Intelligent blind restoration method for blurred star images with low signal-to-noise ratio under high dynamic conditions

Abstract

The present invention relates to an intelligent blind restoration method for low-SNR blurred star maps under high-dynamic conditions, belonging to the field of star map restoration technology. First, from the image perspective, Gaussian filtering is used to denoise the star map, thereby improving the signal-to-noise ratio of the star map; then, a grid clustering method is used to remove noise points, so that low-SNR and broken stars are effectively extracted, and the star point detection rate for large-field star sensors exceeds 90%; finally, a deep learning algorithm is used to achieve end-to-end restoration of local blurred star maps. This invention effectively solves the problems of traditional algorithms that rely on inertial navigation information, have poor real-time performance, and have a single applicable situation when dealing with the problem of restoring high-dynamic and low-SNR blurred star maps, and greatly improves the accuracy, robustness, and reliability of autonomous starlight navigation for near-space aircraft and high-angular-velocity spacecraft.

Description

Technical Field

[0001] The present invention belongs to the technical field of star image restoration and relates to an intelligent blind restoration method for fuzzy star images with low signal-to-noise ratio under high dynamic conditions. Background Art

[0002] During starlight navigation, high-dynamic flight can cause star streaking. This reduced signal-to-noise ratio (SNR) results in the fragmentation of low-brightness star bands, making valid star points difficult to detect. Furthermore, the airborne environment places stringent demands on the real-time and robustness of restoration algorithms. Current research on the restoration of blurred star images falls into two main categories. The first utilizes the vehicle's angular velocity and attitude information provided by the inertial navigation system to determine the star map blur kernel or generate a blur template for correlation matching to determine star point positions. Deconvolution filtering is then used to restore the blurred star map. These algorithms rely on the integration of information from external devices, compromising their autonomy. The second category performs blind restoration without prior information. These algorithms first treat the blurred star streaks as straight lines, estimate the blur length and blur angle using methods such as feature transformation or image enhancement, and perform a rough extraction of the blurred stars. Finally, algorithms such as maximum likelihood estimation are used to accurately estimate the star point centroids. These algorithms are susceptible to noise interference when detecting star points with low SNRs and are computationally complex. Both of the above categories use traditional algorithms to decompose the deblurring problem into two parts: blur parameter estimation and image deblurring. The accuracy of star point centroid extraction depends heavily on the accuracy of the blur parameter estimation. Using an inaccurate blur kernel for deblurring can easily produce ringing and checkerboard artifacts in the image. Furthermore, the two-stage computational process significantly increases the time complexity. In the field of natural image blurred restoration, many researchers have used deep learning techniques to achieve end-to-end blurred restoration. This means that the blurred image is input and the restored image is directly output without estimating the blur parameters. However, for globally inconsistent blur caused by camera rotation, these methods often employ complex network structures and use the entire image as input, resulting in high computational complexity and poor real-time performance, making them unsuitable for missile-borne environments. Therefore, to address the problems of these methods, it is necessary to design an intelligent method that can ensure real-time and robustness while also enabling high-precision end-to-end restoration of blurred star images without the assistance of inertial navigation. Summary of the Invention

[0003] The technical problems to be solved by the present invention are:

[0004] To address the problem of fuzzy star maps, such as low signal-to-noise ratio (PSNR < 2) and global inconsistency, caused by the high dynamic motion of hypersonic vehicles in the absence of prior inertial navigation information. Traditional starlight navigation systems struggle to accurately restore star map information in these situations, requiring extensive computation and processing, resulting in long computational times. This paper combines image filtering, unsupervised clustering, and a deep learning framework to provide an end-to-end intelligent fuzzy star map restoration method based on deep learning.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] An intelligent blind restoration method for a fuzzy star image with a low signal-to-noise ratio under high dynamic conditions, characterized by comprising:

[0007] Simulate fuzzy star images and static star images to generate training samples;

[0008] The local blurred restoration area is determined by the restoration area determination method, and the local blurred star map in the area is intercepted;

[0009] The local blurred star map in the restored area is used as the input data set, and the static star map corresponding to the blurred star map in the restored area is used as the output data set. A local blurred star map restoration network is constructed and trained.

[0010] Input the original blurred star image into the trained network and output the restored image.

