An explainable attribution method, system, device and storage medium based on a star semantic segmentation model

Abstract

The application discloses an explainable attribution method and system based on a star atlas semantic segmentation model, a device and a storage medium, and belongs to the technical field of deep space exploration image processing. The star atlas image is subjected to gray processing and gradient calculation, and the image is divided into super-pixel regions based on a gradient minimum strategy; a regional-level topological structure diagram is constructed, and node links are established based on spatial adjacency and brightness gradient similarity; the structural importance of each region is calculated, a center point of a mask is sampled, a directional elliptical Gaussian perturbation mask set is generated according to a gradient main direction angle; the perturbation mask is applied to the original image, and a semantic segmentation model is input to calculate a saliency score; and saliency information is propagated in the topological structure diagram based on a graph attention mechanism, and a structural saliency explanation diagram is constructed. Through regional division and topological propagation of terrain perception, model explanation of spatial continuity and terrain consistency is realized, and the technical problems that a traditional method ignores the physical characteristics of terrain structure and lacks a propagation path description are solved.

Description

Technical Field

[0001] This invention relates to the field of deep space exploration image processing technology, specifically to an interpretable attribution method, system, device, and storage medium based on a star catalog semantic segmentation model. Background Technology

[0002] Planetary surface semantic segmentation is one of the core technologies of autonomous decision-making systems for deep space exploration. Based on remote sensing images acquired during deep space exploration, it enables pixel-level identification and classification of landforms such as craters, ridges, and gullies, directly impacting landing site selection, rover path planning, and research on geological evolution mechanisms. With the increasing complexity and routine nature of deep space exploration missions, tasks such as planetary surface terrain identification, obstacle avoidance, and scientific target localization urgently require real-time and reliable semantic segmentation algorithms.

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

[0004] Traditional interpretation methods such as Grad-CAM and RISE, while applicable to image-level interpretation, exhibit problems like poor spatial consistency and unreliable attribution in complex terrain backgrounds, weak texture features, and semantically ambiguous boundaries. Existing methods are mostly based on pixel-level masking perturbations, lacking characterization of the physical properties of terrain structures in images. They employ unstructured mask generation methods, failing to effectively align with the structural boundaries of the real terrain, leading to fragmentation of salient regions. DeepLabV3 (Deep-Lab-(Version)-3) is an advanced semantic segmentation deep learning model; Grad-CAM stands for Grad-weighted Class Activation Mapping; and RISE stands for Randomized Input Sampling for Explanation, an interpretation method based on random input sampling. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem addressed by this invention is how to solve the technical issues of existing interpretability methods for star catalog semantic segmentation models, such as neglecting the physical characteristics of terrain structure, lacking structurally aligned region division, and lacking characterization of propagation paths and dependency strength in the interpretation results. Considering the special characteristics of star catalog images, this invention provides an interpretable attribution method that combines terrain perception and topological propagation, achieving model interpretation with spatial continuity and terrain consistency.

[0007] To solve the above technical problems, the present invention provides the following technical solution: an interpretable attribution method based on a star catalog semantic segmentation model, which includes performing grayscale processing and gradient calculation on the star catalog image, adjusting the cluster center based on the gradient minimum strategy, and dividing the star catalog image into superpixel regions.

[0008] 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.

[0009] 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 elliptical covariance matrix, and generate a set of directional elliptical Gaussian perturbation masks.

[0010] 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.

[0011] Based on the graph attention mechanism, saliency information is propagated in the topological graph, and a structural saliency interpretation graph is constructed.

[0012] As a preferred embodiment of the interpretable attribution method based on a star catalog semantic segmentation model described in this invention, wherein dividing the star catalog image into superpixel regions includes:

[0013] 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;

[0014] Pixels are assigned to the cluster center with the smallest distance based on the joint distance, which combines gradient differences and spatial location information.

[0015] As a preferred embodiment of the interpretable attribution method based on a star-table semantic segmentation model described in this invention, the construction of the region-level topology graph includes:

[0016] Design link weight function based on regional spatial distance and gradient differences;

[0017] Set link filtering conditions to establish link connections only when both spatial distance and gradient difference meet the thresholds.

[0018] As a preferred embodiment of the interpretable attribution method based on a star catalog semantic segmentation model described in this invention, wherein: sampling the mask center point based on gradient values ​​within the candidate region includes:

[0019] The regional structural importance is normalized to obtain the probability of each region being selected as a candidate region.

[0020] Sample the mask center point within the candidate region according to the probability distribution of pixel gradient values.

[0021] As a preferred embodiment of the interpretable attribution method based on a star catalog semantic segmentation model described in this invention, wherein: the set of directional elliptical Gaussian perturbation masks includes:

[0022] Construct a two-dimensional rotation matrix based on the principal direction angle of the mask center point;

[0023] Define the perturbation range in the primary and secondary directions, and construct the rotated elliptic covariance matrix;

[0024] The Gaussian distribution determination function is used to determine whether a pixel falls within the masking range.

