A collaborative modeling and detection method for false alarm sources and targets in infrared imaging
Through the coordinated modeling and detection method of false alarm sources and targets of infrared imaging, local structure tensor decomposition and significant feature extraction are used to solve the false alarm and missed alarm problems in infrared weak target detection, and the detection accuracy and false alarm sources are improved. Detection performance.
Patent Information
- Application Number
- CN202210597023.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-30
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-05-30
Smart Images

Figure CN115205696B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of image processing, and specifically designs a method for collaborative modeling and detection of infrared imaging false alarm sources and targets. Background Art
[0002] In remote sensing, telemetry and military applications, infrared small target detection, recognition and tracking play a very important role. Since infrared small target images usually have complex backgrounds, high-brightness and strong-texture false alarm sources, and infrared small targets are point-shaped and lack texture and structural information, the detection of infrared small targets faces huge challenges. How to effectively suppress the background, improve the detection performance of targets and false alarm sources, and reduce missed alarms and false positives is a very critical issue in the field of remote sensing, telemetry and infrared detection.
[0003] In the application of infrared detection system, a large number of false alarm sources coexist with the target. The false alarm source mainly shows two characteristics: one is that it has high scattering and high reflection characteristics, and then shows high brightness; the other is that the false alarm source is affected by the terrain and shows obvious texture and edge characteristics, so the false alarm source block tensor can be regarded as a high brightness, strong texture tensor. Existing detection algorithms often only consider the processing of target, background and noise interference, ignoring the existence and influence of dense high-reflection false alarm sources in the scene, thereby causing the false alarm source to be misdetected as the target during the infrared weak target detection process in complex scenes, resulting in the problems of reduced detection rate and high false alarm rate. For this reason, the present invention adopts the scheme of collaborative modeling and detection of false alarm sources and targets, thereby further improving the detection performance of infrared weak targets in remote sensing detection systems. Summary of the invention
[0004] The purpose of the present invention is to solve the problem of high false positive and false negative rates in detection results of current infrared small target detection algorithms.
[0005] The technical solution adopted by the present invention is as follows:
[0006] A method for collaborative modeling and detection of infrared imaging false alarm sources and targets, the method comprising the following steps:
[0007] Step 1: The original infrared remote sensing image f0∈R m×n Read in, where R represents the Euclidean space, m represents the length of the image, and n represents the width of the image. Generally, the image size is uniformly set to m=256 and n=256;
[0008] Step 2: Construct the local structure tensor and use an h×h window to map the original infrared image f0∈R m×n Perform traversal processing, set the window size to h = 40, and obtain the original infrared image f0∈R m×n The corresponding local structure tensor, that is, the original block tensor D∈R h×h×l, h×h represents the length and width of the traversal window, l represents the original infrared image f0∈R m×n The number of image blocks divided by the window, thus forming a three-dimensional tensor block of h×h×l;
[0009] Step 3: Decomposition of the local structure tensor: According to the infrared block tensor model, the original block tensor constructed by the input image is decomposed in a low-rank, sparse iterative manner to obtain the decomposed image block, which includes the background block tensor B, B∈R h×h×l , target block tensor T, T∈R h×h×l and the false alarm source block tensor A, A∈R h×h×l , where R represents the Euclidean space, h×h represents the length and width of the traversal window, and l represents that the original infrared image is divided into l h×h small structural blocks by the window;
[0010] Step 4: Reconstruction of the local structure block tensor. Reconstruct the background block tensor B, the false alarm source block tensor A, and the target block tensor T. The reconstruction method is opposite to the order of step 2 "constructing the local structure tensor", that is, the divided image blocks are spliced in the order of their original infrared image to obtain the background image f B ∈R m×n , target image f T ∈R m×n and false alarm source f A ∈R m ×n After the reconstruction is completed, the reconstructed image f T and the reconstructed image f A Perform adaptive threshold segmentation to obtain the final detected target T∈R m×n , false alarm source A∈R m×n , where R represents the Euclidean space, m represents the length of the image, and n represents the width of the image.
