A Detection Method for False Alarm Sources of Cirrus Clouds with Multi-Band Sparse Feature Fusion

Abstract

The present invention relates to the technical field of image processing and machine vision, and proposes a method for detecting cirrus-like false alarm sources by fusing multi-band sparse features, aiming to solve the problem that the cirrus-like false alarm source detection system cannot accurately detect cirrus with sparsity. To this end, the present invention proposes a method for detecting cirrus-like false alarm sources by fusing multi-band sparse features, which captures the sparse features and details of cirrus and is closer to the imaging essence of cirrus-like false alarm sources. The main techniques include: S1, reading in multi-band original remote sensing images; S2, constructing tensor blocks of each band by moving a local window; S3, performing low-rank sparse decomposition on the constructed tensor blocks of each band, and transforming the problem of detecting cirrus-like false alarm sources into the problem of recovering and optimizing a cyclic tensor; S4, reconstructing cirrus-like false alarm sources; S5, fusing and binarizing the cirrus detection results of each band to obtain the final cirrus detection result.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method for detecting cirrus-like false alarm sources. Background Art

[0002] In remote sensing and military applications, the detection of small and weak infrared targets plays a very important role. However, when the earth observation system images, clouds will cover the earth's surface, resulting in problems such as cloud occlusion and interference in the images for detecting the area of interest, making it very difficult to detect the area of interest. In addition, high-altitude cirrus clouds will generate strong radiation in earth observation, interfering with the detection of infrared targets. At the same time, cirrus clouds with sizes and gray levels similar to those of targets will also cause serious false alarm problems in infrared target detection. Therefore, cirrus cloud detection has become an important step in remote sensing image analysis and processing.

[0003] Existing algorithms can detect large-area cirrus-like false alarm sources, while cirrus-like false alarm sources with gray levels and sizes similar to those of targets and having sparsity are often regarded as noise and removed, resulting in missed detection of false alarm sources. At the same time, most current cirrus cloud detection algorithms are only performed on single-band infrared images, so when the image source provides multi-band information, the cirrus cloud information in other bands is not fully utilized. For this reason, the present invention adopts a method for detecting cirrus-like false alarm sources by fusing multi-band sparse features, which can accurately detect cirrus-like false alarm sources with sparsity while using multi-band information. Summary of the Invention

[0004] In view of the above technical problems, the present invention provides a method for detecting cirrus-like false alarm sources by fusing multi-band sparse features, aiming to solve the problem of high missed detection rate in the current detection algorithms for cirrus-like false alarm sources.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A method for detecting cirrus-like false alarm sources by fusing multi-band sparse features includes the following steps:

[0007] S1. Read in the multi-band original infrared remote sensing image f i ;

[0008] S2. Traverse and process the multi-band original infrared remote sensing image f i through a window of size ws×ws, divide it into l image blocks, and obtain the structure tensor blocks of each band of the multi-band original infrared remote sensing image f i , that is, form an original block tensor block of size ws×ws×l

[0009] S3. Decomposition of the local structure tensor: According to the infrared block tensor model, decompose the original block tensor blocks constructed in step S2 by means of low-rank and sparse iterative decomposition based on the circular tensor structure to obtain the decomposed image blocks, where the decomposed image blocks include background tensor blocks and cirrus-like false alarm source tensor blocks

[0010] S4. Reconstruction of the local structure block tensor: Reconstruct the background tensor blocks and cirrus-like false alarm source tensor blocks obtained in step S3 for each band respectively. The reconstruction steps are opposite to those of constructing the structure tensor blocks for each band in step S2. Stitch the divided background tensor blocks and cirrus-like false alarm source tensor blocks in the order in the multi-band original infrared remote sensing image f i to obtain the background image B i and cirrus-like false alarm source image C i for each band;

[0011] S5. Perform multi-band fusion based on weighted average on the cirrus-like false alarm source images C i obtained in step S4 for each band to obtain a multi-band fusion image I. Perform adaptive threshold segmentation on the fusion image I to obtain the final cirrus-like false alarm source detection image D.

[0012] Furthermore, the multi-band original infrared remote sensing image the original block tensor blocks for each band the background tensor blocks the cirrus-like false alarm source tensor blocks the background images for each band the cirrus-like false alarm source images the multi-band fusion image the cirrus-like false alarm source detection image

[0013] where, represents the Euclidean space, m represents the length of the original infrared remote sensing image in the i-th band, n represents the width of the original infrared remote sensing image in the i-th band, i represents the band number, ws×ws represents the window size, and l represents the number of image blocks obtained by dividing the original infrared remote sensing image in the i-th band by the window, represents the Euclidean space, and i represents the band number.