[0025] As a preferred embodiment of the interpretable attribution method based on a star catalog semantic segmentation model described in this invention, the calculation of the significance score by combining a distance-weighted attenuation factor includes:

[0026] The absolute difference between the predicted response before and after the perturbation is calculated as the initial significance score;

[0027] An exponential decay factor based on Euclidean distance is introduced for distance-weighted correction.

[0028] As a preferred embodiment of the interpretability attribution method based on a star-table semantic segmentation model described in this invention, the construction of the structural saliency interpretation graph includes:

[0029] Initialize the region saliency features and summarize the saliency scores and region attributes of the pixels within the region;

[0030] 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.

[0031] This invention provides an interpretable attribution system based on a star catalog semantic segmentation model.

[0032] To solve the above technical problems, the present invention provides the following technical solution: an interpretable attribution system based on a star catalog semantic segmentation model, comprising: a region division module, which performs grayscale processing and gradient calculation on the star catalog image, adjusts the cluster center based on the gradient minimum strategy, and divides the star catalog image into superpixel regions;

[0033] 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.

[0034] 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.

[0035] 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.

[0036] 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.

[0037] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor, when executing the computer program, implements the steps of the interpretable attribution method based on a star-list semantic segmentation model.

[0038] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the interpretable attribution method based on a star-list semantic segmentation model.

[0039] The beneficial effects of this invention are as follows: By performing adaptive superpixel region segmentation on star surface images based on a gradient minimum strategy, this invention effectively avoids the problem of traditional pixel-level perturbation methods ignoring the physical characteristics of terrain structure, enabling the segmented regions to align with the boundaries of the real terrain and providing structurally stable basic units for subsequent interpretation. A region-level topology graph is constructed, and a link weight function based on spatial adjacency and structural similarity is designed to establish a high-level information propagation channel, addressing the lack of structural representation capabilities in existing methods. A sampling strategy driven by structural importance is used to generate mask center points, and a terrain-oriented adaptive elliptical perturbation mask is constructed. Addressing the strip-like or crack-like structure of star surface terrain, directional perturbation is implemented to improve the mask's coverage of edge structures and the accuracy of interpretation. Based on a graph attention mechanism, saliency information is propagated in the topology, and the attention weights are adjusted using a structure-space joint guiding term, effectively avoiding problems such as fragmented saliency heatmaps and structural discontinuities caused by traditional methods. Overall, this invention achieves interpretable attribution of a star surface semantic segmentation model with spatial continuity and terrain consistency, providing reliable technical support for model decision tracing and scientific analysis in deep space exploration. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating an interpretable attribution method based on a star-table semantic segmentation model, as provided in one embodiment of the present invention.

[0042] Figure 2 This is a computer device diagram illustrating an interpretable attribution method based on a star-list semantic segmentation model, provided as an embodiment of the present invention.

[0043] Figure 3 This is a schematic diagram illustrating a process of an embodiment of an interpretable attribution system based on a star-table semantic segmentation model, provided as an example of the present invention. Detailed Implementation

[0044] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0045] Example 1, referring to Figures 1-3 This is one embodiment of the present invention, which provides an interpretable attribution method based on a star catalog semantic segmentation model, including:

[0046] S100: Perform grayscale processing and gradient calculation on the star catalog image, adjust the cluster center based on the gradient minimum strategy, and divide the star catalog image into superpixel regions.

[0047] S200: Maps superpixel regions as topological nodes, establishes node links based on spatial adjacency and brightness gradient similarity, and constructs a regional topology graph;

[0048] S300: 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 elliptical covariance matrix, and generate a set of directional elliptical Gaussian perturbation masks.

[0049] S400: Apply the perturbation mask to the original image, input it into the semantic segmentation model to obtain the predicted difference, and calculate the significance score by combining the distance-weighted decay factor;

[0050] S500: Based on the graph attention mechanism, propagate saliency information in the topological graph and construct a structural saliency interpretation graph.

[0051] It should be noted that the core challenges of semantic segmentation of star surfaces lie in the complexity of terrain structure and the opacity of model decisions. Craters, ridges, gullies, and other landforms in star surface images have distinct geometric shapes and directions. Traditional pixel-level perturbation methods cannot effectively align with the structural boundaries of real-world terrain, leading to fragmentation or boundary misalignment of salient regions. Simultaneously, the high noise and weak texture features of the star surface environment cause existing interpretation methods to exhibit poor spatial consistency and unreliable attribution, failing to provide reliable model decision attribution for deep space exploration missions.

