A method for detecting dim space targets in deep space background

Abstract

The present invention relates to a method for detecting faint space targets under a deep space background, which includes the steps of uniform bright star extraction, bright star feature expression and matching, global inter-frame registration, precise matching of dark stars in local areas, deformation field calculation, inter-frame alignment background suppression, and faint target extraction. A uniform bright star extraction strategy is adopted to reduce the influence of dark stars on bright star matching. Feature expression and angular distance constraints can ensure correct matching of bright stars during large inter-frame maneuvers, and global inter-frame registration is effective. A coarse-to-fine inter-frame registration strategy is adopted to update the accuracy of dark star coordinates during local alignment, improve the accuracy of deformation field, reduce background suppression noise interference, and be more robust to large field of view and large edge distortion. The method can identify discordant targets with local inter-frame motion in the field of view at the first time, provide potential faint target positions in each frame for inter-frame trajectory association, and improve the detection rate of faint targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of space target monitoring and relates to a method for detecting dark and weak space targets under a deep space background. Background Art

[0002] With the development of wide-field-of-view space cameras, the ability to perceive space targets has greatly improved. Not only have higher and higher magnitudes been detected, but the ability to stably track targets has also significantly increased due to the wide field of view. The increased tracking arc has significantly improved the precision of space target orbit determination. However, due to the extremely large field of view and the complex stellar background interference, the detection of faint space targets presents significant challenges. This is mainly reflected in the following aspects: the difference in angular velocity of the space target relative to the camera causes a variety of target morphologies within the exposure time, which is intertwined with the dense distribution of strong and weak stars, making it impossible to effectively distinguish them based on morphology alone. Using star catalogs to identify all stars at low magnitudes and eliminate background interference is not a fast solution for dense high-magnitude stars, and it also poses significant challenges to the completeness of the catalog, resulting in the omission of stars. Traditional inter-frame difference background suppression methods are prone to false alarms due to the field distortion caused by the extremely large field of view, causing some stars to be mistakenly detected as space targets. Therefore, achieving inter-frame registration of star charts with large fields of view and high magnitudes is key to extracting faint space targets.

[0003] According to publicly available literature, there are several main methods for inter-frame registration of star sequences: methods based on regional statistics, such as grayscale correlation, mutual information, and Fourier-Mellin transforms based on frequency-domain phase correlation. Grayscale correlation and mutual information methods require a traversal search to determine the optimal registration parameters. The Fourier-Mellin transform, due to the periodicity of rotation coefficients in the frequency domain, also requires multiple judgments, resulting in high computational complexity and generally pixel-level accuracy. Methods based on geometric features, such as star polygon matching and grid matching, typically require constructing feature vectors. The highly variable combinations of star pairs used to generate feature vectors in densely populated stars increase the difficulty of matching and significantly challenge the uniqueness of feature vectors, leading to increased mismatches. The deep learning-based method constructs a training set by pairing adjacent frames and uses the network to learn the transformation parameters of the two. It can be a rigid transformation or a deformable transformation. However, deep learning requires a large number of training samples. When the training set is very different from the actual one, the registration result is not satisfactory and will cause many false alarms, especially for deep space images without texture. Deep learning does not show a strong advantage over traditional methods, but has strict requirements on the training set. Summary of the Invention

[0004] The present invention addresses the technical problem of overcoming the aforementioned shortcomings of the prior art by providing a method for detecting faint space targets in deep space. This method utilizes coarse-to-fine inter-frame matching to obtain the inter-frame deformation field, thereby suppressing stellar background interference. It detects faint targets that are inconsistent with local transformations for inter-frame trajectory correlation, ensuring a high detection rate for faint targets.

