A star image recognition method based on fuzzy neural network

Abstract

The application discloses a star map recognition method based on a fuzzy neural network, which comprises the following steps: 1. screening navigation stars, a screening method for voting the navigation stars by using random visual axes is designed, and a navigation star library uniformly distributed in the whole sky is established; 2. the star closest to the center of a field of view is taken as a main star, a navigation star feature subgraph based on a minimum spanning tree mode is constructed by taking the main star as a root node and the angular distance between stars as the weight of an edge, and the navigation star feature subgraph has scale invariance and rotation invariance; 3. random noise is added to a feature vector to generate a training set; 4. a neural network is constructed and trained; 5. star map recognition is performed, and the binary number of the feature subgraph is output. Compared with the traditional star map recognition method, the star map recognition method based on the fuzzy neural network has high recognition accuracy and fast recognition speed, and has good robustness to star position noise, star loss and pseudo stars by using fuzzy rules for network learning and adjustment.

Description

Technical Field

[0001] The invention belongs to the technical field of celestial navigation and relates to a star map recognition method based on fuzzy neural network. Background Art

[0002] With the continuous advancement of aerospace technology, the accuracy of spacecraft attitude measurement is also continuously improving. Star sensors are attitude measurement devices that use stars as reference objects. They offer advantages such as small size, low power consumption, high accuracy, and the resistance to accumulation of navigation errors over time. They are widely used in the field of astronomical navigation technology. Star pattern recognition algorithms are one of the core technologies of star sensors. By matching the observed stars in the star sensor's field of view with the reference stars in the navigation star library to determine the position of the observed stars in the celestial coordinate system, they directly affect the accuracy of star sensor attitude measurement. Current star pattern recognition algorithms are mainly divided into two categories: subgraph isomorphism algorithms and pattern recognition algorithms.

[0003] Subgraph isomorphism algorithms use star points in a star map as vertices and the angular distances of adjacent star pairs as edges. They construct distinct feature graphs and store them in a navigation feature library. During recognition, the library is searched to find the most similar feature graph in the star map. These algorithms are simple to implement and require a low number of observed stars. A representative example, the triangle algorithm, is currently the most mature and widely used star map recognition method. However, due to the low feature dimensionality, these algorithms require a large navigation feature library, resulting in long recognition times and a tendency to encounter redundant and mismatched matches.

[0004] Pattern recognition algorithms utilize the distribution of star points within a certain neighborhood of each navigation star to construct a unique characteristic pattern. They then perform matching and identification by searching a navigation pattern database for the closest matching pattern to the observed pattern. Grid algorithms are a typical example of this type of algorithm. These algorithms require a high number of star points within the star sensor's field of view, limiting their scope of application. Furthermore, establishing the observation pattern relies on the selection of a reference star, and incorrect selection of a reference star can lead to matching failures.

[0005] Neural network-based star recognition algorithms have garnered widespread attention in recent years. The characteristics of navigational stars are represented by the weight matrix between the hidden layers of the neural network. This parallel structure replaces the need for multiple iterations of matching against a database of patterns, effectively improving star recognition speed. However, due to the complex structure of the neural network, the star recognition rate depends on the selection of the training set. Furthermore, the robustness of the recognition algorithm needs to be further improved in the presence of missing stars or artifacts in the field of view. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing star map recognition algorithms by combining subgraph isomorphism algorithms with neural networks to propose a star map recognition method based on fuzzy neural networks. This method achieves a high recognition rate and rapid recognition speed, addresses the problems of missing and inaccurate star point extraction, and demonstrates robustness to star point position noise.

