obtain the final attention map;

[0085] Among them, the expression of the attention map is:

[0086] ;

[0087] Where ⨁ represents the pixel-by-pixel summation and the channel weight W C Suppresses channel information with less information, spatial weight W S It enhances important and relevant spatial locations on the feature map while ignoring unimportant spatial information. P , which further highlights the feature information in the candidate frame of the target object.

[0088] It should be noted that in the transformer, the self-attention mechanism is calculated on different input channels of multiple instances. This method is called the multi-head attention mechanism. Multi-head attention improves the effectiveness of the attention mechanism by capturing different node relationships. Since the connections between nodes and external units are also diverse, this method is used in this embodiment:

[0089] ;

[0090] Where, h i Indicates the i A head, H is the total number of heads, W0 is the linear transformation matrix, P KN 、 P VN ∈ R s×d It is a shared storage unit of different heads, and the final output is a skip connection, forming an external attention network of the graph.

[0091] Afterwards, in this embodiment, the internal attention of the graph is combined with the external attention of the graph, and the expression is:

[0092] ;

[0093] Where, Represents the first l +1 layer of data points, Represents the first l +1 layer of edges, MPNN represents the message passing GNN network, For the l The data points of the layer, Indicates the l The edge of the layer, A is the connection matrix, Indicates the l The data points of the layer are processed by the attention mechanism to obtain the Transformer layer with self-attention. GEANet represents the external attention of the graph, Indicates the l +1 layer of edges;

[0094] The method further includes updating the node features and edge features, and the expression is:

[0095] ;

[0096] in, 、 、 、 all , l is the layer index, i 、 j is the node index, N ( i ) is the i The neighborhood of a node, f node and f edge They represent the parameters of the learnable function nodes and edge message passing respectively.

[0097] Finally, use the FFN module to aggregate node information and get the next layer of nodes ,at the same time, For the l The edge features of the +1 layer are as follows:

[0098] ;

[0099] Where, 、 all , respectively l The output of the MPNN layer and the transformer, 、 All are l The output of layer GENet.

[0100] Compared with the prior art, the remote sensing scene classification method combining human perception and Lie group features shown in this embodiment has the following beneficial effects:

[0101] This example transforms the human perception process into a feature learning method and proposes a superpixel segmentation algorithm based on Lie group inner mean features to segment samples, preserving features while reducing feature dimensionality. Secondly, the segmentation results are used to construct a graph, and a novel intra-graph and extra-graph attention mechanism is designed. This mechanism first captures the correlations between different graphs through an external attention module, then combines internal attention (Lie group space graph attention) to further aggregate and enhance features, ultimately achieving classification results. This significantly reduces model complexity while maintaining classification accuracy.

[0102] Example 2

[0103] See also Figure 2 A second embodiment of the present invention provides a remote sensing scene classification system that combines human perception and Lie group features, the system comprising:

[0104] The first processing module 10 is used to obtain a remote sensing image corresponding to the target area, and segment the remote sensing image using a superpixel segmentation method based on the inner mean of the Lie group to obtain feature maps corresponding to multiple target objects;

[0105] A second processing module 20 is configured to capture the correlation between the plurality of feature maps obtained by segmentation through an out-of-graph attention mechanism;

[0106] The third processing module 30 is used to aggregate and enhance the scene features of the feature map in combination with the intra-map attention mechanism to obtain a scene classification result of the remote sensing image.

[0107] Compared with the prior art, the remote sensing scene classification system combining human perception and Lie group features shown in this embodiment has the following beneficial effects:

[0108] This example transforms the human perception process into a feature learning method and proposes a superpixel segmentation algorithm based on Lie group inner mean features to segment samples, preserving features while reducing feature dimensionality. Secondly, the segmentation results are used to construct a graph, and a novel intra-graph and extra-graph attention mechanism is designed. This mechanism first captures the correlations between different graphs through an external attention module, then combines internal attention (Lie group space graph attention) to further aggregate and enhance features, ultimately achieving classification results. This significantly reduces model complexity while maintaining classification accuracy.

[0109] Example 3

[0110] A third embodiment of the present invention provides a readable storage medium having computer instructions stored thereon, which implement the steps of the method described in the above embodiment when executed by a processor.

[0111] Example 4

[0112] A fourth embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the above embodiment when executing the program.

[0113] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] Those skilled in the art will appreciate that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device), or for use in conjunction with such instruction execution system, apparatus, or device. For purposes of this specification, a "computer-readable storage medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system, apparatus, or device, or in conjunction with such instruction execution system, apparatus, or device.

