Space-time tensor decomposition model based on sparse target perception and total variation constraint

Abstract

The invention discloses a space-time tensor decomposition model based on sparse target perception and total variation constraint, and relates to the technical field of infrared image processing and target detection, and the model extracts image blocks from a sequence image containing a current frame and two frames before and after the current frame by using a sliding window, and stacks the image blocks to construct a space-time tensor. By analyzing image gradient characteristics, a target prior information graph is calculated by using a multi-direction gray level jump mechanism, and a sparse target perception prior weight tensor is formed after square box filtering and reweighting processing. A tensor truncation nuclear norm is introduced in model construction to more accurately approach the low-rank characteristic of a background component, and total variation constraint is applied to the background component and a target component. And finally, integrating the prior information with various constraints, establishing a unified optimization model, solving by adopting an alternating direction multiplier method, and separating a background tensor, a target tensor and a noise tensor. According to the method, complex background and noise interference can be effectively suppressed, and the detection precision and integrity of the infrared small target are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing and target detection technology, specifically to a spatiotemporal tensor decomposition model based on sparse target perception and total variation constraints. Background Technology

[0002] In infrared search and track systems, the detection of small infrared targets against complex backgrounds plays a crucial role. However, the limited detection range of infrared sensors and interference from background clutter result in low signal-to-noise ratios for infrared targets imaged at long distances. Furthermore, the lack of shape, color, and texture features of the target also presents numerous challenges for detection. Therefore, infrared small target detection algorithms have significant research value and importance.

[0003] Currently, there are four types of infrared small target detection methods: filtering-based methods, local contrast-based methods, deep learning-based methods, and low-rank sparse recovery-based methods. Filtering methods mainly include maximum median / mean filtering, high-pass filtering, and morphological filtering. To simulate the contrast mechanism of the human visual system, many scholars have conducted comprehensive and in-depth research on local contrast-based methods. When the target is very prominent and the background is relatively smooth, local contrast-based detection methods perform well. However, their detection ability is often poor when affected by prominent edges, heterogeneous regions, or dark targets. Furthermore, these methods rarely consider global background features; existing algorithms mainly use ratios and differences, which are ineffective in removing potential interference targets in complex backgrounds. Deep learning-based methods have significant advantages in target detection tasks, but they have not achieved similar success in small target extraction tasks as in other fields. A key reason is that the target scale is too small, making it difficult for convolutional kernels to obtain effective features to represent the target. In addition, small-scale data samples cannot fully utilize the powerful ability of neural networks to represent target features. In the foreseeable future, this makes it difficult for deep neural networks to fully realize their potential.

[0004] With the rise of compressed sensing, low-rank sparse recovery-based methods have attracted widespread attention from researchers, demonstrating their enormous development potential and promising prospects. These methods not only focus on the attributes of the target but also on the self-correlation information of the background, with the Infrared Partition Image (IPI) model being a typical example. Attempts by domestic and international scholars to apply tensor decomposition to video and image processing have validated the powerful representational capabilities of tensors, providing strong technical support for their application in infrared small target detection. Furthermore, high-order data representation based on tensors can not only uncover richer intrinsic features of infrared images but also fuse the temporal and spatial dimensions of infrared image sequences, overcoming the spatiotemporal limitations of traditional single-frame detection. Therefore, many excellent low-rank sparse tensor decomposition models have been proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a spatiotemporal tensor decomposition model based on sparse target perception and total variation constraints to solve the problems proposed in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a spatiotemporal tensor decomposition model based on sparse target perception and total variation constraints, wherein the spatiotemporal tensor decomposition model includes the following steps: Step S100: Select the current frame to be detected and the two frames before and after it from the sequence image, extract image patches using a sliding window, stack them into a three-dimensional small tensor, and merge all the small tensors along the time dimension into a spatiotemporal tensor D. In step S100, selecting a sliding window block with a size of patchsizeR×patchsizeC (psR×psC) from the image patch extraction using a sliding window specifically includes the following steps: Step S101: Traverse the current frame K from left to right and from top to bottom. The step size when traversing from left to right is stepC(sC), and the step size when traversing from top to bottom is stepR(sR). Finally, the current frame is divided into 9 patches. According to the traversal order, each patch is numbered i and denoted as Ki, which represents the i-th patch of the K-th frame image. The i-th value is an integer from 1 to 9. In step S101, psR, psC, sR, and sC are calculated as follows: ; In the formula, R and C represent the height and width of the current frame, respectively. The addition of 6 after rounding is to ensure that there is an overlap between each patch, preventing the target from being split between two patches. Step S102: Traverse the two frames before and after the current frame K using the sliding window block and stride to obtain the patches of each frame. Stack the patches with the same number in all frames to form a small tensor. For tensors Stacking the elements in ascending order of i yields a spacetime tensor of size psR×psC×45. .

