A method for extracting the centroid of star points based on neural network

Abstract

The present invention relates to the technical field of space situation awareness and high-precision applications of star sensors, and particularly relates to a star point centroid extraction method based on a neural network, which includes obtaining apparent magnitude data and generating gray-scale data of star point imaging based on a Gaussian distribution for the neighborhood window of sub-pixel units obtained by subdividing each central pixel unit; using the gray-scale data of star point imaging as training data and the relative sub-pixel unit position of the star point centroid within the central pixel unit as labels to train a multi-layer perceptron neural network model; and using the trained multi-layer perceptron neural network model to extract the star point centroid. The present invention uses a multi-layer perceptron neural network to achieve precise positioning of the star point centroid on the imaging plane through processes such as generating training data, model training, and star point centroid processing, overcoming the disadvantages of poor star point positioning accuracy and large computational amount.

Description

Technical Field

[0001] The present invention relates to the technical field of space situation awareness and high-precision application of star sensors, and particularly relates to a star point centroid extraction method based on a neural network. Background Art

[0002] Star map processing is a key link in the orbit determination and cataloging of space targets. Among them, star point centroid extraction is one of the key technologies in the whole star map processing, which determines the measurement accuracy of star sensors and the monitoring of space-based space targets. Since the imaging resolution of current optoelectronic devices is difficult to achieve the positioning accuracy required for astronomical navigation, it is necessary to obtain sub-pixel centroids from the algorithm level.

[0003] Currently, the commonly used centroid extraction algorithms include the weighted centroid method, the squared weighted centroid method, and the method based on Gaussian surface fitting. Among them, the weighted centroid method and the squared weighted method calculate the weighted average of the centroid with the star point gray value and the gray square value as weights respectively. The gray distribution near the star point center is non-uniform and symmetric, resulting in a large target error. The Gaussian surface fitting method has a large amount of calculation, and when the star point distribution deviates from the ideal shape, the resulting error is also large. Summary of the Invention

[0004] In view of the above deficiencies in the prior art, the present invention provides a star point centroid extraction method based on a neural network. Through a multi-layer perceptron neural network, through processes such as generating training data, model training, and star point centroid processing, the precise positioning of the star point centroid on the imaging plane is realized, overcoming the disadvantages of poor star point positioning accuracy and large amount of calculation.

[0005] In order to achieve the above invention purpose, the technical solution adopted by the present invention is as follows:

[0006] A star point centroid extraction method based on a neural network, comprising the following steps:

[0007] Obtain the apparent magnitude data, and generate the gray data of star point imaging based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit;

[0008] Use the gray data of star point imaging as training data, and use the relative sub-pixel unit position of the star point centroid within the central pixel unit as a label to train a multi-layer perceptron neural network model;

[0009] Use the trained multi-layer perceptron neural network model to extract the star point centroid.

[0010] Optionally, when considering the influence of the movement of the space target, generating the gray data of star point imaging based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit includes:

[0011] Calculate the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data;

[0012] The central pixel unit is subdivided into 9 sub-pixel units, and the centers of each sub-pixel unit are sliced along the horizontal, vertical, and diagonal directions. The sub-pixel units in all regions are traversed, and the center of the sub-pixel unit is used as the centroid of the star point.

[0013] The Monte Carlo sampling method is used to set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution, and the exposure time.

[0014] Traverse the gray window corresponding to the star point centroid, and calculate the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor.

[0015] According to the fact that the number of photons reaching the pixel unit follows a Poisson distribution, the Monte Carlo sampling method is used to set the random value of the number of photons.

[0016] Calculate the gray value of the pixel unit according to the number of photons reaching the pixel unit.

[0017] Optionally, when the influence of the spatial target motion is not considered, the gray data of star point imaging is generated based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit, including:

[0018] Calculate the average arrival rate of photons received by the V-band optical sensor according to the visual magnitude data.

[0019] The central pixel unit is subdivided into 9 sub-pixel units, and the centers of each sub-pixel unit are sliced along the horizontal, vertical, and diagonal directions. The sub-pixel units in the first octant and the second octant regions are traversed, and the center of the sub-pixel unit is used as the centroid of the star point.

[0020] The Monte Carlo sampling method is used to set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution, and the exposure time.

[0021] Traverse the gray window corresponding to the star point centroid, and calculate the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor.

[0022] According to the fact that the number of photons reaching the pixel unit follows a Poisson distribution, the Monte Carlo sampling method is used to set the random value of the number of photons.

