A hybrid network architecture-based all-sky robust high-precision star map star point identification method
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2026-03-17
- Publication Date
- 2026-07-14
Smart Images
Figure CN122391693A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of astronomical navigation and star map recognition technology, specifically relating to a robust and high-precision star map star point recognition method based on a hybrid network architecture for the entire sky. Background Technology
[0002] Astronomical navigation is a navigation method that determines the spatial attitude, position, and other navigation parameters of a spacecraft by observing the known coordinates and motion laws of celestial bodies. Star sensors are a commonly used type of spacecraft attitude sensor. They can determine the three-axis attitude of a spacecraft by observing stars and other celestial bodies without any prior information. Star sensors use specific optical elements and sensors to measure the position and brightness of stars in specific wavelengths, identify and obtain stellar spatial information using star map recognition methods, and finally use attitude calculation algorithms to obtain high-precision three-axis attitude information of the spacecraft. Its working process can be mainly divided into three parts: star point extraction, star map recognition, and attitude calculation. Among these, star map recognition is a key core technology in astronomical navigation and spacecraft attitude determination. Its basic task is to identify the specific stars corresponding to star points in the image based on the starscape image observed by sensors (such as star sensors) and prior information, thereby determining the observation direction and the attitude of the spacecraft.
[0003] As space missions become increasingly diverse and complex, higher demands are placed on star image recognition methods in terms of noise robustness and resistance to environmental interference. Since star sensors rely on imaging stars to determine their attitude, the quality of the star image ultimately affects the attitude determination accuracy of the star sensor. On one hand, the imaging quality of the star image itself directly impacts the star extraction process. For example, in high-dynamic observation scenarios, star images are prone to motion blur, resulting in unclear star images and distorted star shapes, affecting the accurate extraction of star positions. Under strong noise interference, star extraction is also prone to positional deviations, star loss, and numerous "false stars." On the other hand, a decline in star extraction performance significantly impacts subsequent star image recognition, leading to decreased recognition accuracy or even recognition failure, severely affecting the attitude calculation of the star sensor.
[0004] Currently, there are many traditional star map recognition methods based on subgraph isomorphism and pattern recognition, such as the triangle method, pyramid algorithm, and grid method. These methods typically extract invariant features such as angular distances between stars to construct a pattern library, which is then compared with observed patterns. However, they all rely on the retrieval and matching of pre-constructed navigation star maps, and their matching and iteration time is affected by the size of the navigation star map. Large-scale retrieval, especially for all-sky recognition, inevitably slows down the recognition speed. In addition, manually constructed features are very susceptible to noise interference, especially in the case of false stars and missing stars. Although the sensor accuracy of modern star sensors has greatly improved, in high dynamic, noisy, and all-sky recognition scenarios, sufficiently robust recognition algorithms are still needed to reduce hardware costs.
[0005] Deep learning methods, with their powerful feature extraction and noise resistance capabilities, have been widely applied in various fields, such as point cloud recognition and segmentation, and feature point matching. Currently, some works have attempted to introduce deep learning methods into star map recognition, for example, modeling the observed star map as a sparsely connected graph structure and then using a graph attention network for node classification. This model can adaptively learn the topological features of the star map, thus achieving accuracy exceeding traditional algorithms even with missing stars and the presence of false stars. However, its artificially designed sparse connections lead to information loss at the edges of the observation area or in high-noise conditions. Furthermore, some end-to-end algorithms attempt to model star map recognition as an image classification task, but these have a large number of parameters, high computational cost, lack real-time performance, and modeling star distribution pattern classification as an image classification task results in significant information redundancy.
