Inplanatory attribution method, system and device based on star catalogue semantic segmentation model and storage medium

Abstract

The invention discloses an interpretable attribution method, system and device based on a star catalogue semantic segmentation model and a storage medium, and belongs to the technical field of deep space exploration image processing. Graying processing and gradient calculation are performed on a star catalogue image, and the image is divided into super-pixel regions based on a gradient minimum strategy; constructing a region-level topological structure diagram, and establishing a node link based on spatial adjacency and brightness gradient similarity; calculating the structural importance of each region, sampling a mask center point, and generating a directional elliptic Gaussian disturbance mask set according to a gradient principal direction angle; acting the disturbance mask on the original image, and inputting a semantic segmentation model to calculate a significance score; and spreading saliency information in the topological structure diagram based on a diagram attention mechanism, and constructing a structure saliency interpretation diagram. Through regional division and topology propagation of terrain perception, model interpretation of spatial continuity and terrain consistency is realized, and the technical problem that a traditional method neglects physical characteristics of a terrain structure and lacks propagation path description is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep space exploration image processing, and in particular to an explainable attribution method and system based on a star map semantic segmentation model, a device and a storage medium. BACKGROUND

[0002] Planetary surface semantic segmentation is one of the core technologies of deep space exploration autonomous decision-making systems. Based on remote sensing images collected by deep space exploration, pixel-level identification and classification of craters, ridges, gullies and other landforms are realized, which directly affects landing site selection, patrol path planning and geological evolution mechanism research. With the complication and normalization of deep space exploration missions, there is an urgent need for real-time and reliable semantic segmentation algorithms for tasks such as planetary surface terrain identification, obstacle avoidance and scientific target positioning.

[0003] Deep learning models such as DeepLabV3 have shown advantages in star map semantic segmentation by capturing multi-scale context information through dilated convolution and spatial pyramid pooling modules. However, star map images have high noise, large intra-class differences, and high inter-class similarity, and the black box nature of deep learning models hinders their credibility verification and scientific decision-making traceability, severely restricting their application in deep space exploration.

[0004] Traditional explanation methods such as Grad-CAM, RISE, etc. can be used for image-level explanation, but in complex landform backgrounds, weak texture features, and fuzzy semantic boundaries in star map environments, they show poor spatial consistency, untrustworthy attribution, etc. The existing methods are mostly based on pixel-level mask perturbation, lack of description of the physical properties of the terrain structure in the image, and use unstructured mask generation methods, which fail to effectively align the structure boundaries of the real landforms, resulting in fragmentation of the salient regions. DeepLabV3 is an advanced semantic segmentation deep learning model, Grad-CAM is Gradient-weighted-Class-Activation-Mapping, RISE is Randomized-Input-Sampling-for-Explanation. SUMMARY

[0005] In view of the above problems, the present application is proposed.

[0006] Therefore, the technical problem solved by the present application is: how to solve the technical problem that the existing star catalog semantic segmentation model explainability method ignores the physical characteristics of the terrain structure, lacks structure-aligned regional division, and the explanation result lacks the depiction of the propagation path and dependence strength. In view of the particularity of the star catalog image, an explainability attribution method combining terrain perception and topological propagation is provided to realize model explanation with spatial continuity and terrain consistency.

[0007] To solve the above technical problems, the present application provides the following technical solutions: an explainability attribution method based on a star catalog semantic segmentation model, which comprises: performing gray processing and gradient calculation on a star catalog image, adjusting a clustering center based on a gradient minimum strategy, and dividing the star catalog image into superpixel regions; mapping the superpixel regions into topological nodes, establishing node links based on spatial adjacency and brightness gradient similarity, and constructing a regional-level topological structure graph; calculating the structural importance of each region, sampling candidate regions according to a probability distribution, sampling a mask center point in the candidate region based on a gradient value, calculating the gradient main direction angle of the mask center point, constructing a rotation matrix and an elliptical covariance matrix, and generating a directional elliptical Gaussian perturbation mask set; applying the perturbation mask to the original image, inputting the semantic segmentation model to obtain a prediction difference, and calculating a saliency score in combination with a distance weighted decay factor; propagating saliency information in the topological structure graph based on a graph attention mechanism, and constructing a structural saliency explanation graph.

[0008] As a preferred scheme of the explainability attribution method based on a star catalog semantic segmentation model, the star catalog image is divided into superpixel regions by: searching for a pixel point with the minimum gradient value within a 3x3 range around the initial clustering center as a new center position; assigning the pixel point to the clustering center with the minimum distance based on a joint distance, and the joint distance comprehensively considers gradient difference and spatial position information.

[0009] As a preferred scheme of the explainability attribution method based on a star catalog semantic segmentation model, the construction of the regional-level topological structure graph comprises: designing a link weight function based on regional spatial distance and gradient difference; setting a link filtering condition, and establishing a link connection only when the spatial distance and the gradient difference both meet the threshold value.

[0010] As a preferred scheme of the explainability attribution method based on a star catalog semantic segmentation model, the sampling of a mask center point in the candidate region based on a gradient value comprises: The region structure importance is normalized to obtain a probability of each region being selected as a candidate region; A mask center point is sampled according to a probability distribution of pixel point gradient values in the candidate region.

[0011] As a preferred scheme of the attribution method for explainability based on the semantic segmentation model of the star atlas, the generation of the directional elliptical Gaussian perturbation mask set comprises: A two-dimensional rotation matrix is constructed according to the main direction angle of the mask center point; The perturbation range in the main and auxiliary directions is set, and an elliptical covariance matrix after rotation is constructed; It is determined whether a pixel point falls into a mask range based on a Gaussian distribution decision function.