[0023] Calculate the gray value of the pixel unit according to the number of photons reaching the pixel unit.

[0024] Optionally, when the influence of the spatial target motion is not considered and the Gaussian distribution point spread function is isotropic, the gray data of star point imaging is generated based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit, including:

[0025] Calculate the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data;

[0026] Divide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical, and diagonal directions at the centers of each sub-pixel unit, traverse the sub-pixel units in the first octant region, and use the center of the sub-pixel unit as the centroid of the star point;

[0027] Use the Monte Carlo sampling method to set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution, and the exposure time;

[0028] Traverse the gray window corresponding to the centroid of the star point, and calculate the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor;

[0029] According to the fact that the number of photons arriving at the pixel unit satisfies the Poisson distribution, use the Monte Carlo sampling method to set the random value of the number of photons;

[0030] Calculate the gray value of the pixel unit according to the number of photons arriving at the pixel unit.

[0031] Optionally, the calculation formula for calculating the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data is:

[0032]

[0033] where, is the average arrival rate of photons received by the optical sensor, is the average spectral luminous flux density of V-band visible light, is the camera aperture area, is the average optical efficiency of the sensor in the V band, is the average correction coefficient considering the atmospheric influence, is the full width at half maximum of V-band visible light.

[0034] Optionally, the calculation formula for calculating the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor is:

[0035]

[0036] where, is the average arrival rate of photons received by the pixel units in the gray window, is the average arrival rate of photons received by the optical sensor, is the pixel unit in the gray window, is the relative coordinate of the pixel unit in the gray window, is the point spread function of the star point imaging, is the coordinate of the pixel unit.

[0037] Optionally, the calculation formula for calculating the gray value of the pixel unit according to the number of photons reaching the pixel unit is:

[0038]

[0039] where is the gray value of the pixel unit, returns the largest integer not exceeding, is the number of photons reaching the pixel unit, is the quantum efficiency of the sensor, is the full well capacity of the sensor, is the number of bits of the sensor.

[0040] Optionally, using the trained multi-layer perceptron neural network model for star centroid extraction includes:

[0041] Converting the gray data of the star image into a one-dimensional array AL;

[0042] Inputting the one-dimensional array AL into the trained multi-layer perceptron neural network model considering the motion of the space target, and outputting the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit , where is the multi-layer perceptron neural network model considering the motion of the space target;

[0043] Calculating the star centroid position coordinates according to the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit , specifically:

[0044]

[0045]

[0046] where is the coordinate of the central pixel unit of the star image.

[0047] Optionally, using the trained multi-layer perceptron neural network model for star centroid extraction includes:

[0048] Roughly estimating the relative sub-pixel unit position coordinates of the star centroid by using the square weighting method for the gray data of the star image ;

[0049] Judging the region to which the relative sub-pixel unit position coordinates of the star centroid belong, and defining the rotation transformation matrix ;

[0050] If , if it is located in the first octant and the second octant regions, then define ;

[0051] If , if it is located in the third octant and the fourth octant regions, then define ;

[0052] If , if it is located in the fifth octant and the sixth octant regions, then define ;

[0053] If , if it is located in the seventh octant and the eighth octant regions, then define ;

[0054] Use the rotation transformation matrix to transform the gray-scale data of the star point imaging, and obtain the transformed gray-scale data , is the relative coordinate of the pixel unit in the gray-scale window;

[0055] Convert the transformed gray-scale data into a one-dimensional array ;

[0056] Input the one-dimensional array into the trained multi-layer perceptron neural network model that does not consider the spatial target movement, and output the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit , where is the multi-layer perceptron neural network model that does not consider the spatial target movement;

[0057] Use the rotation transformation matrix to transform the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit , and obtain the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit ;

[0058] According to the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit calculate the star point centroid position coordinates , specifically:

[0059]

[0060]

[0061] Among them, is the central pixel unit coordinate of the star point imaging.

