An Infrared Target Detection Method and System Based on Fourth-Order Space-Time Tensor

Abstract

The present invention discloses an infrared target detection method and system based on a fourth-order spatio-temporal tensor, belonging to the technical fields of remote sensing image processing and target detection. The present invention includes obtaining L consecutive original infrared images, constructing a fourth-order spatio-temporal tensor pair for low-rank and sparse decomposition, that is, using the nuclear norms of the three generalized tensor unfolding matrices to characterize the low-rank property of the fourth-order background spatio-temporal tensor, and extracting the spatio-temporal prior of the target in the original infrared image to construct the weight factor of the target tensor, and constructing an objective function; iteratively solving the constructed objective function to obtain a low-rank fourth-order background spatio-temporal tensor and a sparse fourth-order target spatio-temporal tensor, and reconstructing the L-frame background image and target image based on the fourth-order background spatio-temporal tensor and the fourth-order target spatio-temporal tensor according to the reverse process of constructing the fourth-order spatio-temporal tensor in step 1, and the obtained target image is the infrared target. The present invention is used for infrared target detection.

Description

Technical Field

[0001] An infrared target detection method and system based on a fourth-order spatio-temporal tensor, which is used for infrared target detection and belongs to the technical fields of remote sensing image processing and target detection. Background Technique

[0002] Due to the characteristics of strong anti-interference, strong concealment, and all-weather operation of infrared imaging, infrared small target detection technology is widely used in fields such as infrared early warning, precision guidance, and missile tracking systems. However, due to the following reasons, infrared small target detection is still challenging: due to the long detection distance, the target appears as a speckle, resulting in the loss of information such as texture and shape; affected by clutter and atmospheric radiation, the signal-to-clutter ratio of the target is low; the imaging environment of the target is complex, and the high-brightness areas in the background and pixel-level high-brightness noise points will affect the detection performance of infrared targets.

[0003] Current infrared small target detection methods can be divided into sequence-based detection methods and single-frame-based detection methods. The former detects the target based on the motion continuity and trajectory consistency of the target. This type of method requires processing multiple frames of infrared images and has good detection effects in the case of a static background. However, in many cases, the imaging background changes with the movement of the target, resulting in a reduction in the performance of sequence-based detection methods.

[0004] Single-frame-based detection methods can be divided into three categories: background prediction-based methods, target saliency-based methods, and low-rank sparse decomposition-based methods. Background prediction-based methods mainly design algorithms according to the properties of infrared background images to estimate the background component from the original infrared image, then subtract the background from the original image, and finally perform threshold segmentation to detect small targets. Typical algorithms include bilateral filters, two-dimensional least mean square filters, maximum median / mean filters, etc.

[0005] Target saliency-based methods assume that the target is the most prominent object among local regions of the image. The core of this type of method lies in how to effectively evaluate the local gray difference between the target region and its neighborhood. Some methods use the difference between the gray mean value of the target region and the gray mean value of the neighborhood to highlight the target, and some detection algorithms use the ratio of the central gray value to the gray value of the neighborhood as the local contrast. Target saliency-based methods obtain the target saliency map by designing different algorithms to measure the gray difference between the target and its neighborhood. In the target saliency map, the background is suppressed and the target is enhanced, and then threshold segmentation is performed on the target saliency map to obtain the detection result. This type of method includes local contrast measurement, contrast measurement based on multi-scale patches, Gaussian difference filters, etc.

[0006] The method based on low-rank sparse decomposition believes that some image patches in the background are approximately linearly correlated, and the background satisfies low-rank property. While the target only occupies very few pixels in the infrared image, so the target has sparsity. In view of this, the problem of infrared small target detection can be transformed into a mathematical optimization problem of recovering the low-rank component and the sparse component from the original image. Compared with the methods based on background prediction and target saliency, this kind of method utilizes the low-rank property of the background and the global sparsity of the target, and has stronger robustness and better detection performance. Typical algorithms include infrared block image model, total variation principal component pursuit, reweighted infrared block tensor model, etc. This kind of method mainly makes improvements in aspects such as data construction, relaxation of rank function, relaxation of L0 norm, and addition of constraint terms. However, the information available in a single-frame infrared image is limited, lacking prior information such as target shape and motion, which greatly limits the detection performance of small targets.