[0014] Furthermore, step S3 specifically includes the following steps:

[0015] S3.1. The original block tensor blocks for each band of the original infrared remote sensing image f i of The expression of

[0016]

[0017] In formula (1), represents the background class tensor block of the original infrared remote sensing image in the i-th band, represents the cirrus cloud class false alarm source tensor block of the original infrared remote sensing image in the i-th band;

[0018] Based on formula (1), first set the number of iterations or cut-off conditions, and solve the objective function through the iterative optimization method to obtain the background tensor block and the cirrus cloud class false alarm source tensor block The objective function includes:

[0019]

[0020] In formula (2), s.t. is the abbreviation of "subject to" in mathematics, representing "subject to" certain constraint conditions; is the i-th band tensor block of the d-th modality, represents the expansion form based on the cyclic tensor when the i-th band tensor block of the d-th modality is at position L. Among them, K is the total dimension of the tensor block, d is the modality number, L is the cyclic tensor decomposition position, K = 3, d ∈ {1,..., K}, L = 1, 2, 3; * represents the nuclear norm, and the background is suppressed by combining the nuclear norm as a regularization term with the cyclic tensor expansion mode; w represents the constraint regularization factor for the background weight, w d represents the constraint regularization factor for the background weight, w d The value range of is 0 - 1, and at the same time, it needs to satisfy λ d represents the constraint regularization factor for the cirrus cloud class false alarm source, and P is the sampling rate, P = 1, S i is the size of the tensor slice on the d-th modality, S i ∈ {ws, l}, is the maximum value of the size of the tensor slice of the d-th mode tensor, i represents the band number, ws represents the window size, and l represents the number of image blocks into which the original infrared remote sensing image in the i-th band is divided by the window; is the sampling operator, T = S1 ×... × S K , T is the total number of samples, represents the Euclidean space corresponding to the current tensor block;

[0021] The augmented Lagrangian function corresponding to formula (2) is:

[0022]

[0023] In Equation (3), w d represents the constraint regularization factor for the background weight, and λ and σ i represent the penalty factors, and and v i are dual variables, K is the total dimension of the tensor block, F represents the Frobenius norm, and the Frobenius norm is used as the joint regularization term to suppress the background. The value range of σ is from 10e-8 to 10e-3, i represents the background class tensor block of the original infrared remote sensing image in the i-th band, represents the cirrus cloud false alarm source tensor block of the original infrared remote sensing image in the i-th band, * represents the nuclear norm, and l represents the number of image blocks obtained by dividing the original infrared remote sensing image in the i-th band by the window;

[0024] S3.2. Transform the model in Equation (3) into five sub-problems by the alternating direction method of multipliers (ADMM), and finally obtain the background block tensor and the cirrus cloud false alarm source tensor block

[0025] Furthermore, the iterative methods for the five sub-problems are as follows:

[0026] a. Iterative method for variable :

[0027] Fix the background tensor block and the cirrus cloud false alarm source tensor block Convert into the equation for solving :

[0028]

[0029] In Equation (4), w d represents the constraint regularization factor for the background weight, represents the unfolded form of the original block tensor block corresponding to the original infrared remote sensing image in the i-th band based on the cyclic tensor, σ i represents the penalty factor, * represents the nuclear norm, represents the background class tensor block of the original infrared remote sensing image in the i-th band, is the dual variable, F represents the Frobenius norm, and argmin(·) represents the variable value when the minimum value of the objective function is taken;

[0030] This problem has an analytical solution:

[0031] ​

[0032] In Equation (5), H(·) represents the singular value shrinkage operator. represents the background class tensor block of the original infrared remote sensing image in the i-th band. represents the dual variable, and σ i represents the penalty factor, and w d represents the constraint regularization factor for the background weight, and * represents the nuclear norm.

[0033] For a variable Z, the solution formula for its singular value shrinkage operator is:

[0034]

[0035] In Equation (6), H(·) represents the singular value shrinkage operator, S(·) represents the soft threshold shrinkage operator, and UζV T is the singular value decomposition result of the variable Z, and w d represents the constraint regularization factor for the background weight, and σ i is the penalty factor.

[0036] For a variable i, the solution formula for its soft threshold shrinkage operator is:

[0037]

[0038] In Equation (7), S(·) represents the soft threshold shrinkage operator, and w d represents the constraint regularization factor for the background weight, and σ i represents the penalty factor, sign(·) represents the sign function, and max(·) represents taking the maximum value.

