A weak and small target detection method based on super-resolution reconstruction and multi-frame accumulation

Abstract

The application discloses a weak and small target detection method based on super-resolution reconstruction and multi-frame accumulation, which comprises the following steps: S1, directly multi-frame accumulation of input image sequence; S2, setting a segmentation threshold thr1 to segment preselected targets; S3, using two-dimensional Gaussian surface fitting to reconstruct target waveform for all preselected targets, and extracting target mass center in each frame image; S4, deleting all abnormal frames from the image sequence; S5, taking the first frame image as a standard image, setting the mass center of the target in the remaining images as the mass center of the target in the standard image, calculating the gray value of the pixels occupied by the target in each frame image according to a two-dimensional Gaussian function, and realizing sub-pixel registration of the target; S6, accumulating the registered image sequence to obtain an image after second accumulation; S7, setting a segmentation threshold thr2 to segment the target; and S8, image fusion. The application can improve the influence of target frame shaking on accumulation efficiency.

Description

Technical Field

[0001] This invention belongs to the field of ground-based astronomical observation and space exploration, specifically involving a method for detecting weak targets based on super-resolution reconstruction and multi-frame accumulation. Background Technology

[0002] Currently, for multi-frame detection of weak targets, the energy accumulation method is often used. Accumulating along the target's location can increase the energy of the weak target. Multi-frame accumulation includes two types: multi-frame mean accumulation and multi-frame median accumulation ([1] Kong Sijie. Research on the extraction technology of weak targets against dense star background [D]. University of Chinese Academy of Sciences (Institute of Optoelectronics, Chinese Academy of Sciences), 2019.). The background noise of the image after multi-frame mean accumulation and multi-frame median accumulation has the following relationship with the background noise of the single-frame image before accumulation:

[0003]

[0004]

[0005] In the formula: σ individual This represents the cumulative standard deviation of background noise in a single frame of the image.

[0006] σ mean The standard deviation of background noise in the image after averaging multiple frames;

[0007] σ median The standard deviation of background noise in the image after accumulation of median values ​​from multiple frames;

[0008] N is the cumulative number of frames.

[0009] The relationship between the target signal-to-noise ratio before and after accumulation is as follows:

[0010]

[0011]

[0012] Where: SNR individual This is the cumulative target signal-to-noise ratio in the previous single frame image;

[0013] SNR mean The target signal-to-noise ratio is the sum of the averages from multiple frames.

[0014] SNR median The target signal-to-noise ratio is the cumulative median of multiple frames.

[0015] As can be seen, the target signal-to-noise ratio (SNR) is significantly improved after multi-frame averaging and multi-frame median accumulation, with the improvement being even greater after multi-frame averaging. Therefore, the multi-frame averaging algorithm is chosen to accumulate the target energy.

[0016] The multi-frame accumulation effect is affected by several factors, including target frame jitter, target brightness fluctuations, and background undulations. Under ideal conditions where there is no target frame jitter, no target brightness fluctuations, and no background undulations, the signal-to-noise ratio increase after multi-frame averaging accumulation is:

[0017]

[0018] In practice, due to fluctuations in target brightness and background, and random target jitter between frames, the cumulative efficiency of multi-frame averages is significantly reduced. In short exposure scenarios, the impact of target brightness fluctuations and background undulations is negligible; only the impact of target jitter between frames needs to be considered on the cumulative efficiency. When target jitter exists between frames, the formula for calculating the target signal-to-noise ratio increase ΔSNR relative to the number of cumulative frames is as follows:

[0019]

[0020] According to equation (6), the increase in target signal-to-noise ratio ΔSNR can be calculated when the target has different amounts of sway. When the target's inter-frame sway is x = 0.5 pixels, the calculated value is... When the target inter-frame jitter x = 1 pixel, the calculation is as follows: It is evident that as the amount of jitter between target frames increases, the efficiency of multi-frame mean accumulation continuously decreases. This results in a much worse increase in the target signal-to-noise ratio after direct multi-frame accumulation compared to the ideal situation, which may lead to the target not being effectively detected, resulting in a low detection probability for the target detection algorithm. Summary of the Invention

