A Novel Star Chart Recognition Method Based on Lightweight Neural Network

Abstract

The present invention relates to a new star map recognition method based on a lightweight neural network, belonging to the technical field of star map recognition. First, a navigation star catalog is constructed, and the companion stars of all navigation stars in the navigation star catalog are extracted. For the companion star distribution around each navigation star, different levels of noise are added, and the equal-frequency bin radial features of the navigation stars are extracted to construct a training set. A lightweight neural network is built and trained using the training set to obtain a feature classification model. In the process of actually recognizing a star map, first, navigation stars are selected from the recognition area, and the equal-frequency bin radial features of the navigation stars are constructed. Then, the triangle loop matching method is used to verify the classification result of the network. Three stars are cyclically extracted for recognition and verification until the image angular distances and star catalog angular distances of three stars all pass the matching, and it can be considered that the star map has been successfully recognized.

Description

Technical Field

[0001] The present invention relates to the technical field of star map recognition, and particularly to a method for star map recognition, which is a core process during the operation of a star sensor, especially a novel star map recognition method based on a lightweight neural network. Background Art

[0002] A star sensor is the most widely used attitude sensor on current spacecraft, and is an important prerequisite for the spacecraft to fly stably, observe stably, and complete its own tasks. The star sensor uses celestial bodies with relatively stable positions outside the solar system as references for attitude measurement. Its working principle is to use its optical system to image stars, and through preprocessing and centroid extraction of the stars, the image coordinates of the stars are calculated. Then, a star map recognition algorithm is used to match the captured stars with the star points in the navigation star catalog. Based on the image coordinates of the captured stars and the celestial sphere coordinates of the stars in the navigation star catalog, combined with the internal and external parameters of the star sensor, the attitude of the star sensor coordinate system in the celestial sphere coordinate system can be calculated, and then the attitude of the spacecraft can be obtained. The method of star map recognition is one of the most important processes during the operation of a star sensor, and the efficiency and robustness of star map recognition directly affect the accuracy and efficiency of the star sensor in obtaining the attitude.

[0003] Currently, the main star map recognition methods include the triangle algorithm, pyramid algorithm, grid algorithm, polar coordinate transformation algorithm, genetic algorithm, singular value algorithm, neural network algorithm, etc. Traditional algorithms either require a huge recognition database or rely too much on accurate angular distance measurement information. Some algorithms also need to select reference stars. Therefore, the efficiency and the ability to resist noise interference of these algorithms need to be improved. The star map recognition algorithm based on a neural network has the advantages of fast recognition speed and high robustness, and has become a new direction in current star map recognition research. However, these neural network methods generally face the disadvantages of insufficiently robust features and overly complex networks. For star sensors with relatively limited computing resources, there is currently no particularly effective neural network star map recognition solution. Summary of the Invention

[0004] The object of the present invention is to provide a novel star map recognition method based on a lightweight neural network, which solves the problems existing in the prior art such as the inability to effectively resist noise interference, a large recognition database, and easy misrecognition. The present invention has the characteristics of being robust to noise, having a small network scale, high classification efficiency, less resource consumption, and the potential for on-orbit deployment. The present invention achieves the effect of being robust to noise by proposing a robust navigation star feature - equal-frequency binning radial feature as the input data for neural network classification; by constructing a lightweight neural network, the operation efficiency of the algorithm is ensured; finally, the triangle loop matching method is used to make full use of the high-efficiency matching of the neural network and the stability of triangle matching, which can effectively improve the recognition rate and reduce the misrecognition rate. Compared with the traditional star map recognition algorithm, the present invention can make the process of star map recognition more accurate and efficient.