[0039] b. Iterative method for the background tensor block :

[0040] The adjoint matrix of the sampling operator is denoted as and is the auxiliary sampling operator, then is constrained by the following optimization conditions:

[0041]

[0042] In Equation (8), represents the dual variable, ε represents the Identity operator, represents the dual variable, and σ i represents the penalty factor, represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, Denote the background class tensor block of the original infrared remote sensing image in the i-th band;

[0043] For a tensor variable is a tensor with all elements equal to 1, denotes the Hadamard product;

[0044] Its solution is:

[0045]

[0046] In Equation (9), denotes the sampling tensor, denotes element-wise division, and σ i denotes the penalty factor, denotes the dual variable, denotes the cirrus cloud false alarm source tensor block of the original infrared remote sensing image in the i-th band, denotes the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, denotes the tensor block with all elements equal to 1 in the i-th band;

[0047] c. Iterative method for the cirrus cloud false alarm source tensor block is as follows:

[0048] Update based on the following optimization model:

[0049]

[0050] In Equation (10), denotes the dual variable, denotes the sampling operator, denotes the cirrus cloud false alarm source tensor block of the original infrared remote sensing image in the i-th band, denotes the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, denotes the background class tensor block of the original infrared remote sensing image in the i-th band, F represents the Frobenius norm, min(·) represents taking the minimum value, and λ and σ i denote the penalty factors, and l represents the number of image blocks obtained by dividing the original infrared remote sensing image in the i-th band by the window;

[0051] Its optimal solution is:

[0052]

[0053] In Equation (11), denotes the dual variable, S(·) represents the soft thresholding shrinkage operator, is the sampling tensor, denotes the Hadamard product, and λ and σi denotes the penalty factor, denotes the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, denotes the background class tensor block of the original infrared remote sensing image in the i-th band;

[0054] d. Dual variable Iterative method:

[0055]

[0056] In Equation (12), denotes the dual variable, σ i denotes the penalty factor, denotes the background class tensor block of the original infrared remote sensing image in the i-th band;

[0057] e. Dual variable Iterative method:

[0058]

[0059] In Equation (13), represents the dual variable, σ i denotes the penalty factor, is the sampling tensor, denotes the background class tensor block of the original infrared remote sensing image in the i-th band, denotes the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, denotes the cirrus cloud false alarm source tensor block of the original infrared remote sensing image in the i-th band, denotes the Hadamard product;

[0060] The maximum number of iteration steps is set to q = 100, and the cut-off condition is set to κ = 10e-8. When the above iteration steps and cut-off conditions are reached, the obtained is the background tensor block to be obtained and the cirrus cloud false alarm source tensor block

[0061] Furthermore, step S5 specifically includes the following steps:

[0062] S5.1. The formula for the multi-band fusion map I is:

[0063]

[0064] In Equation (14), C i denotes the cirrus cloud false alarm source image in the i-th band, k i (i = 1, 2,..., 6) are the weights corresponding to each band, Let N be the total number of bands, N = 6, m represents the length of the original infrared remote sensing image in the i-th band, and n represents the width of the original infrared remote sensing image in the i-th band;

[0065] S5.2. The multi-band fusion image I is subjected to adaptive threshold segmentation to obtain the final cirrus cloud-like false alarm source detection image D.

[0066] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0067] 1. The present invention aims at the difficulties in detecting complex and weak cirrus cloud-like false alarm sources in infrared remote sensing images, and proposes a method for detecting cirrus cloud-like false alarm sources by multi-band sparse feature fusion, which is used to overcome the current detection algorithms for such problems that only consider the wide coverage area and ignore the cirrus clouds with small occupied areas and sparsity, and most algorithms only use single-band information, resulting in insufficient details in the cirrus cloud detection results.

[0068] 2. The method adopted by the present invention utilizes a circular tensor model constructed based on high-dimensional structural features to transform the problem of detecting cirrus cloud-like false alarm sources into a problem of tensor low-rank recovery of the background image. At the same time, based on weighted average multi-band fusion, the overall effect of cirrus cloud detection is improved, and the imaging details are supplemented. Simulation experiments and results show that compared with similar methods, the detection results of the present invention have higher accuracy and reduce the probability of false alarms and missed detections. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0070] Figure 1 is the overall flowchart of the present invention;

[0071] Figure 2 is the multi-band original infrared remote sensing image of Embodiment 1 of the present invention;