[0012] As a preferred scheme of the attribution method for explainability based on the semantic segmentation model of the star atlas, the calculation of the saliency score in combination with the distance weighting decay factor comprises: The absolute difference value of the predicted response before and after perturbation is calculated as an initial saliency score; An exponential decay factor based on Euclidean distance is introduced for distance weighting correction.

[0013] As a preferred scheme of the attribution method for explainability based on the semantic segmentation model of the star atlas, the construction of the structural saliency explanation graph comprises: The region saliency feature is initialized, and the saliency scores of the pixel points in the region and the region attributes are summarized; The structural saliency feature is propagated by using a structure-enhanced graph attention network, and the attention weight is adjusted by using a structure-space joint guide term.

[0014] The present application provides an attribution system for explainability based on a semantic segmentation model of a star atlas.

[0015] To solve the above technical problems, the present application provides the following technical scheme: an attribution system for explainability based on a semantic segmentation model of a star atlas, comprising: a region division module, which performs gray scale processing and gradient calculation on a star atlas image, adjusts a clustering center based on a gradient minimum strategy, and divides the star atlas image into superpixel regions; A topology construction module maps the superpixel regions into topological nodes, establishes node links based on spatial adjacency and brightness gradient similarity, and constructs a region-level topological structure graph; A mask generation module calculates the structural importance of each region, samples a candidate region according to a probability distribution, samples a mask center point in the candidate region based on a gradient value, calculates the gradient main direction angle of the mask center point, constructs a rotation matrix and an elliptical covariance matrix, and generates a directional elliptical Gaussian perturbation mask set; The saliency calculation module applies the disturbance mask to the original image, inputs a semantic segmentation model to obtain a prediction difference, and calculates a saliency score in combination with a distance weighted attenuation factor; The explanation graph construction module propagates saliency information in the topological structure graph based on a graph attention mechanism to construct a structural saliency explanation graph.

[0016] The application provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the attribution method when executing the computer program.

[0017] The application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to implement the steps of the attribution method.

[0018] The application has the advantages that: the adaptive superpixel region division based on the gradient minimum strategy is performed on the star image, the problem that the traditional pixel-level disturbance method ignores the physical characteristics of the terrain structure is effectively avoided, the divided region can be aligned with the real terrain boundary, a stable basic unit is provided for subsequent explanation, a region-level topological structure graph is constructed, a link weight function based on spatial adjacency and structural similarity is designed, a high-level information propagation channel is established, the defects of the existing methods in lack of structural expression ability are solved, a structure importance driven sampling strategy is used to generate a mask center point, an ellipse disturbance mask that is adaptive to the terrain direction is constructed, the direction disturbance is realized according to the characteristics of the zonal or crack structure of the star terrain, the coverage ability of the mask to the edge structure and the accuracy of the explanation are improved, the saliency information is propagated in the topological structure based on the graph attention mechanism, the attention weight is adjusted by using a structure-space joint guide term, the problems of the saliency heat map fragmentation and structure fragmentation generated by the traditional method are effectively avoided, and overall, the attribution method realizes the star semantic segmentation model explainability attribution with spatial continuity and terrain consistency, and provides reliable technical support for model decision tracing and scientific analysis in deep space exploration. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Fig. 1 A general flowchart of the attribution method based on the star semantic segmentation model is provided for an embodiment of the application.

[0021] Fig. 2 A computer device diagram of an explainability attribution method based on a star map semantic segmentation model is provided for an embodiment of the present application.

[0022] Fig. 3 A process diagram of an embodiment of a star map semantic segmentation model-based explainability attribution system is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0024] Embodiment 1, reference Figs. 1-3 For an embodiment of the present application, the embodiment provides an explainability attribution method based on a star map semantic segmentation model, comprising: S100: performing gray scale processing and gradient calculation on the star map image, adjusting the clustering center based on the gradient minimum strategy, and dividing the star map image into superpixel regions; S200: mapping the superpixel regions to topological nodes, establishing node links based on spatial adjacency and brightness gradient similarity, and constructing a regional-level topological structure graph; S300: calculating the structural importance of each region, sampling candidate regions according to the probability distribution, sampling the mask center point in the candidate region based on the gradient value, calculating the gradient principal direction angle of the mask center point, constructing the rotation matrix and the elliptical covariance matrix, and generating a directional elliptical Gaussian perturbation mask set; S400: applying the perturbation mask to the original image, inputting the semantic segmentation model to obtain the prediction difference, and calculating the saliency score combined with the distance weighted decay factor; S500: propagating saliency information in the topological structure graph based on a graph attention mechanism, and constructing a structural saliency explanation graph.

[0025] It should be noted that the core challenge faced by star map semantic segmentation lies in the complexity of the terrain structure and the opacity of the model decision. The meteorite craters, ridges, gullies and other landforms in the star map image have obvious geometric shapes and directionality. The traditional pixel-level perturbation method cannot effectively align the structural boundaries of the real landforms, resulting in fragmentation or boundary misplacement problems in the saliency region. At the same time, the high noise and weak texture features in the star map environment make the existing explanation methods exhibit poor spatial consistency, untrustworthy attribution, and other problems, which cannot provide reliable model decision traceability for deep space exploration missions.