[0007] Although in the prior art, second-order (such as the content disclosed in CN202211339176.0, a target detection method, device, electronic device and storage medium based on unitary transformation) or third-order tensors are used to detect infrared targets, the following technical problems exist:

[0008] 1. Using second-order or third-order tensors to detect targets does not utilize the time-domain information and the correlation between local block images at the same time, and there are background clutter and strong noise in the original infrared image sequence that have the same high contrast and sparse characteristics as the target, which is likely to cause a high false alarm rate;

[0009] 2. The prior art does not fully utilize the correlation between two tensors, and cannot accurately obtain the background component, thus easily causing the problem of poor accuracy of infrared detection targets. Summary of the Invention

[0010] Aiming at the problems studied above, the purpose of the present invention is to provide an infrared target detection method based on a fourth-order spatio-temporal tensor, to solve the problem that the prior art uses second-order or third-order tensors to detect targets, does not utilize the time-domain information and the correlation between local block images at the same time, and there are background clutter and strong noise in the original infrared image sequence that have the same high contrast and sparse characteristics as the target, which is likely to cause a high false alarm rate.

[0011] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0012] An infrared target detection method based on a fourth-order spatio-temporal tensor, comprising the following steps:

[0013] Step 1, obtain L consecutive original infrared images and construct a fourth-order spatio-temporal tensor

[0014] Step 2, for Perform low-rank and sparse decomposition, that is, use the nuclear norms of the three generalized tensor unfolding matrices of to characterize the low-rank property of the fourth-order background spatio-temporal tensor

[0015] Step 3. Iteratively solve the objective function constructed in Step 2 to obtain a low-rank fourth-order background spatio-temporal tensor and a sparse fourth-order target spatio-temporal tensor

[0016] Step 4. Based on the fourth-order background spatio-temporal tensor and the fourth-order target spatio-temporal tensor Reconstruct them into the background image and the target image of L frames according to the reverse process of constructing the fourth-order spatio-temporal tensor in Step 1. The obtained target image is the infrared target.

[0017] Further, the specific steps of Step 1 are as follows:

[0018] Step 1.1. Obtain continuous L frames of original infrared images;

[0019] Step 1.2. Use a sliding window of size w×w and traverse the L frames of original infrared images with a step size of w. Each frame of the original infrared image is divided into n block images by the sliding window. Denote as the s-th block image obtained by the sliding window in the r-th frame of the original infrared image in the order from left to right and from top to bottom, where represents the Euclidean space, r = 1, 2,...L, s = 1, 2,...n;

[0020] Step 1.3. Stack the block images at the same spatial position in each frame of the original infrared image in chronological order to form a third-order local spatio-temporal tensor, that is, stack P rs in chronological order to form a third-order local spatio-temporal tensor

[0021] Step 1.4. Stack the n local spatio-temporal tensors D s obtained in Step 1.3 in the fourth dimension according to the order of their traversal by the sliding window in a single frame image. After stacking, a fourth-order spatio-temporal tensor is obtained, that is, stack D s as a slice with subscript s in the fourth dimension. Specifically, it is expressed by the following formula:

[0022]

[0023] where \(a = 1, 2, \cdots, w\), \(b = 1, 2, \cdots, w\), \(r = 1, 2, \cdots, L\), \(s = 1, 2, \cdots, n\).