[0072] Figure 3 are the background images of each band obtained in S4 of Embodiment 1 of the present invention;

[0073] Figure 4 are the cirrus cloud-like false alarm source images of each band obtained in S4 of Embodiment 1 of the present invention;

[0074] Figure 5 is the multi-band fusion image obtained in S5 of Embodiment 1 of the present invention;

[0075] Figure 6This is the final cirrus-like false alarm source detection image for Embodiment 1S5 of the present invention; Detailed implementation manners

[0076] The embodiments of the present invention will be described in detail below. Although the present invention will be described and explained in conjunction with some specific implementation manners, it should be noted that the present invention is not limited to these implementation manners only. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0077] In addition, in order to better illustrate the present invention, numerous specific details are given in the following detailed implementation manners. Those skilled in the art will understand that the present invention can also be implemented without these specific details.

[0078] In order to achieve the above object, the present invention adopts the following technical means:

[0079] Embodiment 1

[0080] A cirrus-like false alarm source detection method based on multi-band sparse feature fusion, comprising the following steps:

[0081] S1. Read in the multi-band original infrared remote sensing image f to be processed, where the multi-band original infrared remote sensing image, where represents the Euclidean space, m represents the length of the image, n represents the width of the image, and i represents the band number. The multi-band original infrared remote sensing image f to be processed in this embodiment, as shown, has a length m of 1024 and a width n of 1024; i wherein, wherein, represents the Euclidean space, m represents the length of the image, n represents the width of the image, i represents the band number, and the multi-band original infrared remote sensing image f to be processed in this embodiment, as shown, has a length m of 1024 and a width n of 1024; i such as Figure 2 shown, its length m is 1024 and its width n is 1024;

[0082] S2. Traverse and process the multi-band infrared original remote sensing image f through a window of size 60×60 to obtain the structural tensor blocks of each band of the multi-band infrared original remote sensing image f, that is, the original block tensors, where represents the Euclidean space, ws×ws represents the size of the window, l represents the number of image blocks into which the original infrared remote sensing image of the i-th band is divided by the window, and i represents the band number; i wherein, i wherein, wherein represents the Euclidean space, ws×ws represents the size of the window, l represents the number of image blocks into which the original infrared remote sensing image of the i-th band is divided by the window, and i represents the band number;

[0083] S3. Decomposition of the local structure tensor. According to the infrared block tensor model, perform low-rank and sparse iterative decomposition based on the cyclic tensor structure on the original block tensors constructed in step S2 to obtain the decomposed image blocks, and the decomposed image blocks include the background tensor blocks and the cirrus-like false alarm source tensor blocks. The background tensor blocks, the cirrus-like false alarm source tensor blocks wherein, and the cirrus-like false alarm source tensor blocks The background tensor blocks The cirrus-like false alarm source tensor blocks

[0084] Step S3 specifically includes the following steps:

[0085] S3.1. For the original block tensors of each band of the original infrared remote sensing image f i The expression is: The expression is:

[0086]

[0087] In formula (1), represents the background class tensor block of the original infrared remote sensing image in the i-th band, represents the cirrus cloud class false alarm source tensor block of the original infrared remote sensing image in the i-th band;

[0088] Based on formula (1), first set the number of iterations or the cut-off condition, and solve the objective function through the iterative optimization method to obtain the background tensor block and the cirrus cloud class false alarm source tensor block The objective function includes:

[0089]

[0090] In formula (2), s.t. is the abbreviation of "subject to" in mathematics, representing "subject to" certain constraint conditions; is the tensor block of the i-th band in the d-th modality, represents the expansion form based on the cyclic tensor of the tensor block of the i-th band in the d-th modality at position L, where K is the total dimension of the tensor block K = 3, d ∈ {1,..., K}, L = 1, 2, 3; * represents the nuclear norm, and the background is suppressed by combining the nuclear norm as a regularization term with the cyclic tensor expansion mode; w d represents the constraint regularization factor for the background weight, w d The value range of is 0 - 1, and at the same time, it needs to satisfy λ d represents the constraint regularization factor for the cirrus cloud class false alarm source, and P is the sampling rate, P = 1, S i is the size of the tensor slice on the d-th modality, S i ∈ {ws, l}, is the maximum value of the tensor slice size of the d-th mode tensor, i represents the band number, ws represents the window size, and l represents the number of image blocks into which the original infrared remote sensing image in the i-th band is divided by the window; is the sampling operator, T = S1 ×... × S K, where T is the total number of samples, represents the Euclidean space corresponding to the current tensor block;