[0052] Therefore, to address the aforementioned issues of model interpretability and terrain awareness, a complete interpretability attribution scheme for star catalog semantic segmentation models is constructed through steps S100-S500. S100 achieves precise alignment with the terrain structure through superpixel partitioning using a gradient minimum strategy, providing stable structural units for subsequent analysis. S200 constructs a topological graph to establish spatial adjacency relationships between regions, laying the foundation for saliency propagation. S300's structure importance-driven sampling and directional elliptical masking design ensure the representativeness of perturbation testing and terrain fit. S400's distance-weighted saliency calculation enhances the spatial consistency of scoring. S500's graph attention propagation mechanism establishes causal dependencies between regions. The overall scheme achieves interpretable attribution with spatial continuity and terrain consistency, solving the key technical problems of traditional methods' inability to characterize the physical properties of terrain structures and the lack of propagation paths.

[0053] Example 2, refer to Figures 1-3 As an embodiment of the present invention, based on the previous embodiment, an interpretable attribution method based on a star catalog semantic segmentation model is provided, including: in step S100, the star catalog image is divided into superpixel regions, and step S100 includes the following steps A1-A2:

[0054] A1: 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;

[0055] The star catalog image is initialized, and then the image grayscale is processed and the gradient value is calculated in sequence. The gradient value is calculated using the difference approximation.

[0056] Star catalog image preprocessing: The original star catalog image size is uniformly adjusted to... ,at this time , .

[0057] ① For the original RGB color star chart image Grayscale processing is performed to extract the brightness edge features of the star catalog image. The grayscale processing uses a weighted average formula as follows:

[0058] ,

[0059] in, , , These are the pixel positions in the original image. The red, green, and blue channel values; The corresponding grayscale value; , These represent the height and width of the image, respectively. Let be a real matrix space. A real matrix of dimension 1.

[0060] ②Based on grayscale values The brightness gradient of each pixel in the image is calculated using a difference approximation. To extract edge information and structural strength from the image, the specific difference calculation formula is as follows:

[0061] ,

[0062] in, This represents the brightness gradient of each pixel in the image. and Corresponding pixel positions grayscale value and pixel position at the location The grayscale value at that location; , Corresponding pixel positions grayscale value and pixel position at the location The grayscale value at that location.

[0063] The star catalog image is divided into superpixel regions by performing clustering operations based on the image processed by A1.

[0064] A2: Pixels are assigned to the cluster centers with the smallest distance based on the joint distance, which combines gradient differences and spatial location information.

[0065] ① Set initial cluster centers: Set initial cluster spacing Divide the image into The initial region will be the initial cluster center. Set as the center pixel of each region.

[0066] Specifically, let the initial cluster spacing be... The image is initially divided into There are 1 cluster region, and the initial cluster center is set. The center pixel of each region.

[0067] ② Replace initial cluster centers: Considering that terrain edges in star catalog images often correspond to higher brightness gradient values, a local gradient minimization strategy is introduced to enhance the structural stability of region partitioning and avoid cluster centers being located at edges or in noisy regions. Specifically, around each initial cluster center... Within the specified range, the pixel with the smallest gradient value is selected as the new center location to enhance the structural stability of the region division and avoid cluster centers being located at edges or in noisy regions. The cluster center replacement formula is as follows:

[0068] ,

[0069] in, The adjusted version after the 0th iteration The initial cluster center coordinates corresponding to each cluster region ; Indicates the initial cluster centers Center The set of pixels within the region; Used to calculate the pixel position corresponding to the minimum gradient value.

[0070] ③ Perform pixel clustering: Find each pixel Calculate the pixel value of all cluster centers within the neighborhood. Joint distance to these cluster centers The pixel is then assigned to the cluster center with the smallest joint distance, thus completing one region partitioning operation. The specific joint distance... The calculation formula is:

[0071] ,

[0072] ,

[0073] in, The square of the joint distance; For pixels Coordinates and the Cluster centers of each cluster region The difference in coordinates; For pixels Coordinates and the Cluster centers of each cluster region The difference in coordinates; This represents the initial cluster spacing; For the first The average brightness gradient of the cluster centers of each cluster region; As a balancing factor between gradient and spatial information, ; For the first The coordinates of the cluster centers of each cluster region; Indicates the first The set of pixel coordinates in the cluster centers of each cluster region; Used to count the number of elements in a set. This distance design effectively enhances the resolution of structural boundaries while maintaining a compact spatial layout.

[0074] ④ After each region division, the new location of each cluster center and the new average gradient value of each region are recalculated. The calculation formula is as follows:

[0075] ,

[0076] ,

[0077] Formula (6) represents the calculation Centroid of clustered regions after round clustering As The initial cluster center coordinates for round clustering; Formula (7) is used to calculate regional brightness gradient of the wheel ; Indicates the first 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.

[0078] ⑤ Repeat steps ②-④ until the cluster centers converge, and finally obtain the results. One superpixel region.

[0079] 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 .