[0021] To overcome the shortcomings of existing solutions, this invention provides a weak target detection method based on super-resolution reconstruction and multi-frame accumulation, which avoids the impact of target frame jitter on accumulation efficiency during multi-frame accumulation. This method overcomes the defects of direct multi-frame accumulation, fully utilizes the characteristic that the actual target energy approximates a two-dimensional Gaussian distribution, reconstructs the target waveform, extracts the target centroid, and performs sub-pixel registration. Accumulating the registered image significantly improves accumulation efficiency and enhances the weak target detection capability of the accumulation algorithm.

[0022] The technical solution adopted in this invention is: a method for detecting weak targets based on super-resolution reconstruction and multi-frame accumulation, the implementation steps of which include:

[0023] Step S1: Accumulate the input image sequence directly over multiple frames to obtain the image after the first accumulation;

[0024] Step S2: Set the segmentation threshold thr1 to segment the pre-selected target;

[0025] Step S3: For all pre-selected targets, reconstruct the target waveform using two-dimensional Gaussian surface fitting, and extract the target centroid in each frame image;

[0026] Step S4: In the reconstructed two-dimensional Gaussian function, if the magnitude of the two-dimensional Gaussian function is too small or the standard deviation of the two-dimensional Gaussian function is too large, then mark the frame as an abnormal frame and delete all abnormal frames from the image sequence.

[0027] Step S5: Using the first frame image as the standard image, set the centroid of the target in the remaining images as the centroid of the target in the standard image, and calculate the gray value of the pixels occupied by the target in each frame image at this time according to the two-dimensional Gaussian function to achieve sub-pixel registration of the target.

[0028] Step S6: Accumulate the registered image sequence to obtain the second accumulated image;

[0029] Step S7: Set the segmentation threshold thr2 to segment the target;

[0030] Step S8: Image fusion to obtain the final detection result.

[0031] Furthermore, the preprocessing steps in step S1 are as follows:

[0032] Step 21: Input a sequence f containing N frames of infrared images i For i = 1, 2, ..., N, the image sequence is directly accumulated by multi-frame mean to obtain the first accumulated image Z0. The relationship between the target signal-to-noise ratio before and after accumulation is as follows:

[0033]

[0034] In the formula: σ individual This represents the cumulative standard deviation of background noise in a single frame of the image.

[0035] σ mean The standard deviation of background noise in the image after averaging multiple frames;

[0036] σ median The standard deviation of background noise in the image after accumulation of median values ​​from multiple frames;

[0037] N is the cumulative number of frames.

[0038] Furthermore, the pre-selected target segmentation step in step S2 is as follows:

[0039] Step 31: Based on the required target magnitude, set the pre-selected target segmentation threshold thr1. Since the target has inter-frame jitter which leads to low direct accumulation efficiency, the threshold thr1 should be set to a small value, such as thr1=2.

[0040] Furthermore, the two-dimensional Gaussian surface fitting steps in step S3 are as follows:

[0041] Step 41: Assume f i (x,y) is the image f i The gray values ​​located at (x, y), i = 1, 2, ..., M, where M is the number of pre-selected targets segmented in step S2, (x ij ,y ij Let be the center of the i-th preselected target in the j-th frame image. The Gaussian surface fitting function can be expressed as:

[0042]

[0043] Where: σ ij The mean square error is a Gaussian function and is related to the object distance, focal length, defocusing amount, and aperture size.

[0044] K represents the total energy projected onto the CCD plane by the light target, which is related to the brightness of the light target and the exposure time.