[0091] The augmented Lagrangian function corresponding to Equation (2) is:

[0092]

[0093] In Equation (3), K is the total dimension of the tensor block w d represents the constraint regularization factor for the background weight, and λ and σ i represent the penalty factors, and and v i are dual variables, K is the total dimension of the tensor block, F represents the Frobenius norm, and the Frobenius norm is used as the joint regularization term to suppress the background. The value range of σ is from 10e - 8 to 10e - 3, i represents the background class tensor block of the original infrared remote sensing image in the i - th band, represents the cirrus false - alarm source tensor block of the original infrared remote sensing image in the i - th band, * represents the nuclear norm, and l represents the number of image blocks obtained by window - dividing the original infrared remote sensing image in the i - th band;

[0094] S3.2. Transform the model in Equation (3) into five sub - problems through the alternating direction method of multipliers (ADMM), and finally obtain the background block tensor and the cirrus false - alarm source tensor block

[0095] The iterative methods of the five sub - problems are as follows:

[0096] a. The iterative method of variable :

[0097] Fix the background tensor block and the cirrus false - alarm source tensor block Convert into the equation to solve :

[0098]

[0099] In Equation (4), w d represents the constraint regularization factor for the background weight, represents the unfolded form of the original block tensor block corresponding to the original infrared remote sensing image in the i - th band based on the cyclic tensor, and σ i represents the penalty factor, and * represents the nuclear norm, Denote the background class tensor block of the original infrared remote sensing image in the i-th band, is the dual variable, F represents the Frobenius norm, and argmin(·) represents the variable value when taking the minimum value of the objective function;

[0100] This problem has an analytical solution:

[0101]

[0102] In equation (5), H(·) represents the singular value shrinkage operator, Denote the background class tensor block of the original infrared remote sensing image in the i-th band, Denote the dual variable, σ i Denote the penalty factor, w d Denote the constraint regularization factor for the background weight, and * represents the nuclear norm;

[0103] For a variable Z, the solution formula for its singular value shrinkage operator is:

[0104]

[0105] In equation (6), H(·) represents the singular value shrinkage operator, S(·) represents the soft threshold shrinkage operator, and UζV T is the singular value decomposition result of the variable Z, and w d Denote the constraint regularization factor for the background weight, and σ i is the penalty factor;

[0106] For a variable i, the solution formula for its soft threshold shrinkage operator is:

[0107]

[0108] In equation (7), S(·) represents the soft threshold shrinkage operator, and w d Denote the constraint regularization factor for the background weight, and σ i Denote the penalty factor, sign(·) represents the sign function, and max(·) represents taking the maximum value;

[0109] b. Iterative method for the background tensor block is:

[0110] The adjoint matrix of the sampling operator is denoted as while and is the auxiliary sampling operator, then is constrained by the following optimization conditions:

[0111]

[0112] In Equation (8), K is the tensor block total dimension, represents the dual variable, ε represents the Identity operator, represents the dual variable, σ i represents the penalty factor, represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, represents the background class tensor block of the original infrared remote sensing image in the i-th band;

[0113] For a tensor variable is the all - one tensor, represents the Hadamard product;

[0114] Its solution is:

[0115]

[0116] In Equation (9), K is the tensor block total dimension, represents the sampling tensor, represents element - wise division, σ i represents the penalty factor, represents the dual variable, represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, represents the all - one tensor block in the i-th band;

[0117] c. Iterative method for the cirrus false alarm source tensor block :

[0118] Update based on the following optimization model:

[0119]

[0120] In Equation (10), represents the dual variable, represents the sampling operator, represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, represents the background class tensor block of the original infrared remote sensing image in the i-th band, F represents the Frobenius norm, min(·) represents taking the minimum value, λ and σ irepresents the penalty factor, and \(l\) represents the number of image blocks obtained by window partitioning the original infrared remote sensing image in the \(i\)-th band;

[0121] Its optimal solution is:

[0122]

[0123] In Equation (11), represents the dual variable, and \(S(·)\) represents the soft thresholding shrinkage operator. is the sampling tensor. represents the Hadamard product, and \(\lambda\) and \(\sigma\) i represent the penalty factor. represents the original block tensor block corresponding to the original infrared remote sensing image in the \(i\)-th band. represents the background class tensor block of the original infrared remote sensing image in the \(i\)-th band.

[0124] \(d\), the dual variable The iterative method for:

[0125]

[0126] In Equation (12), represents the dual variable, and \(\sigma\) i represents the penalty factor. represents the background class tensor block of the original infrared remote sensing image in the \(i\)-th band.