[0011] A further technical solution of the present invention is as follows: the locally blurred star map restoration network includes two types: a high-noise star map restoration network and a low-noise star map restoration network. Both network models adopt a densely connected network structure. The processing flow is as follows: first, the star map noise of the input image is calculated and then input into the corresponding network. Then, the input image is processed by the front network and then input into the dense network. The dense network is composed of three dense blocks and two transition layers alternately connected. Each dense block has a 5*5 convolution structure with 16 channels and a 7*7 convolution structure with 8 channels stacked three times. The transition layer uses a 7*7 convolution structure and controls the number of output channels to be halved. After passing through the dense network, the restored image is finally processed by the back network.

[0012] A further technical solution of the present invention is as follows:

[0013] The midpoint of the x and y boundaries of each cluster is taken as the center of the star stripe restoration area (x ci ,y ci );

[0014] Calculate the maximum blur length L of the star streak in the x and y directions x max and L y max Pixels:

[0015]

[0016] Where N x 、N y Respectively represents the resolution of the star map in the x and y directions; Fov x 、Fov y Respectively represent the size of the star sensor's field of view in the x and y directions; ω x max 、ω y max They represent the maximum angular velocity of the star sensor around the x-axis and the y-axis respectively; T represents the exposure time of the star sensor;

[0017] If more than 90% of the star points in each star map are satisfied, the error between the restored center of the star stripes extracted by the grid clustering algorithm and the center of the real star stripes is at least δ x ,δ y ;

[0018] Determine the size of the restoration area as (2δ x +2L x max )×(2δ y +2L y max ), then the i-th star stripe restoration area ROI i for:

[0019]

[0020] A further technical solution of the present invention is as follows: the restoration area determination method further includes the following contents:

[0021] Performing Gaussian smoothing filtering on the original motion blurred star image; performing binarization processing on the filtered star image to obtain a binary image;

[0022] Divide the star map plane S into a grid; determine whether the grid is a high-density grid, if it is a high-density grid, record it as 1, otherwise it is a low-density grid, record it as 0; perform 8-neighborhood connected domain detection on the 0,1 matrix generated by the high and low density grid judgment, and mark it;

[0023] Filter clusters and noise: Define the cluster judgment threshold η, that is, the minimum number of cells contained in the connected domain. If it is greater than the threshold, the connected domain forms a cluster, and all data points contained in the area belong to the same star stripe. Otherwise, it is regarded as noise and eliminated.

[0024] A further technical solution of the present invention is as follows: the Gaussian smoothing filter processing of the original motion blurred star image f(x, y) is performed:

[0025] Calculate the Gaussian filter kernel standard deviation σ and kernel size m and n according to the empirical formula:

[0026]

[0027] m=n=2×ceil(2σ)+1

[0028] Use the designed Gaussian filter kernel to filter and smooth the original star image:

[0029]

[0030]

[0031] Where w(x,y) represents the Gaussian filter kernel; G(x,y) indicates that the filter satisfies the two-dimensional Gaussian distribution.

[0032] A further technical solution of the present invention is as follows: the binarization processing of the filtered star map to obtain a binary image is specifically performed as follows:

[0033]

[0034] Where T is the binarization threshold.

[0035] A further technical solution of the present invention is to divide the star map plane S into grids as follows:

[0036] Divide each dimension of the two-dimensional space S = [a1, b1] × [a2, b2] into M equal parts, generating a total of M 2 The domain of space S is obtained as follows: Given by the formula:

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] Where N x 、N y Respectively represent the resolution of the star map in the x and y directions; a1, a2, b1, b2 represent the coordinate values ​​of the boundaries of the two-dimensional space S in the x and y directions; M represents the number of grids divided in each dimension of the two-dimensional space; d ik Represents the boundary coordinate value of the kth grid in the i-th dimension of the two-dimensional space S;

[0043] The binarized star map c bi The pixel points Z1, Z2, ..., Z in (x, y) are 1 nMapped to the grid, if Z n (x,y)∈C k , then the data point is divided into C k within the grid cell.