[0045] Because the unknowns contained in equation (2) are A ij , σ ij x ij y ij There are 4 in total, therefore, at least four target pixels are needed to calculate the Gaussian surface fitting function;

[0046] Step 42: Taking the logarithm of both sides of equation (2), we have:

[0047]

[0048] Let a = 1 / 2σ ij 2 b = x ij / σ ij 2 c = y ij / σ ij 2 d=(x ij 2 +y ij 2 ) / 2σ ij 2 -lnA ij Then equation (3) can be equivalent to:

[0049] lnf i (x,y)=a·(x 2 +y 2 )+b·x+c·yd (4)

[0050] Assuming the target size is 3×3, with a total of 9 pixels, then we can let:

[0051]

[0052] The system of equations is obtained as follows:

[0053] Y = AB (6)

[0054] The error column vector of the target's 9 pixels is:

[0055] E = Y - AB (7)

[0056] The parameters that minimize the mean square error are obtained using the least squares method:

[0057]

[0058] Perform QR decomposition on A, i.e., A = QR. After decomposition, Q is a 9×9 orthogonal matrix and R is a 9×4 upper triangular matrix. The derivation of E = Y - AB is as follows:

[0059]

[0060] In the formula, Q is an orthogonal matrix, then:

[0061]

[0062] make Where S is a 4-dimensional column vector, T is a 5-dimensional column vector, and R1 is a 4×4 upper triangular matrix, then:

[0063]

[0064] When S = R1B, MSE reaches its minimum value. Therefore, by solving equation (12), the unknown parameters a, b, c, and d can be obtained.

[0065]

[0066] After solving for a, b, c, and d, substitute them into a = 1 / 2σ ij 2 b = x ij / σ ij 2 c = y ij / σ ij 2 d = (x ij 2 +y ij 2 ) / 2σ ij 2 -lnA ij The unknown parameter A in the two-dimensional Gaussian surface function can then be obtained. ijx ij y ij σ ij ;

[0067] Step 43: Calculate the parameters A of the two-dimensional Gaussian surface function obtained in Step 42. ij x ij y ij σ ij Substituting the solution into equation (2), we can obtain the two-dimensional Gaussian surface fitting result of the i-th pre-selected target in the j-th frame image.

[0068] Furthermore, the abnormal frame exclusion step in step S4 is as follows:

[0069] Step 51: For the i-th pre-selected target, after step S3, N different parameters A can be obtained. ij and σ ij , j = 1,...N, set parameter A ij Too small a frame or σ ij Frames that are too large are collectively referred to as abnormal frames. Abnormal frames are deleted from the original image sequence, and the resulting image sequence after deleting abnormal frames is denoted as I. i For i = 1, 2, ..., M, the abnormal frame determination method is as follows:

[0070]

[0071] In the formula: This represents all parameters A obtained by calculating the i-th target in step S3. ij The mean;

[0072] This represents all parameters σ obtained by calculating the i-th target in step S3. ij The mean;

[0073] t1 and t2 are abnormal frame detection parameters, with values ​​ranging from 0 to 1.

[0074] Furthermore, the sub-pixel registration step in step S5 is as follows:

[0075] Step 61: Assuming that after excluding abnormal frames in step S4, the image sequence I of the i-th pre-selected target... i There are N1 frames remaining. (I) i The first frame image is the standard image; the centroid of the target in the remaining images is modified to (x... i1 ,y i1 );

[0076] Step 62: In image I ik The center pixel coordinates are obtained by rounding to the nearest integer (x...). kp ,y kp), k=2,...,N1, Δx=x i1 -x kp Δy=y i1 -y kp This represents the deviation of the pixel coordinates;

[0077] Step 63: Assuming the target size is 3×3, then for the pixel (x) containing the target... ab ,y ab The formula for calculating the grayscale of a given value is:

[0078]

[0079] In the formula: a and b represent the pixel range of the target, a = b = 1, 2, 3;

[0080] Step 64: For the image sequence I of the i-th target i Each frame of the image is processed using steps 62 and 63 to obtain the target centroid, which is (x... i1 ,y i1 Image sequence I after subpixel registration i '.

[0081] Furthermore, the second image accumulation step in step S6 is as follows:

[0082] Step 71: For image I i 'Multi-frame accumulation yields the accumulated image Z' i , i = 1, 2, ..., M.