[0007] To achieve the above object, the technical solution adopted by the present invention is: a star map recognition method based on fuzzy neural network, the method comprising the following steps:

[0008] Step 1: Establish a balanced navigation star library;

[0009] Step 2: Construct a feature vector based on the minimum spanning tree; take each star in the navigation star library as the main star, select three auxiliary stars in its neighborhood, construct a feature subgraph, and generate a feature vector based on the adjacency matrix of the feature subgraph and store it in the feature database;

[0010] Step 3: Construct a training set; add different random noises to the feature vectors in the feature database, expand the number of feature vectors, and obtain a training set;

[0011] Step 4: Construct and train the fuzzy neural network; construct the fuzzy neural network, input the training set generated in step 3 into the fuzzy neural network in a random order for training, and generate a star map recognition model;

[0012] Step 5: Perform star map recognition; extract star points from the star map captured by the star sensor to obtain star point positions. The star point closest to the center of the field of view is taken as the main star, and the feature vector of the navigation star is constructed. The feature vector is input into the trained fuzzy neural network for recognition.

[0013] Furthermore, the method for establishing a navigation star library involves selecting the Tycho-2 star catalog as the original catalog. First, based on the maximum magnitude observable by the star sensor, stars with a magnitude greater than 6.0, variable stars, and stars with larger magnitudes in binary stars are deleted from the catalog. Then, 10,000 fields of view are randomly generated using the Monte Carlo method. The stars within each field of view are sorted in ascending order of magnitude, and a vote is conducted on the top five navigation stars. After the voting is completed, stars with a final vote of 0 are deleted. Finally, the remaining navigation stars are renumbered as star numbers. The navigation star library consists of the star number, magnitude, right ascension, and declination information of the screened navigation stars.

[0014] Furthermore, the construction is based on the eigenvector of the minimum spanning tree: First, the navigation star library is traversed, and each navigation star is taken as the main star S0 in turn, and the angular distance θ from the neighboring star to the main star is calculated. 0i , expressed as:

[0015] θ 0i =arccos(sinδi sinδ0+cosδ i cosδ0cos(α i -α0))

[0016] Among them, i = 1, 2, ..., n, n is the number of neighboring stars around the main star, α0, α i ,δ0,δ i They represent the right ascension and declination of the primary star and the neighboring star respectively.

[0017] Then, the angular distance is located at (θ min ,θ FOV ) are sorted by magnitude, θ min is the minimum angular distance required by the star point extraction algorithm, θ FOV is the field of view of the star sensor, select the 6 stars with the smallest magnitude as the auxiliary stars of the main star, and calculate the angular distance between the auxiliary stars. Taking the main star as the root node, the angular distance between the navigation stars θ ij As the edge weights in the graph, a minimum spanning tree algorithm is used to generate a characteristic subgraph of the navigation star. According to the order in which the minimum spanning tree is generated, the auxiliary stars are named S1, S2, …, S6, and a feature vector pat = [θ1θ2θ3θ4θ5θ6] is generated based on the corresponding star-pair angular distances. This minimum spanning tree-based navigation star pattern feature is rotationally and scale-invariant, describing the distribution of star points within the neighborhood of the primary star and is relatively insensitive to magnitude errors. Finally, the star numbers and feature vectors of the primary and auxiliary stars are stored in a feature database.

[0018] Furthermore, the training set is constructed by adding uniformly distributed random noise to the original feature vector, with the noise intensity being 5% of the maximum component of the feature vector, and generating 100 expanded training samples for each feature vector.

[0019] Furthermore, the fuzzy neural network constructed in the present invention has 6 layers: the first layer is the input layer, which is directly connected to the component x of the main star feature vector to be identified. i The number of input nodes is equal to the dimension of the navigation star feature vector. The second layer is the fuzzification layer. Each input component corresponds to a group of nodes, and each node corresponds to a Gaussian membership function, which is calculated as follows:

[0020]

[0021] Among them, c ij and σ ijRepresenting the center and width of the membership function, respectively; the third layer is the rule calculation layer, whose input is the membership values ​​of different groups in the previous layer; the fourth layer is the normalization layer; and the fifth and sixth layers are both fully connected layers. Unlike common neural networks whose output dimension is equal to the number of navigation stars, the neural network constructed in this invention outputs binary numbers representing the feature subgraphs. Since the equalized navigation star library contains a total of 3569 navigation stars,

[0022] Furthermore, the fuzzy neural network is trained by randomly sorting the feature vectors in the training set generated in step 3, inputting them into the fuzzy neural network in sequence, and using the binary code corresponding to the main star as output for training.