[0024] Further, the specific steps of step 2 are as follows:

[0025] Step 2.1: Expand the fourth-order background spatio-temporal tensor into three generalized tensor expansion matrices according to different dimensionality combinations and where \(n\) 1 \(= [1, 2, 3, 4]\), \(n\) 2 \(= [3, 1, 2, 4]\), \(n\) 3 \(= [2, 3, 1, 4]\) respectively represent three different dimensionality combinations, and are vectors, representing the first half and the second half of \(n\) k respectively, represents the generalized tensor expansion matrix of the fourth-order background spatio-temporal tensor \(B\) 4D , \(k = 1, 2, 3\);

[0026] Step 2.2: Extract the spatio-temporal prior of the target in each infrared original image, and construct the weight factor of the target tensor, including the following steps:

[0027] Step 2.2.1: For the \(r\)-th infrared original image \(D\) r , extract the feature points of \(D\) r and \(D\) r+i . The value range of \(i\) is \((-m, -m + 1, \cdots, m - 1, m)\), then perform pairwise feature point matching, and fit the transformation matrix \(T\) according to the matched feature points. The feature points include SIFT features;

[0028] Step 2.2.2: Transform \(D\) r+i to \(D\) r according to the transformation matrix \(T\), and denote the transformed image as \(D'\) r+i . Calculate the difference between \(D\) r and \(D'\) r+i , and take the mean of the differences of \(2m\) frames as the spatio-temporal prior \(W\) r of \(D\) r ;

[0029] Step 2.2.3: Construct the prior weight r for the obtained \(L\) spatio-temporal priors \(W\) The construction method is the same as that of constructing the fourth-order spatio-temporal tensor ;

[0030] Step 2.3: Use the nuclear norms of \(L1\), \(L2\) and \(L3\) and the prior weight Construct the objective function, the formula is as follows:

[0031]

[0032]

[0033] where N = 3, ||L k || * represents the nuclear norm of the generalized tensor unfolding matrix of L k , λ1 is the balance coefficient, and its value range is 1 to 100. represents the fourth-order target spatio-temporal tensor, s.t. represents the constraint condition. is 's weighting factor, L k is 's k-th generalized tensor unfolding matrix, α k is the weight parameter of L k and satisfies

[0034] Furthermore, the specific steps of step 3 are as follows:

[0035] Step 3.1: Construct the augmented Lagrangian function based on the objective function;

[0036] Step 3.2: Iteratively solve the fourth-order target spatio-temporal tensor and the fourth-order background spatio-temporal tensor

[0037] Step 3.3: When the error between and is less than 10^(-6), stop the iteration; otherwise, update and and execute step 3.3 again. The update formula for the fourth-order background spatio-temporal tensor is:

[0038]

[0039] where N is the number of generalized tensor unfolding matrices of the fourth-order background spatio-temporal tensor , is the value obtained after update, N = 3, γ(L k +C k / μ k ) represents reconstructing the matrix L k +C k / μ k into a fourth-order spatio-temporal tensor. Here, ε and C k are Lagrange multipliers, γ and u k represent penalty factors, k = 1, 2, 3 is the serial number of the generalized tensor unfolding matrix.

[0040] An infrared target detection system based on a fourth-order spatio-temporal tensor, comprising:

[0041] A fourth-order spatio-temporal tensor construction module: obtaining L consecutive frames of original infrared images and constructing a fourth-order spatio-temporal tensor

[0042] A target function construction module: performing low-rank and sparse decomposition on That is, using the nuclear norms of the three generalized tensor unfolding matrices of To characterize the low rank of the fourth-order background spatio-temporal tensor And extracting the spatio-temporal prior of the target in the original infrared image to construct the weight factor of the target tensor, and constructing a target function;

[0043] A solution module: iteratively solving the target function constructed in step 2 to obtain a low-rank fourth-order background spatio-temporal tensor And a sparse fourth-order target spatio-temporal tensor

[0044] An infrared target detection module: reconstructing the fourth-order background spatio-temporal tensor And the fourth-order target spatio-temporal tensor In the reverse order of constructing the fourth-order spatio-temporal tensor into L frames of background images and target images, and the obtained target image is the infrared target.