[0007] Step S200: Calculate the prior information map of the sequence images, construct a tensor by performing box filtering on the prior information map, take the reciprocal of all elements in the tensor, and calculate the perceptual prior weight tensor W of the sparse target. P ; In step S200, the step of calculating the prior information of the target image is as follows: traversing the sequence of image pixels from left to right and from top to bottom, and constructing an 11×11 rectangular region centered on the currently traversed pixel. This specifically includes the following steps: Step S201: Calculate the gray-level transition set of the current center pixel in 8 directions, select the maximum value in the gray-level transition set in each direction to form the maximum gray-level transition set, and use the product of the minimum value and the mean value in the maximum transition set and the square of the product as the prior information of the current pixel. It is stipulated that the number of pixels in each direction is 5. After the pixel traversal calculation of the image is completed, the prior information (STP) map of the target image is obtained. The prior information map of the target image in step S201 is obtained by the following formula: ; In the formula, STP(x,y) represents the prior information of the current pixel, and GSMAX... d GJS represents the maximum value in the set of grayscale jumps along direction d, CM represents the mean of the current center pixel and its 8 neighboring pixels, and GJS represents the mean of the grayscale jumps along direction d. d Let f(x,y) represent the set of grayscale jumps along direction d. The value of η is 0, 1, 2, 3, or 4. When η=0, f(x,y) represents the center pixel of an 11×11 rectangular region. Step S202: Perform a box filter operation on the STP image to obtain the target image weight map w. T ; In step S202, the box filtering operation is specifically as follows: ; In the formula, box(·) represents box filtering, and 1 represents the radius of the box; Step S203: Obtain the target image weight map w of the 5 frames. T Then, it is converted into tensor form W according to the spacetime tensor construction method in step S100. T ; Step S204: By adjusting W T The prior weight tensor W is obtained by taking the reciprocal of all elements in the set. P ; Step S205: For W P Perform reweighting to obtain the reweighted target image enhancement weights W. rP .

[0008] In step S205, the reweighted target enhancement weight W rP Specifically: ; In the formula, W r It is a coefficient-enhanced weighting. It is the Hadamard product. Let n represent the iteration number. In the (n+1)th iteration, W r n+1 It can be defined as: ; In the formula, This is a preset constant used to prevent the denominator from being 0; in this paper, it is set to 1. This represents the target tensor obtained in the nth iteration.

[0009] Step S300: Introduce a tensor truncation kernel norm to approximate the background rank. Based on the correlation between the background component and the target component in different dimensions, apply total variation constraints to them respectively to more effectively separate the target from the background. In step S300, to more accurately approximate the background rank, the definition of the tensor truncation norm is proposed for the background tensor B. For a tensor... tensor truncation kernel norm The definition is as follows: ; In the formula, P and Q are obtained by performing singular value decomposition on X to obtain two orthogonal tensors U and V, respectively. J is the unit tensor. The tensor obtained by retaining only the first r columns of the second dimension of the orthogonal tensors while keeping the other dimensions unchanged is specifically represented as follows: ; Set the truncated singular value size to 0.1 to obtain the tensor truncated rank r, which is to minimize the sum of all singular values ​​less than 0.1; The total variation constraints for the background and target components are defined as follows: ; In the formula, express Difference operators in the horizontal and vertical directions, express Difference operators in the horizontal direction and the third dimension; i, j, k represent the positions of pixels in the three dimensions, and B and T are the background tensor and target tensor to be decomposed by the model.

[0010] Step S400: Convert the prior weight tensor W P Integrate the constraints and use the spatiotemporal tensor D as the data basis to construct a new optimization model; The expression for the new optimization model in step S400 is: ; In the formula, R psR×psC×45 , representing the background tensor, target tensor, and noise tensor respectively, with the balance between the components determined by coefficients. control.