[0044] A further technical solution of the present invention is to determine whether a grid is a high-density grid by:

[0045] If the number of data points contained in the grid is n c >ζ, then the grid C k For high density grid, it is marked as 1, otherwise the grid C k The low-density grid is recorded as 0; where ζ is the grid density threshold of space S.

[0046] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0047] A computer-readable storage medium is characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0048] The beneficial effects of the present invention are:

[0049] The present invention provides an intelligent blind restoration method for low-SNR fuzzy star maps under high-dynamic conditions. First, from an image perspective, Gaussian filtering is used to denoise the star map, improving the signal-to-noise ratio. A grid clustering method is then used to remove noise points, effectively extracting low-SNR and broken stars. The star point detection rate for wide-field star sensors exceeds 90%. Finally, a deep learning algorithm is used to achieve end-to-end restoration of locally fuzzy star maps. When the star point peak signal-to-noise ratio is less than 2, the average star point centroid positioning error for the dimmest star visible to the star sensor, a magnitude 6 star, can be reduced to within 1.5 pixels. This invention effectively addresses the problems of traditional algorithms in restoring high-dynamic, low-SNR fuzzy star maps, such as reliance on inertial navigation information, poor real-time performance, and limited applicability. It significantly improves the accuracy, robustness, and reliability of autonomous starlight navigation for near-space aircraft and high-angular-velocity spacecraft.

[0050] Compared with traditional methods, this method does not require precise fuzzy parameter estimation, thus avoiding the impact of parameter estimation errors on the results. It can cluster and extract broken star belts and achieve high-precision center of mass positioning, providing reliable star point coordinate information for subsequent star map recognition and attitude solution, greatly improving the autonomy, real-time performance, robustness and anti-interference ability of the navigation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0052] Figure 1 Binarized fracture star map;

[0053] Figure 2 Densely connected network structure diagram;

[0054] Figure 3 Flowchart of the intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions;

[0055] Figure 4 Original fuzzy star map;

[0056] Figure 5 Star map after Gaussian smoothing filtering;

[0057] Figure 6 Schematic diagram of grid division of binary star map;

[0058] Figure 7 Schematic diagram of grid clustering;

[0059] Figure 8 Schematic diagram of the local restoration area;

[0060] Figure 9 Schematic diagram of the end-to-end restoration network for locally blurred star images based on dense connections;

[0061] Figure 10 Partially blurred star map;

[0062] Figure 11 Restoring star map from local blur;

[0063] Figure 12 Original blurred star image (triple grayscale);

[0064] Figure 13 Star target clustering;

[0065] Figure 14 Restore star map;

[0066] Figure 15 Distribution map of star spot code values ​​after restoration. DETAILED DESCRIPTION

[0067] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0068] A method for determining the restoration area of ​​celestial targets based on grid clustering

[0069] Under high-dynamic conditions, if the star sensor rotates around the optical axis, the trailing directions of stars at different locations in the star map may be inconsistent. Using deep learning methods to restore the global image, for globally inconsistent, blurred star maps from large-field-of-view star sensors with high resolution, the network input is large, making the network structure overly complex. Therefore, we use local regions instead of global images for restoration to simplify input and output. The selection of the restoration region is a key step. The restoration region must contain complete star trails within the region, and the region must be as small as possible. On the one hand, if the region is small enough, the blur caused by rotation around the optical axis can be considered consistent within the region; on the other hand, because the restoration process in the background region without star trails is meaningless, reducing the restoration region can reduce the time complexity of the restoration calculation and improve computational efficiency. Therefore, the present invention proposes a method for determining the restoration region of a star target based on grid clustering.

[0070] (1) Gaussian smoothing filter denoising

[0071] To study motion-blur restoration of star images, we must first understand the process by which the image is blurred and establish an image degradation model. Assume that the original clear image g(x,y) is input to the system. After undergoing a degradation process h and being disturbed by external noise n(x,y), the resulting blurred image f(x,y) can be expressed as:

[0072]