[0127] \(e\), the dual variable The iterative method for:

[0128]

[0129] In Equation (13), represents the dual variable, and \(\sigma\) i represents the penalty factor. is the sampling tensor. represents the background class tensor block of the original infrared remote sensing image in the \(i\)-th band. represents the original block tensor block corresponding to the original infrared remote sensing image in the \(i\)-th band. represents the cirrus cloud false alarm source tensor block of the original infrared remote sensing image in the \(i\)-th band. represents the Hadamard product;

[0130] The maximum number of iteration steps is set to \(q = 100\), and the cut-off condition is set to \(\kappa = 10e - 8\). When the above iteration steps and cut-off conditions are reached, the obtained is the required background tensor block and the cirrus cloud false alarm source tensor block

[0131] S4, Reconstruction of the local structural block tensors: For the background tensor blocks of each band obtained in step S3 and the cirrus false alarm source tensor blocks perform reconstruction respectively. The reconstruction steps are opposite to those of constructing the structural tensor blocks of each band in step S2. The divided background tensor blocks and the cirrus false alarm source tensor blocks are stitched together in the order in the multi-band original infrared remote sensing image f i to obtain the background image B of each band i and the cirrus false alarm source image C i . The background image The cirrus false alarm source image The background image B of each band in this embodiment i As Figure 3 shown, the cirrus false alarm source image C of each band i As Figure 4 shown;

[0132] S5. Perform multi-band fusion based on weighted average on the cirrus false alarm source image C of each band obtained in S4 i to obtain the multi-band fusion image I. As Figure 5 shown, perform adaptive threshold segmentation on the fusion image I to obtain the final cirrus false alarm source detection image D. The multi-band fusion image As Figure 6 shown, the cirrus false alarm source detection image

[0133] Step S5 specifically includes the following steps:

[0134] S5.1. The formula for the multi-band fusion map I is:

[0135]

[0136] In formula (14), C i represents the cirrus false alarm source image of the i-th band, and k i (i = 1, 2,..., 6) are the corresponding weights of each band, N is the total number of bands, N = 6, in this embodiment represents the length of the original infrared remote sensing image of the i-th band, and n represents the width of the original infrared remote sensing image of the i-th band. As Figure 5 shown;

[0137] S5.2. The multi-band fusion map I is subjected to adaptive threshold segmentation to obtain the final cirrus false alarm source detection image D. As Figure 6 shown.

[0138] Under the condition of filtering out background clutter well, the present invention can preferably detect sparse cirrus-like false alarm sources, reduce the probability of missed reports and false alarms during the detection process, and can also utilize multi-band information to enrich the imaging details of cirrus clouds, further improving the cirrus cloud detection performance in the remote sensing detection system.

[0139] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting false alarm sources of cirrus clouds by fusing multi-band sparse features, characterized in that: Including the following steps: S1. Read in the multi-band original infrared remote sensing image f i ; S2. Traverse and process the multi-band infrared original remote sensing image f through a window of size ws×ws i to divide it into l image patches, obtaining the structure tensor patches of each band of the multi-band infrared original remote sensing image f i , that is, forming the original patch tensor block O of size ws×ws×l i ; S3. Decomposition of the local structure tensor: According to the infrared block tensor model, decompose the original block tensor block O constructed in step S2 i by means of low-rank and sparse iterative decomposition based on a cyclic tensor structure to obtain the decomposed image block, where the decomposed image block includes a background tensor block B i and a cirrus-like false alarm source tensor block T i ; S4. Reconstruction of local structural block tensors: For the background tensor blocks B of each band obtained in step S3 i and the cirrus false alarm source tensor blocks T i perform reconstruction respectively. The reconstruction steps are opposite to those of constructing the structural tensor blocks of each band in step S2. Concatenate the divided background tensor blocks B i and the cirrus false alarm source tensor blocks T i in the order in which they appear in the multi-band original infrared remote sensing image f i to obtain the background images B of each band i and the cirrus false alarm source images C i ; S5. For the cirrus-like false alarm source images C of each band obtained in S4 i perform multi-band fusion based on weighted average to obtain a multi-band fused image I, and perform adaptive threshold segmentation on the fused image I to obtain the final cirrus-like false alarm source detection image D.