[0062] Optionally, using the trained multi-layer perceptron neural network model for star centroid extraction includes:

[0063] Coarsely estimate the relative sub-pixel unit position coordinates of the star centroid for the grayscale data of the star image using the square weighting method ;

[0064] Judge the region to which the relative sub-pixel unit position coordinates of the star centroid belong, and define the rotation transformation matrix ;

[0065] If , and it is located in the first octant region, then define ;

[0066] If , and it is located in the second octant region, then define ;

[0067] If , and it is located in the third octant region, then define ;

[0068] If , and it is located in the fourth octant region, then define ;

[0069] If , and it is located in the fifth octant region, then define ;

[0070] If , and it is located in the sixth octant region, then define ;

[0071] If , and it is located in the seventh octant region, then define ;

[0072] If , and it is located in the eighth octant region, then define ;

[0073] Use the rotation transformation matrix to transform the grayscale data of the star image to obtain the transformed grayscale data , is the relative coordinate of the pixel unit in the grayscale window;

[0074] Convert the transformed grayscale data into a one-dimensional array ;

[0075] The one-dimensional array Input the trained multi-layer perceptron neural network model that does not consider the influence of the spatial target motion and has an isotropic Gaussian distribution point spread function, and output the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit. , where is a multi-layer perceptron neural network model that does not consider the influence of the spatial target motion and has an isotropic Gaussian distribution point spread function;

[0076] Use the rotation transformation matrix to transform the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit, and obtain the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit;

[0077] Calculate the star centroid position coordinates according to the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit, specifically:

[0078]

[0079]

[0080] where is the central pixel unit coordinate of the star imaging.

[0081] The present invention has the following beneficial effects:

[0082] The present invention constructs the extraction of star points by using a multi-layer neuron network method. The training data generation method used has universality, making the scheme universal and applicable to various star map resolutions and noise distribution situations, and has high accuracy; the trained model has no complex numerical calculations, high efficiency, and can realize real-time star centroid extraction. Description of the Drawings

[0083] Figure 1 is a schematic flow chart of a star centroid extraction method based on a neural network;

[0084] Figure 2 is a schematic diagram of the symmetric partitions of the sub-pixel unit in the horizontal, vertical, and diagonal directions;

[0085] Figure 3 is a schematic diagram of the gray level window corresponding to the sub-pixel unit;

[0086] Figure 4 is a schematic diagram of the sub-pixel unit;

[0087] Figure 5 is a schematic diagram of the imaging when the spatial target moves along the θ direction;

[0088] Figure 6 is the centroid of the star point corresponding Schematic diagram of the grayscale window;

[0089] Figure 7 is the training schematic diagram of the multi-layer perceptron neural network for the grayscale window data. Specific implementation manner

[0090] The following describes the specific implementation manner of the present invention so that those skilled in the art of the present technology can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation manner. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions made using the concept of the present invention are within the scope of protection.

[0091] As Figure 1 shown, a method for extracting the centroid of a star point based on a neural network provided by an embodiment of the present invention includes the following steps S1 to S3:

[0092] S1. Obtain the apparent magnitude data, and generate grayscale data of star imaging based on a Gaussian distribution for the neighborhood window of sub-pixel units obtained by subdividing each central pixel unit;

[0093] In this embodiment, step S1 adopts the following data generation process based on the different ranges of traversal of training sample-related sub-pixel units caused by different assumptions of the point spread function.

[0094] When step S1 considers the influence of the spatial target movement, generating grayscale data of star imaging based on a Gaussian distribution for the neighborhood window of sub-pixel units obtained by subdividing each central pixel unit includes:

[0095] Set the apparent magnitude interval as required , and the sub-pixel subdivision parameter ; Based on , the apparent magnitude traverses its value range ;

[0096] Calculate the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data;

[0097] Subdivide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical, and diagonal directions for the center of each sub-pixel unit, and traverse to all sub-pixel units in the area , that is , take the center of the sub-pixel unit as the centroid of the star point, that is ; As Figure 2 shown;

[0098] Set the motion direction angle using the Monte Carlo sampling method , the mean square deviation of the Gaussian distribution and the exposure time as random values;

[0099] Traverse the gray window corresponding to the star centroid, that is , as Figure 3 shown; Calculate the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor;

[0100] Set the random value of the number of photons using the Monte Carlo sampling method according to the fact that the number of photons reaching the pixel unit follows a Poisson distribution;

[0101] Calculate the gray value of the pixel unit according to the number of photons reaching the pixel unit.