[0026] Therefore, in order to solve the above-mentioned problems of model explainability and terrain perception, a complete set of star atlas semantic segmentation model explainability attribution scheme is constructed through the steps of S100-S500. S100 realizes accurate alignment with the terrain structure through superpixel division by gradient minimization strategy, providing stable structure units for subsequent analysis; S200 establishes the spatial adjacency relationship between regions by constructing a topological structure graph, laying the foundation for saliency propagation; S300 ensures the representativeness and terrain fitting of perturbation test by designing structure importance driven sampling and directional elliptical mask; S400 enhances the spatial consistency of the score by distance weighted saliency calculation; S500 establishes the causal dependence relationship between regions by the graph attention propagation mechanism. The overall scheme realizes explainability attribution with spatial continuity and terrain consistency, solving the key technical problems that traditional methods cannot depict the physical characteristics of terrain structure and lack of propagation path.

[0027] Embodiment 2, refer to Figs. 1-3 For an embodiment of the present application, based on the above embodiment, a star atlas semantic segmentation model based explainability attribution method is provided, comprising: step S100 divides the star atlas image into superpixel regions, step S100 includes steps A1-A2: A1: search for the pixel point with the minimum gradient value within the 3x3 range around the initial cluster center as the new center position; The star atlas image is initialized, and the image gray processing and gradient value calculation are performed in turn. The gradient value calculation adopts differential approximation calculation.

[0028] Star atlas image preprocessing: uniformly adjust the size of the original star atlas image to At this time , .

[0029] ①The original RGB color star atlas image is grayed to extract the brightness edge features of the star atlas image. The gray processing adopts the weighted average formula as follows: , Among them, , , are the red, green and blue channel values of the pixel position in the original image; is the corresponding gray value; , respectively represent the height and width of the image, is a real matrix space, a real matrix with dimension.

[0030] ②According to the gray value , the luminance gradient of each pixel of the image is calculated using a differential approximation to extract edge information and structural strength in the image, and the specific differential calculation formula is as follows: , wherein, is the luminance gradient of each pixel in the image; and correspond to the gray value at pixel position and the gray value at pixel position respectively; , correspond to the gray value at pixel position and the gray value at pixel position respectively.

[0031] Star image region division: according to the image processed by A1, a clustering operation is performed to divide the star image into super-pixel regions.

[0032] A2: Assigning pixel points to the cluster center with the smallest joint distance, which integrates gradient difference and spatial position information.

[0033] ① Setting initial cluster centers: setting the initial cluster distance , dividing the image into initial regions, and setting the initial cluster center as the center pixel point of each region.

[0034] Specifically, setting the initial cluster distance , preliminarily dividing the image into cluster regions, and setting the initial cluster center as the center pixel point of each region.

[0035] ② Replacing the initial cluster center: considering that the topographic edge in the star image often corresponds to a higher luminance gradient value, a local minimum gradient strategy is introduced to enhance the structural stability of region division and avoid the cluster center being located at the edge or noise region. Specifically, within a range of around each initial cluster center, the pixel point with the minimum gradient value is searched as the new center position to enhance the structural stability of region division and avoid the cluster center being located at the edge or noise region. The cluster center replacement formula is as follows: , wherein, is the coordinate of the initial cluster center corresponding to the th cluster region after adjustment in the 0th iteration, ; indicates the center of the initial cluster center ​ a set of pixels in a region; for calculating the position of the pixel corresponding to the minimum gradient value.

[0036] ③ Realize pixel clustering: find each pixel point all cluster centers in the neighborhood range, calculate the joint distance between the pixel point and these cluster centers , and assign the pixel point to the cluster center with the minimum joint distance, thus completing one region division. The specific joint distance The calculation formula is: , , Wherein, is the square of the joint distance; is the difference between the coordinates of the pixel point and the cluster center of the first cluster region; is the difference between the coordinates of the pixel point and the cluster center of the first cluster region; is the initial cluster spacing; is the average brightness gradient of the cluster center of the first cluster region; is the balance factor between gradient and spatial information, ; is the cluster center coordinate of the first cluster region; represents the set of pixel coordinates in the cluster center of the first cluster region; is used to calculate the number of elements in the set. This distance design can effectively enhance the resolution of structural boundaries and keep the region space compact. ④ After each region division, the new position of each cluster center and the new average gradient value of each region are recalculated, and the calculation formula is as follows: , ,

[0037] Wherein, formula (6) represents the calculation of the centroid of the cluster region after round clustering , which is used as the initial cluster center coordinate of round clustering; formula (7) is used to calculate the region brightness gradient of round; represents the set of pixel coordinates in the cluster center of the first cluster region; represents the set of pixel coordinates in the cluster center of the first cluster region; represents the set of pixel coordinates in the cluster center of the first cluster region; represents the set of pixel coordinates in the cluster center of the first cluster region; represents the set of pixel coordinates in the cluster center of the first cluster region; represents the set of pixel coordinates in the cluster center of the first cluster region; The cluster region in the th... The set of pixel coordinates contained in the wheel; Indicates the first The cluster region in the th... The number of pixels contained in the wheel.

[0038] ⑤ Repeat steps ②-④ until the cluster centers converge, and finally obtain the results. One superpixel region.

[0039] And by setting the maximum number of iterations to 10, we finally obtained... The star catalog image region. Among them, the first Each star catalog image region is used The set of regions is represented as , In this embodiment, the numerical region set is represented as follows: , , No. The coordinates of the cluster centers corresponding to each cluster region are: ;No. The average brightness gradient of each cluster region is .