[0080] 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.

[0081] In this embodiment, the construction of the regional topology diagram in step S200 includes the following steps B1-B2:

[0082] B1: Design link weight function based on regional spatial distance and gradient difference;

[0083] 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:

[0084] 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 set of nodes is denoted as . .

[0085] in, A set of nodes; For each node; For the first Superpixel region The corresponding topology node;

[0086] Link topology design: To quantify the connection strength between image regions, a link weighting function is designed to measure the structural similarity and spatial proximity between regions. This involves assigning weights to any two regions. , The link weights between them are defined as follows:

[0087] ,

[0088] ,

[0089] in, The link weight between the two nodes; This is for exponential function operations; Represents any pair of image regions , The structural-spatial joint guiding term; , Represents any pair of image regions , The coordinates of the cluster centers ; Total number of regions; For any pair of image regions The average brightness gradient; For any pair of image regions The average brightness gradient is calculated using formula (7); This indicates the spatial distance between calculated coordinates; This indicates the calculation of absolute value; This is a spatial distance weighting adjustment constant used to control the degree of influence of regional spatial proximity on link weights; This is a structural similarity weight adjustment constant used to control the influence of regional gradient differences on link weights. It controls the ratio between the two components: regional spatial proximity and structural similarity.

[0090] B2: Set link filtering conditions to establish link connections only when both spatial distance and gradient difference meet the threshold.

[0091] Topology Link Optimization: To prevent excessive graph complexity or the introduction of meaningless connections, this invention sets link filtering conditions, the specific formulas of which are as follows:

[0092] ,

[0093] Formula (10) represents the expression for any two image regions , If its spatial distance Less than or equal to the maximum distance And the average brightness gradient difference Less than or equal to the structural threshold Then it is considered that the two image regions , There must be a link between them, otherwise there must not be; This indicates that the value is 0 under other circumstances, meaning that there is no link between the two regions when the above conditions are not met. This represents the maximum permissible spatial distance threshold. The adjacent range coefficient; The initial clustering distance is set during superpixel partitioning.

[0094] in, Used to determine any two regions , Is there a link between them? , is the adjacent range coefficient; The initial cluster spacing is set in step A2.

[0095] Establish the star catalog image topology: based on the node set set in step S100. Steps B1 and B2 define the optimized topology links, and the set of links is denoted as . The final topology diagram is formed. .

[0096] In this embodiment, step S300, sampling the mask center point based on the gradient value within the candidate region, includes the following steps C1-C5:

[0097] C1: Normalize the regional structural importance to obtain the probability of each region being selected as a candidate region;

[0098] To improve the structural alignment capability of saliency analysis, this invention uses the topological structure diagram obtained in step S200. A star catalog image masking method combining the importance of star catalog topographic structure and gradient information was designed. This method was then used to generate a set of star catalog image masks, which were used to perturb the model in physically significant key regions, ensuring the spatial representativeness of the saliency interpretation results. The specific steps are as follows:

[0099] 1. Mask center point sampling:

[0100] ① Sampling of candidate regions for mask center point: Based on the region division results in S300, sample candidate regions for mask center point, and sample a mask center point in each candidate region.

[0101] A) Based on the region division results in step S300, set the star catalog image region. Structural importance Based on this, representative structural regions are selected as candidate regions for the masking center, with structural importance... The calculation formula is as follows:

[0102] ,

[0103] in, For structural importance; Indicates the region Corresponding node The set of adjacent nodes; express A certain adjacent node; The link weight between the two nodes is calculated using formula (8).

[0104] B) Structural importance of each region After normalization, the probability of each region being selected as a candidate region for the mask center point is:

[0105] ,

[0106] in, For the region The probability of being selected as a candidate region; Total number of regions; For the region The structural importance is determined by formula (11).

[0107] C) Based on this probability distribution From regional sets Random sampling without replacement is performed to select... The regions are selected as candidate mask center points, resulting in the final set of candidate regions. :

[0108] ,

[0109] in, This forms the final set of candidate regions. Represents probability distribution Sampling without replacement One region; For a set of regions; , is the set of candidate regions, with the number of elements being . .

[0110] C2: Sample the mask center point within the candidate region according to the probability distribution of pixel gradient values.

[0111] ② Pixel sampling within the candidate region: The set of candidate regions obtained in step C1 ① In the middle, for each region Internal pixels are sampled from the center of the mask, with the number of elements being... .

[0112] A) To improve the mask's coverage of edge structures, this invention uses the brightness gradient value within the region as the sampling basis to define the region. Inner pixel The probability of being sampled as the center of the mask as follows:

[0113] ,

[0114] in, To define the region Inner pixel position The probability of being sampled as the center of the mask; For pixels The corresponding brightness gradient value; pixel position The corresponding brightness gradient value is solved by formula (2).