2. The method for detecting cirrus-like false alarm sources with multi-band sparse feature fusion according to claim 1, wherein: The multi-band original infrared remote sensing image f i ∈R m×n , where i = 1, 2, 3... 6, the original block tensor block O of each band i ∈R ws×ws×l , where i = 1, 2, 3... 6, the background tensor block B i ∈R ws×ws×l , where i = 1, 2, 3... 6, the cirrus-like false alarm source tensor block T i ∈R ws×ws×l , where i = 1, 2, 3... 6, the background image B of each band i ∈R m×n , where i = 1, 2, 3... 6, the cirrus-like false alarm source image C i ∈R m×n , where i = 1, 2, 3... 6, the multi-band fusion image I ∈ R m×n , the cirrus-like false alarm source detection image D ∈ R m×n ; Wherein, R represents the Euclidean space, m represents the length of the original infrared remote sensing image in the i-th band, n represents the width of the original infrared remote sensing image in the i-th band, i represents the band number, ws×ws represents the window size, and l represents the number of image blocks obtained by dividing the original infrared remote sensing image in the i-th band with the window.

3. A method for detecting cirrus-like false alarm sources with multi-band sparse feature fusion according to claim 2, characterized in that: Specifically, step S3 includes the following steps: S3.

1. The original block tensor block O i of each band of the original infrared remote sensing image f i is expressed as: O i = B i + T i (1) In formula (1), B i represents the background class tensor block of the original infrared remote sensing image in the i-th band, and T i represents the cirrus cloud class false alarm source tensor block of the original infrared remote sensing image in the i-th band; Based on Equation (1), first set the number of iterations or termination conditions, and solve the objective function by an iterative optimization method to obtain the background tensor block B i and the cirrus-like false alarm source tensor block T i , and the objective function includes: s.t.F O (B i +T i )=F O (O i ), In formula (2), min(·) represents taking the minimum value, and s.t. is the abbreviation of "subject to" in mathematics, representing "subject to" certain constraint conditions; X i d is the i-th band tensor block of the d-th mode, represents the unfolded form based on the circular tensor when the i-th band tensor block of the d-th mode is at position L. Among them, K is the total dimension of the tensor block X i The total dimension, d is the mode number, L is the circular tensor decomposition position, K = 3, d ∈ {1,..., K}, L = 1, 2, 3; * represents the nuclear norm, and the background is suppressed by combining the nuclear norm as a regularization term with the circular tensor unfolding mode; w d represents the constraint regularization factor for the background weight, w d The value range of is 0 - 1, and at the same time, it needs to satisfy λ d represents the constraint regularization factor for the cirrus false alarm source, and P is the sampling rate, P = 1, S i is the size of the tensor slice on the d-th mode, S i ∈ {ws, l}, is the maximum value of the tensor slice size of the d-th mode tensor, i represents the band number, ws represents the window size, and l represents the number of image blocks into which the original infrared remote sensing image of the i-th band is divided by the window; F0 is the sampling operator, T = S1×...×S K , T is the total number of samples, represents the Euclidean space corresponding to the current tensor block; The augmented Lagrangian function \(L\) corresponding to formula (2) β is as follows: In formula (3), K is the tensor block X i total dimension, w d represents the constraint regularization factor for the background weight, λ and σ i represent the penalty factors, and Z i d and v i are dual variables, K is the tensor block X i total dimension, F represents the Frobenius norm, and the Frobenius norm is used as the joint regularization term to suppress the background, σ i ranges from 10e-8 to 10e-3, B i represents the background class tensor block of the original infrared remote sensing image in the i-th band, T i represents the cirrus cloud false alarm source tensor block of the original infrared remote sensing image in the i-th band, * represents the nuclear norm, and l represents the number of image blocks obtained by window-dividing the original infrared remote sensing image in the i-th band; S3.

2. Transform the model in Equation (3) into five sub-problems by the alternating direction method of multipliers, and finally obtain the background block tensor B i and the cirrus-like false alarm source tensor block T i .