[0006] Traditional star map recognition methods primarily rely on retrieval strategies based on subgraph isomorphism and pattern matching. However, these methods exhibit inherent limitations when facing highly dynamic and noisy complex spatial environments. On one hand, they depend on manually defined geometric features (such as star diagonal distances) for pattern construction and matching, making them extremely sensitive to star position noise, loss, and false stars, resulting in severely insufficient robustness. On the other hand, the matching process requires frequent retrieval of large-scale navigation star catalogs, leading to high computational costs and poor real-time performance in all-sky recognition. Furthermore, they typically require complex pre-processing and post-processing procedures to handle interference, increasing the system's uncertainty and complexity. Although deep learning methods introduced in recent years have shown potential in feature extraction and noise resistance, existing methods still have significant limitations. Some end-to-end models simply model the recognition task as image classification, resulting in a large number of model parameters and high computational costs, making it difficult to meet the real-time deployment requirements of onboard embedded platforms. Moreover, redundant image processing fails to fully and efficiently utilize the precise geometric relationships between stars, resulting in continued poor robustness. Other graph neural network-based models, while considering the topological structure of stars, often employ manually designed sparse connections. These connections are easily disrupted by noise interference, leading to the destruction of the global correlation between real stars. This is especially true when stars are missing or pseudo-stars are present, resulting in distortion of subgraph patterns and affecting recognition robustness. Furthermore, their node features still rely on handcrafted features such as angular distances, limiting their representational capabilities and restricting the model's ability to learn more robust and discriminative features.
[0007] In summary, existing technologies struggle to simultaneously ensure high recognition accuracy, robustness against complex noise, high efficiency in all-sky recognition, and real-time deployability in resource-constrained environments. Summary of the Invention
[0008] In view of this, the present invention provides a robust and high-precision star map star point recognition method based on a hybrid network architecture for the entire sky. This method has both high robustness and high recognition efficiency.
[0009] The technical solution for implementing the present invention is as follows:
[0010] A robust, high-precision star point identification method for all-sky star maps based on a hybrid network architecture includes: Training data generation: Simulated observation data for the entire sky is generated based on the star catalog. Star point data augmentation is performed on each set of star points in the observation data. From the augmented observation data, any star is selected as the principal star, and the closest star to the principal star is chosen. Each star point forms a fully connected subgraph; the pseudo right ascension and pseudo declination of each star point in the fully connected subgraph are extracted as initial features; Hybrid architecture neural network construction: The construction includes a shared weight high-dimensional mapping network based on MLP, a stacked layer of multi-head Transformer encoders without position encoding, and a neural network with feature aggregation and star point recognition layers; Network training and recognition: The neural network is trained using the training data, and the trained neural network is used to recognize star points; a pseudo-star point classification auxiliary network is added during training to assist in training.
[0011] Optionally, the data augmentation described in this invention includes random rotation, random position offset, star point loss, and pseudo star point addition; The random rotation is: according to a set probability. Rotate the stars in each set of stars by the specified angle. from Uniform sampling in the middle; The random position offset is determined according to a set probability. For each set of stars, the stars are offset in position, starting from a mean of 0 and a variance of . Sampling is performed from a Gaussian distribution, and the sample is truncated to the maximum absolute value. ; Star point loss is determined by: according to a set probability. Several real stars are removed from the star observation set; Add pseudo-star points according to the set probability. Pseudo-star points are added to the star point set, and these pseudo-star points are uniformly sampled from within the pixel plane.
[0012] Optionally, the present invention selects the star closest to the center of the image as the primary star. .
[0013] Optionally, the extraction process of pseudoright ascension and pseudo declination for each star point in this invention is as follows: First, the three-dimensional pointing vector of the star point is obtained based on the pinhole camera model. , These are the projected coordinates of the star points in the pixel plane. Camera intrinsic parameter matrix; subscript Indicates the number of observations, subscript This indicates a specific star observed, indicated by the subscript. Indicates the first The first observation (i.e., a certain image) Each observation yields several star points; Secondly, a local celestial coordinate system is defined based on the principal star vector and the optical axis vector, where the North Pole points to the principal star vector. The axis is perpendicular to the plane formed by the optical axis and the principal star vector. Let the original optical axis vector be... Based on geometric relationships, the representations of the three axes of the local coordinate system in the camera coordinate system are obtained, thereby determining the coordinate transformation matrix:
[0014] in, This represents the three-dimensional pointing vector of the primary star; Finally, the vector representation of the star points in the local coordinate system is obtained using the coordinate transformation matrix. The pseudo-right ascension and declination in the local celestial coordinate system are calculated as the initial features of the star points:
[0015] in, , and These represent pseudoright ascension and pseudo declination, respectively.