[0005] The above object of the present invention is achieved by the following technical solutions:

[0006] A novel star map recognition method based on a lightweight neural network, which constructs a navigation star table and extracts the neighboring stars of all navigation stars in the navigation star table; for the companion star distribution around each navigation star, different levels of noise are added, and the equal-frequency binning radial feature of the navigation star is extracted to construct a training set; a lightweight neural network is built and trained using the training set to obtain a feature classification model; in the process of actually recognizing a star map, first select navigation stars from the recognition area and construct the equal-frequency binning radial feature of the navigation stars; then use the triangle loop matching method to verify the classification result of the network, and loop to extract three stars for recognition and verification until the image angular distance and the star table angular distance of three stars both pass the matching, then it can be considered that the star map has been successfully recognized. The specific steps are as follows:

[0007] Step 1: According to the limiting magnitude detection ability M of the star sensor, intercept the magnitude from the original star table to extract the navigation star table, and the intercepted magnitude value needs to be slightly greater than the limiting magnitude detection ability of the star sensor;

[0008] Step 2: Extract the equal-frequency binning radial feature EFB-RF of the navigation stars in the navigation star table. First, determine the size of the feature generation domain FGD. FGD is a circular area centered on the navigation star with a radius of r FGD The companion stars within this circular area participate in the generation of the EFB-RF of the navigation star; in order to enable enough stars to generate a complete EFB-RF when recognizing a star map and ensure that there are enough neighboring stars in the FGD to generate unique features, r FGD is set to r FOV which is the radius of the largest inscribed circle of the star sensor's field of view;

[0009] After equal-frequency binning, the FGD is divided into k annular intervals with equal areas; when the FGD ranges from S0 to Sk-1 After being divided into k intervals, the area of the intervals will satisfy formula (1.1);

[0010]

[0011] where Area(S i ) is the area of the i-th interval, and Area(FGD) is the area of the entire FGD; from formula (1.1), the outer circle radius of the i-th interval can be obtained For a companion star with an angular distance d from the navigation star GN , for the companion star, the angular distance d GN has the following relationship with the interval number j where the companion star is located:

[0012]

[0013] From formula (1.2), it can be obtained

[0014]

[0015] where Area(d GN ) is the area of a circle with the navigation star as the center and d GN as the radius; calculate all the companion stars within the FGD of the navigation star, calculate the angular distance, obtain the interval where each companion star is located; count the number of companion stars in each interval to obtain k integer values, forming a one-dimensional vector;

[0016] Step 3: Extract the companion stars within the FGD of all star points in the navigation star table; extract the equal-frequency binning radial features of the companion stars, and add various levels of noise, including false stars and position noise, to form a training set;

[0017] Step 4: Construct a lightweight neural network for feature classification, and use the prepared training set to train the network to obtain a feature classification model;

[0018] Step 5: During the training process, the neural network will learn how to correctly identify EFB-RF from samples with different noise levels and output the correct classification results; the loss function used for training is the cross-entropy loss function, and the parameters in the network model are optimized using stochastic gradient descent to minimize the input cross-entropy loss;

[0019] Step 6: Use triangular cyclic verification to verify the recognition results of the neural network.

[0020] The lightweight neural network for feature classification described in Step 4 contains two hidden layers, and there is a batch normalization layer (BN) behind each hidden layer to accelerate convergence and improve the generalization ability of the network. At the same time, a Relu activation function is added to each hidden layer to enable the network to have the ability of non-linear mapping. Finally, a Softmax activation function is added behind the output layer to normalize the output of the network into probabilities. The output of the network is a probability distribution P 0-N , P i represents the probability that the network identifies the sample as the star with the serial number i in the navigation star catalog, and

[0021] The verification of the recognition result of the neural network using triangular loop verification described in Step 6 is as follows:

[0022] During the process of recognizing a star chart, with the principal point of the image as the center, the radius of the recognition area is defined as At this time, it can be ensured that the complete EFB-RF of the stars within the area can be extracted. Select the stars within the area for recognition; sort the stars within the recognition area by magnitude, and sequentially extract the EFB-RF of the first three unrecognized stars, and use the trained feature classification model for recognition; the stars used for triangular verification must first use the classification probability threshold p th to eliminate part of the recognition results; for the classification probability p output by the network, only when p ≥ p th can it be used for triangular verification; then calculate the angular distance between these three stars in the star chart

[0023] For two stars x(X x , Y x ) and y(X y , Y y ), their angular distance is expressed by the following formula:

[0024]

[0025]

[0026]

[0027] According to the recognized star ID x , ID y and ID z , extract the celestial coordinates (α x , β x ), (α y , β y ) and (α z , β z), and then calculate the star catalog angular distance according to formula (1.7).