[0005] The technical solution of the present invention is: a method for detecting dim space targets under a deep space background, comprising:

[0006] Obtain star observation sequence images; perform large-scale median filtering on each frame of star image to obtain the background light of the star image, and subtract the background light from each frame of star image to obtain the star image with clean background I i ;

[0007] Divide the star map into u×v different regions R evenly according to the image size uv The variance of each region is used as the noise variance σ, which is binarized with 3σ as the threshold, and the connected domain and the peak value in the connected domain are obtained. The connected domain with multiple peaks is deleted to obtain the single-peak connected domain; the sum of the DN values ​​of the single-peak connected domain is used as the star point energy; for two close star points, the relatively dark one needs to be deleted to ensure that the distance between two stars in the star map is more than γ pixels, and the γ size is larger than the connected domain where the brightest star is located; all detected star points are sorted from large to small according to energy;

[0008] The stars with the highest energy ranking of Y% are regarded as bright stars. For each bright star s k , calculate the neighborhood feature expression vector of the m star points ranked top by energy in the ρ degree field of view neighborhood, that is, the m companion stars;

[0009] Based on the neighborhood feature expression vectors of all bright stars in the star map of the i-th frame and the i+1-th frame, the corresponding possible matching point set is found according to the principle of shortest Euclidean distance and mutual matching of the neighborhood feature expression vectors, that is, the initial matching point set {p}i of the star map of the i-th frame and the point set {p′}i of the star map of the i+1-th frame are obtained. +1 ;

[0010] Introducing angular distance constraints on {p} i and {p′} i+1 Screen and get the correct matching point set M i，i+1 ;

[0011] The correct matching point set M i，i+1 The transformation relationship from the star image of the i-th frame to the star image of the i+1-th frame is calculated and represented by a 3x3 single response matrix model H;

[0012] The star image of frame i i After the homography transformation and the i+1 star map I i+1 Align and get the transformed star map I i';

[0013] After transformation, star map I i ′ and the star image I of the i+1th frame i+1 The aligned parts are evenly divided into U sub-areas {b k}, k=1, ..., U and {c k}, k=1, ..., U, obtain U sub-regions to be matched, and find the star point set of each sub-region to be matched from the obtained energy-ranked star points;

[0014] Each matching area b k and c k The star point set {p} detected in k i and{q} k i+1 , confirm the matching point pairs through cross-correlation, and calculate the homography transformation model H of each area to be matched k , and adjust the positions of the matching point pairs confirmed by cross-correlation to obtain the precise matching point set of all the areas to be matched;

[0015] The point set S of the i-th frame i ′ and the previous homography transformation H are inversely transformed to obtain the point set of the star map of the i-th frame

[0016] All the precise matching point sets S i and S i+1 Calculate the star map I of the i-th frame i To the i+1th frame star map I i+1 The deformation field T;

[0017] Using the deformation field T, the star image I of the i-th frame i The pixel coordinates (x i ,y i ) are calculated to get the position of the star map in the i+1 frame (x i+1 ,y i+1 ), use interpolation to get the transformed image I i ″;

[0018] Using conventional background suppression techniques to process I i "And I i+1 , suppress the same areas between the two, and obtain the different areas, that is, the potential dark target areas. Then, through multiple frames, the potential dark target areas are associated according to the trajectory association constraints to further confirm the dark target.

[0019] The Y=5, ρ is set to one tenth of the star map field angle, and m is greater than 10.

[0020] For each bright star s k, calculate the neighborhood feature expression vector of the m star points ranked top by energy in the ρ degree field of view neighborhood, that is, the m companion stars, including:

[0021] s k Establish the x-axis with the closest companion star, and calculate the distance between all companion stars and s k The distance between {d j}, j = 1, ..., m, and s k The line connecting the companion star and the x-axis forms an angle {θ j}, j = 1, ..., m;

[0022] With Δθ as the step size, a circle is divided into 360 / Δθ intervals, and the sum of the distances of the companion stars in each interval is counted to obtain a histogram and normalize it as s k The neighborhood feature expression vector of .