[0011] In the above technical solution, the step 3 specifically includes the following steps:
[0012] Step 3.1, the original infrared block tensor is expressed as follows:
[0013] D=B+T+A+N,
[0014] Among them, D, B, T, A and N refer to the original infrared block tensor, background block tensor, target block tensor, false alarm source block tensor and random noise and other interference items;
[0015] Step 3.2: The false alarm source has significant features such as high brightness and strong texture. In order to further improve the detection rate of the false alarm source, significant features of A are extracted, which can be expressed as S(X), where the independent variable X represents the original infrared image f0∈R to be processed. m×n, since the subsequent tensor decomposition belongs to the minimization problem, and the purpose of this application is to make the visual saliency features of the area where the false alarm source is located as large as possible, it is necessary to convert S(X) into the inverse form:
[0016]
[0017] Among them, ||·|| S(X) The form of representation that indicates the distinctive features;
[0018] By introducing the significant features of false alarm sources, the following objective function for tensor sparse and low-rank decomposition is constructed:
[0019]
[0020] By setting the number of iterations or the cutoff condition and solving the above objective function through the iterative optimization method, the background block tensor B∈R can be obtained. h×h×l , target block tensor T∈R h×h×l And the false alarm source block tensor A∈R h×h×l ;
[0021] in δ represents the regularization factor for background and target constraints respectively, st is the abbreviation of "subject to" in mathematics, which means "satisfied with" certain constraints, F represents the Frobenius norm, and the Frobenius norm is used as a joint regularization term to suppress noise, λ is the balance factor, μ is the regularization factor, and β is the regularization factor. The value range of L is 0.5-1.5, the value range of β is 0.02-0.1, and the value range of μ is 100-300. The objective function is further rewritten as follows:
[0022]
[0023] The corresponding Lagrangian function L is:
[0024]
[0025] Among them, γ is the penalty factor; ρ is the auxiliary variable. Finally, the augmented Lagrangian function of the above model is solved by the alternating direction multiplier method (ADMM), and the background block tensor B∈R h×h×l , target block tensor T∈λ h×h×l and the false alarm source tensor A∈R h×h×l ;
[0026] Three types of tensor iteration methods for different objects can be expressed as:
[0027] The iterative formula of the false alarm source tensor A is:
[0028] Ap+1 =D 1 / (β+ρ) (Y)
[0029] Where p is the number of iterations, D(·) represents the singular value contraction operator.
[0030] The iterative formula of the background block tensor B is:
[0031] B p+1 D 1 / (μ+ρ) (X)
[0032] Where p is the number of iterations, D(·) represents the singular value contraction operator. For a variable Y, the solution formula of its singular value contraction operator is:
[0033] D 1 / (μ+ρ) (Y)=US 1 / (u+ρ) (ζ)V T
[0034] Where S(·) represents the soft threshold shrinkage operator, UζV T is the singular value decomposition result of Y.
[0035] The iterative formula of the target block tensor T is:
[0036] T p+1 =S λ (DB p+1 -A p+1 )
[0037] Where p is the number of iteration steps, S(·) represents the soft threshold shrinkage operator, and for a variable x, the solution formula of the soft threshold shrinkage operator is:
[0038] S λ (x)=sign(x)·max(|x|-λ, 0);
[0039] Among them, μ is the regularization factor, β is the regularization factor, the iteration of the background block tensor, the iteration of the background block tensor, and the iteration of the target block tensor are processed simultaneously;
[0040] The maximum number of iteration steps is set to p = 300, and the cutoff condition is set to ||B p+1 -B p || F / ||B p || F <ε,ε=10e-5. When the above iteration steps and cutoff conditions are reached, the obtained A p+1 , B p+1 、T p+1 That is the desired false alarm source block tensor A∈R h×h×l , background patch tensor B∈Rh×h×l and the target block tensor T∈R h×h×l .