[0028]

[0029] Compare the corresponding star chart angular distance with the star catalog angular distance to check whether the star catalog angular distance and the star chart angular distance match; the matching rule is that the difference between the star chart angular distance and the true angular distance of two stars is less than Δd, that is, Only when all three corresponding angular distances match, is this star chart considered successfully recognized; otherwise, star points still need to be selected from the star chart for recognition verification until the star chart is successfully recognized; when all the stars in the recognition area of this star chart have been recognized once and the verification has not been successful at this time, it means that this star chart recognition has failed.

[0030] The beneficial effects of the present invention are as follows:

[0031] The present invention uses the currently popular neural network to design a star chart recognition method, implicitly stores the features of navigation stars, and calculates the probability that the features most likely belong to a navigation star through forward deduction, which can get rid of the database matching process in traditional star chart recognition and can achieve efficient star chart recognition.

[0032] The equal-frequency binning radial feature specially designed for the neural network, which is constructed using the companion star number distribution, can resist a large number of false star interferences. At the same time, the position offset of the star point can be limited within the interval, with little impact on the feature, making the algorithm relatively robust to position noise.

[0033] The triangle loop matching mechanism makes full use of the fast matching speed of the neural network and combines the principle of stable triangle matching to verify the results output by the feature classification network, which can greatly improve the recognition rate of star chart recognition and reduce the misrecognition rate.

[0034] Compared with the traditional star chart recognition method, this method has the advantages of fast recognition speed, high recognition rate, robustness to noise, and the ability to adapt to large-field-of-view star sensors, providing a stable and reliable core algorithm for the stable operation of satellites. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and the schematic examples and descriptions of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0036] Figure 1 Schematic diagram of the extraction method of EFB-RF of the present invention;

[0037] Figure 2 Schematic diagram of the noise of the present invention;

[0038] Figure 3Structural diagram of the lightweight neural network of the present invention;

[0039] Figure 4 Flowchart of the triangular loop matching verification of the present invention;

[0040] Figure 5 Schematic diagram of the operation of the star sensor of the present invention;

[0041] Figure 6 Flowchart of the neural network star chart recognition algorithm of the present invention. Detailed implementation manners

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0043] Refer to Figures 1 to 6 As shown, the novel star chart recognition method based on a lightweight neural network of the present invention constructs a navigation star table and extracts the companion stars of all navigation stars in the navigation star table. For each navigation star, different levels of noise are added, and the equal-frequency binning radial features of the navigation star are extracted to construct a training set. A lightweight neural network is built and trained using the training set to obtain a feature classification model. Then, the triangular loop matching method is used to verify the classification result of the network. Three stars are cyclically extracted for recognition and verification until the image angular distance and the star table angular distance of three stars both pass the matching, and then it can be considered that this star chart has been successfully recognized. The specific steps are as follows:

[0044] Step 1: According to the limiting magnitude detection ability M of the star sensor, the navigation star table is intercepted from the original star table by extracting the magnitude.

[0045] Among them, regarding the construction of the navigation star table GCat, the selected original star table is the Tycho-2 star table. The intercepted magnitude value should be slightly greater than the limiting magnitude detection ability of the star sensor.

[0046] In a specific embodiment, when the limiting magnitude detection ability of the star sensor is M and the intercepted magnitude value is g, then the magnitude m of any star s in the intercepted navigation star table s should be less than M + C, that is, for there is m s < M + C, where C is a constant greater than zero, and at this time g = M + C.