[0102] When step S1 does not consider the influence of the spatial target motion, the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit generates gray data of star imaging based on the Gaussian distribution, including:

[0103] Set the visual magnitude interval as needed , and the sub-pixel subdivision parameter ; Based on , the visual magnitude traverse its value range ;

[0104] Calculate the average arrival rate of photons received by the V-band optical sensor according to the visual magnitude data;

[0105] Subdivide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical and diagonal directions for the centers of each sub-pixel unit, traverse the sub-pixel units in the first octant A and the second octant B regions, and use the center of the sub-pixel unit as the star centroid;

[0106] Set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution and the exposure time using the Monte Carlo sampling method;

[0107] Traverse the gray window corresponding to the star centroid and calculate the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor;

[0108] Set the random value of the number of photons using the Monte Carlo sampling method according to the fact that the number of photons reaching the pixel unit follows a Poisson distribution;

[0109] Calculate the gray value of the pixel unit according to the number of photons reaching the pixel unit.

[0110] When step S1 does not consider the influence of the spatial target motion and the Gaussian distribution point spread function is isotropic, the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit generates the gray-scale data of star imaging based on the Gaussian distribution, including:

[0111] Calculate the average arrival rate of photons received by the V-band optical sensor according to the visual magnitude data;

[0112] Subdivide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical and diagonal directions at the center of each sub-pixel unit, traverse the sub-pixel units in the first octant A area, and use the center of the sub-pixel unit as the centroid of the star point;

[0113] Use the Monte Carlo sampling method to set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution and the exposure time;

[0114] Traverse the gray-scale window corresponding to the star point centroid, and calculate the average arrival rate of photons received by the pixel units in the gray-scale window according to the average arrival rate of photons received by the V-band optical sensor;

[0115] According to the fact that the number of photons reaching the pixel unit satisfies the Poisson distribution, use the Monte Carlo sampling method to set the random value of the number of photons;

[0116] Calculate the gray-scale value of the pixel unit according to the number of photons reaching the pixel unit.

[0117] This embodiment refers to the data of the Hipparcos Catalog. Among them, the central wavelength of the V-band visible light is , the full width at half maxima (FWHM) is , the spectral luminous flux density of the zero magnitude star is , then the average spectral luminous flux density (the dimension is of the star with a visual magnitude of , where represents the number of photons), that is,

[0118]

[0119] Among them are the Planck constant and the speed of light respectively.

[0120] For a star with a visual magnitude of , the average arrival rate of photons received by the V-band optical sensor is

[0121]

[0122] Among them is the camera aperture area ( ), is the average optical efficiency of the sensor in the V band, is the average correction coefficient considering the influence of the atmosphere. For a specific star sensor, is known, is also known as a calibration parameter, and in the on-orbit case, it is considered that , then combined with the data of the Hipparcos catalog, it can be known that the variable in the above formula is only the apparent magnitude .

[0123] China's deep space exploration takes the lunar exploration project as a breakthrough. The limit of the apparent magnitude that the star sensors of the "Chang'e" series of the lunar exploration project can observe is +5.5. In addition, the apparent magnitude of Sirius, the brightest star (except the sun) that can be observed currently, is -1.45, and the apparent magnitude of the brightest artificial satellite, BlueWalker 3, can reach 0.4. Therefore, traversing the following apparent magnitude intervals at a certain interval in the generation of training data can meet the actual needs (for example, with an interval of 0.05, 140 apparent magnitude data can be obtained from the interval given by the following formula).

[0124]

[0125] In this embodiment, the pixel unit where the centroid of the star point is located can be easily determined according to the gray-scale data of the star point imaging. The goal of the centroid extraction algorithm is to determine its sub-pixel position. Then, each pixel unit is further divided into sub-pixel units, and the centers of the sub-pixel units obtained by traversing the subdivision are used to generate training data; Figure 4 In, the central pixel unit is divided into 9 sub-pixel units.

[0126] Since when the contrast data in the simulation star map is applied with the weighted centroid method, the squared weighted method, and the Gaussian fitting method, etc., the sub-pixel centroid accuracy of the three is not higher than pixels; considering that the signal-to-noise ratio of the actual star point imaging is indeed relatively low, according to the sampling theorem, it can be known that: when the number of subdivisions of the pixel unit in each direction satisfies the following formula condition, the sub-pixel centroid accuracy of the training data can be higher than the commonly used star point centroid extraction algorithm; in addition, for the sake of convenience is taken as an odd number:

[0127]

[0128] If the influence of the moving direction of the space target is considered, the point spread function of the star point imaging approximated by the Gaussian distribution is

[0129]

[0130] where The distribution is The mean square error in the direction; is the centroid position of the target imaging; Denotes the motion direction angle of the space target, such as Figure 5 shown, and its value range is .