[0083] Furthermore, the target segmentation steps in step S7 are as follows:

[0084] Step 81: Set the pre-selected target segmentation threshold thr2. Since the target waveform was reconstructed, the target centroid was extracted, and sub-pixel registration was performed after two-dimensional Gaussian fitting, the accumulation efficiency of the second multi-frame accumulation was significantly improved. The threshold thr2 should be set to a value larger than thr1, such as thr2=3.

[0085] Furthermore, the image fusion step in step S8 is as follows:

[0086] Step 91: Process the M pre-selected targets segmented in step S2 using steps S3-S7. This will result in M ​​binary images. By overlaying these M binary images, the final detection result can be obtained.

[0087] The principle of this invention is as follows: the actual target energy distribution approximates a two-dimensional Gaussian distribution. The target waveform can be reconstructed using a two-dimensional Gaussian surface fitting, the centroid extracted, and sub-pixel registration performed on the target can mitigate the impact of inter-frame jitter on multi-frame accumulation efficiency. This invention's weak target detection method based on super-resolution reconstruction and multi-frame accumulation fully utilizes the characteristic that the target energy distribution approximates a two-dimensional Gaussian distribution, achieving sub-pixel registration of weak targets and improving the ability of the accumulation algorithm to detect weak targets.

[0088] The advantages of this invention compared to the prior art are:

[0089] (1) The weak target detection algorithm based on super-resolution reconstruction and multi-frame accumulation in this invention overcomes the defect of direct multi-frame accumulation being severely affected by the shaking between target frames, and improves the detection probability of target detection.

[0090] (2) The weak target detection algorithm based on super-resolution reconstruction and multi-frame accumulation in this invention makes full use of the characteristic that the target energy distribution is approximately a two-dimensional Gaussian distribution, realizes sub-pixel matching of the target, and improves the efficiency of multi-frame accumulation.

[0091] (3) The weak target detection algorithm based on super-resolution reconstruction and multi-frame accumulation of this invention can detect darker targets compared with the direct multi-frame accumulation algorithm after improving the efficiency of multi-frame accumulation. Attached Figure Description

[0092] Figure 1 This is a flowchart of a weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to the present invention;

[0093] Figure 2 This is a schematic diagram of the original image of the present invention;

[0094] Figure 3 This is a schematic diagram of the direct cumulative results of the present invention;

[0095] Figure 4 This is a schematic diagram of the pre-selected target segmentation results of the present invention;

[0096] Figure 5 This is a schematic diagram of the target waveform after reconstruction by two-dimensional Gaussian surface fitting according to the present invention;

[0097] Figure 6 This is a schematic diagram of the second multi-frame accumulation result of the present invention;

[0098] Figure 7 This is a schematic diagram of the target segmentation result of the present invention;

[0099] Figure 8 This is a schematic diagram of the fused image of the present invention; Detailed Implementation

[0100] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0101] like Figure 1 As shown, this invention provides a weak target detection method based on super-resolution reconstruction and multi-frame accumulation, applicable to target imaging by ground-based telescopes, reducing the impact of target frame jitter on multi-frame accumulation efficiency. Specifically, the weak target detection method based on super-resolution reconstruction and multi-frame accumulation includes the following steps:

[0102] Step S1: Accumulate the input image sequence directly over multiple frames to obtain the image after the first accumulation;

[0103] Step S2: Set the segmentation threshold thr1 to segment the pre-selected target;

[0104] Step S3: For all pre-selected targets, reconstruct the target waveform using two-dimensional Gaussian surface fitting, and extract the target centroid in each frame image;

[0105] Step S4: In the reconstructed two-dimensional Gaussian function, if the magnitude of the two-dimensional Gaussian function is too small or the standard deviation of the two-dimensional Gaussian function is too large, then mark the frame as an abnormal frame and delete all abnormal frames from the image sequence.