[0115] B) Based on probability distribution A pixel is randomly sampled from each candidate region as the mask center:

[0116] ,

[0117] in, It is a probability distribution; For the region; Indicates the first Candidate regions The center point of the mask sampled from the sample. These are the coordinates corresponding to the center point of the mask; The set of all candidate regions is solved by step ①, where n is the number of regions. All parameters involving n in the following text represent the number of regions.

[0118] C) Finally, we obtain the set of coordinates of all mask center points:

[0119] ,

[0120] in, The set of coordinates of all mask center points; The center point of each mask; To indicate the first The coordinates of the mask center point obtained by sampling from each candidate region.

[0121] Each point in this set will serve as the center of the perturbation mask, driving the construction of subsequent directional perturbation masks, thereby improving the interpretability of the saliency map to perceive the edge regions.

[0122] The set of directional elliptical Gaussian perturbation masks includes:

[0123] Since the surface topography of celestial bodies is mainly characterized by banded or fissure-like structures, this invention designs the mask shape as elliptical. However, due to the varying orientations of natural terrain structures on the celestial surface, including fissures and bands, the elliptical mask needs to be oriented. Therefore, this invention constructs a directional elliptical mask to meet the above requirements. The specific steps are as follows:

[0124] C3: Construct a two-dimensional rotation matrix based on the principal direction angle of the mask center point;

[0125] ① Construct the mask rotation matrix:

[0126] 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:

[0127] ,

[0128] ,

[0129] 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).

[0130] 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 :

[0131] ,

[0132] 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.

[0133] C4: Define the perturbation range in the main and secondary directions, and construct the rotated elliptic covariance matrix;

[0134] ② Construct the covariance matrix of the rotated ellipse:

[0135] 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:

[0136] ,

[0137] in, The covariance matrix is ​​an unrotated elliptic matrix; The range of disturbance in the main direction; This represents the disturbance range in the secondary direction.

[0138] Based on this, we obtain the result in the first... The covariance matrix at the center of each mask, rotated along the gradient direction:

[0139] ,

[0140] 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).

[0141] C5: Determines whether a pixel falls within the masking range based on the Gaussian distribution determination function.

[0142] 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:

[0143] ,

[0144] in, It is a pixel offset vector; and Each pixel Coordinates and the individual mask center Coordinate difference, pixel points Coordinates and the individual mask center The difference in coordinates; Indicates the first The center point of the mask The coordinates.

[0145] 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. Did it fall into the first The specific formula for determining the masking range at the center of the mask is as follows:

[0146] ,

[0147] 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.

[0148] Based on the above results, define a set of perturbation masks. for:

[0149] ,

[0150] 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.

[0151] In this embodiment, step S400, which combines the distance-weighted attenuation factor to calculate the significance score, includes the following steps D1-D2:

[0152] 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.

[0153] D1: Calculate the absolute difference between the predicted response before and after the disturbance as the initial significance score;

[0154] Generate a perturbation image and obtain the perturbation output.

[0155] 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 :

[0156] ,

[0157] 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.

[0158] Will pass through the first Image after masking perturbation Input semantic segmentation model Obtain the predicted perturbation response for the target category:

[0159] ,

[0160] in, Indicates the first A masked perturbation image, input Obtain pixels The corresponding disturbance prediction response output.

[0161] ① Calculate the initial significance score: Using the original star catalog image... enter Obtain pixels Unperturbed predictive response Predict the response based on the disturbance. and unperturbed predictive response Calculate the first The response change caused by each mask to each pixel:

[0162] ,

[0163] in, For the first The response change caused by each mask to each pixel; For unperturbed predictive response; For disturbance prediction response; This indicates the calculation of absolute value.

[0164] The response changes are mapped to the corresponding masked regions to obtain the initial saliency scores for the pixels:

[0165] ,

[0166] in, Represents pixels In the Initial significance score under the perturbation of the mask; For the first The response change caused by each mask to each pixel; For the first A mask at pixel position The value at the given location is a binary mask of 0 or 1, which is solved by formula (23).

[0167] D2: Introduce an exponential decay factor based on Euclidean distance for distance-weighted correction.

[0168] ② Introducing Distance-Weighted Attenuation: Since the mask center is the focal point of the disturbance, the influence of pixels farther from the center is less reliable. This invention introduces a distance-weighted attenuation factor to enhance the spatial consistency of the interpretation score for the disturbance region. The formula for calculating the distance-weighted attenuation factor is:

[0169] ,

[0170] ,

[0171] in, To adjust the parameters, It can be set between 0.01 and 0.05, with α=0.02. This value can ensure the weight of the central region while avoiding the excessive decay of the weight of the peripheral region, and control the decay rate of the significance score due to distance. This is the distance-weighted attenuation factor; Represents pixels To the individual mask center Euclidean distance; Indicates the first The center point of the mask The coordinates.