[0131] Due to the randomness of the motion direction of the space target, at this time, generating training data must traverse all sub-pixel units to ensure completeness, as Figure 2 shown.

[0132] If the target motion is not considered (for example, only considering stellar targets), then can be taken, and at this time, the point spread function of the star imaging is

[0133]

[0134] Since the PSF given by the above formula is axisymmetric about the axis passing through the centroid , only traverse the sub-pixel units in Figure 2 the and two regions in total to ensure the completeness of the training data. At this time, the training data is reduced to of all.

[0135] If the target motion is not considered and at the same time is considered, at this time, the point spread function of the star imaging is

[0136]

[0137] Since the PSF given by the above formula is axisymmetric about the axis passing through the centroid and the diagonal, only traverse the sub-pixel units in Figure 2 the region in to ensure the completeness of the training data. At this time, the training data is reduced to

[0138] This embodiment is based on the defocusing technology. The image spot diffusion of star imaging is generally in the to neighborhood window, and the range of the mean square error of the corresponding Gaussian distribution point spread function is

[0139]

[0140] Then the gray window size of the training data is taken as ; as Figure 6 shown.

[0141] For the pixel unit in the gray window, where is the relative coordinate with the central pixel unit as the reference, using the point spread function The average arrival rate of photons received by the V-band optical sensor can be calculated, and its average arrival rate of received photons is

[0142]

[0143] Given the exposure time , the number of photons arriving at the pixel unit satisfies the following Poisson distribution, that is

[0144]

[0145] where .

[0146] Considering the faint characteristics of space targets and avoiding overexposure, the exposure time for imaging them generally ranges from

[0147]

[0148] For example, the typical exposure times of the visible light sensors carried by the American Midcourse Space Experiment (MSX) satellite are 0.4, 0.625, 1.0, 1.6 (s).

[0149] Based on the number of arriving photons , the gray value of the pixel unit can be obtained as

[0150]

[0151] where the function returns the largest integer not exceeding , and are the quantum efficiency and full well capacity of the sensor respectively (both are inherent performance parameters of the sensor), is the number of bits of the sensor (for space-based space targets, 16-bit gray data is generally used, that is ).

[0152] In this embodiment, in the generation of training data, the target motion direction angle , the mean square deviation of the Gaussian distribution PSF, the exposure time and the number of arriving photons Perform random settings. By introducing randomness, the applicability of the trained model can be effectively improved, especially for scenes with optical distortion and trailing. In these scenes, the accuracy of common centroid extraction algorithms often cannot be guaranteed. The sampling Monte Carlo sampling method requires a pre-given distribution function of random variables. Here, it is assumed that the target motion direction angle , the mean square deviation of the Gaussian distribution PSF , and the exposure time all satisfy the uniform distribution.

[0153] S2. Use the gray-scale data of star imaging as training data, and use the relative sub-pixel unit position of the star centroid within the central pixel unit as the label to train a multi-layer perceptron neural network model;

[0154] In this embodiment, step S2 uses the gray-scale data of star imaging generated in step S1 as training data, and uses the relative sub-pixel unit position of the star centroid within the central pixel unit as the label, and divides it into a training set and a validation set according to a ratio of 7:3.

[0155] The training of the multi-layer perceptron neural network based on the gray-scale window data is as Figure 7 shown. First, the gray-scale window data is sorted into a one-dimensional array in the order from bottom to top and from left to right. The position in the one-dimensional array can be obtained by the following formula; then connect the input layer of the multi-layer perceptron neural network; the input layer is fully connected to the hidden layer, between adjacent hidden layers, and between the hidden layer and the output layer through activation functions.

[0156]

[0157] The basic parameters of the model are shown in Table 1.

[0158] Table 1

[0159]

[0160] S3. Use the trained multi-layer perceptron neural network model to extract the star centroid.

[0161] In this embodiment, step S3 is based on the three types of training data sets generated in step S2, and three star centroid extraction models are correspondingly trained.

[0162] Step S3 uses the trained multi-layer perceptron neural network model to extract the star centroid, including:

[0163] Convert the gray-scale data of star imaging into a one-dimensional array AL;

[0164] Input the one-dimensional array AL into the trained multi-layer perceptron neural network model considering the spatial target motion, and output the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit. , where is the multi-layer perceptron neural network model considering the spatial target motion;

[0165] Calculate the star centroid position coordinates based on the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit , specifically:

[0166]

[0167]

[0168] where is the central pixel unit coordinate of the star imaging.