[0047] Step 2: Extract the Equal-Frequency Binning Radial Feature (EFB-RF) of the navigation stars in the navigation star catalog.

[0048] The radial feature is a feature that converts the two-dimensional geometric distribution of the companion stars around the navigation star into a one-dimensional radial distribution. This feature has rotational invariance and is one of the more ideal star map recognition features. Equal-frequency binning is a discretization method aimed at keeping the statistical quantity of data within the divided intervals consistent. EFB-RF is based on the traditional radial feature. By non-uniformly dividing the intervals, the probability that star points fall into different intervals is made the same, thus constituting EFB-RF. The equal-frequency division of the radial intervals can be achieved by using the equal-area division method.

[0049] To extract EFB-RF, it is first necessary to determine the size of the Feature Generation Domain (FGD). The FGD is a circular area centered on the navigation star, and the companion stars within this area will participate in the generation of the EFB-RF of the navigation star. The FGD is a circular area centered on the navigation star with a radius of r FGD The companion stars within this circular area participate in the generation of the EFB-RF of the navigation star;. To enable enough stars to generate a complete EFB-RF when recognizing a star map, while also ensuring that there are enough companion stars within the FGD to generate unique features, r FGD needs to be set to r FOV which is the radius of the largest inscribed circle of the star sensor's field of view.

[0050] Figure 1 shows a schematic diagram of extracting EFB-RF. The FGD after equal-frequency binning is divided into k annular intervals with equal areas. When the FGD goes from S0 to S k-1 is divided into k intervals, the area of the intervals will satisfy formula (1.1).

[0051]

[0052] where Area(S i ) is the area of the i-th interval, and Area(FGD) is the area of the FGD. From (1.1), the outer circle radius of the i-th interval can be obtained For a star with an angular distance d GN from the navigation star, there is the following relationship between the angular distance d GN and the interval number j where the companion star is located:

[0053]

[0054] From (1.2), it can be obtained

[0055]

[0056] where Area(d GN ) is the area of a circle with the navigation star as the center and d GN as the radius. Calculate all neighboring stars within the FGD of the navigation star, calculate the angular distance, and obtain the interval where each companion star is located. Count the number of companion stars in each interval to obtain k integer values, forming a one-dimensional vector.

[0057] Step 3: Extract the companion stars within the FGD of all star points in the navigation star table. Extract the equal-frequency binning radial features of the companion stars, and add various levels of noise (false stars, position noise) to form a training set. Figure 3 Shows the schematic diagrams of these two types of noise.

[0058] The specific noise generation methods are as follows:

[0059] Position noise: The distortion of the optical system and the error in star point centroid extraction will cause the image coordinates of the star points to shift, which is the position noise. The specific method for generating position noise training samples is to add random offsets to the image coordinates of the companion stars of the navigation star. The random offsets added to the X coordinate and the Y coordinate both follow a normal distribution with a mean of 0 and a standard deviation of σ, that is, (X + p x , Y + p y ), p x ~N(0, σ 2 ), p y ~N(0, σ 2 ).

[0060] For each navigation star, the standard deviation of the added position noise ranges from 0.1 pixel to 1.6 pixels, with a total of 16 levels, and each level differs by 0.1 pixel. There are ten different samples in each level, that is, for one navigation star, there are 160 position noise samples in total.

[0061] False stars: The sources of false stars include exoplanets, asteroids, high-energy particles, background white noise, or spacecraft, etc. Their imaging characteristics are similar to those of star points and cannot be removed through the star map preprocessing stage. False stars have a greater impact on traditional star map recognition algorithms. To reduce the impact of false stars, when constructing the training set, false stars with random positions are added to the FGD of the navigation star. Considering that in reality, different types of noise do not appear alone, the addition of false stars is carried out on the basis of the position noise samples. The number of false stars added is determined by the percentage of companion stars. For each sample, 0% to 30% of false stars are randomly added. Each false star first randomly generates a direction angle θ~U(0°, 360°), and then generates a random distance r~U(0, r FGD ).