[0016] Optionally, the shared weight high-dimensional mapping network based on MLP described in this invention is used to map the initial two-dimensional features of the input. The mapping yields 512-dimensional star point embedding features. ; A multi-head Transformer encoder layer without positional encoding, using an attention mechanism to embed features into fully connected subgraphs. Deep interaction is used to achieve global feature aggregation, resulting in star-shaped features rich in contextual information. ; Feature aggregation and recognition layer, based on star point features Perform max pooling and output the probability distribution of bars belonging to each class through a classification head.
[0017] Optionally, for each layer of the Transformer network, the input features of this invention are... After linear projection, the Query vector, Key vector, and Value vector are obtained separately, respectively. Calculate the attention weight matrix Aggregated features are obtained through the attention weight matrix. ,in Aggregate the features from each attention head. The features of each attention point are concatenated; finally, residual connections, layer normalization, and a feedforward network are applied to obtain the final star point feature representation of the layer. :
[0018]
[0019] in, Indicates the number of floors. This indicates the total number of layers in the Transformer network.
[0020] Optionally, during training, the neural network of the present invention adds a pseudo-star classification auxiliary network, which classifies star features... The classification probability is obtained through the auxiliary binary classifier. Supervision information is This indicates whether each star point is a real star point; the model's loss function consists of a star point category classification loss and a pseudo-star point classification auxiliary loss.
[0021] in, This refers to batch size, which is the number of data points processed at one time during neural network training. Each data point consists of N two-dimensional features. It is the weight of binary classification loss. This represents the classification probability of a hybrid neural network architecture. This represents the supervisory information for the sample data.
[0022] Optionally, during training, this invention generates a one-dimensional binary mask vector. Its length is equal to the number of stars. According to probability Randomly set non-primary star points as This does not participate in subsequent feature extraction and recognition. The one-dimensional binary mask is transposed, and then a two-dimensional attention mask is obtained through matrix multiplication.
[0023] in, For the inverse operation, Represents star points For star points The embedded feature contribution weight is 0. After calculating the dot product weight using the mask vector, the corresponding position is set to negative infinity, thus making it pass through... The weight is 0 afterward.
[0024] Using the Used to replace the attention weight matrix The description is: Add a star mask during training, with a certain probability. Since the star points are actively discarded, the corresponding discarded star points should also be removed when calculating attention weights, global feature aggregation, and loss.
[0025] Optionally, this invention removes corresponding star points when performing global feature aggregation and calculating classification auxiliary loss, i.e., sets it to... Non-primary stars.
[0026] Beneficial effects: First, a hybrid architecture model combining a high-dimensional mapping network based on shared weights and a Transformer feature extraction network is adopted to achieve end-to-end mapping from original coordinate features to primary star IDs. This method avoids the traditional matching retrieval process, directly using a deep learning model to extract robust key features for identification, simplifying the system pipeline and making it more suitable for embedded deployment and real-time inference. The proposed shared weight network independently processes each star point, and then a position-free Transformer encoder extracts robust features, conforming to the constraint that the order of star points is irrelevant. Max pooling is used to aggregate subgraph features. Max pooling allows the model to ignore extreme features caused by certain noise, making the network more robust to various noises.
[0027] Second, a fully connected subgraph is used to model the neighborhood relationships of the main stars, establishing a local celestial coordinate system to better represent the initial features of the stars. This avoids manual extraction of complex features and facilitates the extraction of robust features by the hybrid architecture network. The fully connected subgraph allows information to flow freely throughout the subgraph, enabling the Transformer encoder to fully utilize the information of the entire image and learn multiple effective star patterns, rather than relying solely on certain key stars for recognition. This avoids significant performance degradation under complex noise conditions. Using subgraph features for star image recognition can fully utilize the model capacity, achieving excellent performance in all-sky recognition scenarios. Furthermore, it allows for easy improvement of model capacity, better addressing the challenges of complex scenarios with an increased number of stars.
[0028] Third, the proposed local celestial coordinate system decouples the actual spatial location from the star sensor parameters, making the model input independent of the sensor. The initial characteristics of the pseudo-right ascension and declination representation, in addition to the "angular distance" in traditional methods, also allow the model to directly obtain the relative azimuth relationships between star points.