[0169] Step S3 for star centroid extraction using the trained multi-layer perceptron neural network model includes:

[0170] Coarsely estimate the relative sub-pixel unit position coordinates of the star centroid for the gray data of the star imaging using the square weighting method ;

[0171] Judge the region to which the relative sub-pixel unit position coordinates of the star centroid belong, and define the rotation transformation matrix ;

[0172] If , and it is located in the first and second octant regions, then define ;

[0173] If , and it is located in the third and fourth octant regions, then define ;

[0174] If , and it is located in the fifth and sixth octant regions, then define ;

[0175] If , and it is located in the seventh and eighth octant regions, then define ;

[0176] Use the rotation transformation matrix to transform the gray data of the star imaging to obtain the transformed gray data , is the relative coordinate of the pixel unit in the gray window;

[0177] The transformed gray data Convert to a one-dimensional array ;

[0178] Input the one-dimensional array into the trained multi-layer perceptron neural network model that does not consider the motion of the spatial target, and output the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit , where is the multi-layer perceptron neural network model that does not consider the motion of the spatial target;

[0179] Use the rotation transformation matrix to transform the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit to obtain the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit ;

[0180] Calculate the star centroid position coordinates according to the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit , specifically:

[0181]

[0182]

[0183] where is the central pixel unit coordinate of the star imaging.

[0184] Step S3 for star centroid extraction using the trained multi-layer perceptron neural network model includes:

[0185] Coarsely estimate the relative sub-pixel unit position coordinates of the star centroid for the grayscale data of the star imaging using the square weighting method ;

[0186] Judge the region to which the relative sub-pixel unit position coordinates of the star centroid belong, and define the rotation transformation matrix ;

[0187] If , and it is in the first octant region, then define ;

[0188] If , and it is in the second octant region, then define ;

[0189] If , and it is in the third octant region, then define ;

[0190] If , if it is located in the fourth octant region, then define ;

[0191] If , if it is located in the fifth octant region, then define ;

[0192] If , if it is located in the sixth octant region, then define ;

[0193] If , if it is located in the seventh octant region, then define ;

[0194] If , if it is located in the eighth octant region, then define ;

[0195] Use the rotation transformation matrix to transform the gray-scale data of the star point imaging, and obtain the transformed gray-scale data , is the relative coordinate of the pixel unit in the gray-scale window;

[0196] Convert the transformed gray-scale data into a one-dimensional array ;

[0197] Input the one-dimensional array into the trained multi-layer perceptron neural network model that does not consider the influence of the spatial target movement and has an isotropic Gaussian distribution point spread function, and output the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit , where is the multi-layer perceptron neural network model that does not consider the influence of the spatial target movement and has an isotropic Gaussian distribution point spread function;

[0198] Use the rotation transformation matrix to transform the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit , and obtain the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit ;

[0199] According to the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit calculate the star point centroid position coordinates , specifically:

[0200]

[0201]

[0202] Among them, is the coordinate of the central pixel unit for star point imaging.

[0203] In this embodiment, the given gray-scale data of the star point spot , and the coordinate of its central pixel unit is ( is a non-negative integer), that is, the integer pixel coordinate of the centroid position of the star corresponding to this star spot is . After obtaining the relative sub-pixel position of the star point centroid within the central pixel unit from the trained model, the sum of the two gives the exact position of the star point centroid on the imaging plane.

[0204] Starting from the assumption that the star point imaging spot approximately follows a Gaussian distribution, the present invention generates gray-scale data based on the Gaussian distribution for the neighborhood windows of each sub-pixel point that meets the accuracy requirements, and introduces randomness during the generation of gray-scale data through Monte Carlo sampling. The obtained data is used as the training data for a multi-layer perceptron neural network to obtain a neural network model with the window gray-scale data as the input and the relative position of the sub-pixel centroid as the output. The system has the advantages of good real-time performance and accuracy, and due to the randomness of the training data, it can also be applied to the situation where there are distortions and tails in star point imaging, with stronger versatility. In addition, if the anisotropy of the star point imaging point spread function is restricted, the data symmetry can be utilized to ensure the completeness of the samples while reducing the training sample data. In this way, the generality can be partially reduced at the cost of reducing the model training cost.