[0040] This clustering process enables each region to be more consistent in terms of physical geomorphological structure, thereby providing a more stable structural unit for subsequent interpretation.

[0041] In this embodiment, the construction of the regional topology diagram in step S200 includes the following steps B1-B2: B1: Design link weight function based on regional spatial distance and gradient difference; To enhance the structural expressive power of saliency interpretation, this invention further constructs a region-level topological structure map based on image processing and region segmentation in S100, establishing a high-level information channel for subsequent saliency propagation. The specific steps are as follows: Topology node settings: Configure each region obtained in step S100. Mapped to a node Node information includes: the coordinates of the cluster center for each region. and the average brightness gradient of each region All of these are obtained from step A2. The node set is denoted as . .

[0042] in, A set of nodes; For each node; For the first Superpixel region The corresponding topology node; Topology link design: In order to quantify the connection strength between image regions, the link weight function is designed to measure the structural similarity and spatial proximity between regions. The link weight between any two regions 、 is defined as follows: , , wherein, is the link weight between two nodes; is the exponential function operation; represents the structure-space joint guide item of any pair of image regions 、 , represents the cluster center coordinates of any pair of image regions 、 , is the total number of regions; is the average brightness gradient of any pair of image regions ; is the average brightness gradient of any pair of image regions , which is solved by formula (7); represents the calculation of the spatial distance between coordinates; represents the calculation of the absolute value; is the spatial distance weight adjustment constant, which is used to control the influence degree of regional spatial proximity in the link weight;

[0043] B2: Set link filtering conditions, only when the spatial distance and gradient difference meet the threshold value, establish link connection.

[0044] Topology link optimization: In order to prevent the graph structure from being too complex or introducing meaningless connections, the invention sets link filtering conditions, and the specific condition formula is as follows: , Formula (10) represents any two image regions 、 , if the spatial distance is less than or equal to the maximum distance , and the average brightness gradient difference is less than or equal to the structural threshold , then the two image regions ,​​​ There is a link between the two regions, otherwise there is no link; 0, that is, there is no link between the two regions when the above conditions are not met, wherein is the maximum allowed spatial distance threshold; is the adjacent range coefficient; is the initial clustering distance, which is set in the superpixel division.

[0045] wherein is used to determine whether there is a link between any two regions , ; is the adjacent range coefficient; is the initial clustering distance, which is set in step A2.

[0046] Establish a star atlas image topology: according to the node set set in step S100 , steps B1 and B2 define the optimized topology link, and the link set is denoted as . Finally, the topology graph is formed.

[0047] In this embodiment, the sampling of the mask center point in the candidate region based on the gradient value in step S300 includes the following steps C1-C5: C1: Normalize the region structure importance to obtain the probability of each region being selected as a candidate region; To improve the structure alignment capability of the saliency analysis, the present application designs a star atlas image mask method combining the star atlas terrain structure importance and gradient information according to the topology graph obtained in step S200 . And use this method to generate a star atlas image mask set, so as to disturb the model in the key region with physical meaning, and ensure that the saliency interpretation result has representativeness in space. The specific steps are as follows: 1. Mask center point sampling: ①Sampling of the candidate region of the mask center point, according to the region division result in S300, sampling the candidate region of the mask center point, and sampling a mask center point in each candidate region.

[0048] A) According to the region division result in step S300, set the structure importance of the star atlas image region , and select a representative structure region as the candidate region of the mask center according to the structure importance . The calculation formula of the structure importance is as follows: , wherein is the structure importance; representative region corresponding node adjacent node set of representative a certain adjacent node of is the link weight between two nodes, which is solved by formula (8).

[0049] B) structural importance of each region is normalized to obtain the probability of each region being selected as a candidate region for the mask center point: , wherein, is the probability of region being selected as a candidate region; is the total number of regions; is the structural importance of region , which is solved by formula (11).

[0050] C) according to the probability distribution , non-replacement random sampling is performed from the region set to select regions as mask center point candidate regions, and the final candidate region set is : , wherein, is the final candidate region set; represents that regions are sampled without replacement according to the probability distribution . is the region set; is the candidate region set, and the number of elements is .

[0051] C2: sampling the mask center point according to the probability distribution of the pixel point gradient value in the candidate region.

[0052] ② sampling of pixel points in the candidate region: in the candidate region set obtained in step C1, the mask center is sampled for each pixel point in the region , and the number of elements is .

[0053] A) in order to improve the covering ability of the mask on the edge structure, the present application uses the brightness gradient value in the region as the sampling basis, and defines the probability of the pixel point in the region being sampled as the mask center as follows: , wherein, is the defined region is the pixel position is sampled as the probability of the mask center; is the pixel point corresponding luminance gradient value; is the pixel position corresponding luminance gradient value, solved by formula (2).

[0054] B) According to the probability distribution , a pixel point is randomly sampled from each candidate region as the mask center: , wherein, is the probability distribution; is the region; represents the candidate region from which the mask center point is sampled, is the coordinate corresponding to the mask center point; is the set of all candidate regions, solved by step ①, n is the number, and the parameters related to n below represent the number.

[0055] C) Finally, the set of all mask center point coordinates is obtained: , wherein, is the set of all mask center point coordinates; is each mask center point; represents the mask center point coordinate sampled from the candidate region.

[0056] Each point in the set will serve as the center of the perturbed mask, driving the subsequent directional perturbed mask construction, thereby improving the perceptual coverage ability of the explanation saliency map on the edge region.