[0029] Fourth, this invention introduces a collaborative design of star point mask training mechanism and comprehensive data augmentation, enabling the model to learn incomplete star point patterns from the essence, and exhibiting excellent robustness to position noise, star point loss and pseudo-star point interference. This significantly reduces the reliance on complex preprocessing and post-processing error correction mechanisms, and improves the overall efficiency and reliability of the system. Attached Figure Description
[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0031] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A schematic diagram for generating simulated observation data; Figure 3 This is a schematic diagram of data augmentation methods; Figure 4 shows the model structure diagram; Figure 5 All-day area recognition performance; Figure 6 The effect of star point recognition under noisy conditions. Detailed Implementation
[0032] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0033] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.
[0034] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.
[0035] This application provides an embodiment of a robust, high-precision star map star point identification method based on a hybrid network architecture, comprising: Training data generation: Simulated observation data for the entire sky is generated based on the star catalog. Star point data augmentation is performed on each set of star points in the observation data. From the augmented observation data, any star is selected as the principal star, and the closest star to the principal star is chosen. Each star point forms a fully connected subgraph; the pseudo right ascension and pseudo declination of each star point in the fully connected subgraph are extracted as initial features; Neural network construction: The neural network includes a shared weight high-dimensional mapping network based on MLP, a stacked layer of multi-head Transformer encoders without position encoding, and a feature aggregation and star point recognition layer; Neural network training and recognition: The neural network is trained using the training data, and the trained neural network is used to recognize star points; a pseudo-star point classification auxiliary network is added during training to assist in training.
[0036] In this embodiment, the observation data can be divided into training and testing sets for training and testing the neural network. Data augmentation algorithms are used to simulate various noises that may exist in real application scenarios, such as star point position errors, star point loss, and pseudo-star points. A fully connected subgraph representation of the main star point is constructed as the input of the neural network. A hybrid architecture network is constructed by stacking a shared weight high-dimensional mapping network based on MLP (Multilayer Perceptron) and a multi-head position-free encoder layer of Transformer (a neural network model based on self-attention mechanism). The proposed hybrid architecture network is used to extract star point features, and max pooling is used to aggregate subgraph features. A pseudo-star point classification auxiliary network is used to further improve the robustness of the network and train the star point recognition network.
[0037] The specific implementation process of each of the above steps will be explained in detail below, such as... Figure 1 As shown: Step 1: Generate training data from real star catalogs Set the star sensor field of view to The resolution is The star catalog in the J2000 coordinate system is as follows ,in, It is the total number of stars in the star catalog. These are the right ascension, declination, and name of the star. The camera intrinsic parameter matrix can be obtained from the field of view and resolution.
[0038] like Figure 2 As shown, by sampling several viewing directions for each star, and based on a standard pinhole camera model, the pixel coordinates of the star set observed from the current viewing angle can be obtained. Sampling the entire sky area yields training data that does not include rotation transformations. ,in It is the set of stars from the current perspective. These are the projected coordinates of the star points in the pixel plane. It is the name or label corresponding to the primary star in the i-th training sample. It is the total number of training data. It is the total number of star points observed in the i-th training sample.
[0039] Step Two: Small-Batch Data Construction and Data Augmentation like Figure 3 As shown, star data augmentation is performed for each set of stars in the batch. Star data augmentation includes the following four types: A. Random rotation. For example... Figure 3 The diagram in the lower left corner is based on probability. Rotate each set of stars by the specified angle. from Uniform sampling in the middle:
[0040] B. Random position offset. For example... Figure 3 The dashed star points in the lower right diagram represent the probability. The positions of the stars in each star set are offset to simulate the error when a real star sensor extracts the coordinates of a star's centroid. The position offset starts with a mean of 0 and a variance of... Sampling is performed from a Gaussian distribution, and the sample is truncated to the maximum absolute value. :
[0041] C. Star points are missing. For example... Figure 3 The dashed star points connected by the dashed lines in the lower right diagram are represented by probabilities. Several real stars are removed from the star observation set to simulate the loss of stars during the star sensor observation process due to occlusion and star extraction.
[0042] D. False star points. For example... Figure 3 The densely dotted star-shaped dots in the lower right diagram are represented by probabilities. A certain proportion of pseudo-star points are added to the star point set. The pseudo-star points are sampled uniformly from within the pixel plane.