[0205] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified function in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0206] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the specified function in Figure 1 one process or multiple processes and / or blocks Figure 1The functions specified in one or more boxes.

[0207] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing the steps of the functions specified in one or more processes and / or boxes Figure 1 One process or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes.

[0208] In the present invention, specific embodiments are used to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0209] Those of ordinary skill in the art will realize that the embodiments described herein are for helping readers understand the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A method for extracting the centroid of star points based on a neural network, characterized in that, It includes the following steps: Obtain the apparent magnitude data, and generate the gray-scale data of star imaging based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit; Use the gray-scale data of star imaging as the training data, and use the relative sub-pixel unit position of the star centroid within the central pixel unit as the label to train a multi-layer perceptron neural network model; The star centroid extraction using the trained multi-layer perceptron neural network model includes: Convert the gray-scale data of star imaging into a one-dimensional array AL; Input the one-dimensional array AL into the trained multi-layer perceptron neural network model that takes into account the motion of the spatial target, and the relative sub-pixel unit position coordinates (x c , y c ) = F m (AL), where F m is the multi-layer perceptron neural network model that takes into account the motion of the spatial target; Calculate the centroid position coordinates of the star point according to the relative sub-pixel unit position coordinates of the star point centroid within the central pixel unit Specifically: Among them, (x0, y0) are the coordinates of the central pixel unit where the star point is imaged; Use the trained multi-layer perceptron neural network model to extract the star centroid.

2. The method for extracting the centroid of star points based on a neural network according to claim 1, characterized in that, When considering the influence of the spatial target motion, generating the gray-scale data of star imaging based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit, including: Calculate the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data; Subdivide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical, and diagonal directions for the center of each sub-pixel unit, traverse the sub-pixel units in all regions, and use the center of the sub-pixel unit as the star centroid; Use the Monte Carlo sampling method to set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution, and the exposure time; Traverse the gray-scale window corresponding to the star centroid, and calculate the average arrival rate of photons received by the pixel units in the gray-scale window according to the average arrival rate of photons received by the V-band optical sensor; According to the fact that the number of photons arriving at the pixel unit satisfies the Poisson distribution, use the Monte Carlo sampling method to set the random value of the number of photons; Calculate the gray-scale value of the pixel unit according to the number of photons arriving at the pixel unit.

3. A method for extracting the centroid of star points based on a neural network according to claim 1, characterized in that, When not considering the influence of the spatial target motion, generating the gray-scale data of star imaging based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit, including: Calculate the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data; Subdivide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical, and diagonal directions for the center of each sub-pixel unit, traverse the sub-pixel units in the first octant and the second octant regions, and use the center of the sub-pixel unit as the star centroid; Use the Monte Carlo sampling method to set the random values of the motion direction angle, the mean square deviation of the Gaussian distribution, and the exposure time; Traverse the gray-scale window corresponding to the star centroid, and calculate the average arrival rate of photons received by the pixel units in the gray-scale window according to the average arrival rate of photons received by the V-band optical sensor; According to the fact that the number of photons arriving at the pixel unit satisfies the Poisson distribution, use the Monte Carlo sampling method to set the random value of the number of photons; Calculate the gray-scale value of the pixel unit according to the number of photons arriving at the pixel unit.

4. A method for extracting the centroid of star points based on a neural network according to claim 1, characterized in that When not considering the influence of the spatial target motion and the Gaussian distribution point spread function is isotropic, generating the gray-scale data of star imaging based on the Gaussian distribution for the neighborhood window of the sub-pixel units obtained by subdividing each central pixel unit, including: Calculate the average arrival rate of photons received by the V-band optical sensor according to the apparent magnitude data; Subdivide the central pixel unit into 9 sub-pixel units, cut along the horizontal, vertical, and diagonal directions for the center of each sub-pixel unit, traverse the sub-pixel units in the first octant region, and use the center of the sub-pixel unit as the star centroid; Set random values for the motion direction angle, the mean square deviation of the Gaussian distribution, and the exposure time using the Monte Carlo sampling method; Traverse the gray window corresponding to the star centroid, and calculate the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor; Set a random value for the number of photons using the Monte Carlo sampling method according to the fact that the number of photons reaching the pixel unit follows a Poisson distribution; Calculate the gray value of the pixel unit according to the number of photons reaching the pixel unit.