[0023] The introduction of angular distance constraint {p} i and {p′} i+1 Screen and get the correct matching point set M i，i+1 ,include:

[0024] For {p} i Each point s k , in {p} i Find all the distances between the points [τ min , τ max ] degrees, select the N star points with the highest energy ranking p 1,...N ; At the same time, we also found {p′} i+1 The corresponding point s′ in the initial matching k and N star points p′ 1，...N ;

[0025] Assume s k The initial confidence is 0; p1 and p2 and s k To form a triangle, find the angular distances between the three vertices At the same time, p′1 and p′2 are k To form a triangle, find the angular distances between the three vertices. If their corresponding angular distances are all less than ∈ instantaneous field of view, then s k The confidence level increases by 1, and the ∈ is greater than the maximum distortion of the star map; for the same reason p2 and p3 are related to s k Form a triangle to verify the angular distance until p N-1 and p N With s k Form a triangle to verify the angular distance and get the final s kIf the confidence is greater than 1, it means that at least one quadrilateral angular distance constraint is satisfied, and the point is considered to be the correct matching point M(s k , s′ k );

[0026] Check in sequence and finally get the correct matching point set M i，i+1 ={M(s k , s′ k )}.

[0027] The τ min Set to one tenth of the star map field of view; τ max Set it to one-third of the star map field of view, and N is greater than 10.

[0028] The correct matching point set M i，i+1 The transformation relationship from the star image of the i-th frame to the star image of the i+1-th frame is calculated, including:

[0029] The correct matching point set M of the star image of the i-th frame and the star image of the i+1-th frame i，i+1 The pixel coordinates are normalized so that the mean of the normalized pixel coordinate distribution is 0 and the variance is 1;

[0030] Use RANSAC strategy to calculate the homography matrix H';

[0031] The homography matrix H' is denormalized to obtain the homography matrix transformation model H from the star image of the i-th frame calculated by the bright star to the star image of the i+1-th frame.

[0032] Each matching area b k and c k The star point set {p} detected in k i and{q} k i+1 , confirm the matching point pairs through cross-correlation, and calculate the homography transformation model H of each area to be matched k , and adjust the positions of the matching point pairs confirmed by cross-correlation to obtain the precise matching point set of all the areas to be matched, including:

[0033] From sub-area b k Get the a×a window area w centered at each star point p p i , from c k Get the adjacent a×a window area w centered at each point q q i ; Since it has been roughly aligned, w p i Can find its neighboring w q iPerform cross-correlation operation, then p and w with the largest correlation value q i The corresponding q is a pair of matching points; thus we get b k and c k The set of candidate matching points between ; a is a positive integer;

[0034] Use RANSAC to determine the homography transformation model H for the candidate matching point set k , and satisfy H k Correct some candidate matching point sets of the relationship to obtain the coordinates of the precise matching point set;

[0035] The precise matching point set of all areas to be matched Combined into the i-th frame I i ′'s point set To the i+1th frame I i+1 Point set

[0036] The deformation field T(a0, ..., a9, b0, ..., b9, c0, ..., c9) satisfies the following relationship:

[0037]

[0038] Among them, (x i ,y i ) is S i The pixel coordinates of the elements in (x i+1 ,y i+1 ) is S i+1 The pixel coordinates of the elements in the deformation field are obtained by iterative method to obtain the deformation field model parameters a0, ..., a9, b0, ..., b9, c0, ..., c9; the initial value of the iteration a1 = h 11 , a2=h 12 , a0=h 13 , a4~a9 are all 0, c0=h 33 , c1=h 31 , c2=h 32 , c4~c9 are all 0, b0=h 23 , b1=h 21 , b2=h 22 , b4~b9 are all 0, among which h ij is the element in the i-th row and j-th column of the homography matrix H.

[0039] The beneficial effects of the present invention compared with the prior art are:

[0040] (1) The present invention adopts a uniform bright star screening mode and feature expression method, and introduces angular distance constraints to ensure accurate matching of bright stars between frames even in the case of large inter-frame maneuvers. The calculation is simple and fast.

[0041] (2) After the bright stars are roughly aligned between frames, the present invention performs confirmation of the cross-correlation matching point set in the local area. Since the alignment is basically complete, there is no need to adjust the time-consuming cross-correlation registration parameters. Confirmation is performed only through the maximum correlation value, ensuring the one-to-one correspondence between the dark stars. The local transformation relationship is calculated and the precise coordinate position of the matching point set is updated, thereby improving the accuracy of the subsequent deformation field.