[0045] Furthermore, the specific implementation steps of the fourth-order spatio-temporal tensor construction module are as follows:

[0046] Step 1.1: Obtain L consecutive frames of original infrared images;

[0047] Step 1.2: Use a sliding window of size w×w and traverse the L frames of original infrared images with a step size of w. Each frame of the original infrared image is divided into n block images by the sliding window. Denote As the s-th block image obtained by the sliding window in the r-th frame of the original infrared image in the order from left to right and from top to bottom, where Represents the Euclidean space, r = 1, 2,...L, s = 1, 2,...n;

[0048] Step 1.3: Stack the block images at the same spatial position in each frame of the original infrared image in chronological order into a third-order local spatio-temporal tensor, that is, stack P rs Into a third-order local spatio-temporal tensor in chronological order

[0049] Step 1.4: Stack the n local spatio-temporal tensors D s Obtained in step 1.3 in the fourth dimension in the order of their traversal by the sliding window in a single frame image. After stacking, a fourth-order spatio-temporal tensor Coming soon A slice with subscript s in the fourth dimension is expressed as follows:

[0050]

[0051] Among them, a=1,2,...,w,b=1,2,...,w,r=1,2,...L, s=1,2,...,n.

[0052] Furthermore, the specific implementation steps of the objective function construction module are:

[0053] Step 2.1: Fourth-order background space-time tensor According to different dimensional combinations, it is expanded into three generalized tensor expansion matrices and where n 1 =[1, 2, 3, 4], n 2 =[3,1,2,4],n 3 =[2, 3, 1, 4] represent three different dimension combinations, and are vectors, representing n k The first and second half of Represents the fourth-order background space-time tensor B 4D The generalized tensor expansion matrix of , k = 1, 2, 3;

[0054] Step 2.2, extracting the spatial and temporal priors of the target in each infrared original image and constructing the weight factor of the target tensor, including the following steps:

[0055] Step 2.2.1: For the infrared original image D of the rth frame r , extract D r With D r+i The value range of i is (-m, -m+1, ..., m-1, m), and then the feature points are matched in pairs, and the transformation matrix T is fitted according to the matched feature points. The feature points include SIFT features;

[0056] Step 2.2.2: According to the transformation matrix T, D r+i Transform to D r , the transformed image is recorded as D r+i , calculate D r With D r+i The difference of 2m frames is taken as the average value of D r The space-time prior W r ;

[0057] Step 2.2.3: Get the L frame space-time prior W r Constructing prior weights The construction method is the same as that for constructing a fourth-order spatio-temporal tensor ;

[0058] Step 2.3: Construct an objective function using the nuclear norms and prior weights of L1, L2, and L3 The formula is as follows:

[0059]

[0060]

[0061] where N = 3, ||L k || * denotes the nuclear norm of the generalized tensor expansion matrix of L k , λ1 is a balance coefficient with a value range of 1 to 100, denotes the fourth-order target spatio-temporal tensor, s.t. represents the constraint condition, is 's weighting factor, L k is 's k-th generalized tensor expansion matrix, α k is the weight parameter of L k and satisfies

[0062] Furthermore, the specific implementation steps of the solving module are as follows:

[0063] Step 3.1: Construct an augmented Lagrangian function based on the objective function;

[0064] Step 3.2: Iteratively solve the fourth-order target spatio-temporal tensor and the fourth-order background spatio-temporal tensor

[0065] Step 3.3: When the error between and is less than 10^(-6), stop the iteration; otherwise, update and and execute Step 3.3 again. The update formula for the fourth-order background spatio-temporal tensor is:

[0066]

[0067] where N is the number of generalized tensor expansion matrices of the fourth-order background spatio-temporal tensor , is the value obtained after update, N = 3, γ(L k +C k / μ k ) represents the matrix L k +C k / μ k Reconstructed into a fourth-order spatio-temporal tensor, where ε and C k are Lagrange multipliers, γ and u k represent penalty factors, and k = 1, 2, 3 is the serial number of the generalized tensor expansion matrix.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] First, by constructing a fourth-order spatio-temporal tensor and taking the time domain alone as a dimension of the spatio-temporal tensor, the present invention overcomes the defects in the third-order tensor that the time domain information cannot be effectively utilized and the local correlation between spatial domains cannot be effectively utilized. The present invention can simultaneously utilize spatial correlation and temporal correlation, effectively restore and suppress the background, and retain the target details;

[0070] Second, the generalized tensor expansion matrix of the fourth-order spatio-temporal tensor of the present invention can fully explore the correlation between different modes in the fourth-order spatio-temporal tensor, making the restored background closer to the real background, thereby improving the detection accuracy of the target;

[0071] Third, by extracting the spatio-temporal prior of the target, the present invention effectively reduces the influence of background changes on target detection, can significantly suppress background clutter and strong noise with high contrast and sparse characteristics, and reduces the false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for use 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.