[0057] Generating a set of directional elliptical Gaussian perturbed masks includes: Since the star table topography mainly presents a strip or crack structure, the mask shape is designed as an ellipse in the present application. However, due to the differences in the arrangement direction of the natural terrain structure of the star table, the natural terrain structure includes cracks and strips, and the elliptical mask needs to be directionally set. Therefore, the present application constructs an elliptical mask with direction to meet the above needs, and the specific steps are as follows: C3: Constructing a two-dimensional rotation matrix according to the mask center point main direction angle; ① Constructing a mask rotation matrix: For each mask center point obtained in step C1 ①, calculate its gradient principal direction angle in radians. The specific calculation formula is as follows: , , in, Indicates the first The center point of the mask The principal direction angle; and These are the center points of the mask ( exist gradient value in direction, mask center point exist Gradient value in the direction; Indicates the first The center point of the mask The coordinates; , Used to solve pixel points Gradient value in the direction; The corresponding gray value is represented by formula (1).

[0058] Based on the main direction angle of the mask center point Rotate the mask to conform to the terrain orientation, constructing the first The two-dimensional rotation matrix corresponding to the center point of the mask : , in, It is a two-dimensional rotation matrix; and These are the principal direction angles. The negative and positive sine values, principal direction angles The sine value; Main direction angle The cosine value.

[0059] C4: Define the perturbation range in the main and secondary directions, and construct the rotated elliptic covariance matrix; ② Construct the covariance matrix of the rotated ellipse: First, define the scale difference in the principal and secondary directions, and set the perturbation range of the mask in the principal direction as follows: The disturbance range in the secondary direction is Construct the unrotated elliptic covariance matrix: , in, The covariance matrix of the ellipse is unrotated; The range of disturbance in the main direction; This represents the disturbance range in the secondary direction.

[0060] Based on this, we obtain the result in the first... The covariance matrix at the center of each mask, rotated along the gradient direction: , Formula (21) is used to represent the first Each mask center corresponds to the mask information under the local gradient direction, including the shape, orientation and spatial coverage of the perturbation mask, to ensure that the subsequently generated mask can better fit the actual terrain structure; The covariance matrix is ​​rotated along the gradient direction; T is the matrix transpose symbol. No. The two-dimensional rotation matrix corresponding to each mask center point is solved by formula (19).

[0061] C5: Determines whether a pixel falls within the masking range based on the Gaussian distribution determination function.

[0062] ③ Define a pixel offset vector: to represent the pixel points in the star catalog image Compared to the first Set the pixel offset vector at the center position of the mask. for: , in, It is a pixel offset vector; and Each pixel Coordinates and the individual mask center The difference in coordinates, pixels Coordinates and the individual mask center The difference in coordinates; Indicates the first The center point of the mask The coordinates.

[0063] ④ Construct a mask set: Based on the mask information in step C4 ② and the pixel offset vector in step C5 ③ At the center point of each mask At this point, a Gaussian distribution determination function is set to represent the pixel position. Whether it falls into the first The specific formula for determining the masking range at the center of the mask is as follows: , Formula (23) defines an elliptical mask, representing the first... The principal axis direction of the mask and its corresponding mask center point The gradient principal directions are consistent, and the range is from control; The Gaussian response threshold is used to define the mask boundaries; For the first A mask at pixel position The value at that location is a binary mask of 0 or 1; This indicates that when the Gaussian response value is greater than or equal to the threshold... The value is 1, where "1" represents a pixel. Falling within the range of the nth elliptical mask; "0" represents a pixel. Not within the range of the nth elliptical mask; "" indicates all other cases except those that meet the Gaussian response threshold condition.

[0064] Based on the above results, define a set of perturbation masks. for: , Each mask A separate structure-sensitive perturbation region is defined on the star catalog image, which will be used as a perturbation input to evaluate the model's significance response.

[0065] In this embodiment, step S400, which combines the distance-weighted attenuation factor to calculate the significance score, includes the following steps D1-D2: To measure the impact of local perturbations on model predictions, this invention uses the perturbation mask set obtained in step C5. By combining a distance-weighted mechanism, the predictive impact of each perturbation mask on the semantic segmentation model is evaluated.

[0066] D1: Calculate the absolute difference between the predicted response before and after the disturbance as the initial significance score; Generate a perturbation image and obtain the perturbation output.

[0067] Based on the perturbation mask set generated in step S300, the original star catalog image is... Each pixel in the image is related to the first pixel. A mask The elements in the array are multiplied one by one to generate the product of the first element. Image after masking perturbation : , in, For the first The image after masking perturbation; This represents the Hadamard product, which is the element-wise multiplication of matrices of the same shape to produce a third matrix of the same dimensions.

[0068] Will pass through the first Image after masking perturbation , inputting the semantic segmentation model , obtaining a perturbation prediction response of the target class: , wherein, represents inputting the first mask perturbation image, , obtaining a perturbation prediction response output corresponding to the pixel point .

[0069] ①Calculate the initial saliency score: input the original star atlas image into , obtain the non-perturbation prediction response of the pixel point . According to the perturbation prediction response and the non-perturbation prediction response , calculate the response change caused by the first mask to each pixel point: , wherein, is the response change caused by the first mask to each pixel point; is the non-perturbation prediction response; is the perturbation prediction response; represents calculating the absolute value.

[0070] Map the response change to the corresponding mask region to obtain the initial saliency score of the pixel point: , wherein, represents the initial saliency score of the pixel point under the perturbation of the first mask; is the response change caused by the first mask to each pixel point; is the value of the first mask at the pixel position , a binary mask of 0 or 1, solved by formula (23).