[0172] ③ Calculate the final significance score: use the distance-weighted decay factor Introducing initial significance scores , obtained the Distance-weighted significance score under masking perturbation:

[0173] ,

[0174] in, For the first Distance-weighted significance score under masking perturbation; This is the initial significance score; This is the distance-weighted attenuation factor.

[0175] Will The significance scores under each masking perturbation are fused and weighted to obtain the final significance score:

[0176] ,

[0177] in, The final significance score is defined in step S300. For the first The distance-weighted significance score under each masking perturbation is solved by formula (31); Indicates the first The mask is solved by formula (23); The distance-weighted attenuation factor is calculated using formula (29).

[0178] In this embodiment, the construction of the structural saliency interpretation map in step S500 includes the following steps E1-E2:

[0179] The final significance score obtained in step S400 Based on the topology diagram in step S200 The graph attention mechanism is used to achieve directed propagation and spatial completion of saliency information, thereby improving the structural continuity of saliency scores and the ability to interpret the whole graph.

[0180] E1: Initialize the region saliency feature and summarize the saliency score of the pixels in the region and the region attributes;

[0181] Initialize the saliency characteristics of the region: Summarize the region The initial saliency feature of the region is determined by the saliency score of all pixels within the region. for:

[0182] ,

[0183] in, This is the initial significance feature; Indicates the first The set of pixels contained in the cluster center of each cluster region; Represents pixels The significance score; Indicates the first The average brightness gradient of the cluster centers of each cluster region Solve using formula (5); Indicates the first The coordinates of the cluster center points of each cluster region; It is a 4-dimensional real vector space.

[0184] The initial saliency characteristics of the five regions are as follows:

[0185] ,

[0186] ,

[0187] ,

[0188] ,

[0189] ,

[0190] in, This is the salient feature vector of the first region at layer 0 (initialization); This is the salient feature vector of the second region at layer 0 (initialization); This represents the salient feature vector of the third region at layer 0 (initialization). This is the salient feature vector of the 4th region at layer 0 (initialization); This is the salient feature vector of the 5th region at layer 0 (initialization).

[0191] E2: Employ the propagation saliency features of the structure-enhanced graph attention network and adjust the attention weights using the structure-space joint guiding term.

[0192] The saliency information graph propagation mechanism employs a structure-enhanced graph attention network in the topology graph. Above, the significant characteristics of the propagation area. (The following is a list of points, likely related to the spread of the virus.) The formula for updating the features of layer nodes is:

[0193] ,

[0194] ,

[0195] in, For the first Layer nodes Features; Indicates the first Layer nodes Features; For shared feature transformation matrices; The characteristic transformation matrix With the Layer nodes eigenvectors Matrix product; Represents a node The set of adjacent nodes; Represents a node For nodes The attention weight determines the direction and intensity of the propagation path; Represents a learnable attention weight vector; This represents a non-linear activation function with a slope of 0.2. This represents a vector concatenation operation; The structure-space joint guidance term for a pair of adjacent image regions is used to adjust attention preference and is solved by formula (9).

[0196] Output structural saliency propagation results: propagation After layering, the final set of salient features for all regions is obtained. Set of information propagation paths and dependency strengths between nodes The specific formula is as follows:

[0197] ,

[0198] ,

[0199] in, This is the final set of salient features; To indicate the first Layer nodes Features; The total number of regions refers to the total number of nodes in this invention; This indicates the existence of slave nodes. To the node Information dissemination paths; This is a set of information propagation paths and dependency strengths between nodes; This represents the attention weights along the path, characterizing the strength of the dependency. This represents the set of links, which is solved in step E2.

[0200] Based on all the information above, construct a complete structural saliency interpretation diagram:

[0201] ,

[0202] in, A diagram illustrating structural saliency; This represents the set of nodes, as defined in step S200; The information propagation path between nodes and their dependency strength are represented by formula (37); The final saliency characteristics of all regions are represented by formula (36).

[0203] Compared to other inventions that perform interpretable analysis of semantic segmentation, this invention conforms to the continuous distribution characteristics of extraterrestrial surface landforms. By identifying terrain structures, it enhances the spatial continuity of heat maps and establishes structural interpretation paths.

[0204] Example 3, referring to Figure 1 and Figure 3 This is one embodiment of the present invention, which provides an interpretable attribution method based on a star table semantic segmentation model. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0205] Table 1 shows the results generated based on steps A1-A2 in Example 2, specifically derived from the cluster center replacement in formula (3) and the average brightness gradient calculation in formula (5). The cluster centers are adjusted by searching for the minimum gradient value within a 3×3 range to ensure that the center points are located in stable terrain areas rather than edge noise areas. Due to the large amount of data, the cluster center coordinates and average brightness gradient values ​​are calculated for the first five regions as shown in Table 1.