[0073] Figure 1 is the flowchart of the present invention;

[0074] Figure 2 is the infrared original image containing the target and its three-dimensional distribution diagram of the present invention;

[0075] Figure 3 is for the present invention Figure 2 detection result diagram and three-dimensional distribution diagram;

[0076] Figure 4 is the detection result diagram and three-dimensional gray scale diagram of the MPCM method for Figure 2 ;

[0077] Figure 5 is the detection result diagram and three-dimensional gray scale diagram of the IPI method for Figure 2 ;

[0078] Figure 6The detection result diagram and three-dimensional grayscale diagram of the PSTNN method for Figure 2 Specific implementation mode

[0079] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation modes.

[0080] An infrared target detection method based on a fourth-order spatio-temporal tensor includes the following steps:

[0081] Step 1, obtain L consecutive frames of original infrared images and construct a fourth-order spatio-temporal tensor The specific steps are as follows:

[0082] Step 1.1, obtain L consecutive frames of original infrared images;

[0083] Step 1.2, use a sliding window of size w×w and traverse the L frames of original infrared images with a step size of w. Each frame of the original infrared image is divided into n block images by the sliding window. Denote as the s-th block image obtained by the sliding window in the r-th frame of the original infrared image in the order from left to right and from top to bottom, where represents the Euclidean space, r = 1, 2,...L, s = 1, 2,...n;

[0084] Step 1.3, stack the block images at the same spatial position in each frame of the original infrared image in chronological order to form a third-order local spatio-temporal tensor, that is, stack P rs in chronological order to form a third-order local spatio-temporal tensor

[0085] Step 1.4, stack the n local spatio-temporal tensors D s obtained in Step 1.3 in the fourth dimension according to the order of their traversal by the sliding window in a single-frame image. After stacking, a fourth-order spatio-temporal tensor is obtained, that is, use D s as a slice with subscript s in the fourth dimension, which is specifically expressed by the following formula:

[0086]

[0087] where a = 1, 2,...w, b = 1, 2,...w, r = 1, 2,...L, s = 1, 2,...n.

[0088] Step 2, perform low-rank sparse decomposition on , that is, use the nuclear norms of the three generalized tensor unfolding matrices of ​The low-rank property is utilized, and the spatio-temporal prior of the target in the infrared original image is extracted to construct the weight factor of the target tensor, and the objective function is constructed. The specific steps are as follows:

[0089] Step 2.1, Fourth-order background spatio-temporal tensor Expand it into three generalized tensor expansion matrices according to different dimension combinations and where n 1 =[1, 2, 3, 4], n 2 =[3, 1, 2, 4], n 3 =[2, 3, 1, 4] respectively represent three different dimension combinations, and are vectors, representing the first half and the second half of n k respectively, represents the generalized tensor expansion matrix of the fourth-order background spatio-temporal tensor B 4D , k = 1, 2, 3;

[0090] Step 2.2, Extract the spatio-temporal prior of the target in each infrared original image, and construct the weight factor of the target tensor, including the following steps:

[0091] Step 2.2.1, For the r-th frame of the infrared original image D r , extract the feature points of D r and D r+i . The value range of i is (-m, -m + 1,..., m - 1, m), and then perform pairwise feature point matching. According to the matched feature points, fit the transformation matrix T. The feature points include SIFT features;

[0092] Step 2.2.2, According to the transformation matrix T, transform D r+i to D r . Denote the transformed image as D r+i , calculate the difference between D r and D r+i . Take the mean of the 2m-frame differences as the spatio-temporal prior W r of D r ;