[0062] Each navigation star has 160 noise samples of different levels, so there are a total of 160 * N samples. These samples are all with accurate star IDs and constitute the training set.

[0063] Step 4: Construct a lightweight neural network for feature classification, and use the prepared training set to train the network to obtain the feature classification network.

[0064] The structure of the lightweight neural network is as Figure 3 shown. The lightweight neural network for feature classification contains two hidden layers, and there is a batch normalization layer (Batch Normalization, BN) behind each hidden layer to accelerate convergence and improve the generalization ability of the network. At the same time, a Relu activation function is added to each hidden layer to enable the network to have the ability of non-linear mapping. Finally, a Softmax activation function is added behind the output layer to normalize the output of the network to probabilities.

[0065] The parameters of each layer are shown in Table 1:

[0066] Table 1 Parameters of Each Layer of the Lightweight Classification Network

[0067]

[0068] The output of the network is a probability distribution P 0-N , P i represents the probability that the network recognizes the sample as the star with the serial number i in the navigation star table, and

[0069] Step 5: In this step, the network will learn how to correctly identify EFB-RF from samples with different noise levels and output the correct classification results. The loss function used for training is the cross-entropy loss function, and the parameters in the network model are optimized using stochastic gradient descent to minimize the input cross-entropy loss. The hyperparameters of network training are shown in Table 2:

[0070] Table 2 Training Parameters of the Lightweight Classification Network

[0071]

[0072] The neural network star chart recognition algorithm belongs to a type of pattern recognition algorithm. It only considers the recognition of a single navigation star, and there will also be cases of misrecognition. When using a neural network to recognize a star, a probability distribution will be obtained, and the category with the highest probability is used as the recognition result. However, when the highest recognition probability is very low, it means that the network cannot determine which star it is, and the recognition result at this time is not reliable. Moreover, using the recognition result of a single navigation star as the recognition result of a frame of star chart is very irresponsible and will not have excellent performance. Therefore, when using a neural network for star chart recognition, it is necessary to add a verification step to screen the recognition results and eliminate misrecognition.

[0073] Step 6: Use triangular loop verification to verify the recognition result of the neural network. This step combines traditional triangular matching and repeatedly uses the high-efficiency navigation star matching efficiency of the neural network, which can quickly and accurately verify the recognition result, improve the recognition rate, and reduce the misrecognition rate.

[0074] Figure 4 Schematic diagram of the triangular loop verification process. During the process of recognizing a star chart, with the principal point of the image as the center, the radius of the recognition area is delimited as At this time, it can ensure that the complete EFB-RF of the stars within the area can be extracted. Select the stars within the area for recognition; sort the stars within the recognition area by magnitude, and sequentially extract the EFB-RF of the first three unrecognized stars, and use the trained feature classification model for recognition. The stars used for triangular verification must first use the classification probability threshold p th to eliminate a part of the recognition results; for the classification probability p output by the network, only when p ≥ p th can it be used for triangular verification; then calculate the star chart angular distance between these three stars in the star chart

[0075] For two stars x(X x , Y x ) and y(X y , Y y ), their star chart angular distance can be expressed by the following formula:

[0076]

[0077]

[0078]

[0079] The stars used for triangular verification must first use the classification probability threshold p th to eliminate a part of the recognition results. For the classification probability p output by the network, only when p ≥ p th can it be used for triangular verification.