[0106] Step S5: Using the first frame image as the standard image, set the centroid of the target in the remaining images as the centroid of the target in the standard image, and calculate the gray value of the pixels occupied by the target in each frame image at this time according to the two-dimensional Gaussian function to achieve sub-pixel registration of the target.

[0107] Step S6: Accumulate the registered image sequences to obtain the second accumulated image;

[0108] Step S7: Set the segmentation threshold thr2 to segment the target;

[0109] Step S8: Image fusion to obtain the final detection result.

[0110] The process of the method of the present invention is as follows: Figure 1 As shown. Specifically, the weak target detection method based on super-resolution reconstruction and multi-frame accumulation of the present invention includes the following steps:

[0111] 1. Original infrared image as follows Figure 2 As shown, N consecutive infrared images are taken, and the average of the multiple frames is directly accumulated to obtain the accumulated image Z0, as shown. Figure 3 As shown.

[0112] 2. Set the pre-selected target segmentation threshold thr1. Due to inter-frame jitter in the target, direct accumulation is inefficient; therefore, the threshold thr1 should be set to a small value, such as thr1 = 2. The pre-selected target segmentation result is as follows: Figure 4 As shown in the figure, the targets marked with white boxes are the pre-selected targets, including real targets and noise.

[0113] 3. For the j-th pre-selected target segmented in step 2, reconstruct the target waveform using a two-dimensional Gaussian surface fitting method and extract the target centroid. The reconstructed target waveform is as follows: Figure 5 As shown.

[0114] 4. For the j-th pre-selected target, after step 3, N different parameters A can be obtained. ij and σ ij j = 1,...N. The parameter A... ij Too small a frame or σ ij Frames that are too large are collectively referred to as abnormal frames. Abnormal frames are deleted from the original image sequence, and the resulting image sequence after deleting abnormal frames is denoted as I. i Let i = 1, 2, ..., M. The method for determining abnormal frames is as follows:

[0115]

[0116] In the formula: A i This represents all parameters A obtained by calculating the i-th target in step S3. ij The mean;

[0117] This represents all parameters σ obtained by calculating the i-th target in step S3. ij The mean;

[0118] t1 and t2 are abnormal frame detection parameters, with values ​​ranging from 0 to 1.

[0119] 5. Assuming that after excluding abnormal frames in step 4, the image sequence I of the i-th pre-selected target is... i There are N1 frames remaining. (I) i The first frame image is the standard image; the centroid of the target in the remaining images is modified to (x... i1 ,y i1 In image I ik The center pixel coordinates are obtained by rounding to the nearest integer (x...). kp ,y kp ), k=2,...,N1, Δx=x i10 -x kp Δy=y i10 -y kp , which represents the deviation of the pixel coordinates. Assuming the target size is s×s, then for the pixel (x, y) containing the target... ab ,y ab The formula for calculating the grayscale of a given value is:

[0120]

[0121] In the formula: a and b represent the pixel range of the target, a = b = 1, 2, ..., s.

[0122] Image sequence I for the i-th target i Perform the above processing on each frame of the image to obtain a target centroid that is (x... i1 ,y i1 Image sequence I after subpixel registration i '.

[0123] 6. For image sequence I i Accumulation is performed to obtain the second accumulated image Z. i ,like Figure 6 As shown.

[0124] 7. Set the segmentation threshold thr2. Since steps 3 to 5 have improved the efficiency of multi-frame accumulation, the segmentation threshold thr2 should be set to a value larger than thr1, such as thr2 = 3. The segmentation result is as follows. Figure 7 As shown in the figure, the targets marked with white boxes are the segmented targets.

[0125] 8. After processing the M pre-selected targets through the above steps, M target segmentation binary images will be obtained. By overlaying these M binary images, the final detection result can be obtained, such as... Figure 8 As shown in the figure, the targets marked with white boxes are the finally detected targets.