[0042] (3) The present invention adopts a coarse-to-fine deformation field calculation method, which can not only effectively solve the large maneuvering situation between frames, but also adapt to the situation of large edge distortion of large field of view due to the fitting of local area deformation, and can detect potential dark and weak targets in the first time. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of a method for detecting dim and weak space targets in a deep space background according to the present invention. DETAILED DESCRIPTION

[0044] like Figure 1 As shown, a method for detecting dim space targets in a deep space background includes:

[0045] Step (1) Obtain a star viewing sequence image; perform a large-scale median filter on each frame of the star image to obtain the background light of the star image, and subtract the background light from each frame of the star image to obtain a star image with a clean background I i .

[0046] Step (2) Divide the star map into u times v different regions R according to the image size uv The variance calculated for each region is used as the noise variance σ. This is binarized using a threshold of 3σ. Connected domains and peak values ​​within these domains are obtained. Connected domains with multiple peaks are deleted to obtain single-peak connected domains. The sum of the DN values ​​of these single-peak connected domains is calculated as the star energy. For two close-knit stars, the dimmer one is deleted. Ensure that the distance between any two stars in the star map exceeds γ pixels. γ should be larger than the size of the connected domain containing the brightest star, typically between 30 and 100 pixels. Sort all detected star points by energy, from highest to lowest.

[0047] Step (3): The stars ranked in the top 5% of energy are regarded as bright stars. For each bright star s k , calculate the neighborhood feature expression vector composed of the m star points (companion stars) ranked by energy in the neighborhood of the ρ degree field of view, ρ is set to one-tenth of the star map field of view angle, m is generally greater than 10, usually 20. The entire step (3) process is as follows:

[0048] 1)s k Establish the x-axis with the closest companion star, and calculate the distance between all companion stars and s k The distance between {dj}, j = 1, ..., m, and s k The line connecting the companion star and the x-axis forms an angle {θ j}, j = 1, ..., m.

[0049] 2) Using Δθ as the step size, divide a circle into 360 / Δθ intervals, and count the sum of the distances to the companion stars in each interval to obtain a histogram, recorded as D k , and normalized as s k The neighborhood feature expression vector of , Δθ is set to 10 degrees.

[0050] Step (4): Based on the neighborhood feature expression vectors of all bright stars in the star image of the i-th frame and the i+1-th frame, find the corresponding possible matching point set according to the principle of shortest Euclidean distance and mutual matching of the neighborhood feature expression vectors, that is, obtain the initial matching point set {p} of the star image of the i-th frame. i and the star map point set {p′} of the i+1th frame i+1 .

[0051] Step (5): Introduce angular distance constraint {p} i and {p′} i+1 Screening, the whole process of step (five) is as follows:

[0052] 1) For {p} i Each point s k , in {p} i Find all the distances between the points [τ min , τ max ] degrees, select the N star points with the highest energy ranking p 1，...N ; At the same time, we also found {p′} i+1 The corresponding point s′ in the initial matching k and N star points p′ 1，...N τ min Set to one tenth of the star map field of view, τ max Set it to one-third of the star map field of view, N should be greater than 10, usually set to 15.

[0053] 2) Let s k The initial confidence is 0. k To form a triangle, find the angular distances between the three vertices At the same time, p′1 and p′2 are k To form a triangle, find the angular distances between the three vertices. If their corresponding angular distances are all less than ∈ instantaneous field of view, then s k The confidence level of ∈ is increased by 1, and ∈ is slightly larger than the maximum distortion of the star map (unit is pixel); similarly, p2 and p3 can be used to calculate the value of s kForm a triangle to verify the angular distance until p N-1 and p N With s k Form a triangle to verify the angular distance and get the final s k If the confidence is greater than 1, it means that at least one quadrilateral angular distance constraint is satisfied, and the point is considered to be the correct matching point M(s k , s′ k ).