[0023] Furthermore, the star map recognition is performed as follows: first, the star point is extracted from the star map taken by the star sensor, the distance between each star and the center of the star map is calculated, and the observed star with the smallest distance is selected as the main star. Then, the direction vector r of each observed star in the star map in the star sensor coordinate system is calculated. i , expressed as:

[0024]

[0025] Among them, (u i , v i ) is the coordinate of the observed star in the star map, and f is the focal length of the star sensor lens. i With star point S j The cosine of the angular distance θ ij =r i T r j , the relative magnitude of the star point is:

[0026]

[0027] Among them, m max is the magnitude threshold that the star sensor can capture, and g0 is the grayscale value of the center pixel of the star point. Construct the feature vector according to step 2. Finally, input the feature vector into the trained fuzzy neural network for recognition, and obtain the star number of each star in the feature subimage based on the output binary number.

[0028] Compared with the prior art, the present invention has the following characteristics:

[0029] (1) After training, the star map recognition speed of the present invention is fast and the recognition accuracy is high.

[0030] (2) The present invention combines the subgraph isomorphism algorithm and utilizes all the relative position information of the star points in the feature subgraph, so the recognition result has good reliability.

[0031] (3) The present invention has good robustness against star position noise, magnitude noise, star point missing and pseudo stars.

[0032] (4) The algorithm has a low requirement on the number of star points in the star map, which expands the scope of application of the star sensor. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the star map recognition method of the present invention.

[0034] Figure 2 The distribution of navigation stars after Monte Carlo voting screening

[0035] Figure 3 Schematic diagram for constructing the characteristic vector of the observed star.

[0036] Figure 4 This is the structure diagram of the fuzzy neural network. DETAILED DESCRIPTION

[0037] The present invention will be described in further detail below with reference to the accompanying drawings.

[0038] Combine Figure 1 The specific steps of the star map recognition method of the present invention include:

[0039] 1. Establish a navigation star library. The present invention selects the Tycho-2 star catalog as the original star catalog. First, according to the maximum magnitude that can be observed by the star sensor, the stars with a magnitude greater than 6.0, variable stars, and stars with larger magnitudes in binary stars in the star catalog are deleted. Then, 10,000 fields of view are randomly generated according to the Monte Carlo method, and the stars in each field of view are arranged in ascending order of magnitude. The top 5 navigation stars are voted. After the voting is completed, the stars with a final vote of 0 are deleted. Finally, the remaining 3569 navigation stars (such as Figure 2 The navigation star library consists of the star number, magnitude, right ascension and declination information of the selected navigation stars.

[0040] 2. Construct the eigenvector based on the minimum spanning tree. First, traverse the navigation star library, take each navigation star as the main star S0, and calculate the angular distance θ from the neighboring star to the main star. 0i , expressed as:

[0041] θ 0i =arccos(sinδ i sinδ0+cosδ i cosδ0cos(α i -α0))

[0042] Among them, i = 1, 2, ..., n, n is the number of neighboring stars around the main star, α0, α i ,δ0,δ iThey represent the right ascension and declination of the primary star and the neighboring star respectively.

[0043] Then, the angular distance is located at (θ min ,θ FOV ) are sorted by magnitude, θ min is the minimum angular distance required by the star point extraction algorithm, θ FOV is the field of view of the star sensor, select the 6 stars with the smallest magnitude as the auxiliary stars of the main star, and calculate the angular distance between the auxiliary stars. Taking the main star as the root node, the angular distance between the navigation stars θ ij As the weight of the edge in the graph, the minimum spanning tree algorithm is used to generate the characteristic subgraph of the navigation star, such as Figure 3 As shown. According to the generation order of the minimum spanning tree, the auxiliary stars are named S1, S2, ..., S6, and the feature vector pat = [θ1θ2θ3θ4θ5θ6] is generated based on the corresponding star-pair angular distance. This navigation star pattern feature based on the minimum spanning tree is rotationally invariant and scale-invariant, describing the star point distribution characteristics within the neighborhood of the primary star and is relatively insensitive to magnitude errors. Finally, the star numbers and feature vectors of the primary and auxiliary stars are stored in the feature database;

[0044] 3. Construct a training set. By adding uniformly distributed random noise to the original feature vector, the noise intensity is 5% of the maximum component of the feature vector, and 25 training samples are generated for each feature vector.