[0206] Table 1. Cluster center coordinates and average brightness gradient values ​​of the first five regions.

[0207]

[0208] As shown in Table 1, the cluster center coordinates of the five regions exhibit a regular distribution, with the y-coordinate all being 16, indicating that these regions are located on the same horizontal band; the average brightness gradient varies between 0.22 and 0.37. The minimum gradient value in the region is 0.22. The maximum regional gradient value is 0.37, reflecting the differences in terrain complexity across different regions. Compared to traditional fixed grid partitioning, the adaptive partitioning of this invention can dynamically adjust regional boundaries based on terrain features, resulting in better structural consistency within each region.

[0209] Table 2 shows the topology construction process based on steps B1-B2 in Example 2, where link weights are calculated using formulas (8)-(10) and filtering conditions are set. The link weights take into account both spatial distance and structural similarity, and only regions that simultaneously satisfy the distance threshold ≤ 48 and the gradient difference threshold ≤ 0.2 are connected. Similarly, taking five links as examples, the starting regions and link weights are shown in Table 2.

[0210] Table 2 shows the starting regions and link weights of five links.

[0211]

[0212] Table 2 shows five links with spatial distances ranging from 28.7 to 46.1, structural differences ranging from 0.09 to 0.18, and corresponding link weights ranging from 0.611 to 0.713. It can be observed that... ~ The link has the highest weight of 0.713, corresponding to the smallest structural difference of 0.09, verifying that the weight function designed in this invention can accurately quantify the similarity between regions. Compared with simple Euclidean distance connections, the dual constraint mechanism of this invention ensures both spatial adjacency and structural consistency, providing a more reliable topological foundation for subsequent saliency propagation.

[0213] Table 3 is derived from the mask generation process in steps C1-C5 of Example 2. The mask center is determined by the structural importance-driven sampling formulas (11)-(16), and the principal orientation angle is obtained by calculating the gradient direction using formulas (17)-(18). This design ensures that the mask is located in the critical area of ​​the structure and conforms to the actual orientation of the terrain. Similarly, taking the first five masks as examples, the calculated mask center coordinates and principal orientation angles are shown in Table 3.

[0214] Table 3 Center coordinates and principal orientation angles of the first five masks

[0215]

[0216] The principal orientation angles of the five masks in Table 3 are concentrated in the range of 1.47–1.62 radians. This relative concentration of angles indicates that the terrain in this area has similar dominant orientation characteristics. The center coordinates of the masks are relatively dispersed. arrive This demonstrates the effectiveness of structural importance sampling, enabling the identification of key terrain features in different regions. Compared to traditional random or uniform sampling, the importance-driven sampling of this invention ensures that the masking is concentrated in the key areas that have the greatest impact on model decisions, improving the targeting and effectiveness of perturbation testing.

[0217] Table 4 shows the significance calculation process based on steps D1-D2 in Example 2, completing the full calculation chain from perturbation input to final score through formulas (25) to (32). The difference between the unperturbed output and the perturbed output reflects the degree of influence of the masking region on the model prediction, and the distance-weighted correction enhances the spatial consistency of the score. Taking the first five masks as examples, the significance scores are shown in Table 4.

[0218] Table 4 shows the starting regions and link weights of five links.

[0219]

[0220] Table 4 shows that the output without perturbation is between 0.88 and 0.93, and the output after perturbation drops to 0.76 to 0.83. The predicted difference is converted into an initial significance score of 0.065 to 0.086, and the final score after distance weighting correction is 0.079 to 0.112. The masking region had the highest final significance score of 0.112, indicating that it had the greatest impact on the model's decision. Compared to the simple difference calculation of traditional methods, the distance-weighted mechanism introduced in this invention effectively enhances the spatial continuity of the significance score and avoids the interference of edge effects on the results.

[0221] Table 5 is derived from the graph attention propagation mechanism in steps E1-E2 of Example 2, where attention weights between nodes are calculated using formulas (34) and (35). Joint guiding term Combining spatial distance and structural similarity, attention weights determine the intensity and direction of salient information propagation in the topological network. (Based on region...) For example, its adjacent area , , The attention weights are shown in Table 5.

[0222] Table 5 shows the starting regions and link weights of five links.

[0223]

[0224] Assuming that in formula (34) - formula (35) Given the identity matrix, the updated features are:

[0225] ,

[0226] in, This represents the salient feature vector of the first region after propagation in the first layer; This is the initial salient feature vector for the second region; This is the initial salient feature vector for the 17th region.

[0227] Table 5 Taking a region as an example, its relationship with adjacent regions , , The joint guiding terms were 0.78, 0.65, and 0.60, respectively, with corresponding attention weights of 0.412, 0.336, and 0.252. The weight distribution was positively correlated with the size of the guiding terms. - The highest attention weight of 0.412 indicates that saliency propagation between these two regions is most active. Compared to traditional methods that only output saliency heatmaps, the graph attention mechanism of this invention can clearly characterize the dependencies and propagation paths between regions, providing a quantitative basis for causal analysis of model decisions and significantly improving the scientific validity and credibility of interpretability analysis.