[0054] 3) Check in sequence and finally get the correct matching point set M i，i+1 ={M(s k , s′ k )}.

[0055] Step (six), from the correct matching point set M i，i+1 , we can calculate the transformation relationship from the star image of the i-th frame to the star image of the i+1-th frame, and the transformation relationship adopts the 3x3 homograph matrix model H. The specific process of the whole step (six) is:

[0056] 1) The correct matching point set M of the star image of the i-th frame and the star image of the i+1-th frame i，i+1 The pixel coordinates are normalized so that the mean of the normalized pixel coordinate distribution is 0 and the variance is 1;

[0057] 2) Use the RANSAC strategy to calculate the homography matrix H';

[0058] 3) Denormalize the homography matrix H' to obtain the homography matrix transformation model H from the star image of the i-th frame calculated by the bright star to the star image of the i+1-th frame.

[0059] Step (seven): the star map of frame i i After the homography transformation and the i+1 star map I i+1 Align and get the transformed star map I i ′.

[0060] Step (eight), transform the star map I i ′ and the star image I of the i+1th frame i+1 The aligned parts are evenly divided into U sub-areas {b k}, k=1, ..., U and {c k}, k=1, ..., U, obtain U sub-regions to be matched, and find the star point set of each sub-region to be matched from the energy-ranked star points obtained in step (2).

[0061] Step (9): each matching area b k and c k The star point set {p} detected in k i and{q}k i+1 , confirm the matching point pairs through cross-correlation, and calculate the homography transformation model H of each area to be matched k , and adjust the positions of the matching point pairs confirmed by cross-correlation to obtain the precise matching point set of all the areas to be matched. The specific process of the overall step (9) is as follows:

[0062] 1) From sub-area b k Get the a×a window area w centered at each star point p p i , from c k Get the adjacent a×a window area w centered at each point q q i . As the rough alignment has been completed, w p i You can find its neighboring w q i Perform cross-correlation operation, then p and w with the largest correlation value q i The corresponding q is a pair of matching points. Thus we can get b k and c k The set of candidate matching points between . a is usually set to 31.

[0063] 2) Use RANSAC to determine the homography transformation model H for the candidate matching point set k , and satisfy H k The partial candidate matching point sets of the relationship are modified to obtain the coordinates of the precise matching point set.

[0064] 3) The precise matching point set of all areas to be matched Combined into the i-th frame I i ′'s point set To the i+1th frame I i+1 Point set

[0065] Step (10): From the point set S of the i-th frame i ′ and the previous homography transformation H are inversely transformed to obtain the point set of the star map of the i-th frame

[0066] Step (11): All the precise matching point sets are matched to S i and S i+1 Calculate the star map I of the i-th frame i To the i+1th frame star map I i+1 The deformation field T(a0, ..., a9, b0, ..., b9, c0, ..., c9) satisfies the following relationship:

[0067]

[0068] Among them, (x i ,y i ) is S i The pixel coordinates of the elements in (x i+1 ,y i+1 ) is S i+1 The pixel coordinates of the elements in the deformation field are obtained by iterative method to obtain the deformation field model parameters a0, ..., a9, b0, ..., b9, c0, ..., c9. The initial value of the iteration a1 = h 11 , a2=h 12 , a0=h 13 , a4~a9 are all 0, c0=h 33 , c1=h 31 , c2=h 32 , c4~c9 are all 0, b0=h 23 , b1=h 21 , b2=h 22 , b4~b9 are all 0, among which h ij is the element in the i-th row and j-th column of the homography matrix H.

[0069] Step (12): Based on the deformation field T, the star image I of the i-th frame is i The pixel coordinates (x i ,y i ) can calculate the position of the star map in the i+1 frame (x i+1 ,y i+1 ), the transformed image I can be obtained by interpolation i ″, compared to I i ′,I i ″The i+1 frame star map I can be well aligned locally i+1 .

[0070] Step (thirteen), using conventional background suppression technology to process I i "And I i+1 , the same areas of the two are suppressed, and the different areas are obtained, that is, the potential dark target areas. Then, the potential dark target areas are associated with each other according to the trajectory association constraints through multiple frames, and the dark target can be further confirmed.