[0080] Then, according to the identified asterisk IDs x , ID y and ID z , extract the celestial coordinates (α x , β x ) of the stars corresponding to these three asterisks from the navigation star catalog, (α y , β y ) and (α z , β z ), and then calculate the catalog angular distance according to (1.7)

[0081]

[0082] Compare the corresponding star chart angular distance with the catalog angular distance to check whether the catalog angular distance and the star chart angular distance match; the matching rule is that the difference between the star chart angular distance and the true angular distance of two stars is less than Δd, that is, Only when all three corresponding angular distances match is the star chart considered successfully recognized; otherwise, star points still need to be selected from the star chart for recognition verification until the star chart is successfully recognized. When all the stars in the recognition area of this star chart have been recognized once and the verification has not been successful at this time, it means that this star chart recognition has failed.

[0083] Embodiment:

[0084] When implementing this method specifically, it is necessary to solve the selection of some variables, the constant C, the number of interval divisions k, the classification probability threshold p th and the angular distance difference Δd. And r FGD and the size of the recognition area in the star chart are related to the optical parameters of the actually used star sensor.

[0085] For example, the optical parameters of a star sensor are shown in Table 3:

[0086] Table 3 Optical parameters of a star sensor

[0087]

[0088] At this time, it can be set that the radius r of the recognition area iden is also 10°. The constant C can be selected as 0.5, and at this time, the star magnitude g of the intercepted star catalog is 6.5Mv.

[0089] The selection of the classification probability threshold p th needs to be appropriate. Too small p th will reduce the effect of the rejection result and increase the number of triangle verifications, while too large p thMany originally correctly recognized results will be discarded, which will greatly reduce the number of stars available for triangle verification in a complex noise environment, increase the recognition time, and even reduce the recognition success rate. Take p th = 0.5 is a reasonable choice for most cases.

[0090] Considering the existence of position noise, the selection of the angular distance difference Δd cannot be too small, otherwise it will be difficult to complete the angular distance matching. According to the optical parameters of the star sensor during specific implementation, the image distance between two pixels is converted into an angular distance and set as the value of the angular distance difference. The angular distance difference in this example

[0091] The determination of the number k of interval divisions needs to be determined through experiments according to the optical parameters of the star sensor implemented currently. The size of the number k of interval divisions is related to the input layer of the network and will affect the complexity of the network model. Therefore, the smallest division number is selected as the value of k for the complete method while ensuring the recognition rate. To determine the size of k, it is necessary to determine the approximate range of k in advance (10 to 50), select the value of k within this range, construct the training set and neural network according to the above steps and train the model, and compare the recognition results of the model for lower-level noise samples (the proportion of false stars is 5%, and the position noise is 0.1 pixel) under different values of k. Among a series of k values with better recognition results, select the smallest k as the value of k for the complete method. In this embodiment, the value of k is determined to be 26 through experiments.

[0092] The above are only the preferred examples of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made to the present invention shall be included within the protection scope of the present invention.