[0228] Example 4, refer to Figure 1 and Figure 3 This is an embodiment of the present invention, which provides an interpretable attribution system based on a star catalog semantic segmentation model, including a region division module, which performs grayscale processing and gradient calculation on the star catalog image, adjusts the cluster center based on the gradient minimum strategy, and divides the star catalog image into superpixel regions.

[0229] 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.

[0230] 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.

[0231] 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.

[0232] 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.

[0233] This embodiment also provides an electronic device applicable to an interpretable attribution method based on a star-list 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 implement an interpretable attribution method based on a star-list semantic segmentation model as proposed in the above embodiment.

[0234] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements an interpretable attribution method based on a star catalog semantic segmentation model as proposed in the above embodiments.

[0235] The storage medium proposed in this embodiment and the interpretable attribution method based on the star table semantic segmentation model proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0236] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0237] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An explainability attribution method based on a star semantic segmentation model, characterized in that: The method comprises the steps of: The star atlas image is subjected to grayscale processing and gradient calculation, the clustering center is adjusted based on a gradient minimum strategy, and the star atlas image is divided into super-pixel regions; The super-pixel regions are mapped into topological nodes, node links are established based on spatial adjacency and brightness gradient similarity, and a regional topological structure graph is constructed; The structural importance of each super-pixel region in the regional topological structure graph is calculated, candidate regions are sampled according to a probability distribution, a mask center point is sampled in the candidate region based on a gradient value, a gradient main direction angle of the mask center point is calculated, a rotation matrix and an elliptical covariance matrix are constructed, and a directional elliptical Gaussian perturbation mask set is generated; The step of sampling the mask center point in the candidate region based on the gradient value comprises the steps of: The step of generating the directional elliptical Gaussian perturbation mask set comprises the steps of: A two-dimensional rotation matrix is constructed according to the main direction angle of the mask center point; A perturbation range in the main and secondary directions is set, and a rotated elliptical covariance matrix is constructed; It is determined whether a pixel point falls into a mask range based on a Gaussian distribution determination function; 2. The method of claim 1, wherein: The perturbation mask is applied to the original image, a predicted difference is obtained by inputting a semantic segmentation model, a saliency score is calculated by combining a distance weighted decay factor; Saliency information is propagated in the topological structure graph based on a graph attention mechanism, and a structural saliency interpretation graph is constructed. The step of dividing the star atlas image into super-pixel regions comprises the steps of:

3. The method of claim 2, wherein: A pixel point with the minimum gradient value in a 3*3 range around the initial clustering center is searched as a new center position; A pixel point is assigned to a clustering center with the minimum distance based on a joint distance, and the joint distance comprehensively considers gradient difference and spatial position information. The step of constructing the regional topological structure graph comprises the steps of:

4. The method of claim 3, wherein: A link weight function is designed based on regional spatial distance and gradient difference; A link filtering condition is set, and a link connection is established only when the spatial distance and the gradient difference both satisfy a threshold value. The step of calculating the saliency score by combining the distance weighted decay factor comprises the steps of:

5. The method of claim 4, wherein: An absolute difference value of a predicted response before and after perturbation is calculated as an initial saliency score; A Euclidean distance-based exponential decay factor is introduced for distance weighted correction. The step of constructing the structural saliency interpretation graph comprises the steps of:

6. A system for explainable attribution based on a star semantic segmentation model, applying a method for explainable attribution based on a star semantic segmentation model according to any one of claims 1 to 5, characterized in that, An initial regional saliency feature is initialized, and the saliency scores of the pixel points in the region and the region attributes are summarized; A structure-enhanced graph attention network is adopted to propagate the saliency feature, and a structure-space joint guide item is used to adjust the attention weight. The method comprises the steps of: A regional division module is configured to perform grayscale processing and gradient calculation on the star atlas image, adjust the clustering center based on a gradient minimum strategy, and divide the star atlas image into super-pixel regions; A topological construction module is configured to map the super-pixel regions into topological nodes, establish node links based on spatial adjacency and brightness gradient similarity, and construct a regional topological structure graph; A mask generation module is configured to calculate the structural importance of each super-pixel region in the regional topological structure graph, sample candidate regions according to a probability distribution, sample a mask center point in the candidate region based on a gradient value, calculate a gradient main direction angle of the mask center point, construct a rotation matrix and an elliptical covariance matrix, and generate 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. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method for explainable attribution based on a star semantic segmentation model according to any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method for explainable attribution based on a star semantic segmentation model according to any one of claims 1 to 5.

Images