4. A method for detecting cirrus-like false alarm sources by multi-band sparse feature fusion according to claim 3, characterized in that: The iterative methods for the five sub-problems are respectively as follows: a. Variable X i Iterative method: Fixed background tensor block B i and cirrus-like false alarm source tensor block T i , convert L β to solution X i d of the equation: In formula (4), w d represents the constraint regularization factor for the background weight, represents the unfolded form of the original block tensor block corresponding to the original infrared remote sensing image in the i-th band based on the cyclic tensor, σ i represents the penalty factor, * represents the nuclear norm, B i represents the background class tensor block of the original infrared remote sensing image in the i-th band, Z i d is the dual variable, F represents the Frobenius norm, and arg min(·) represents the variable value when taking the minimum value of the objective function; This problem has an analytical solution: In formula (5), H(·) represents the singular value shrinkage operator, and B i represents the background class tensor block of the original infrared remote sensing image in the i-th band, and Z i d represents the dual variable, σ i represents the penalty factor, w d represents the constraint regularization factor for the background weight, and * represents the nuclear norm; For a variable Z, the solution formula of its singular value shrinkage operator is: In Equation (6), H(·) represents the singular value shrinkage operator, S(·) represents the soft threshold shrinkage operator, and UζV T is the singular value decomposition result of variable Z, and w d represents the constraint regularization factor for the background weight, and σ i is the penalty factor; For a variable i, the solution formula of its soft threshold shrinkage operator is: In formula (7), S(·) represents the soft thresholding operator, and w d represents the constraint regularization factor for the background weight, and σ i represents the penalty factor, sign(·) represents the sign function, and max(·) represents taking the maximum value; b. Iterative method for background tensor block B i : Sampling operator F O The adjoint matrix of is denoted as F O * , and F O * (v i ) = P O (V i ), and F O * F O = P O , P O is an auxiliary sampling operator, then B i is constrained by the following optimization conditions: In formula (8), K is the tensor block X i total dimension, V i represents the dual variable, E represents the Identity operator, Z i (d) represents the dual variable, σ i represents the penalty factor, T i represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, O i represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, B i represents the background class tensor block of the original infrared remote sensing image in the i-th band; For a tensor variable W, E(W) = I#W, where I is the all-ones tensor, and # represents the Hadamard product; Its solution is: In formula (9), K is the tensor block X i total dimension, P i represents the sampling tensor, % represents element-wise division, σ i represents the penalty factor, Z i d represents the dual variable, T i represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, O i represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, I i represents the tensor block with all elements set to 1 in the i-th band; c. Iterative method for cirrus-like false alarm source tensor block T i : T i Updated based on the following optimized model: In formula (10), V i represents the dual variable, F o represents the sampling operator, T i represents the cirrus false alarm source tensor block of the original infrared remote sensing image in the i-th band, O i represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, B i represents the background class tensor block of the original infrared remote sensing image in the i-th band, F represents the Frobenius norm, min(·) represents taking the minimum value, and λ and σ i represent the penalty factors, and l represents the number of image blocks obtained by dividing the original infrared remote sensing image in the i-th band with a window; Its optimal solution is: In formula (11), V i represents the dual variable, S(·) represents the soft thresholding shrinkage operator, and P i is the sampling tensor, # represents the Hadamard product, and λ and σ i represent the penalty factors, O i represents the original block tensor block corresponding to the original infrared remote sensing image of the i-th band, and B i represents the background class tensor block of the original infrared remote sensing image of the i-th band; d. Dual variable Z i d Iterative method: Z i d = Z i d + σ i (X i d - B i ) (12) In formula (12), Z i d represents the dual variable, and σ i represents the penalty factor, and B i represents the background class tensor block of the original infrared remote sensing image in the i-th band; e. Iterative method for dual variable V i : V i = V i + σ i P i #(T i + B i - O i ) (13) In formula (13), V i represents the dual variable, σ i represents the penalty factor, P i is the sampling tensor, B i represents the background class tensor block of the original infrared remote sensing image in the i-th band, O i represents the original block tensor block corresponding to the original infrared remote sensing image in the i-th band, T i represents the cirrus cloud class false alarm source tensor block of the original infrared remote sensing image in the i-th band, # represents the Hadamard product; The maximum number of iterative steps is set to q = 100, and the termination condition is set to ||B i q+1 -B i q || F / ||B i q || F <κ, where κ = 10e - 8. After reaching the above number of iterative steps and termination condition, the obtained B i q+1 、T i q+1 is the desired background tensor block B i and the cirrus - like false - alarm source tensor block T i .

5. The cirrus-like false alarm source detection method based on multi-band sparse feature fusion according to claim 1, wherein: Step S5 specifically includes the following steps: S5.

1. The formula for the multi-band fusion map I is: In formula (14), C i represents the cirrus cloud type false alarm source image of the i-th band, and k i is the corresponding weight for each band, where i = 1, 2, 3... 6, N is the total number of bands, N = 6, m represents the length of the original infrared remote sensing image of the i-th band, and n represents the width of the original infrared remote sensing image of the i-th band; S5.

2. The multi-band fusion map I is subjected to adaptive threshold segmentation to obtain the final cirrus cloud type false alarm source detection image D.

Images