[0045] 4. Construct and train fuzzy neural network. Figure 4 As shown in the figure, the fuzzy neural network for star image recognition has 6 layers: the first layer is the input layer, which is directly connected to the component x of the feature vector of the main star to be identified. i The number of input nodes is equal to the dimension of the navigation star feature vector. The second layer is the fuzzification layer. Each input component corresponds to a group of nodes, and each node corresponds to a Gaussian membership function, which is calculated as follows:

[0046]

[0047] Among them, c ij and σ ij They represent the center value and width value of the membership function respectively; the third layer is the rule calculation layer, and the input is the membership value of different groups in the previous layer; the fourth layer is the normalization layer; the fifth and sixth layers are both fully connected layers, and the output is the binary number representing the feature subgraph.

[0048] 5. Perform star map recognition. First, extract star points from the star map captured by the star sensor, calculate the distance between each star and the center of the star map, and select the observed star with the smallest distance as the main star. Then, calculate the direction vector r of each observed star in the star map in the star sensor coordinate system. i , expressed as:

[0049]

[0050] Among them, (u i , v i ) is the coordinate of the observed star in the star map, and f is the focal length of the star sensor lens. i With star point S j The cosine of the angular distance θ ij =r i T r j , the relative magnitude of the star point is:

[0051]

[0052] Among them, m max is the magnitude threshold that the star sensor can capture, and g0 is the grayscale value of the center pixel of the star point. Construct the feature vector according to step 2. Finally, input the feature vector into the trained fuzzy neural network for recognition, and obtain the star number of each star in the feature subimage based on the output binary number.

[0053] In summary, compared with traditional methods, the star map recognition based on fuzzy neural network proposed in this invention has the following advantages:

[0054] 1. Using the Monte Carlo method to vote and select navigation stars that are evenly distributed on the celestial sphere, while ensuring the minimum number of identifiable navigation stars in the field of view, dark stars with a low probability of forming a characteristic sub-image are eliminated, effectively reducing the storage capacity of the navigation star library;

[0055] 2. The characteristic pattern corresponding to the star map is generated based on the minimum spanning tree algorithm. The principle is simple and easy to implement. The generated navigation star characteristic pattern is unique, which improves the matching success rate.

[0056] 3. Using fuzzy rules for network learning and adjustment has high recognition accuracy and fast recognition speed, and is robust to star position noise, star missing and pseudo stars.

[0057] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A star map recognition method based on fuzzy neural network, characterized in that: The method comprises the following steps: Step 1: Establish a balanced navigation star library; Step 2: Construct a feature vector based on the minimum spanning tree: Take each star in the navigation star library as the main star, select 6 auxiliary stars in its neighborhood, construct a feature subgraph, and generate a feature vector based on the adjacency matrix of the feature subgraph and store it in the feature database; Step 3: Construct a training set: add different noises to the feature vectors in the feature database, expand the number of feature vectors, and obtain a training set; Step 4: Construct and train the fuzzy neural network: Construct the fuzzy neural network and input the training set generated in step 3 into the fuzzy neural network in random order for training to generate a star map recognition model; Step 5: Perform star map recognition: Extract star points from the star map captured by the star sensor to obtain star point positions. The star point closest to the center of the field of view is taken as the main star, and the feature vector of the navigation star is constructed. The feature vector is input into the trained fuzzy neural network for recognition.