[0073] Where, Represents the convolution operation. The noise n(x,y) mainly refers to the noise of the star sensor imaging device, including photon shot noise, dark current noise, etc., which is usually assumed to be Gaussian noise and salt and pepper noise. Under static conditions, the star point energy is concentrated in 3×3 or 5×5 pixels, and the grayscale value is much larger than the noise size, so denoising can be ignored. However, under high dynamic conditions, it is essential to remove noise from blurred star images with low signal-to-noise ratio. The Gaussian smoothing filter algorithm can effectively suppress the Gaussian noise and salt and pepper noise of the image, and it is a linear symmetric filter to ensure that the center of mass will not be affected after processing. The linear spatial filtering of the original star image using a Gaussian filter kernel of size m×n can be expressed as:

[0074]

[0075] Where w represents the filter kernel, and m = 2a + 1 and n = 2b + 1, where a and b are non-negative integers; c(x, y) is the denoised star image after Gaussian smoothing; and σ is the standard deviation of the two-dimensional Gaussian filter kernel. The standard deviation σ and the filter kernel sizes m and n are calculated from each pixel of the original blurred star image using an empirical formula. The empirical formula is as follows:

[0076]

[0077] (2) Grid clustering

[0078] After Gaussian filtering, the noise in the blurred star image has been smoothed, and the signal-to-noise ratio of the star points has been significantly improved. At this time, the threshold method is used to classify the star targets and the background. The binary threshold T is selected as 12 based on experience. The calculation formula is as follows:

[0079]

[0080] Depend on Figure 1 It can be seen that after binarization, images of stars with larger magnitudes exhibit discontinuities due to the grayscale of some pixels being less than the noise level. Using a traditional connected domain algorithm for extraction would result in a star being segmented into distinct stars based on its discontinuities. This invention employs clustering, treating star streaks as clusters, and achieves rapid star extraction in an unsupervised manner. The specific method is as follows:

[0081] Define the star map plane as a two-dimensional space S, establish a two-dimensional coordinate axis with the lower left corner of the image as the origin, and set the domain of the i-th dimension of space S to be [a i ,b i ], where i = 1, 2, then S = [a1, b1] × [a2, b2]. Divide each dimension of S into M equal parts, and a total of M 2 grid cells. The kth partition area of ​​the data space in the i-th dimension is denoted as [d ik ,d ik+1 ). This results in a square area is called a grid cell. ik It can be expressed as:

[0082]

[0083] The pixel points Z1, Z2, ..., Z with a gray value of 1 in the star image after binarization are n As n data points in the space, if Z n (x,y)∈C k , then the data point is divided into C k In the grid unit, x, y are the pixel coordinates corresponding to the pixel point. To avoid the data point falling into the common boundary of two grid units or the outermost boundary of the outermost grid, the space S is set at the original star map resolution N.x ×N y Expand and move based on size, that is:

[0084]

[0085] Assume that the grid C k (k=1,2,…,M 2 ) contains n c data points, let the grid density threshold of space S be ζ. c >ζ, then the grid C k is a high-density grid, otherwise the grid C k = is a low-density grid. Each grid cell is traversed and tested for high-density grids, with a value of 1 if it is high and 0 if it is not. This forms a binary matrix. This binary matrix is ​​then tested for connected domains in its 8-neighborhood and marked. A cluster determination threshold η is defined, which is the minimum number of cells in a connected domain. If the threshold is exceeded, the connected domain forms a cluster, and all data points within the region belong to the same star streak. Otherwise, the data points are considered noise and removed.

[0086] (3) Determine the restoration area

[0087] After the above clustering process, the approximate area of ​​the star stripes has been determined, and the midpoint of the x and y boundaries of each cluster is used as the center of the star stripe restoration area (x c ,y c ). Considering that there is an error between the restored center and the true star point mass center position, and the restored area needs to contain the complete star stripes, the error δ between the restored center and the true star point mass center position is determined. x and δ y , and calculate the maximum blur length L of the star streak in the x and y directions x max and L y max Pixels, determine the size of the restored area is (2δ x +2L x max )×(2δ y +2L y max ). Where δ x ,δ y Through experimental statistics under the conditions of large angular velocity (6-10° / s) and high noise (8-18 grayscale), more than 90% of the star points in each star map are extracted using the grid clustering algorithm to extract the error range between the restored center of the star stripes and the real star stripe center. x max ,L y max It is estimated by the following formula.

[0088]

[0089] Where, ωx max ,ω y max Respectively represent the maximum angular velocity of the star sensor around the x and y axes; Fov x ,Fov y Respectively represent the size of the field of view in the x-axis and y-axis directions; T is the exposure time.