[0071] D2: Introduce an exponential decay factor based on Euclidean distance for distance weighted correction.

[0072] ②Introduce distance weighted decay: since the mask center is the perturbation focus, the pixels farther from the center are less credible, and the present application introduces a distance weighted decay factor to enhance the spatial consistency of the explanation score to the perturbation region. The distance weighted decay factor calculation formula is: , , wherein, is an adjustment parameter, which can be set between 0.01-0.05, α=0.02, which can avoid the rapid decay of the edge region weight while ensuring the center region weight, and control the decay speed of the distance near and far on the saliency score; is a distance weighted decay factor; represents the Euclidean distance from the pixel point to the mask center ; represents the coordinates of the mask center point .

[0073] ③Calculate the final saliency score: introduce the initial saliency score into the distance weighted decay factor , to get the distance weighted saliency score under the mask perturbation: , wherein, is the distance weighted saliency score under the mask perturbation; is the initial saliency score; is the distance weighted decay factor.

[0074] Fuse the saliency scores under the mask perturbation, and perform weighted average processing to get the final saliency score: , wherein, is the final saliency score; defined in step S300; is the distance weighted saliency score under the mask perturbation, solved by formula (31); represents the mask, solved by formula (23); is the distance weighted decay factor, solved by formula (29).

[0075] In this embodiment, the structural saliency interpretation graph is constructed in step S500, including the following steps E1-E2: On the basis of the final saliency score obtained in step S400, according to the topological structure graph in step S200, the directed propagation and spatial completion of saliency information are realized by using the graph attention mechanism, so as to improve the structural continuity and full graph interpretation ability of the saliency score.

[0076] E1: Initialize region saliency feature, aggregate the saliency score of all pixels in the region and the region attribute; Initialize region saliency feature: aggregate the saliency score of all pixels in the region, then the initial saliency feature of the region is: is the initial saliency feature; represents the pixel set contained in the cluster center of the i-th cluster region; represents the saliency score of the pixel; represents the average value of the brightness gradient of the cluster center of the i-th cluster region , which is solved by formula (5); represents the coordinate of the cluster center point of the i-th cluster region; is a 4-dimensional real number vector space. The initial saliency features of the five regions are:

[0077] is the saliency feature vector of the first region at the 0th layer (initialization); is the saliency feature vector of the second region at the 0th layer (initialization); is the saliency feature vector of the third region at the 0th layer (initialization); is the saliency feature vector of the fourth region at the 0th layer (initialization); is the saliency feature vector of the fifth region at the 0th layer (initialization).

[0078] E2: Propagate saliency features using structure-enhanced graph attention network, and adjust attention weights using structure-spatial joint guide term.

[0079] Saliency information graph propagation mechanism: use structure-enhanced graph attention network to propagate region saliency features on the topological structure graph The formula for updating the features of the nodes at the i-th layer is: ​​​​​​​​​​​​​​​ , , where, is the feature of the layer node ; denotes the feature of the layer node ; is the shared feature transformation matrix; is the feature transformation matrix of the layer node ; is the matrix product of the feature vector denotes the set of adjacent nodes of node ; denotes the attention weight of node to node , which determines the direction and strength of the propagation path; denotes the learnable attention weight vector; denotes the nonlinear activation function with a slope of 0.2; denotes the vector concatenation operation; denotes the structure-spatial joint guidance term for a pair of adjacent image regions, which is used to adjust the attention preference and is calculated by equation (9).

[0080] Output the structural saliency propagation result: After the propagation layer, the final saliency feature set of all regions, the information propagation path and dependency strength set between nodes are obtained, and the specific formula is as follows: , , where, is the final saliency feature set; is the feature of the layer node ; is the total number of regions, and the total number of nodes in the present application; denotes the existence of an information propagation path from node to node ; is the information propagation path and dependency strength set between nodes; denotes the attention weight on the path, which characterizes the dependency strength; denotes the link set, which is calculated by step E2.

[0081] According to all the above information, a complete structural saliency interpretation graph is constructed: , wherein, is a structural saliency map; represents a node set defined by step S200; represents a path of information propagation between nodes and its dependency strength, solved by formula (37); represents the final saliency feature of all regions, solved by formula (36).

[0082] Compared with other inventions for explainable analysis of semantic segmentation, the present invention conforms to the continuous distribution characteristics of extraterrestrial topography, enhances the spatial continuity of the heat map by identifying the terrain structure, and establishes a structural explanation path.

[0083] Embodiment 3, refer to Fig. 1 and Fig. 3 , an embodiment of the present invention provides an explainable attribution method based on a star table semantic segmentation model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0084] Table 1 is based on the results generated in steps A1-A2 of embodiment 2, specifically from the cluster center replacement of formula (3) and the average brightness gradient calculation of formula (5). The cluster center is adjusted by searching for the minimum value of the gradient within a 3x3 range, ensuring that the center point is located in a stable terrain area rather than at the edge of noise. Due to the large amount of data, the first five regions are taken as examples to calculate the cluster center coordinates and average brightness gradient values, as shown in Table 1.

[0085] Table 1 Cluster center coordinates and average brightness gradient values of the first five regions

[0086] As can be seen from Table 1, the cluster center coordinates of the five regions show a regular distribution, with y coordinates all being 16, indicating that these regions are located on the same horizontal band; the average brightness gradient values vary between 0.22 and 0.37, among which the gradient value of region 1 is the smallest at 0.22, the gradient value of region 5 is the largest at 0.37, reflecting the differences in terrain complexity of different regions. Compared with traditional fixed grid division, the adaptive division of the present invention can dynamically adjust the region boundary according to the terrain characteristics, making each region have better structural consistency.