[0093] Step 2.2.3, For the obtained L-frame spatio-temporal priors W r , construct the prior weight . The construction method is the same as that for constructing the fourth-order spatio-temporal tensor ;

[0094] Step 2.3, Utilize the nuclear norms of L1, L2, and L3 and the prior weight to construct the objective function. The formula is as follows:

[0095]

[0096]

[0097] Among them, N = 3, ||L k || + represents L k the nuclear norm of the generalized tensor unfolding matrix, λ1 is the balance coefficient, and its value range is 1 to 100. represents the fourth-order target spatio-temporal tensor, and s.t. represents the constraint condition. is the weighting factor of, L k is the k-th generalized tensor unfolding matrix of, α k is L k the weight parameter of, satisfying

[0098] Step 3: Iteratively solve the objective function constructed in Step 2 to obtain the low-rank fourth-order background spatio-temporal tensor and the sparse fourth-order target spatio-temporal tensor The specific steps are as follows:

[0099] Step 3.1: Construct the augmented Lagrangian function based on the objective function;

[0100] Step 3.2: Iteratively solve the fourth-order target spatio-temporal tensor and the fourth-order background spatio-temporal tensor

[0101] Step 3.3: When and the error between them is less than 10^(-6), stop the iteration, otherwise update and and execute Step 3.3 again. Among them, the update formula of the fourth-order background spatio-temporal tensor is:

[0102]

[0103] Among them, N is the number of generalized tensor unfolding matrices of the fourth-order background spatio-temporal tensor is the value obtained after updating, N = 3, γ(L k + C k / μ k ) represents reconstructing the matrix L k + C k / μ k into a fourth-order spatio-temporal tensor. Among them, ε and C k are Lagrange multipliers, γ and u k represent the penalty factors, k = 1, 2, 3, which are the serial numbers of the generalized tensor unfolding matrices.​

[0104] Step 4: Based on the fourth-order background spatio-temporal tensor and the fourth-order target spatio-temporal tensor Reconstruct the background image and the target image of L frames according to the reverse process of constructing the fourth-order spatio-temporal tensor in Step 1 (i.e., the reverse process of Steps 1.1 to 1.4). The obtained target image is the infrared target.

[0105] Figure 2 In the infrared image shown, the target is small and has low contrast. In its background, in addition to woods and roads, there are also many clutter points similar to the target, which inevitably interfere with the detection of the target. Figure 3 The figure shows the target image and the three-dimensional distribution map separated by the present invention. It can be seen that the present invention can well detect the real target and suppress the complex background. While the MPCM, IPI, and PSTNN methods still contain a large amount of background residuals and false alarms. By constructing a fourth-order spatio-temporal tensor and integrating time-domain information, the generalized tensor expansion matrix can fully explore the correlation between different modalities in the fourth-order spatio-temporal tensor, obtain a more refined background, and assist in the detection of small targets by extracting the spatio-temporal prior of the target, reducing the influence of noise and high-brightness regions on the target image.

[0106] The above are only representative embodiments within the many specific application scopes of the present invention, and do not constitute any limitation to the protection scope of the present invention. Any technical solutions formed by transformation or equivalent replacement fall within the scope of the rights protection of the present invention.

Claims

1. An infrared target detection method based on a fourth-order spatio-temporal tensor, characterized in that, Including the following steps: Step 1. Obtain consecutive raw infrared images of frames and construct a fourth-order spatio-temporal tensor ; Step 2. Perform low-rank sparse decomposition on , that is, use the nuclear norms of the three generalized tensor unfolding matrices of to characterize the low-rank property of the fourth-order background spatio-temporal tensor , and extract the spatio-temporal domain prior of the target in the original infrared image to construct the weight factor of the target tensor and construct the objective function. Step 2.1, Fourth-order background spatio-temporal tensor Expand it into three generalized tensor expansion matrices according to different dimensional combinations , and , where respectively represent three different dimensional combinations, and are vectors, representing the first half and the second half of respectively, represents the generalized tensor expansion matrix of the fourth-order background spatio-temporal tensor , ; Step 2.2: Extract the spatio-temporal prior of the target in each original infrared image, and construct the weight factor of the target tensor, including the following steps: Step 2.2.1: For the original infrared image of the r-th frame , extract and feature points. The value range of i is (-m, -m + 1, …, m - 1, m). Then, pairwise feature point matching is performed, and the transformation matrix T is fitted according to the matched feature points. The feature points include SIFT features; Step 2.2.2: According to the transformation matrix T, is transformed to . Denote the transformed image as . Calculate the difference between and . Take the mean of the differences of 2m frames as the spatio-temporal prior of ; Step 2.2.