[0126] To compare the advantages of the weak target detection algorithm based on super-resolution reconstruction and multi-frame accumulation of this invention compared with the direct multi-frame accumulation algorithm, target detection was performed on the same image data using both the algorithm of this invention and the direct multi-frame accumulation algorithm. The changes in the mean signal-to-noise ratio (MSNR) of the target before and after processing by the two algorithms are shown in Table 1:

[0127] Table 1. Comparison of target MSNR before and after processing by the algorithm of this invention and the direct multi-frame accumulation algorithm.

[0128]

[0129] As shown in Table 1, after processing with the weak target detection algorithm based on super-resolution reconstruction and multi-frame accumulation of this invention, the mean signal-to-noise ratio of the target increased from 0.3974 to 3.1823. However, without further processing of the original image, direct multi-frame accumulation resulted in a target signal-to-noise ratio that increased from 0.3974 to 2.1594. Comparing the algorithm of this invention with the direct multi-frame accumulation algorithm, the algorithm of this invention can improve the low accumulation efficiency caused by inter-frame jitter, increasing the accumulation efficiency by approximately 50%, and greatly enhancing the ability of the accumulation algorithm to detect weak targets.

Claims

1. A method for detecting weak targets based on super-resolution reconstruction and multi-frame accumulation, characterized in that, The steps to implement this method include: Step S1: Accumulate the input image sequence directly over multiple frames to obtain the image after the first accumulation; Step S2: Set the segmentation threshold thr1 to segment the pre-selected target; Step S3: For all pre-selected targets, reconstruct the target waveform using two-dimensional Gaussian surface fitting, and extract the target centroid in each frame image; Step S4: In the reconstructed two-dimensional Gaussian function, if the magnitude of the two-dimensional Gaussian function is too small or the standard deviation of the two-dimensional Gaussian function is too large, then mark the frame as an abnormal frame and delete all abnormal frames from the image sequence. Step S5: Using the first frame image as the standard image, set the centroid of the target in the remaining images as the centroid of the target in the standard image, and calculate the gray value of the pixels occupied by the target in each frame image at this time according to the two-dimensional Gaussian function to achieve sub-pixel registration of the target. Step S6: Accumulate the registered image sequences to obtain the second accumulated image; Step S7: Set the segmentation threshold thr2 to segment the target; Step S8: Image fusion to obtain the final detection result; The steps for fitting the two-dimensional Gaussian surface in step S3 are as follows: Step 41: Assume f i (x,y) is the image f i The gray values ​​located at (x, y), i = 1, 2, ..., M, where M is the number of pre-selected targets segmented in step S2, (x ij ,y ij Let be the center of the i-th pre-selected target in the j-th frame image, and the Gaussian surface fitting function is expressed as: (2) Where: σ ij The mean square error is a Gaussian function and is related to the object distance, focal length, defocusing amount, and aperture size. K is the total energy projected onto the CCD plane by the light target, which is related to the brightness of the light target and the exposure time. Because the unknowns contained in equation (2) are A ij , σ ij x ij y ij There are 4 in total, therefore, at least four target pixels are needed to calculate the Gaussian surface fitting function; Step 42: Taking the logarithm of both sides of equation (2), we have: (3) Let a = 1 / 2σ ij 2 b=x ij / σ ij 2 c=y ij / σ ij 2 d=(x ij 2 +y ij 2 ) / 2σ ij 2 -lnA ij Then equation (3) can be equivalent to: (4) Assuming the target size is 3×3, with a total of 9 pixels, then we can let: ， ， (5) The system of equations is obtained as follows: (6) The error column vector of the target's 9 pixels is: (7) The parameters that minimize the mean square error are obtained using the least squares method: (8) Perform QR decomposition on A, i.e., A = QR. After decomposition, Q is a 9×9 orthogonal matrix and R is a 9×4 upper triangular matrix. The derivation of E = Y - AB is as follows: (9) In the formula, Q is an orthogonal matrix, then: (10) make , Where S is a 4-dimensional column vector, T is a 5-dimensional column vector, and R1 is a 4×4 upper triangular matrix, then: (11) When S=R1B, MSE reaches its minimum value. Therefore, by solving equation (12), the unknown parameters a, b, c, and d can be obtained. (12) After solving for a, b, c, and d, substitute them into a = 1 / 2σ ij 2 b=x ij / σ ij 2 c=y ij / σ ij 2 d=(x ij 2 +y ij 2 ) / 2σ ij 2 -lnA ij The unknown parameter A in the two-dimensional Gaussian surface function can then be obtained. ij x ij y ij σ ij ; Step 43: Calculate the parameters A of the two-dimensional Gaussian surface function obtained in Step 42. ij x ij y ij σ ij Substituting the solution into equation (2), we can obtain the two-dimensional Gaussian surface fitting result of the i-th pre-selected target in the j-th frame image.

2. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The preprocessing steps in step S1 are as follows: Step 21: Input a sequence f containing N frames of infrared images i For i=1,2,...,N, the image sequence is directly accumulated by multi-frame mean to obtain the first accumulated image Z0. The relationship between the target signal-to-noise ratio before and after accumulation is as follows: (1) Where: SNR individual This is the cumulative target signal-to-noise ratio in the previous single frame image; SNR mean The target signal-to-noise ratio in the image after averaging multiple frames; N is the cumulative number of frames.

3. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The pre-selected target segmentation steps in step S2 are as follows: Step 31: Based on the required target magnitude, set the pre-selected target segmentation threshold thr1. Since the target has inter-frame jitter, the direct accumulation efficiency is low, so set the threshold thr1=2.

4. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The abnormal frame exclusion steps in step S4 are as follows: Step 51: For the i-th pre-selected target, after step S3, N different parameters A can be obtained. ij and σ ij , j=1,...N, set parameter A ij Too small a frame or σ ij Frames that are too large are collectively referred to as abnormal frames. Abnormal frames are deleted from the original image sequence, and the resulting image sequence after deleting abnormal frames is denoted as I. i For i=1,2,...,M, the abnormal frame determination method is as follows: (13) In the formula: i This represents all parameters A obtained by calculating the i-th target in step S3. ij The mean; i This represents all parameters σ obtained by calculating the i-th target in step S3. ij The mean; t1 and t2 are abnormal frame detection parameters, with values ​​ranging from 0 to 1.

5. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The sub-pixel registration steps in step S5 are as follows: Step 61: Assuming that after excluding abnormal frames in step S4, the image sequence I of the i-th pre-selected target... i There are N1 frames remaining; with I i The first frame image is the standard image; the centroid of the target in the remaining images is modified to (x... i1 ,y i1 ); Step 62: In image I ik The center pixel coordinates are obtained by rounding to the nearest integer (x...). kp ,y kp ), k=2,...,N1, Δx = x i1 - x kp Δy = y i1 - y kp This represents the deviation of the pixel coordinates; Step 63: Assuming the target size is 3×3, then for the pixel (x) containing the target... ab ,y ab The formula for calculating the grayscale of a given value is: (14) In the formula: a and b represent the pixel range of the target, a=b=1,2,3; Step 64: For the image sequence I of the i-th target i Each frame of the image is processed using steps 62 and 63 to obtain the target centroid, which is (x... i1 ,y i1 Image sequence I after subpixel registration i '.

6. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The second image accumulation step in step S6 is as follows: Step 71: For image I i 'Multi-frame accumulation yields the accumulated image Z' i , i=1,2,...,M.

7. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The target segmentation steps in step S7 are as follows: Step 81: Set the pre-selected target segmentation threshold thr2. Since the target waveform was reconstructed, the target centroid was extracted, and sub-pixel registration was performed after two-dimensional Gaussian fitting, the accumulation efficiency of the second multi-frame accumulation was significantly improved. The threshold thr2 should be set to a value larger than thr1, such as thr2=3.

8. The weak target detection method based on super-resolution reconstruction and multi-frame accumulation according to claim 1, characterized in that, The image fusion steps in step S8 are as follows: Step 91: Process the M pre-selected targets segmented in step S2 using steps S3-S7. This will result in M ​​binary images. By overlaying these M binary images, the final detection result can be obtained.

Images