[0071] The above description is only the best specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for detecting dim space targets in deep space, characterized in that: include: Acquire stargazing sequence images; Perform large-scale median filtering on each frame of star image to obtain the background light of the star image, and subtract the background light from each frame of star image to obtain the star image with clean background I i ; Divide the star map into u×v different regions R evenly according to the image size uv The variance of each region is used as the noise variance σ, which is binarized with 3σ as the threshold, and the connected domain and the peak value in the connected domain are obtained. The connected domain with multiple peaks is deleted to obtain the single-peak connected domain; the sum of the DN values ​​of the single-peak connected domain is used as the star point energy; for two close star points, the relatively dark one needs to be deleted to ensure that the distance between two stars in the star map is more than γ pixels, and the γ size is larger than the connected domain where the brightest star is located; all detected star points are sorted from large to small according to energy; The stars with the highest energy ranking of Y% are regarded as bright stars. For each bright star s k , calculate the neighborhood feature expression vector of the m star points ranked top by energy in the ρ degree field of view neighborhood, that is, the m companion stars; Based on the neighborhood feature expression vectors of all bright stars in the star image of the i-th frame and the i+1-th frame, the corresponding possible matching point set is found according to the principle of shortest Euclidean distance and mutual matching of the neighborhood feature expression vectors, that is, the point set {p} of the i-th frame star image is obtained. i and the star map point set {p′} of the i+1th frame i+1 ; Introduction Angular distance constraint pair {p} i and {p′} i+1 Screen and get the correct matching point set M i，i+1 ; The correct matching point set M i，i+1 The transformation relationship from the star image of the i-th frame to the star image of the i+1-th frame is calculated and represented by a 3x3 single response matrix model H; The star image of frame i i After the homography transformation and the i+1 star map I i+1 Align and get the transformed star map I i '; After transformation, star map I i ′ and the star image I of the i+1th frame i+1 The aligned parts are evenly divided into U sub-areas {b k }, k=1, ..., U and {c k }, k=1, ..., U, obtain U sub-regions to be matched, and find the star point set of each sub-region to be matched from the obtained energy-ranked star points; Each matching area b k and c k The star point set {p} detected in k i and{q} k i+1 , confirm the matching point pairs through cross-correlation, and calculate the homography transformation model H of each area to be matched k , and adjust the positions of the matching point pairs confirmed by cross-correlation to obtain the precise matching point set of all the areas to be matched; The point set S of the i-th frame i ′ and the previous homography transformation H are inversely transformed to obtain the point set of the star map of the i-th frame All the precise matching point sets S i and S i+1 Calculate the star map I of the i-th frame i To the i+1th frame star map I i+1 The deformation field T; Using the deformation field T, the star image I of the i-th frame i The pixel coordinates (x i ,y i ) are calculated to get the position of the star map in the i+1 frame (x i+1 ,y i+1 ), use interpolation to get the transformed image I i ″; Using conventional background suppression techniques to process I i "And I i+1 , suppress the same areas between the two, and obtain the different areas, that is, the potential dark target areas. Then, through multiple frames, the potential dark target areas are associated according to the trajectory association constraints to further confirm the dark target.

2. The method for detecting faint space targets in deep space background according to claim 1, characterized in that: The Y=5, ρ is set to one tenth of the star map field of view, and m is greater than 10.

3. The method for detecting faint space targets in deep space background according to claim 1, characterized in that: For each bright star s k , calculate the neighborhood feature expression vector of the m star points ranked top by energy in the ρ degree field of view neighborhood, that is, the m companion stars, including: s k Establish the x-axis with the closest companion star, and calculate the distance between all companion stars and s k The distance between {d j }, j = 1, ..., m, and s k The line connecting the companion star and the x-axis forms an angle {θ j }, j = 1, ..., m; With Δθ as the step size, a circle is divided into 360 / Δθ intervals, and the sum of the distances of the companion stars in each interval is counted to obtain a histogram and normalize it as s k The neighborhood feature expression vector of .