[0090] End-to-end blind restoration algorithm for locally blurred star images based on densely connected networks

[0091] We have roughly extracted the positions of star streaks and determined the restoration area of ​​the star target. In this section, we design a neural network to achieve end-to-end restoration of locally blurred star images.

[0092] (1) Network structure

[0093] Because low-signal-to-noise-ratio blurred star images have limited available information and relatively sparse features, to avoid significant loss of spatial detail during neural network propagation, the end-to-end restoration network for locally blurred star images adopts a densely connected network architecture. This architecture directly links input and output along the channel dimension, allowing each layer to connect to the previous layer. This enhances feature reuse, reduces the number of parameters, mitigates vanishing / exploding gradients and network degradation, and makes the network easier to train. The densely connected network primarily consists of dense blocks and transition layers. A dense block consists of several densely connected units, each of which contains two convolutional structures. Each convolutional structure is composed of a batch normalization layer, a Reinforced Luminance (ReLU) activation layer, and a convolutional layer. Their input is the concatenation of the outputs of all preceding convolutional structures, thus enabling information sharing and transmission. Because blurred star images have limited available information, pooling would result in information loss. Therefore, pooling is not performed in the transition layer to reduce the feature map dimensionality. Instead, a single convolutional structure controls the number of channels in the feature map. First, the input image is processed by the front network and then input into the dense network. The dense network is composed of three dense blocks and two transition layers connected alternately. After passing through the dense network, the back network finally processes and outputs the restored image. The overall structure is composed of Figure 2 As shown:

[0094] (2) Network training

[0095] Considering that the training dataset contains low-signal-to-noise ratio star images with varying noise levels, the standard deviation of the noise ranges from 8 to 18 grayscales. Star images with low and high noise exhibit different characteristics. Using a single network model for training is prone to overfitting, making it difficult to achieve high-precision restoration. Therefore, we propose a network model based on star image noise classification. Specifically, we train two restoration networks, one for low noise (σ < 12 grayscales) and one for high noise (σ ≥ 12 grayscales), both using the densely connected network structure described above. When restoring locally blurred star images, we first calculate the standard deviation of the blurred star image, determine whether it is greater than or less than 12 grayscales, and input it into the corresponding network for restoration.

[0096] The proposed algorithm needs to be trained and learned from a star map database under dynamic conditions. Since real star maps are difficult to obtain, training data is generated by simulating star maps. In the process of constructing the data set, 100 optical axis directions are randomly selected. For each optical axis direction, different rotation angular velocities and noise levels are set to simulate fuzzy star maps and static star maps. The x-axis and y-axis rotation angular velocities ω x and ω y The combined angular velocity is selected as 6° / s, 8° / s, and 10° / s for 3 groups, each group is separated by 15°, and there are 12 speed combinations in total; the angular velocity ω z The selected values ​​are 0° / s and ±10° / s, for a total of three groups. The standard deviation of Gaussian white noise added to the star image is selected based on the network type: 8, 10, 12, and 14 grayscale for low-noise networks, and 12, 14, 16, and 18 grayscale for high-noise networks, for a total of four groups. Therefore, each optical axis orientation generates 3×12×3×4=432 blurred star images. For each blurred star image, the local blurred star image within the region is intercepted as the input dataset based on the restoration region determination method described in Section 2.1. The static star image corresponding to the blurred star image within the region is used as the output dataset. The datasets are randomly divided into training and test sets in an 8:2 ratio. The training loss function is the L2 (MSE) loss function, and the optimizer uses the Adam optimizer. The initial learning rate is set to 1e-4, and the learning rate is adjusted using cosine annealing, namely:

[0097]

[0098] Among them, η t is the current learning rate; η min is the minimum learning rate; η max is the maximum learning rate; T cur is the current epoch; T max The training is stopped when overfitting is obvious.

[0099] 1. Implementation methods

[0100] The specific implementation process of the intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions is as follows:

[0101] (1) Perform Gaussian smoothing filtering on the original motion blurred star image f(x,y)

[0102] a) Calculate the Gaussian filter kernel standard deviation σ and kernel sizes m and n according to the empirical formula.