Claims

1. A novel star chart recognition method based on a lightweight neural network, characterized in that: Construct a navigation star catalog, and extract the neighboring stars of all navigation stars in the navigation star catalog; for the companion star distribution around each navigation star, add noises of different levels, and extract the equal-frequency binning radial features of the navigation stars to construct a training set; build a lightweight neural network and train it using the training set to obtain a feature classification model; During the actual process of identifying a star chart, first select navigation stars from the identification area and construct the equal-frequency binning radial features of the navigation stars; Then use the method of triangular loop matching to verify the classification results of the network. Loop to extract three stars for identification and verification until the image angular distances and catalog angular distances of three stars both pass the matching, then it can be considered that the star chart has been successfully identified; including the following steps: Step 1: According to the limiting magnitude detection ability M of the star sensor, intercept the magnitude from the original star catalog to extract the navigation star catalog, and the intercepted magnitude value needs to be slightly greater than the limiting magnitude detection ability of the star sensor; Step 2: Extract the equal-frequency binning radial feature EFB-RF of the navigation stars in the navigation star catalog. First, determine the size of the feature generation domain FGD. FGD is a circular area centered on the navigation star with a radius of r FGD such that the companion stars within this circular area participate in the generation of the EFB-RF of the navigation star. To ensure that when identifying a star map, enough stars can generate a complete EFB-RF and at the same time ensure that there are enough neighboring stars within the FGD to generate unique features, set r FGD to r FOV which is the radius of the largest inscribed circle of the star sensor's field of view; After equal-frequency binning, the FGD is divided into k annular intervals with equal areas; when the FGD ranges from S0 to S k-1 After being divided into k intervals, the area of the intervals will satisfy formula (1.1); where Area(S i ) is the area of the i-th interval, and Area(FGD) is the area of the entire FGD; The outer circle radius of the $i$-th interval can be obtained from Equation (1.1). For a companion star with an angular distance $d$ from the navigation star GN , the angular distance $d$ GN has the following relationship with the interval number $j$ where the companion star is located: It can be obtained from Equation (1.2) where Area(d GN ) is the area of a circle with the navigation star as the center and d GN as the radius; calculate all the companion stars within the FGD of the navigation star, calculate the angular distance, and obtain the interval where each companion star is located; count the number of companion stars in each interval to obtain k integer values, forming a one-dimensional vector; Step 3: Extract the companion stars within the FGD of all star points in the navigation star catalog; extract the equal-frequency binning radial features of the companion stars and add noises of various levels, including false stars and position noises, to form a training set; Step 4: Construct a lightweight neural network for feature classification, and train the network using the prepared training set to obtain a feature classification model; Step 5: The neural network will learn how to correctly identify EFB-RF from samples with different noise levels during training and output the correct classification results; the loss function used for training is the cross-entropy loss function, and the parameters in the network model are optimized using stochastic gradient descent to minimize the input cross-entropy loss; Step 6: Use triangular loop verification to verify the recognition results of the neural network.

2. The novel star chart recognition method based on a lightweight neural network according to claim 1, wherein: The lightweight neural network for feature classification described in Step 4 contains two hidden layers, and there is a batch normalization layer BN behind each hidden layer to accelerate convergence and improve the generalization ability of the network. At the same time, a Relu activation function is added to each hidden layer to enable the network to have the ability of non-linear mapping. Finally, a Softmax activation function is added behind the output layer to normalize the output of the network into probabilities. The output of the network is a probability distribution P 0-N , P i represents the probability that the network recognizes the sample as the star with the serial number i in the navigation star catalog, and 3. The novel star chart recognition method based on a lightweight neural network according to claim 1, characterized in that: The use of triangular loop verification to verify the recognition results of the neural network described in Step 6 is specifically: In the process of identifying a star chart, with the principal point of the image as the center, the radius of the identification area is At this time, it can be ensured that the stars within the area can extract complete EFB-RF. Select the stars within the area for identification; sort the stars within the identification area by magnitude, and sequentially extract the EFB-RF of the first three unrecognized stars, and use the trained feature classification model for identification; the stars used for triangle verification must first use the classification probability threshold p th to eliminate part of the recognition results; for the classification probability p output by the network, only when p ≥ p th can it be used for triangle verification; then calculate the star chart angular distance between these three stars in the star chart For two stars x(X x ,Y x ) and y(X y ,Y y ), their angular distance is represented by the following formula: According to the recognized asterisk ID x , ID y and ID z , extract the celestial coordinates (α x , β x ) of the stars corresponding to these three asterisks from the navigation star catalog, and then calculate the catalog angular distance according to formula (1.7) y , β y ), (α z , β z ), and (α Compare the corresponding angular distances on the star chart and in the star catalog to check whether the angular distances in the star catalog and on the star chart match; the matching rule is that the difference between the angular distance on the star chart and the true angular distance of two stars is less than Δd, that is, Only when all three corresponding angular distances match is the star chart considered successfully recognized; otherwise, star points still need to be selected from the star chart for recognition verification until the star chart is successfully recognized; when all the stars in the recognition area of this star chart have been recognized once and the verification has not been successful yet, it means that the recognition of this star chart has failed.

Images