[0011] Step S500: Decompose the new optimization model into subproblems using the Alternating Direction Multiplier Method (ADMM), and solve each subproblem iteratively to obtain the target tensor. Background tensor and noise tensor ; Step S500 uses the Alternating Direction Multiplier Method (ADMM) to solve the optimization model from step S400, specifically including the following steps: Step S501: Introduce auxiliary variables The optimization model in step S400 is rewritten; In step S501, the rewritten optimization model is as follows: ; Step S502 expresses the rewritten model using the augmented Lagrange equation; In step S502, the augmented Lagrange equation is expressed as: ; In the formula, arrive Represents the Lagrange multipliers. It is a punishment factor; Step S503: Decompose the augmented Lagrange equation into subproblems and solve them. When updating one variable, fix the other variables and update all variables alternately. In step S503, assuming k is the iteration number, the subproblems and solution methods of the augmented Lagrange equation in step S502 can be described as follows: The subproblems are: ; Solve the above equation using the partial singular value thresholding operator, assuming X∈R. n1×n2 , Y∈R n1×n2 l = min(n1, n2), N ≤ l, Y can be expressed as the sum of two matrices, i.e., Y = Y1 + Y2, and Y1 = (U Y1 )D Y1 (V Y1 ) T Y2=(U Y2 )D Y2 (V Y2 ) T , among which, U Y1 and V Y1 Corresponding to the first N singular values ​​of Y, U Y2 and V Y2 For the remaining singular values, a minimization problem is defined as follows: ; In the formula, The singular value threshold operator is defined as: (This represents the sum of some singular values.) ; The optimal solution to the above equation is: ; In the formula, , , ; The subproblems are: ; The equivalent system of linear equations for the above expression is: ; In the formula, Indicates matrix transpose. , , , ; The closed-form solution to the above equivalent linear system of equations is: ; In the formula, These represent complex conjugate, fast Fourier transform, and inverse Fourier transform, respectively. The subproblems are: ; The above equation is solved using the soft threshold operator, specifically: ; Assume τ>0, X∈R n1×n2 , Y∈R n1×n2 Define a The norm minimization problem is as follows: ; The above equation has a closed-form solution S. τ ( Y ), where S τ The soft threshold operator is defined as follows: ; The subproblems are: ; The equivalent system of linear equations for the above expression is: ; In the formula, , ; The closed-form solution to the above equivalent linear system of equations is: ; The subproblems are: ; The above equation is solved using the soft threshold operator, and the result is: ; The subproblems are: ; The solution to the above equation is: ; Step S504: Update the Lagrange multipliers and penalty factors; In step S504, the Lagrange multipliers are updated. arrive The expression is: ; The updated penalty factor expression is: ; In the formula, Indicates the penalty factor The growth rate will be used to set the stopping condition as follows: ; In the formula, It is a constant representing the minimum acceptable error when iteration stops; this paper sets it to 2 × 10⁻⁶. -3+ .

[0012] Step S600: After optimization, the optimized target tensor Background tensor and noise tensor The target enhancement images are reconstructed by reversing the steps of step S100, resulting in 5 frames of target enhancement images. One frame is taken as the detection response map of the current frame, and the target detection result map, background image, and noise image are obtained.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1. High detection accuracy and complete target recovery: This invention calculates prior information about the target through a multi-directional grayscale transition mechanism and combines it with a reweighting strategy to enhance the saliency of the real target while accurately suppressing sparse interference such as edges and corners in the background. This enables the model to recover targets with more complete outlines and higher signal-to-noise ratios even in complex backgrounds.

[0014] 2. Strong background suppression and low false alarm rate: By introducing a tensor truncated kernel norm to approximate the background rank, the low-rank characteristics of the background are accurately described, solving the problem of insufficient contribution of large singular values ​​in traditional kernel norm minimization. At the same time, a total variation constraint is applied to the background component to enhance the spatial smoothness of the background, resulting in stronger suppression of complex texture backgrounds and noise, and reducing false alarms.