[0043] Step 3: Constructing a star-point representation The observed distribution of star points on the pixel plane is as follows Figure 3 As shown in the top left image. After data augmentation in step two, any star is selected from the image as the principal star, and its location can be identified by the distribution of surrounding stars. For convenience, this step selects the star closest to the image center as the principal star. The image uses solid markers. The closest star to the primary star is selected. Each node forms a fully connected subgraph to represent the main star, where each node is a star.
[0044] The three-dimensional pointing vector of the star point is obtained based on the pinhole camera model. A local celestial coordinate system is defined based on the principal star vector and the optical axis vector, where the North Pole points to the principal star vector. The axis is perpendicular to the plane formed by the optical axis and the principal star vector. Let the original optical axis vector be... Based on geometric relationships, the representations of the three axes of the local coordinate system in the camera coordinate system can be obtained, thereby determining the coordinate transformation matrix:
[0045] Using coordinate transformation matrices to obtain vector representations of star points in the local coordinate system The pseudo-right ascension and declination calculated in the local celestial coordinate system are used as the initial features of the star points, which are independent of the star sensor parameters:
[0046] Therefore, the initial feature of each star point is represented as: Each fully connected subgraph is represented as ,in It is the number of stars in the subgraph, let it be. The initial characteristics of the main star.
[0047] like Figure 4 As shown, the hybrid architecture network in this embodiment consists of a shared weight high-dimensional mapping network based on MLP (Multi-Layer Perceptron), stacked multi-head positionless encoder Transformer (a neural network model based on self-attention mechanism), and a neural network with feature aggregation and star point recognition layers.
[0048] Step 4: High-dimensional mapping In this embodiment, the primary star and the initial two-dimensional features of the star points defined in step one are... First, it passes through a high-dimensional mapping layer of a shared-weight MLP, such as... Figure 4 As shown, the 512-dimensional star point embedding features are obtained. A shared-weight network is used to treat all stars equally, while maintaining the order of the stars as irrelevant. A non-linear network is used to map the initial features from 2D to 512D, which is beneficial for the Transformer encoder to extract features and also helps to increase the model capacity.
[0049] Step 5: Feature Extraction A multi-head Transformer encoder without positional encoding is used to perform deep interaction on the fully connected subgraph embedding features. Since the star observation data is order-independent, positional encoding in the standard Transformer is omitted. Specifically, L=4 Transformer encoder layers are stacked, with each layer employing... The multi-head self-attention architecture has a hidden layer dimension of 512 in its feedforward network. The Transformer allows star features in a subgraph to interact with each other; self-attention calculation enables the model to assign different weights to different star points, and multiple attention heads allow the model to focus on different topological features. This invention removes the position encoding layer from the original Transformer encoder because the star data is independent of its input order.
[0050] Specifically, for each layer of the Transformer network, the input features After linear projection, the Query vector, Key vector, and Value vector obtained from the separate vectors are respectively... Calculate the attention weight matrix Aggregated features are obtained through the attention weight matrix. ,in Aggregate the features from each attention head. The features of each attention point are concatenated. Finally, residual connections, layer normalization, and a feedforward network are applied to obtain the final star feature representation of the layer. :
[0051]
[0052] Step 6: Global Feature Aggregation and Star Point Recognition go through After processing by the Transformer layer, star-shaped features rich in contextual information are obtained. To obtain a global representation of the entire subgraph for recognition, max pooling is used to refine the star features. Maximize pooling along the dimension to obtain a length of Subgraph eigenvectors Max pooling adaptively extracts the most salient features from the subgraph and is inherently robust to variations in the number of stars (e.g., loss) and noise interference. Finally, the subgraph features are processed by a classification head to output the probability distribution of the primary star belonging to each category. The classification head network is a two-layer MLP with ReLU activation function, and the hidden layer dimension is equal to the input.