[0041] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:
[0042] Aiming at the difficulties of complex background and dense distribution of false alarm sources in infrared remote sensing images, the present invention proposes a method for collaborative modeling and detection of false alarm sources and targets in infrared imaging, which is used to overcome the problem that the current detection algorithms for such problems often only detect targets and ignore the influence of false alarm sources, thus leading to problems such as missed detection and false detection. The method adopted by the present invention not only filters out background clutter, but also makes full use of local structure tensor recovery and maximum significance constraint to realize collaborative modeling and detection of infrared weak targets and dense high-reflection false alarm sources. 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
[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.
[0044] Figure 1 This is a general flow chart of Embodiment 1 of the present invention;
[0045] Figure 2 is the original input image of the infrared weak target;
[0046] Figure 3 The final weak target image in step 4 of embodiment 1 of the present invention;
[0047] Figure 4 The final background image of step 4 of embodiment 1 of the present invention;
[0048] Figure 5 The final false alarm source diagram of step 4 of embodiment 1 of the present invention; DETAILED DESCRIPTION
[0049] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions made to the present invention should all be included in the scope of the claims of the present invention.
[0050] In addition, in order to better illustrate the present invention, numerous specific details are given in the following specific embodiments. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.
[0051] In order to achieve the above purpose, the present invention adopts the following technical means:
[0052] Example 1
[0053] A method for collaborative modeling and detection of infrared imaging false alarm sources and targets, the method comprising the following steps:
[0054] Step 1: The original infrared remote sensing image f0∈R m×n Read in, where R represents the Euclidean space, m represents the length of the image, and n represents the width of the image. Generally, the image size is uniformly set to m=256 and n=256;
[0055] Step 2: Construct the local structure tensor and use an h×h window to map the original infrared image f0∈R m×n Perform traversal processing, set the window size to h = 40, and obtain the original infrared image f0∈R m×n The corresponding local structure tensor, that is, the original block tensor D∈R h×h×l , h×h represents the length and width of the traversal window, l represents the original infrared image f0∈R m×n The number of image blocks divided by the window, thus forming a three-dimensional tensor block of h×h×l;
[0056] Step 3: Decomposition of the local structure tensor: According to the infrared block tensor model, the original block tensor constructed by the input image is decomposed in a low-rank, sparse iterative manner to obtain the decomposed image block, which includes the background block tensor B, B∈R h×h×l , target block tensor T, T∈R h×h×l and the false alarm source block tensor A, A∈R h×h×l , where R represents the Euclidean space, h×h represents the length and width of the traversal window, and l represents that the original infrared image is divided into l h×h small structural blocks by the window;
[0057] Step 4: Reconstruction of the local structure block tensor. Reconstruct the background block tensor B, the false alarm source block tensor A, and the target block tensor T. The reconstruction method is opposite to the order of step 2 "constructing the local structure tensor", that is, the divided image blocks are spliced in the order of their original infrared image to obtain the background image f B ∈R m×n , target image f T ∈R m×n and false alarm source f A ∈R m ×n After the reconstruction is completed, the reconstructed image f Tand the reconstructed image f A Perform adaptive threshold segmentation to obtain the final detected target T∈R m×n , false alarm source A∈R m×n , where R represents the Euclidean space, m represents the length of the image, and n represents the width of the image.