5. A method for extracting the centroid of star points based on a neural network according to claim 2, 3 or 4, characterized in that, The calculation formula for the average arrival rate of photons received by the V-band optical sensor based on the apparent magnitude data is as follows: where is the average arrival rate of photons received by the optical sensor, is the average spectral luminous flux density of visible light in the V band, AP is the camera aperture area, is the average optical efficiency of the sensor in the V band, is the average correction factor considering the atmospheric influence, and Δλ0 is the full width at half maximum of visible light in the V band.

6. A method for extracting the centroid of star points based on a neural network according to claim 2, 3 or 4, characterized in that, The calculation formula for calculating the average arrival rate of photons received by the pixel units in the gray window according to the average arrival rate of photons received by the V-band optical sensor is: Among them, is the average arrival rate of photons received by the pixel unit in the grayscale window, is the average arrival rate of photons received by the optical sensor, I(i, j) is the pixel unit in the grayscale window, (i, j) is the relative coordinate of the pixel unit in the grayscale window, g(x, y) is the point spread function of the star imaging, and (x, y) is the pixel unit coordinate.

7. A method for extracting star centroid based on neural network according to claim 2, 3 or 4, characterized in that The calculation formula for calculating the gray value of the pixel unit according to the number of photons reaching the pixel unit is: Among them, ADU (i，j) is the gray value of the pixel unit, floor returns the largest integer not exceeding, n (i，j) is the number of photons reaching the pixel unit, QE is the quantum efficiency of the sensor, FWC is the full well capacity of the sensor, N bits is the number of bits of the sensor.

8. A method for extracting the centroid of star points based on a neural network according to claim 1, characterized in that, Using the trained multi-layer perceptron neural network model for star centroid extraction includes: The gray data of the star imaging is used to roughly estimate the relative sub-pixel unit position coordinates of the star centroid by the square weighting method Determine the position coordinates of the centroid of the star point relative to the sub-pixel unit The area to which it belongs, and define the rotation transformation matrix H; If it is located in the first and second octant regions, then define If it is located in the third and fourth octant regions, then define If it is located in the fifth and sixth octant regions, then define If it is located in the seventh and eighth octant regions, then define Use the rotation transformation matrix H to transform the grayscale data of the star point imaging to obtain the transformed grayscale data is the relative coordinate of the pixel unit in the grayscale window; Convert the transformed grayscale data into a one-dimensional array Input the one-dimensional array into the trained multi-layer perceptron neural network model that does not consider the spatial target motion, and output the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit where F s is the multi-layer perceptron neural network model that does not consider the spatial target motion; Use the rotation transformation matrix H to transform the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit to obtain the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit Based on the relative sub-pixel unit position coordinates (x c , y c ) of the star centroid within the central pixel unit, calculate the position coordinates of the star centroid Specifically: Among them, (x0, y0) is the coordinate of the central pixel unit of the star point imaging.

9. A method for extracting the centroid of star points based on a neural network according to claim 1, characterized in that, Using the trained multi-layer perceptron neural network model for star centroid extraction includes: Coarsely estimate the relative sub-pixel unit position coordinates of the star centroid by using the square weighting method for the gray-scale data of star imaging Determine the position coordinates of the centroid of the star point relative to the sub-pixel unit The area to which it belongs, and define the rotation transformation matrix H; If it is located in the first octant region, then define If it is located in the second octant region, then define If it is located in the third octant region, then define If it is located in the fourth octant region, then define If it is located in the fifth octant region, then define If it is located in the sixth eighth - circle region, then define If it is located in the seventh and eighth octant regions, then define If it is located in the eighth octant region, then define The gray-scale data of star imaging is transformed using the rotation transformation matrix H to obtain the transformed gray-scale data is the relative coordinate of the pixel unit in the gray-scale window; Convert the transformed grayscale data into a one-dimensional array The one-dimensional array is input into the trained multi-layer perceptron neural network model that does not consider the influence of the spatial target motion and has an isotropic Gaussian distribution point spread function, and the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit are output where F i is the multi-layer perceptron neural network model that does not consider the influence of the spatial target motion and has an isotropic Gaussian distribution point spread function; Use the rotation transformation matrix H to transform the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit to obtain the relative sub-pixel unit position coordinates of the star centroid within the central pixel unit According to the relative sub-pixel unit position coordinates (x c , y c ) of the star centroid within the central pixel unit, calculate the position coordinates of the star centroid Specifically: Among them, (x0, y0) is the coordinate of the central pixel unit of the star point imaging.

Images