[0103]

[0104] m=n=2×ceil(2σ)+1

[0105] b) Use the designed Gaussian filter kernel to filter and smooth the original star image.

[0106]

[0107]

[0108] Where w(x,y) represents the Gaussian filter kernel; G(x,y) indicates that the filter satisfies the two-dimensional Gaussian distribution.

[0109] (2) Binarize the filtered star map c(x,y) to obtain the binary image c bi (x,y), the binarization threshold T is 12.

[0110]

[0111] (3) Gridding the star map plane S

[0112] a) Divide each dimension of the two-dimensional space S = [a1, b1] × [a2, b2] into M equal parts, generating a total of M 2 grid cells.

[0113] The domain of space S is obtained according to the following formula: the division area of ​​the grid unit Given by Eq.

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] Where N x 、Ny Respectively represent the resolution of the star map in the x and y directions; a1, a2, b1, b2 represent the coordinate values ​​of the boundaries of the two-dimensional space S in the x and y directions; M represents the number of grids divided in each dimension of the two-dimensional space; d ik Represents the boundary coordinate value of the kth grid in the i-th dimension of the two-dimensional space S.

[0120] b) Binarized star map c bi The pixel points Z1, Z2, ..., Z in (x, y) are 1 n Mapped to the grid, if Z n (x,y)∈C k , then the data point is divided into C k within the grid cell.

[0121] (4) Determine whether the grid is a high-density grid. That is, let the grid density threshold of space S be ζ. If the number of data points n in the grid is c >ζ, then the grid C k For high density grid, it is marked as 1, otherwise the grid C k For low density grid, it is recorded as 0.

[0122] (5) Perform 8-neighborhood connected domain detection on the 0,1 matrix generated by high and low density grid judgment and mark it out.

[0123] (6) Filtering clusters and noise. Define the cluster judgment threshold η, which is the minimum number of cells contained in the connected domain. If it is greater than the threshold, the connected domain forms a cluster, and all data points contained in the region belong to the same star stripe. Otherwise, it is considered as noise and eliminated.

[0124] (7) Determine the restoration area

[0125] a) The midpoint of the x and y boundaries of each cluster is taken as the center of the star stripe restoration area (x ci ,y ci ).

[0126] b) Calculate the maximum blur length L of the star streak in the x and y directions x max and L y max Pixel

[0127]

[0128] Where N x 、N y Respectively represents the resolution of the star map in the x and y directions; Fov x 、Fov y Respectively represent the x and y field angles of the star sensor; ω x max 、ω y maxThey represent the maximum angular velocity of the star sensor around the x-axis and the y-axis respectively; T represents the exposure time of the star sensor.

[0129] c) Experimental statistics show that if more than 90% of the star points in each star map meet the requirement, the error between the restored center of the star stripes extracted by the grid clustering algorithm and the center of the real star stripes is at least δ x ,δ y .

[0130] d) Determine the size of the restoration area as (2δ x +2L x max )×(2δ y +2L y max ), then the i-th star stripe restoration area ROI i for:

[0131] ROI i ={(x,y)|x∈[x ci ∈(δ x +L x max ),x ci +(δ x +L x max )],

[0132] y∈[y ci -(δ y +L y max ),y ci +(δ y +L y max )]}

[0133] (8) Training the local blurred star map restoration network

[0134] a) Generate training sample set

[0135] Randomly select 100 optical axis directions, and set different rotation angular velocities and noise levels for the star map under each optical axis direction to simulate fuzzy star maps and static star maps. x and ω y The combined angular velocity is selected as 6° / s, 8° / s, and 10° / s for 3 groups, each group is separated by 15°, and there are 12 speed combinations in total; the angular velocity ω zThe values ​​of 0° / s and ±10° / s are selected, for a total of 3 groups. The standard deviation of Gaussian white noise added to the star map is selected according to the network category, with the values ​​of 8, 10, 12, and 14 for low-noise networks and 12, 14, 16, and 18 for high-noise networks, for a total of 4 groups. Therefore, 3×12×3×4=432 blurred star maps can be generated for each optical axis direction. For each blurred star map, the local blurred star map in the area is intercepted as the input data set according to the restoration area determination method (1) to (7). The static star map corresponding to the blurred star map in the area is used as the output data set. The data sets are randomly divided into training and test sets in a ratio of 8:2.