[0058] In the above technical solution, the step 3 specifically includes the following steps:
[0059] Step 3.1, the original infrared block tensor is expressed as follows:
[0060] D=B+T+A+N,
[0061] Among them, D, B, T, A and N refer to the original infrared block tensor, background block tensor, target block tensor, false alarm source block tensor and random noise and other interference items;
[0062] Step 3.2: The false alarm source has significant features such as high brightness and strong texture. In order to further improve the detection rate of the false alarm source, significant features are extracted from A, which can be expressed as S(X), where the independent variable X represents the original infrared image f0∈R to be processed. m×n , since the subsequent tensor decomposition belongs to the minimization problem, and the purpose of this application is to make the visual saliency features of the area where the false alarm source is located as large as possible, it is necessary to convert S(X) into the inverse form:
[0063]
[0064] Among them, ||·|| S(X) The form of representation that indicates the distinctive features;
[0065] By introducing the significant features of false alarm sources, the following objective function for tensor sparse and low-rank decomposition is constructed:
[0066]
[0067] By setting the number of iterations or the cutoff condition and solving the above objective function through the iterative optimization method, the background block tensor B∈R can be obtained. h×h×l , target block tensor T∈R h×h×l And the false alarm source block tensor A∈R h×h×l ;
[0068] in δ represents the regularization factor for background and target constraints respectively, st is the abbreviation of "subject to" in mathematics, which means "satisfied with" certain constraints, F represents the Frobenius norm, and the Frobenius norm is used as a joint regularization term to suppress noise, λ is the balance factor, μ is the regularization factor, and β is the regularization factor. The value range of L is 0.5-1.5, the value range of β is 0.02-0.10, and the value range of μ is 100-300. The objective function is further rewritten as follows:
[0069]
[0070] The corresponding Lagrangian function L is:
[0071]
[0072] Among them, γ is the penalty factor; ρ is the auxiliary variable. Finally, the augmented Lagrangian function of the above model is solved by the alternating direction multiplier method (ADMM), and the background block tensor B∈R h×h×l , target block tensor T∈R h×h×l and the false alarm source tensor A∈R h×h×l ;
[0073] Three types of tensor iteration methods for different objects can be expressed as:
[0074] The iterative formula of the false alarm source tensor A is:
[0075] A p+1 D 1 / (β+ρ) (Y)
[0076] Where p is the number of iterations, D(·) represents the singular value contraction operator.
[0077] The iterative formula of the background block tensor B is:
[0078] B p+1 D 1 / (μ+ρ) (X)
[0079] Where p is the number of iterations, D(·) represents the singular value contraction operator. For a variable Y, the solution formula of its singular value contraction operator is:
[0080] D 1 / (μ+ρ) (Y)=US 1 / (μ+ρ) (ζ)V T
[0081] Where S(·) represents the soft threshold shrinkage operator, UζV T is the singular value decomposition result of Y.
[0082] The iterative formula of the target block tensor T is:
[0083] T p+1 =S λ (DB p+1 -A p+1 )
[0084] Where p is the number of iteration steps, S(·) represents the soft threshold shrinkage operator, and for a variable x, the solution formula of the soft threshold shrinkage operator is:
[0085] S λ (x)=sign(x)·max(|x|-λ, 0);
[0086] Among them, μ is the regularization factor, β is the regularization factor, the iteration of the background block tensor, the iteration of the background block tensor, and the iteration of the target block tensor are processed simultaneously;
[0087] The maximum number of iteration steps is set to p = 300, and the cutoff condition is set to ||B p+1 -B p || F / ||B p || F <ε,ε=10e-5. When the above iteration steps and cutoff conditions are reached, the obtained A p+1 , B p+1 、T p+1 That is the desired false alarm source block tensor A∈R h×h×l , background patch tensor B∈R h×h×l and the target block tensor T∈R h×h×l .
[0088] This patent can not only filter out background clutter very well, but also detect weak infrared targets and dense high-reflection false alarm sources well, reducing the probability of missed alarms and false alarms during the detection process, thereby further improving the detection performance of weak infrared targets in remote sensing detection systems.