[0015] 3. Full utilization of spatiotemporal information and stable detection results: A spatiotemporal tensor is constructed using multiple frames of images, overcoming the limitations of single-frame detection. The spatial structure and temporal information of the image sequence are utilized. A total variational constraint in the temporal dimension is applied to the target component, ensuring the continuity of target motion in consecutive frames and effectively combating inter-frame flicker noise.

[0016] 4. Good model generalization and strong adaptability: This invention does not rely on specific features such as the shape and texture of the target, but rather on its general sparsity and difference from the background for detection. Therefore, this model can exhibit stable detection performance for small infrared targets in different scenarios, including dark targets and moving targets. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the spatiotemporal tensor decomposition model based on sparse target perception and total variation constraints of the present invention. Figure 2 This is a flowchart illustrating the spatiotemporal tensor decomposition model based on sparse target perception and total variation constraints of the present invention. Figure 3 This is a schematic diagram illustrating the calculation of multi-directional grayscale jumps within an 11×11 rectangular region in an embodiment of the spatiotemporal tensor decomposition model based on sparse target perception and total variation constraints of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Figures 1-3 As shown, this invention provides a technical solution based on a sparse target perception and a spatiotemporal tensor decomposition model with total variational constraints. First, the infrared image sequence to be detected is read (taking an image with a resolution of 256×256 as an example). At the start of detection, the current frame to be detected, plus the two frames before and after it (a total of 5 frames) are selected. Initial parameters are set, including the image patch size (92 × 92), the sliding step size (stepR=83, stepC=83), and tradeoff parameters. Punishment factor =0.02, truncation of singular values ​​is 0.1, and the minimum error of the iteration is 0.002.

[0020] Step S100: Select the current frame to be detected and the two frames before and after it from the sequence image, extract image patches using a sliding window, stack them into a three-dimensional small tensor, and merge all the small tensors along the time dimension into a spatiotemporal tensor D. In step S100, selecting a sliding window block with a size of patchsizeR×patchsizeC (psR×psC) from the image patch extraction using a sliding window specifically includes the following steps: Step S101: Traverse the current frame K from left to right and from top to bottom. The step size when traversing from left to right is stepC(sC), and the step size when traversing from top to bottom is stepR(sR). Finally, the current frame is divided into 9 patches. According to the traversal order, each patch is numbered i and denoted as Ki, which represents the i-th patch of the K-th frame image. The i-th value is an integer from 1 to 9. In step S101, psR, psC, sR, and sC are calculated as follows: ; In the formula, R and C represent the height and width of the current frame, respectively. The addition of 6 after rounding is to ensure that there is an overlap between each patch, preventing the target from being split between two patches. Step S102: Traverse the two frames before and after the current frame K using the sliding window block and stride to obtain the patches of each frame. Stack the patches with the same number in all frames to form a small tensor. For tensors Stacking the elements in ascending order of i yields a spacetime tensor of size psR×psC×45. .

[0021] Example 1: Traverse the current frame K using a sliding window of psR×psC from left to right and top to bottom. Number each patch i according to the traversal order, denoted as Ki. Stack patches with the same number in all frames to form a small tensor. Stack these patches in ascending order of i to obtain a spatiotemporal tensor of size psR×psC×45. .

[0022] Step S200: Calculate the prior information map of the sequence images, construct a tensor by performing box filtering on the prior information map, take the reciprocal of all elements in the tensor, and calculate the perceptual prior weight tensor W of the sparse target. P ; In step S200, the step of calculating the prior information of the target image is as follows: traversing the sequence of image pixels from left to right and from top to bottom, and constructing an 11×11 rectangular region centered on the currently traversed pixel. This specifically includes the following steps: Step S201: Calculate the gray-level transition set of the current center pixel in 8 directions, select the maximum value in the gray-level transition set in each direction to form the maximum gray-level transition set, and use the product of the minimum value and the mean value in the maximum transition set and the square of the product as the prior information of the current pixel. It is stipulated that the number of pixels in each direction is 5. After the pixel traversal calculation of the image is completed, the prior information (STP) map of the target image is obtained. The prior information map of the target image in step S201 is obtained by the following formula: ; In the formula, STP(x,y) represents the prior information of the current pixel, and GSMAX... d GJS represents the maximum value in the set of grayscale jumps along direction d, CM represents the mean of the current center pixel and its 8 neighboring pixels, and GJS represents the mean of the grayscale jumps along direction d. d Let f(x,y) represent the set of grayscale jumps along direction d. The value of η is 0, 1, 2, 3, or 4. When η=0, f(x,y) represents the center pixel of an 11×11 rectangular region. Step S202: Perform a box filter operation on the STP image to obtain the target image weight map w. T ; In step S202, the box filtering operation is specifically as follows: ; In the formula, box(·) represents box filtering, and 1 represents the radius of the box; Step S203: Obtain the target image weight map w of the 5 frames. T Then, it is converted into tensor form W according to the spacetime tensor construction method in step S100. T ; Step S204: By adjusting W T The prior weight tensor W is obtained by taking the reciprocal of all elements in the set. P ; Step S205: For W P Perform reweighting to obtain the reweighted target image enhancement weights W. rP .