[0136] b) Train the high-noise star map restoration network and the low-noise star map restoration network separately

[0137] The network includes two types: high noise star map restoration network and low noise star map restoration network. Both network models adopt the following Figure 2 The densely connected network structure shown.

[0138] First, the input image is processed with star map noise and then input into the corresponding network. It is then processed by the front network and input into the dense network. The dense network consists of three dense blocks alternately connected with two transition layers. Each dense block has a 5*5 convolution structure with 16 channels and a 7*7 convolution structure with 8 channels stacked three times. The transition layer uses a 7*7 convolution structure and controls the number of output channels to be halved. After passing through the dense network, the back network finally processes and outputs the restored image.

[0139] Table 1 Densely connected network model

[0140]

[0141] The Adam optimizer was used for training. The initial learning rate was set to 1e-4. The cosine annealing was used to adjust the learning rate. The minimum learning rate was set to 1e-6. A total of 120 epochs were trained. When obvious overfitting occurred, the training was stopped.

[0142] (9) Restoration of partially blurred star maps

[0143] The original blurred star map is input according to the process (1) to (7) to determine the local blurred restoration area. Calculate the ROI of each area i The standard deviation σ of all pixels in the original blurred star image i , according to σ i Select the network with the corresponding size. Input the locally blurred image into the corresponding network and output the restored image.

[0144] 2. Implementation Examples

[0145] The embodiments of the present invention are as follows:

[0146] (1) Implementation conditions

[0147] The star sensor parameter settings are shown in Table 2. The optical axis direction is randomly generated, the combined angular velocity of the star sensor's x- and y-axes is set in the range of [6, 10]° / s, the combined velocity direction is randomly selected, the angular velocity of the z-axis is set in the range of [-10, 10]° / s, and the standard deviation of the white noise grayscale of the star image is set in the range of [8, 18] to simulate a dynamic blurred star image.

[0148] Table 2 Star sensor parameters

[0149] Resolution 2048 pixels × 2048 pixels Pixel size 5.5μm×5.5μm Field of view 25°×25° focal length 25mm Exposure time 0.1s Maximum magnitude 6.0MV

[0150] (2) Implementation results

[0151] The number of grid divisions in the grid clustering algorithm is set to 256×256, the grid density threshold is ζ=3, and the cluster judgment threshold η=4. The experiment is designed with the angular velocity of the star sensor around the x-axis and y-axis and the star image noise as variables. The angular velocity is selected as 6° / s, 8° / s, and 10° / s, and the noise size is selected as 8, 10, 12, 14, 16, and 18 grayscale groups. Each group includes 1000 frames of simulated star images under dynamic conditions, and the optical axis of each frame is randomly generated. The detection rate of star points is defined as follows: the pixel error in the x and y directions of all star points in a simulated blurred star map extracted by the grid clustering method and the center of the real star stripe is δ x ,δ y The ratio of inner stars is δ. x =δ y = The detection rate of star points under 10, 15, and 20 pixels. From the results, it can be concluded that in δ x =δ y = 20 pixels, except for a few cases, almost all star maps can achieve a star point detection rate of more than 90%. x ,δ y It is set to 20 pixels. Combined with the star point trailing length when the angular velocity is 10° / s, the restoration area size is finally set to 150 pixels × 150 pixels.

[0152] Taking 12:00 Beijing time on September 1, 2022 as the reference, the optical axis direction is selected as right ascension 109.91° and declination 11.80°, the three-axis angular velocity of the star sensor is [9.9493, 1.0060, 10.0000]° / s, and the standard deviation of the Gaussian noise grayscale is 12. The restoration process of the blurred star map is as follows: Figure 13-14 The centroid extraction results of some star points are shown in Table 2. It can be seen that the invention can effectively increase the signal-to-noise ratio of the fuzzy star image. Figure 15It can be seen from the restoration results and the centroid extraction results in Table 2 that under high dynamic and noise interference, the present invention successfully extracts almost all star points except the edge of the CCD plane, and the star point centroid coordinate error can meet the actual application requirements.