[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for collaborative modeling and detection of infrared imaging false alarm sources and targets, characterized in that: The method comprises the following steps: Step 1: The original infrared remote sensing image to be processed Read in, where represents the Euclidean space, m represents the length of the image, and n represents the width of the image; Step 2: Construct the local structure tensor and use an h×h window to map the original infrared image. Perform traversal processing, set the window size to h, and get the original infrared image The corresponding local structure tensor, that is, the original block tensor h×h represents the length and width of the traversal window, l represents the original infrared image The number of image blocks divided by the window, thus forming a three-dimensional tensor block of h×h×l; Step 3: Decomposition of the local structure tensor: According to the infrared block tensor model, the original block tensor constructed by the input image is decomposed into low-rank, sparse iterative decomposition to obtain the decomposed image blocks, which include the background block tensor. Target block tensor and false alarm source block tensor in represents the Euclidean space, h×h represents the length and width of the traversal window, and l represents that the original infrared image is divided into l h×h small structural blocks by the window; Step 4: Reconstruction of the local structure block tensor and the background block tensor False alarm source block tensor and the target block tensor Reconstruction is performed in the opposite order to step 2 "constructing the local structure tensor", that is, the divided image blocks are spliced in the order of their original infrared image to obtain the background image Target image and false alarm sources After the reconstruction is completed, the reconstructed image f T and the reconstructed image f A Perform adaptive threshold segmentation to obtain the final detected target False Alarm Source in represents the Euclidean space, m represents the length of the image, and n represents the width of the image; The step 3 specifically comprises the following steps: Step 3.1: Original infrared block tensor It is expressed as: in and It refers to the original infrared block tensor, background block tensor, target block tensor, false alarm source block tensor and random noise interference term; Step 3.2: In order to further improve the detection rate of false alarm sources, To extract significant features, it can be expressed as S(X), where the independent variable X represents the original infrared image to be processed. Convert S(X) to reciprocal form: Among them, ||·|| S(X) The form of representation that indicates the distinctive features; By introducing the significant features of false alarm sources, the following objective function for tensor sparse and low-rank decomposition is constructed: By setting the number of iterations or the cutoff condition and solving the above objective function through iterative optimization method, the background block tensor can be obtained. Target block tensor and false alarm source block tensor in δ represents the regularization factor for background and target constraints respectively, st is the abbreviation of "subject to" in mathematics, which means "satisfied with" certain constraints, F represents the Frobenius norm, and the Frobenius norm is used as a joint regularization term to suppress noise, λ is the balance factor, μ is the regularization factor, and β is the regularization factor. The value range of L is 0.5-1.5, the value range of β is 0.02-0.1, and the value range of μ is 100-300. The objective function is further rewritten as follows: The corresponding Lagrangian function L is: Among them, γ is the penalty factor; ρ is the auxiliary variable. Finally, the augmented Lagrangian function of the above model is solved by the alternating direction multiplier method (ADMM), and the background block tensor can be obtained Target block tensor and false alarm source tensor Three types of tensor iteration methods for different objects can be expressed as: False Alarm Source Tensor The iteration formula is: Where p is the number of iterations, D(·) represents the singular value contraction operator; Background block tensor The iteration formula is: Where p is the number of iterations, D(·) represents the singular value contraction operator. For a variable Y, the solution formula of its singular value contraction operator is: D 1 / (+μρ) (Y)=US 1 / (μ+ρ) (ζ)V T Where S(·) represents the soft threshold shrinkage operator, UζV T is the singular value decomposition result of Y; Target block tensor The iteration formula is: Where p is the number of iteration steps, S(·) represents the soft threshold shrinkage operator, and for a variable x, the solution formula of the soft threshold shrinkage operator is: S λ (x)=sign(x)·max(|x|-λ,0); Among them, μ is the regularization factor, β is the regularization factor, the iteration of the background block tensor, the iteration of the background block tensor, and the iteration of the target block tensor are processed simultaneously; The maximum number of iteration steps is set to p = 300, and the cutoff condition is set to When the above iteration steps and cutoff conditions are reached, the obtained That is the desired false alarm source block tensor Background block tensor and the target block tensor