[0023] In step S205, the reweighted target enhancement weight W rP Specifically: ; In the formula, W r It is a coefficient-enhanced weighting. It is the Hadamard product. Let n represent the iteration number. In the (n+1)th iteration, W r n+1 It can be defined as: ; In the formula, This is a preset constant used to prevent the denominator from being 0; in this paper, it is set to 1. This represents the target tensor obtained in the nth iteration.

[0024] Example 2: Using the current pixel as the center, construct an 11×11 rectangular region, and calculate the grayscale transition set GJSd (d=1,2,3,4,5,6,7,8) of the center pixel of the rectangular region in 8 directions, as shown below. Figure 3 The image shows details of an 11×11 rectangular area. Each small rectangle represents a pixel in its corresponding direction, and the upper left curve represents a grayscale jump value. The maximum grayscale jump value in each direction is calculated. , The prior information for the current pixel is: In the formula, CM represents the average of the center pixel of the rectangular region and its 8 neighboring pixels, which is used to suppress noise from isolated highlights; the fourth power is to further improve the difference between the target and the edge, thereby increasing the salience of the target.

[0025] After traversing all the original image pixels, the STP image obtained is subjected to box filtering to obtain the target weight map.

[0026] Step S300: Introduce a tensor truncation kernel norm to approximate the background rank. Based on the correlation between the background component and the target component in different dimensions, apply total variation constraints to them respectively to more effectively separate the target from the background. In step S300, to more accurately approximate the background rank, the definition of the tensor truncation norm is proposed for the background tensor B. For a tensor... tensor truncation kernel norm The definition is as follows: ; In the formula, P and Q are obtained by performing singular value decomposition on X to obtain two orthogonal tensors U and V, respectively. J is the unit tensor. The tensor obtained by retaining only the first r columns of the second dimension of the orthogonal tensors while keeping the other dimensions unchanged is specifically represented as follows: ; Set the truncated singular value size to 0.1 to obtain the tensor truncated rank r, which is to minimize the sum of all singular values ​​less than 0.1; The total variation constraints for the background and target components are defined as follows: ; In the formula, express Difference operators in the horizontal and vertical directions, express Difference operators in the horizontal direction and the third dimension; i, j, k represent the positions of pixels in the three dimensions, and B and T are the background tensor and target tensor to be decomposed by the model.

[0027] Step S400: Convert the prior weight tensor W P Integrate the constraints and use the spatiotemporal tensor D as the data basis to construct a new optimization model; The expression for the new optimization model in step S400 is: ; In the formula, R psR×psC×45 , representing the background tensor, target tensor, and noise tensor respectively, with the balance between the components determined by coefficients. control.