4. The method for detecting faint space targets in deep space background according to claim 1, characterized in that: The introduction of angular distance constraint {p} i and {p′} i+1 Screen and get the correct matching point set M i，i+1 ,include: For {p} i Each point s k , in {p} i Find all the distances between the points [τ min , τ max ] degrees, select the N star points with the highest energy ranking p 1，...N ; At the same time, we also found {p′} i+1 The corresponding point s′ in the initial matching k and N star points p′ 1，...N ; Assume s k The initial confidence is 0; p1 and p2 and s k To form a triangle, find the angular distances between the three vertices At the same time, p′1 and p′2 are k To form a triangle, find the angular distances between the three vertices. If their corresponding angular distances are all less than ∈ instantaneous field of view, then s k The confidence level increases by 1, and the ∈ is greater than the maximum distortion of the star map; for the same reason p2 and p3 are related to s k Form a triangle to verify the angular distance until p N-1 and p N With s k Form a triangle to verify the angular distance and get the final s k If the confidence is greater than 1, it means that at least one quadrilateral angular distance constraint is satisfied, and the point is considered to be the correct matching point M(s k , s′ k ); Check in sequence and finally get the correct matching point set M i，i+1 ={M(s k , s′ k )}.

5. The method for detecting faint space targets in deep space background according to claim 4, characterized in that: The τ min Set to one tenth of the star map field of view; τ max Set it to one-third of the star map field of view, and N is greater than 10.

6. The method for detecting faint space targets in deep space background according to claim 4, characterized in that: The correct matching point set M i，i+1 The transformation relationship from the star image of the i-th frame to the star image of the i+1-th frame is calculated, including: The correct matching point set M of the star image of the i-th frame and the star image of the i+1-th frame i，i+1 The pixel coordinates are normalized so that the mean of the normalized pixel coordinate distribution is 0 and the variance is 1; Use RANSAC strategy to calculate the homography matrix H'; The homography matrix H' is denormalized to obtain the homography matrix transformation model H from the star image of the i-th frame calculated by the bright star to the star image of the i+1-th frame.

7. The method for detecting faint space targets in deep space background according to claim 6, characterized in that: Each matching area b k and c k The star point set {p} detected in k i and{q} k i+1 , confirm the matching point pairs through cross-correlation, and calculate the homography transformation model H of each area to be matched k , and adjust the positions of the matching point pairs confirmed by cross-correlation to obtain the precise matching point set of all the areas to be matched, including: From sub-area b k Get the a×a window area w centered at each star point p p i , from c k Get the adjacent a×a window area w centered at each point q q i ; Since it has been roughly aligned, w p i Can find its neighboring w q i Perform cross-correlation operation, then p and w with the largest correlation value q i The corresponding q is a pair of matching points; thus we get b k and c k The set of candidate matching points between ; a is a positive integer; Use RANSAC to determine the homography transformation model H for the candidate matching point set k , and satisfy H k Correct some candidate matching point sets of the relationship to obtain the coordinates of the precise matching point set; The precise matching point set of all areas to be matched Combined into the i-th frame I i ′'s point set To the i+1th frame I i+1 Point set 8. The method for detecting faint space targets in deep space background according to claim 7, characterized in that: The deformation field T(a0, ..., a9, b0, ..., b9, c0, ..., c9) satisfies the following relationship: Among them, (x i ,y i ) is S i The pixel coordinates of the elements in (x i+1 ,y i+1 ) is S i+1 The pixel coordinates of the elements in the deformation field are obtained by iterative method to obtain the deformation field model parameters a0, ..., a9, b0, ..., b9, c0, ..., c9; the initial value of the iteration a1 = h 11 , a2=h 12 , a0=h 13 , a4~a9 are all 0, c0=h 33 , c1=h 31 , c2=h 32 , c4~c9 are all 0, b0=h 23 , b1=h 21 , b2=h 22 , b4~b9 are all 0, among which h ij is the element in the i-th row and j-th column of the homography matrix H.

Images