3. For the obtained frame spatio-temporal prior construct a prior weight , and the construction method is the same as that for constructing the fourth-order spatio-temporal tensor ; Step 2.

3. Using , and 's nuclear norm and prior weight to construct an objective function, the formula is as follows: Among them, N = 3, denotes the nuclear norm of the generalized tensor unfolding matrix, is the balance coefficient, and its value range is 1 to 100, denotes the fourth-order target spatio-temporal tensor, and s.t. denotes the constraint condition, is the weighting factor of is the th generalized tensor unfolding matrix of is the weight parameter of ; Step 3: Iteratively solve the objective function constructed in Step 2 to obtain a low-rank fourth-order background spatio-temporal tensor and a sparse fourth-order target spatio-temporal tensor ; Step 4: Based on the fourth-order background spatio-temporal tensor and the fourth-order target spatio-temporal tensor Reconstruct them into the background image and the target image of the frame in the reverse process of constructing the fourth-order spatio-temporal tensor in Step 1 The obtained target image is the infrared target 2. The infrared target detection method based on a fourth-order spatio-temporal tensor according to claim 1, wherein The specific steps of the said Step 1 are: Step 1.1, obtain consecutive frames of original infrared images; Step 1.

2. Use a sliding window of size × to traverse the original infrared images frame by frame with a step size of . Each original infrared image frame is divided into block images by the sliding window. Denote as the s-th block image obtained by the sliding window in the r-th original infrared image frame in the order from left to right and from top to bottom, where represents the Euclidean space, , , . Step 1.3: Stack the block images at the same spatial position in each frame of the original infrared image in chronological order to form a third-order local spatio-temporal tensor, that is, Stack them in chronological order to form a third-order local spatio-temporal tensor ; Step 1.4: Stack the local spatio-temporal tensors in the fourth dimension in the order of their traversal of the sliding window in the single-frame image. After stacking, a fourth-order spatio-temporal tensor is obtained, that is, is used as a slice with subscript in the fourth dimension. Specifically, it is expressed by the following formula: Among them, , , , .

3. The infrared target detection method based on a fourth-order spatio-temporal tensor according to claim 2, characterized in that, The specific steps of the said Step 3 are: Step 3.1: Construct an augmented Lagrangian function based on the objective function; Step 3.2: Iteratively solve the fourth-order target spatio-temporal tensor and the fourth-order background spatio-temporal tensor based on the augmented Lagrangian function ; ; Step 3.

3. When the error between the fourth-order spatio-temporal tensor and is less than 10^(-6), stop the iteration; otherwise, update and and execute Step 3.3 again. The update formula for the fourth-order background spatio-temporal tensor is as follows: Among them, is a fourth-order background spatio-temporal tensor The number of generalized tensor unfolding matrices, is the value obtained after updating, = 3, indicates that the matrix is reconstructed into a fourth-order spatio-temporal tensor, where and are Lagrange multipliers, and represent the penalty factor, k = 1, 2, 3, which is the serial number of the generalized tensor unfolding matrix.