2. The star map recognition method based on fuzzy neural network according to claim 1, characterized in that: The method for establishing a balanced navigation star library comprises the following steps: first, selecting the Tycho-2 star catalog as the original star catalog, deleting stars with a magnitude greater than 6.0, variable stars, and stars with larger magnitudes in binary stars from the catalog based on the maximum magnitude observable by the star sensor; then, randomly generating 10,000 fields of view using the Monte Carlo method, arranging the stars within each field of view in ascending order of magnitude, and voting for the top five navigation stars; after the voting is completed, deleting stars with a final number of votes of 0; and finally, renumbering the remaining navigation stars as star numbers; and the navigation star library comprises the star number, magnitude, right ascension, and declination information of the screened navigation stars.

3. The star map recognition method based on fuzzy neural network according to claim 1, characterized in that: The construction is based on the eigenvector of the minimum spanning tree: First, traverse the navigation star library, take each navigation star as the main star S0 in turn, and calculate the angular distance θ from the neighboring star to the main star 0i , expressed as: i 0i =arccos(sinδ i sinδ0+cosδ i cosδ0cos(α i -a0)) Among them, i=1,2,…,n, where n is the number of neighboring stars around the main star, α0, α i Denote the right ascension of the primary star and the neighboring star, δ0, δ i denote the declination of the primary and neighboring stars respectively; Then, the angular distance θ 0i Located at (θ min ,θ FOV ) are sorted by magnitude, θ min is the minimum angular distance required by the star point extraction algorithm, θ FOV is the field of view of the star sensor, and the six stars with the smallest magnitude are selected as the auxiliary stars of the main star; with the main star as the root node, the angular distance between the navigation stars is θ ij As the weight of the edge in the graph, the minimum spanning tree algorithm is used to generate the characteristic subgraph of the navigation star; according to the generation order of the minimum spanning tree, the auxiliary stars are named S1, S2, ..., S6, and the characteristic vector pat = [θ1θ2θ3θ4θ5θ6] is generated according to the corresponding star-pair angular distance; finally, the star signs and characteristic vectors of the main and auxiliary stars are stored in the feature database.

4. The star map recognition method based on fuzzy neural network according to claim 1, characterized in that: The training set is constructed by adding uniformly distributed random noise to the original feature vector, with the noise intensity being 5% of the maximum component of the feature vector, and generating 100 expanded training samples for each feature vector.

5. The star map recognition method based on fuzzy neural network according to claim 1, characterized in that: The fuzzy neural network is constructed as follows: The fuzzy neural network includes 6 layers: the first layer is the input layer, which is directly connected to the component x of the feature vector of the main star to be identified. i Connected, the number of input nodes is equal to the dimension of the navigation star feature vector; The second layer is the fuzzification layer, where each input component corresponds to a set of nodes, and each node corresponds to a Gaussian membership function μ ij , calculated as follows: Among them, c ij and σ ij They represent the center value and width value of the membership function respectively; the third layer is the rule calculation layer, and the input is the membership value of different groups in the previous layer; the fourth layer is the normalization layer; the fifth and sixth layers are both fully connected layers, and the output is the binary number representing the feature subgraph.

6. The star map recognition method based on fuzzy neural network according to claim 3, characterized in that: The star map recognition is performed: First, the star points are extracted from the star map captured by the star sensor, the distance between each star and the center of the star map is calculated, and the observed star with the smallest distance is selected as the main star; Then, calculate the direction vector r of each observed star in the star map in the star sensor coordinate system i , expressed as: Among them, (u i ,v i ) is the coordinate of the observed star in the star map, f is the focal length of the star sensor lens; star point S i With star point S j The angular distance θ ij =arccos(r i T r j ), the relative magnitude of the star point is: Among them, m max is the magnitude threshold that can be captured by the star sensor, g0 is the grayscale value of the center pixel of the star point; construct the feature vector based on the minimum spanning tree; Finally, the feature vector is input into the trained fuzzy neural network for recognition, and the star number of each star in the feature sub-graph is obtained according to the output binary number.

Images