[0028] Step S500: Decompose the new optimization model into subproblems using the Alternating Direction Multiplier Method (ADMM), and solve each subproblem iteratively to obtain the target tensor. Background tensor and noise tensor ; Step S500 uses the Alternating Direction Multiplier Method (ADMM) to solve the optimization model from step S400, specifically including the following steps: Step S501: Introduce auxiliary variables The optimization model in step S400 is rewritten; In step S501, the rewritten optimization model is as follows: ; Step S502 expresses the rewritten model using the augmented Lagrange equation; In step S502, the augmented Lagrange equation is expressed as: ; In the formula, arrive Represents the Lagrange multipliers. It is a punishment factor; Step S503: Decompose the augmented Lagrange equation into subproblems and solve them. When updating one variable, fix the other variables and update all variables alternately. In step S503, assuming k is the iteration number, the subproblems and solution methods of the augmented Lagrange equation in step S502 can be described as follows: The subproblems are: ; Solve the above equation using the partial singular value thresholding operator, assuming X∈R. n1×n2 , Y∈R n1×n2 l = min(n1, n2), N ≤ l, Y can be expressed as the sum of two matrices, i.e., Y = Y1 + Y2, and Y1 = (U Y1 )D Y1 (V Y1 ) T Y2=(U Y2 )D Y2 (V Y2 ) T , among which, U Y1 and V Y1 Corresponding to the first N singular values ​​of Y, U Y2 and V Y2 For the remaining singular values, a minimization problem is defined as follows: ; In the formula, The singular value threshold operator is defined as: (This represents the sum of some singular values.) ; The optimal solution to the above equation is: ; In the formula, , , ; The subproblems are: ; The equivalent system of linear equations for the above expression is: ; In the formula, Indicates matrix transpose. , , , ; The closed-form solution to the above equivalent linear system of equations is: ; In the formula, These represent complex conjugate, fast Fourier transform, and inverse Fourier transform, respectively. The subproblems are: ; The above equation is solved using the soft threshold operator, specifically: ; Assume τ>0, X∈R n1×n2 , Y∈R n1×n2 Define a The norm minimization problem is as follows: ; The above equation has a closed-form solution S. τ ( Y ), where S τ The soft threshold operator is defined as follows: ; The subproblems are: ; The equivalent system of linear equations for the above expression is: ; In the formula, , ; The closed-form solution to the above equivalent linear system of equations is: ; The subproblems are: ; The above equation is solved using the soft threshold operator, and the result is: ; The subproblems are: ; The solution to the above equation is: ; Step S504: Update the Lagrange multipliers and penalty factors; In step S504, the Lagrange multipliers are updated. arrive The expression is: ; The updated penalty factor expression is: ; In the formula, Indicates the penalty factor The growth rate will be used to set the stopping condition as follows: ; In the formula, It is a constant representing the minimum acceptable error when iteration stops; this paper sets it to 2 × 10⁻⁶. -3 .

[0029] Example 3: The above constraints are integrated into the following optimization problem: ; An iterative solution is performed using the alternating direction multiplier method (ADMM), with auxiliary variables introduced. Construct an augmented Lagrangian function and update the variables alternately in sequence.

[0030] renew Solving using the Partial Singular Value Thresholding (PSVT) operator: ; renew Solving a system of linear equations using Fourier transform: ; renew Solving using the soft threshold operator: ; renew Solving a system of linear equations using Fourier transform: ; renew Closed-form solution for updating the noise tensor: ; renew Solving using the soft threshold operator: ; Update the Lagrange multipliers and penalty factor. The iteration termination condition is: ; In the formula, It is a constant representing the minimum acceptable error when iteration stops; this invention sets it to 2 × 10⁻⁶. -3 .

[0031] Step S600: After optimization, the optimized target tensor Background tensor and noise tensor The target enhancement images are reconstructed by reversing the steps of step S100, resulting in 5 frames of target enhancement images. One frame is taken as the detection response map of the current frame, and the target detection result map, background image, and noise image are obtained.

[0032] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A spatio-temporal tensor decomposition model based on sparse target perception and total variation constraint, characterized in that: The spatio-temporal tensor decomposition model comprises the following steps: Step S100: selecting a current frame to be detected and two frames before and after the current frame in the sequence image, extracting image blocks using a sliding window, stacking the image blocks into three-dimensional small tensors, and combining all the small tensors along a time dimension into a spatio-temporal tensor D; Step S200: calculate the prior information map of the sequence image, construct a tensor after box filtering the prior information map, take the inverse of all elements in the tensor, and calculate to obtain the perception prior weight tensor W of the sparse target P ; Step S300: introducing a tensor truncated kernel norm to approximate a background rank, and performing total variation constraint on the background component and the target component according to the correlation of the background component and the target component in different dimensions; Step S400: integrate the prior weight tensor W P and the constraints, and construct a new optimization model with the spatiotemporal tensor D as the data basis; Step S500: decompose the new optimization model into sub-problems by using alternating direction method of multipliers (ADMM), and solve the sub-problems respectively to obtain a target tensor , a background tensor , and a noise tensor ; Step S600: after optimization, the target tensor obtained after optimization , background tensor and noise tensor , respectively, according to the inverse step of the step S100 construction process, each obtains 5 frames of target enhanced images. Take one of them as the detection response map of the current frame, and obtain the detection result map of the target, the background image and the noise image.