[0087] Table 2 shows the link weights of the five links and the starting regions. The link weights are calculated by formula (8) - formula (10) based on the topology construction process of steps B1-B2 in example 2, and the filtering conditions are set. The link weights take into account both spatial distance and structural similarity. Only the region pairs that meet the distance threshold ≤ 48 and the gradient difference threshold ≤ 0.2 at the same time are connected. Similarly, taking the first five links as an example, the starting region and the link weight are shown in table 2.

[0088] Table 2 shows the link weights of the five links and the starting regions. The link weights are calculated by formula (8) - formula (10) based on the topology construction process of steps B1-B2 in example 2, and the filtering conditions are set. The link weights take into account both spatial distance and structural similarity. Only the region pairs that meet the distance threshold ≤ 48 and the gradient difference threshold ≤ 0.2 at the same time are connected. Similarly, taking the first five links as an example, the starting region and the link weight are shown in table 2.

[0089] Among the five links shown in table 2, the spatial distance is between 28.7 and 46.1, the structural difference is between 0.09 and 0.18, and the corresponding link weight is between 0.611 and 0.713. It can be observed that The link with the highest weight 0.713 corresponds to the smallest structural difference 0.09, verifying that the weight function designed by the application can accurately quantify the similarity between regions. Compared with simple Euclidean distance connection, the double constraint mechanism of the application ensures spatial adjacency and structural consistency, providing a more reliable topological basis for subsequent saliency propagation.

[0090] Table 3 is derived from the mask generation process of steps C1-C5 in example 2. The mask center is determined by the structural importance driven sampling formula (11) - formula (16), and the main direction angle is obtained by calculating the gradient direction formula (17) - formula (18). This design ensures that the mask is located in the key structural region and can conform to the actual trend of the terrain. Similarly, taking the first five masks as an example, the calculated mask center coordinates and main direction angles are shown in table 3.

[0091] Table 3 shows the center coordinates and main direction angles of the first five masks.

[0092] The main direction angles of the five masks in table 3 are concentrated in the range of 1.47 to 1.62 radians, and the relatively concentrated angles indicate that the terrain in this region has similar dominant direction characteristics. The mask center coordinates are distributed relatively dispersedly from to

[0093] ​​Table 4 shows the complete calculation chain from perturbation input to final score, which is completed by formula (25) - formula (32) based on the significance calculation process of step D1-D2 in example 2. The difference between the non-perturbation output and the perturbation output reflects the degree of influence of the mask area on the model prediction, and the distance weighted correction enhances the spatial consistency of the score. Taking the first five masks as an example, the significance score is shown in table 4.

[0094] Table 4 shows the starting area and link weight of the five links

[0095] Table 4 shows that the non-perturbation output is between 0.88 and 0.93, and the output after perturbation decreases to 0.76-0.83, the prediction difference is converted to the initial significance score of 0.065-0.086, and the final score is 0.079-0.112 after distance weighted correction. The mask has the highest final significance score of 0.112, indicating that this area has the greatest influence on the model decision. Compared with the simple difference calculation of the traditional method, the distance weighted mechanism introduced in the present application effectively enhances the spatial continuity of the significance score and avoids the interference of edge effect on the result.

[0096] Table 5 shows the attention propagation mechanism of graph derived from step E1-E2 in example 2, which calculates the attention weight between nodes by formula (34) - formula (35). Joint guide item Combining spatial distance and structural similarity, the attention weight determines the propagation strength and direction of significant information in the topological network. Taking the area as an example, its adjacent areas , , , the attention weight is shown in table 5.

[0097] Table 5 shows the starting area and link weight of the five links

[0098] Assuming that is the unit matrix in formula (34) - formula (35), the updated features are: , wherein, is the significance feature vector of the first area after the first layer propagation; is the initial significance feature vector of the second area; is the initial significance feature vector of the 17th area.

[0099] Table 5 takes the area as an example, its adjacent areas , , The joint guide items are 0.78, 0.65, and 0.60, respectively, and the corresponding attention weights are 0.412, 0.336, and 0.252, and the weight distribution is positively correlated with the guide item size. - The two regions have the highest attention weight of 0.412, indicating that the significance propagation between the two regions is the most active. Compared with the traditional method of only outputting a significance heat map, the graph attention mechanism of the present application can clearly depict the dependence relationship and propagation path between regions, provide a quantitative basis for causal analysis of model decision-making, and significantly improve the scientificity and credibility of explainable analysis.

[0100] Embodiment 4, refer to Fig. 1 and Fig. 3 As an embodiment of the present application, the embodiment provides an explainable attribution system based on a star atlas semantic segmentation model, comprising: a region division module, which performs gray scale processing and gradient calculation on a star atlas image, adjusts a clustering center based on a gradient minimum strategy, and divides the star atlas image into superpixel regions; a topology construction module, which maps the superpixel regions into topology nodes, establishes node links based on spatial adjacency and brightness gradient similarity, and constructs a region-level topology structure graph; a mask generation module, which calculates the structural importance of each region, samples candidate regions according to a probability distribution, samples a mask center point in the candidate region based on a gradient value, calculates a gradient main direction angle of the mask center point, constructs a rotation matrix and an elliptical covariance matrix, and generates a set of directional elliptical Gaussian perturbation masks; a saliency calculation module, which applies the perturbation masks to the original image, inputs a semantic segmentation model to obtain a prediction difference, and calculates a saliency score in combination with a distance weighted attenuation factor; an explanation graph construction module, which propagates saliency information in the topology structure graph based on a graph attention mechanism, and constructs a structural saliency explanation graph.