[0053] Step 7: Model Training During model training, the following three strategies are used. Strategy 1: In this embodiment, a pseudo-star classification auxiliary network is added during training to further improve the robustness of the network, such as... Figure 4 As shown. Star-shaped features. It should be able to aggregate information from the entire subgraph and use it for pseudo-star point classification. Optimizing the auxiliary classification loss helps improve the overall network performance. The star point features are then passed through an auxiliary binary classification head to obtain classification probabilities. Supervision information is Indicates whether each star point is a real star point:
[0054] The model's loss function consists of a star-point category classification loss and a pseudo-star-point classification auxiliary loss:
[0055] in, This refers to batch size. This means that in order to calculate the average loss, since B data points are processed at a time during training, a loss can be calculated for each data point. This indicates the weights for binary classification loss.
[0056] Strategy Two: Adjust data augmentation parameters and network hyperparameters for network training. During training, this invention introduces a star-shaped masking mechanism to help the network learn noise-robust topological features. Specifically, a one-dimensional binary mask vector is generated during the training phase. Its length is equal to the number of stars. According to probability Randomly set non-primary star points as This does not participate in subsequent feature extraction and recognition. The one-dimensional binary mask is transposed, and then a two-dimensional attention mask is obtained through matrix multiplication.
[0057] in, For the inverse operation, Represents star points For star points The embedded feature contribution weight is 0. After calculating the dot product weights using the mask vector, the corresponding positions are set to negative infinity, thus making it pass through... The weight is 0 afterward.
[0058] That is, the result In the matrix, each corresponds to The matrix is The position is set to negative infinity. Matrix and The size of the matrix is N rows and N columns.
[0059] Strategy 3: Also remove the corresponding star points during global feature aggregation and classification auxiliary loss calculation in step 6, i.e., set them as follows: Non-primary stars.
[0060] The method of this invention has high robustness, high recognition efficiency and strong engineering applicability. It can make full use of the geometric and topological relationships between stars to achieve star map recognition that is fast and accurate across the entire sky. It can overcome the shortcomings of existing star map recognition methods, such as sensitivity to noise interference, low recognition efficiency across the entire sky, and complex models that are not conducive to real-time embedded deployment.
[0061] Figure 5 The relationship between all-sky area recognition accuracy and noise intensity is listed, and the navigation star list selected includes stars with an apparent magnitude of 6.2 from the SAO star list. The results show that the present invention has excellent robustness to noise interference. Figure 6 The visualization results of some tests on the SAO star catalog are listed, including position offsets, pseudo-stars, and star point loss noise: star points are located at... Within the pixel plane, the confidence level of the prediction is represented by grayscale values. Correctly identified results are labeled in the format of "star point number / confidence level", incorrectly identified results are labeled in the format of "E / confidence level", false stars are labeled with triangles and "F", missing stars are labeled with small rectangles and "D", and the center of the star map is labeled with "X". Except for a few edge stars that were incorrectly identified with low confidence, most stars were correctly identified with high confidence.
[0062] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A robust, high-precision star point identification method for all-sky star maps based on a hybrid network architecture, characterized in that, include: Training data generation: Generate full-sky simulated observation data based on the star catalog, and perform star point data augmentation on each set of star points in the observation data; From the enhanced observational data, an arbitrary star is selected as the primary star, and the one closest to the primary star is chosen. Each star point forms a fully connected subgraph; the pseudo right ascension and pseudo declination of each star point in the fully connected subgraph are extracted as initial features; Hybrid architecture neural network construction: The construction includes a shared weight high-dimensional mapping network based on MLP, a stacked layer of multi-head Transformer encoders without position encoding, and a neural network with feature aggregation and star point recognition layers; Network training and recognition: The neural network is trained using the training data, and the trained neural network is used to recognize star points; a pseudo-star point classification auxiliary network is added during training to assist in training.
2. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky as described in claim 1, characterized in that, The data augmentation includes random rotation, random position offset, star point loss, and pseudo star point addition; The random rotation is: according to a set probability. Rotate the stars in each set of stars by the specified angle. from Uniform sampling in the middle; The random position offset is determined according to a set probability. For each set of stars, the stars are offset in position, starting from a mean of 0 and a variance of . Sampling is performed from a Gaussian distribution, and the sample is truncated to the maximum absolute value. ; Star point loss is determined by: according to a set probability. Several real stars are removed from the star observation set; Add pseudo-star points according to the set probability. Pseudo-star points are added to the star point set, and these pseudo-star points are uniformly sampled from within the pixel plane.
3. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky as described in claim 1, characterized in that, Select the star closest to the center of the image as the primary star. .
4. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky as described in claim 3, characterized in that, The process of extracting the pseudoright ascension and pseudo declination of each star point is as follows: First, the three-dimensional pointing vector of the star point is obtained based on the pinhole camera model. , These are the projected coordinates of the star points in the pixel plane. This is the camera intrinsic parameter matrix; Secondly, a local celestial coordinate system is defined based on the principal star vector and the optical axis vector, where the North Pole points to the principal star vector. The axis is perpendicular to the plane formed by the optical axis and the principal star vector. Let the original optical axis vector be... Based on geometric relationships, the representations of the three axes of the local coordinate system in the camera coordinate system are obtained, thereby determining the coordinate transformation matrix: in, This represents the three-dimensional pointing vector of the primary star; Finally, the vector representation of the star points in the local coordinate system is obtained using the coordinate transformation matrix. The pseudo-right ascension and declination in the local celestial coordinate system are calculated as the initial features of the star points: in, , and These represent pseudoright ascension and pseudo declination, respectively.
5. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky according to claim 4, characterized in that, The MLP-based shared-weight high-dimensional mapping network is used to map the initial two-dimensional features of the input. The mapping yields 512-dimensional star point embedding features. ; A multi-head Transformer encoder layer without positional encoding, using an attention mechanism to embed features into fully connected subgraphs. Deep interaction is used to achieve global feature aggregation, resulting in star-shaped features rich in contextual information. ; Feature aggregation and recognition layer, based on star point features Perform max pooling and output the probability distribution of bars belonging to each class through a classification head.
6. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky according to claim 5, characterized in that, For each layer of the Transformer network, the input features After linear projection, the Query vector, Key vector, and Value vector are obtained separately, respectively. Calculate the attention weight matrix Aggregated features are obtained through the attention weight matrix. ,in Aggregate the features from each attention head. The features of each attention point are concatenated; finally, residual connections, layer normalization, and a feedforward network are applied to obtain the final star point feature representation of the layer. : in, Indicates the number of floors. This indicates the total number of layers in the Transformer network.
7. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky according to claim 6, characterized in that, During training, a pseudo-star classification auxiliary network is added to the neural network. This auxiliary classification network classifies the star features... The classification probability is obtained through the auxiliary binary classifier. Supervision information is This indicates whether each star point is a real star point; the model's loss function consists of a star point category classification loss and a pseudo-star point classification auxiliary loss. in, This refers to batch size. It is the weight of binary classification loss. This represents the classification probability of a hybrid neural network architecture. This represents the supervisory information for the sample data.
8. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky according to claim 7, characterized in that, During training, a one-dimensional binary mask vector is generated. Its length is equal to the number of stars. According to probability Randomly set non-primary star points as This does not participate in subsequent feature extraction and recognition. The one-dimensional binary mask is transposed, and then a two-dimensional attention mask is obtained through matrix multiplication. in, For the inverse operation, Represents star points For star points The embedded feature contribution weight is 0. After calculating the dot product weight using the mask vector, the corresponding position is set to negative infinity, thus making it pass through... The weight is 0 afterward. Using the Used to replace the attention weight matrix .
9. The robust high-precision star map star point identification method based on a hybrid network architecture for the entire sky according to claim 8, characterized in that, When performing global feature aggregation and calculating the classification auxiliary loss, the corresponding star points are removed, which is set as follows: Non-primary stars.
10. A robust, high-precision star map star point identification device for the entire sky based on a hybrid network architecture, obtained using any one of the methods in claims 1-9, characterized in that... include: A shared-weight high-dimensional mapping network based on MLP is used to map the initial two-dimensional features of the input. The mapping yields 512-dimensional star point embedding features. ; A multi-head Transformer encoder layer without positional encoding, using an attention mechanism to embed features into fully connected subgraphs. Deep interaction is used to achieve global feature aggregation, resulting in star-shaped features rich in contextual information. ; Feature aggregation and recognition layer, based on star point features Perform max pooling and output the probability distribution of bars belonging to each class through a classification head.