[0153] Table 3. Results of centroid extraction of some star points

[0154]

[0155] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. An intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions, characterized by include: Simulate fuzzy star images and static star images to generate training samples; A restoration area determination method is used to determine the local fuzzy restoration area, and the local fuzzy star map within the area is intercepted. The restoration area determination method is specifically as follows: Each cluster x 、 y The midpoint of the two directional boundaries is taken as the center of the star stripe restoration area ; Calculate the streak of stars in x 、 y Maximum directional blur length and Pixels: Where, 、 Respectively represent the resolution of the star map in the x and y directions; 、 Respectively represent the size of the star sensor's field of view in the x and y directions; 、 Respectively represent the maximum angular velocity of the star sensor around the x-axis and y-axis; T Indicates the exposure time of the star sensor; If more than 90% of the star points in each star map are satisfied, the error between the restored center of the star stripes extracted by the grid clustering algorithm and the center of the real star stripes is at least ; Determine the size of the recovery area , then i Star streak restoration area for: The local blurred star map in the restoration area is used as the input data set, and the static star map corresponding to the blurred star map in the restoration area is used as the output data set to construct a local blurred star map restoration network and train it. The local blurred star map restoration network includes two types: a high-noise star map restoration network and a low-noise star map restoration network. Both network models adopt a dense connection network structure. The processing flow is as follows: first, the input image is input into the corresponding network after calculating the star map noise, and then it is processed by the front network and then input into the dense network. The dense network is composed of three dense blocks and two transition layers alternately connected, wherein each dense block has a 5*5 convolution structure with 16 channels and a 7*7 convolution structure with 8 channels stacked three times. The transition layer uses a 7*7 convolution structure and controls the number of output channels to be halved. After passing through the dense network, the output restored image is finally processed by the back network; Input the original blurred star image into the trained network and output the restored image.

2. The intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions according to claim 1, characterized in that: The restoration area determination method also includes the following contents: Perform Gaussian smoothing filtering on the original motion blurred star image; Binarization is performed on the filtered star image to obtain a binary image; Divide the star map plane S into grids; determine whether the grid is a high-density grid, if it is a high-density grid, it is recorded as 1, otherwise it is a low-density grid, it is recorded as 0; Perform 8-neighborhood connected domain detection on the 0,1 matrix generated by high and low density grid judgment and mark it out; Filtering clusters and noise: Defining cluster determination thresholds , that is, the minimum number of cells contained in the connected domain. If it is greater than the threshold, the connected domain forms a cluster, and all data points contained in the area belong to the same star stripe. Otherwise, it is regarded as noise and eliminated.

3. The intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions according to claim 2, characterized in that: The original motion blurred star map The Gaussian smoothing filter processing is as follows: Calculate the Gaussian filter kernel standard deviation according to the empirical formula and kernel size m 、 n : Among them, X and Y are 、 the total number of Use the designed Gaussian filter kernel to filter and smooth the original star image: Where, represents the Gaussian filter kernel; Indicates that the filter satisfies a two-dimensional Gaussian distribution.

4. The intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions according to claim 3, characterized in that: The binarization process of the filtered star map to obtain a binarized image is specifically as follows: in, T is the binarization threshold.

5. The intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions according to claim 4, characterized in that: The star map plane S is divided into grids as follows: Two-dimensional space Each dimension is divided into M Equally divided, a total of The domain of space S is obtained according to the following formula: the division area of ​​the grid unit Given by the formula: Where, 、 Respectively represent the resolution of the star map in the x and y directions; 、 、 、 Respectively represent the coordinate values ​​of the boundaries of the two-dimensional space S in the x and y directions; M Indicates the number of grids divided in each dimension of the two-dimensional space; Represents the two-dimensional space S i In the first dimension k The boundary coordinates of the grid; The binary star map Pixels with a value of 1 Mapped to the grid, that is, if , then the data point is divided into within the grid cell.

6. The intelligent blind restoration method for low signal-to-noise ratio blurred star images under high dynamic conditions according to claim 2, characterized in that: To determine whether a grid is a high-density grid: If the number of data points contained in the grid , then the grid For high density grid, it is marked as 1, otherwise the grid For low-density grids, it is recorded as 0; is the grid density threshold of space S.

7. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

8. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images