[0101] The embodiment also provides an electronic device suitable for a case of an explainable attribution method based on a star atlas semantic segmentation model, comprising: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize an explainable attribution method based on a star atlas semantic segmentation model as proposed in the above embodiment.

[0102] The embodiment also provides a storage medium having a computer program stored thereon, which is executed by a processor to realize an explainable attribution method based on a star atlas semantic segmentation model as proposed in the above embodiment.

[0103] The storage medium proposed in the embodiment belongs to the same inventive concept as the implementable explainability attribution method based on the star table semantic segmentation model proposed in the above embodiment. Technical details not described in the embodiment can be seen from the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software and necessary general hardware, and of course can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a floppy disk, a read-only memory (ROM), a random access memory (RAM), a FLASH, a hard disk or an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the method of each embodiment of the present application.

[0105] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. An interpretable attribution method based on a star catalog semantic segmentation model, characterized in that: include, The star catalog image is grayscaled and gradient is calculated. The cluster centers are adjusted based on the gradient minimization strategy to divide the star catalog image into superpixel regions. Superpixel regions are mapped to topological nodes, and node links are established based on spatial adjacency and brightness gradient similarity to construct a regional topological structure graph. Calculate the structural importance of each region, sample candidate regions according to probability distribution, sample the mask center point based on gradient value within the candidate region, calculate the gradient principal direction angle of the mask center point, construct the rotation matrix and elliptic covariance matrix, and generate a set of directional elliptic Gaussian perturbation masks. The perturbation mask is applied to the original image, input into the semantic segmentation model to obtain the predicted difference, and the significance score is calculated by combining the distance-weighted decay factor. Based on the graph attention mechanism, saliency information is propagated in the topological graph, and a structural saliency interpretation graph is constructed.

2. The interpretable attribution method based on a star catalog semantic segmentation model as described in claim 1, characterized in that: The process of dividing the star catalog image into superpixel regions includes: Search for the pixel with the smallest gradient value within a 3×3 area around the initial cluster center and use it as the new center position; Pixels are assigned to the cluster center with the smallest distance based on the joint distance, which combines gradient differences and spatial location information.

3. The interpretable attribution method based on a star catalog semantic segmentation model as described in claim 2, characterized in that: The construction of the regional-level topology diagram includes: Design link weight function based on regional spatial distance and gradient differences; Set link filtering conditions to establish link connections only when both spatial distance and gradient difference meet the thresholds.

4. The interpretable attribution method based on a star catalog semantic segmentation model as described in claim 3, characterized in that: The step of sampling the mask center point based on gradient values ​​within the candidate region includes: The regional structural importance is normalized to obtain the probability of each region being selected as a candidate region. Sample the mask center point within the candidate region according to the probability distribution of pixel gradient values.

5. The interpretable attribution method based on a star catalog semantic segmentation model as described in claim 4, characterized in that: The set of generated directional elliptical Gaussian perturbation masks includes: Construct a two-dimensional rotation matrix based on the principal direction angle of the mask center point; Define the perturbation range in the primary and secondary directions, and construct the rotated elliptic covariance matrix; The Gaussian distribution determination function is used to determine whether a pixel falls within the masking range.

6. The interpretable attribution method based on a star catalog semantic segmentation model as described in claim 4, characterized in that: The significance score calculated using the combined distance-weighted attenuation factor includes: The absolute difference between the predicted response before and after the perturbation is calculated as the initial significance score; An exponential decay factor based on Euclidean distance is introduced for distance-weighted correction.

7. The interpretable attribution method based on a star catalog semantic segmentation model as described in claim 4, characterized in that: The constructed saliency interpretation diagram includes: Initialize the region saliency features and summarize the saliency scores and region attributes of the pixels within the region; We employ the saliency features of attention network propagation using structure-enhanced graphs and utilize the structure-space joint guidance term to adjust attention weights.

8. An interpretable attribution system based on a star catalog semantic segmentation model, employing the interpretable attribution method based on a star catalog semantic segmentation model as described in any one of claims 1 to 7, characterized in that, include: The region segmentation module performs grayscale processing and gradient calculation on the star catalog image, adjusts the cluster centers based on the gradient minimum strategy, and divides the star catalog image into superpixel regions. The topology construction module maps superpixel regions to topology nodes, establishes node links based on spatial adjacency and brightness gradient similarity, and constructs a region-level topology graph. The mask generation module calculates the structural importance of each region, samples candidate regions according to probability distribution, samples the mask center point based on gradient value within the candidate region, calculates the gradient principal direction angle of the mask center point, constructs the rotation matrix and elliptical covariance matrix, and generates a set of directional elliptical Gaussian perturbation masks. The saliency calculation module applies a perturbation mask to the original image, inputs it into the semantic segmentation model to obtain the predicted difference, and calculates the saliency score by combining the distance-weighted attenuation factor. The explanatory graph construction module propagates saliency information in the topological graph based on the graph attention mechanism, and constructs a structural saliency explanatory graph.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the interpretable attribution method based on the star catalog semantic segmentation model according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the interpretable attribution method based on the star catalog semantic segmentation model according to any one of claims 1 to 7.

Images