4. An infrared target detection system based on a fourth-order spatio-temporal tensor, characterized in that, Including: Fourth-order spatio-temporal tensor construction module: Obtain consecutive frame original infrared images and construct a fourth-order spatio-temporal tensor ; Objective function construction module: Perform low-rank sparse decomposition on That is, use the nuclear norms of the three generalized tensor unfolding matrices of to characterize the low-rank property of the fourth-order background spatio-temporal tensor and extract the target spatio-temporal domain prior in the original infrared image to construct the weight factor of the target tensor, and construct the objective function; The specific implementation steps of the said objective function construction module are: Step 2.1, Fourth-order background spatio-temporal tensor Expand it into three generalized tensor expansion matrices according to different dimensional combinations , and , where represent three different dimensional combinations respectively, and are vectors, representing the first half and the second half of respectively, represents the generalized tensor expansion matrix of the fourth-order background spatio-temporal tensor , ; Step 2.2: Extract the spatio-temporal prior of the target in each original infrared image, and construct the weight factor of the target tensor, including the following steps: Step 2.2.1: For the original infrared image of the r-th frame , extract and feature points. The value range of i is (-m, -m + 1, …, m - 1, m). Then, pairwise feature point matching is performed, and the transformation matrix T is fitted according to the well-matched feature points. The feature points include SIFT features; Step 2.2.2: According to the transformation matrix T, transform to . Denote the transformed image as . Calculate the difference between and . Take the mean of the differences of 2m frames as the spatio-temporal prior of ; Step 2.2.

3. For the obtained frame spatio-temporal prior construct a prior weight , and the construction method is the same as that for constructing the fourth-order spatio-temporal tensor ; Step 2.3: Using , and 's nuclear norm and prior weights to construct an objective function, the formula is as follows: Among them, N = 3, denotes the nuclear norm of the generalized tensor unfolding matrix, is the balance coefficient, and its value range is 1 to 100, denotes the fourth-order target spatio-temporal tensor, and s.t. denotes the constraint condition, is the weighting factor of is the th generalized tensor unfolding matrix of is the weight parameter of ; Solution module: Iteratively solve the constructed objective function to obtain a low-rank fourth-order background spatio-temporal tensor and a sparse fourth-order target spatio-temporal tensor ; Infrared target detection module: based on the fourth-order background spatio-temporal tensor and the fourth-order target spatio-temporal tensor Reconstructed in the reverse order of the process of constructing the fourth-order spatio-temporal tensor into the background image and the target image of the frame, and the obtained target image is the infrared target.

5. The infrared target detection system based on a fourth-order spatio-temporal tensor according to claim 4, wherein The specific implementation steps of the said fourth-order spatio-temporal tensor construction module are: Step 1.1, obtain consecutive frames of original infrared images; Step 1.

2. Use a sliding window of size × to traverse the original infrared images frame by frame with a step size of . Each original infrared image frame is divided by the sliding window into block images. Denote as the s-th block image obtained by the sliding window in the r-th original infrared image frame in the order from left to right and from top to bottom. Among them, represents the Euclidean space, , , ; Step 1.3: Stack the block images at the same spatial position in each frame of the original infrared image in chronological order to form a third-order local spatio-temporal tensor, that is, Stack them in chronological order to form a third-order local spatio-temporal tensor ; Step 1.4: Stack the local spatio-temporal tensors in the fourth dimension according to the order of their sliding window traversal in a single-frame image. After stacking, a fourth-order spatio-temporal tensor is obtained, that is, is used as a slice with subscript in the fourth dimension. Specifically, it is expressed by the following formula: Among them, , , , .

6. The infrared target detection system based on a fourth-order spatio-temporal tensor according to claim 5, characterized in that The specific implementation steps of the said solution module are: Step 3.1: Construct an augmented Lagrangian function based on the objective function; Step 3.2: Iteratively solve the fourth-order target spatio-temporal tensor and the fourth-order background spatio-temporal tensor based on the augmented Lagrangian function ; ; Step 3.

3. When and the error between them is less than 10^(-6), stop the iteration, otherwise update and and execute Step 3.3 again. The update formula for the fourth-order background space-time tensor is: Among them, is a fourth-order background spatio-temporal tensor The number of generalized tensor expansion matrices, is the value obtained after updating, = 3, indicates that the matrix is reconstructed into a fourth-order spatio-temporal tensor, where and are Lagrange multipliers, and represent penalty factors, k = 1, 2, 3, which are the sequence numbers of the generalized tensor expansion matrices.

Images