[0115] More specific examples (a non-exhaustive list) of computer-readable storage media include the following: an electrical connection with one or more wires (electronic device), a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable storage medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0116] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the aforementioned embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or a combination of the following technologies known in the art may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0117] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0118] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A remote sensing scene classification method combining human perception and Lie group characteristics, characterized by: The method comprises: Acquire a remote sensing image corresponding to the target area, segment the remote sensing image using a superpixel segmentation method based on the inner mean of a Lie group, and obtain feature maps corresponding to multiple target objects; Capturing the correlation between the multiple feature maps obtained by segmentation through the out-of-graph attention mechanism; The scene features of the feature map are aggregated and enhanced by combining the in-graph attention mechanism to obtain the scene classification result of the remote sensing image; The step of capturing the correlation between the plurality of feature maps obtained by segmentation through an out-of-graph attention mechanism includes: Based on the out-of-graph attention mechanism, the node features of each feature graph and the edge features between any multiple feature graphs are calculated respectively; Calculating the correlation between nodes in different feature graphs based on the node features of each feature graph and the edge features between any plurality of feature graphs; The step of calculating the correlation between nodes in different feature graphs based on the node features of each feature graph and the edge features between any plurality of feature graphs is expressed as follows: ； Where, A E is the similarity between the input graph nodes and the external units, Norm represents the normalization operation, F is a data sample with a size of n × d ,in n is the number of nodes, d For the dimension, P To learn the parameters, T represents the transpose operation, R is a real number, s is the number of external nodes; The step also includes normalizing the similarities between the input graph nodes and the external units in terms of rows and columns, respectively, as follows: ； The steps further include: Two independent sets of external units are used for keys and values respectively, and additional external units are added for edge features. The relationship between edge features and node features is determined by shared units. The expression is ; Where, S EN Represents a shared unit for storing connections, with dimensions of d × d , P KN 、 P VN is the external key-value pair of the node, P KE 、 P VE is the external key-value pair of the edge, T Represents a transpose operation; The step of aggregating and enhancing the scene features of the feature graph in combination with the in-graph attention mechanism to obtain a scene classification result of the remote sensing image includes: Based on the out-of-graph attention mechanism, combined with the in-graph attention mechanism, global average pooling is performed on each of the feature maps along the X-axis and the Y-axis; Merge all feature maps after global average pooling to obtain a process map; The process graph is passed through a deep convolution layer to perform deep feature extraction, and is activated through an activation function.

2. The remote sensing scene classification method combining human perception and Lie group characteristics according to claim 1 is characterized in that: The process graph is subjected to deep feature extraction through a deep convolutional layer, and activation processing is performed through an activation function, including: Perform 1×1 depth convolution operations along the X-axis and Y-axis respectively, use SeLU activation function for activation processing, scale the activation value to the range of [0,1], and obtain the spatial weight W S , feature map P The size is h×w×c, where h represents height, w represents width, and c represents the number of channel layers. It is activated by the SeLU activation function and then processed by the Lie group Sigmoid kernel function to obtain the channel weight. W C , the channel weight W C With feature map P Multiply to get a new feature map P C , the spatial weight W S With feature map P Multiply to get a new feature map P S , the feature map P C With feature map P S Perform pixel-by-pixel summation to obtain the final attention map; Among them, the expression of the attention map is: ； Where, Represents pixel-by-pixel summation, channel weight W C Used to suppress some channel information, spatial weight W S Used to enhance the spatial information in some feature maps.

3. The remote sensing scene classification method combining human perception and Lie group characteristics according to claim 2 is characterized in that: The expression combining out-of-graph attention with in-graph attention is: ； Where, Represents the first l +1 layer of data points, Represents the first l +1 layer of edges, MPNN represents the message passing GNN network, For the l The data points of the layer, Indicates the l The edge of the layer, A is the connection matrix, Indicates the l The data points of the layer are processed by the attention mechanism to obtain the Transformer layer with self-attention. represents the external attention of the graph, Indicates the l +1 layer of edges; The method further includes updating the node features and edge features, and the expression is: ； in, 、 、 、 all , l is the layer index, i 、 j is the node index, N ( i ) is the i The neighborhood of a node, f node and f edge They represent the parameters of the learnable function nodes and edge message passing respectively.

4. A remote sensing scene classification system combining human perception and Lie group characteristics, characterized by: The method according to any one of claims 1 to 3, wherein the system comprises: The first processing module is used to obtain a remote sensing image corresponding to the target area, and segment the remote sensing image using a superpixel segmentation method based on the inner mean of the Lie group to obtain feature maps corresponding to multiple target objects; A second processing module is configured to capture the correlation between the plurality of feature maps obtained by segmentation through an out-of-graph attention mechanism; The third processing module is used to aggregate and enhance the scene features of the feature map in combination with the in-graph attention mechanism to obtain the scene classification result of the remote sensing image.

5. A readable storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 3 are implemented.

Images