2. The sparsity target-aware and total variation constraint based spatio- temporal tensor decomposition model according to claim 1, wherein, In the step S100, a sliding window block with a size of patchsizeR×patchsizeC (psR×psC) is selected in the extraction of the image blocks by using the sliding window, and the step specifically comprises the following steps: Step S101: traversing the current frame K from left to right and from top to bottom, the step length being stepC (sC) when traversing from left to right and the step length being stepR (sR) when traversing from top to bottom, and finally dividing the current frame into nine patches, and numbering each patch as i, denoted as Ki, indicating the i-th patch of the K-th frame image, wherein i is an integer from 1 to 9; Step S102: using the sliding window block and step size to traverse the two frames before and after the current frame K, obtaining the patches of each frame, and stacking the patches with the same number in all frames to form a small tensor , and stacking the tensors in the order of i from small to large to obtain a space-time tensor with a size of psRxpsCx45 . .

3. The spatio-temporal tensor decomposition model based on sparse target perception and total variation constraint according to claim 2, characterized in that, In the step S101, the calculation method of psR, psC, sR and sC is as follows: ； In the formula, R and C represent the height and width of the current frame respectively.

4. The spatio-temporal tensor decomposition model based on sparse target perception and total variation constraint according to claim 3, characterized in that, In the step S200, the step of calculating the target image prior information comprises the following steps: Step S201: calculating a gray level jump set of the current center pixel in eight directions, selecting the maximum value in the gray level jump set in each direction to form a maximum gray level jump set, multiplying the minimum value and the average value in the maximum jump set and then squaring to obtain the prior information of the current pixel, and defining the number of pixels in each direction as 5, and obtaining the prior information (STP) map of the target image after the pixel traversal calculation of the image is completed; Step S202: performing a box filter operation on the STP image to obtain a target image weight map w T ; Step S203: obtaining a target image weight map w of the 5 frames of images T and then converting it into a tensor form W according to the spatiotemporal tensor construction method in step S100 T ; Step S204: Obtain the prior weight tensor W by taking the inverse of all elements of W T P ;​ Step S205: reweighting the W P to obtain the reweighted target image enhancement weight W rP .

5. The spatio-temporal tensor decomposition model based on sparse target perception and total variation constraint according to claim 4, characterized in that, In the step S201, the prior information map of the target image is obtained by the following formula: ； In the formula, STP(x, y) represents the prior information of the current pixel, GSMAX d represents the maximum value in the gray level jump set along the direction d, CM represents the average of the current center pixel and its 8 neighboring pixels, GJS d represents the gray level jump set along the direction d, and η takes values of 0, 1, 2, 3, and 4. When η = 0, f(x, y) represents the center pixel of the 11x11 rectangular region. In the step S202, the operation of the box filter is specifically as follows: ； In the formula, box(·) represents the box filter, and 1 represents the radius of the box; In the step S205, the target enhancement weight W rP Specifically: ； In the equation, W r is a coefficient enhancement weight, is a Hadamard product, where n represents the number of iterations, and at the n+1 iteration, W r n+1 may be defined as: ； In the formula, is a preset constant, denotes the target tensor solved in the nth iteration. 6.The sparsity target-aware and total variation constraint based spatio-temporal tensor decomposition model according to claim 1, wherein, In the step S300, to more accurately approximate the rank of the background, the definition of the tensor truncated kernel norm is proposed for the background tensor B, and for a tensor The definition of the tensor truncated kernel norm of is as follows: ； In the formula, P and Q are two orthogonal tensors U and V obtained by performing tensor singular value decomposition on X, J is a unit tensor, the second dimension of the orthogonal tensor is only kept in the first r columns, and the remaining dimensions are kept unchanged to obtain a tensor, and the tensor is specifically represented as follows: ； The truncated singular value is set to 0.1 to obtain the tensor truncated rank r, that is, the sum of all singular values less than 0.1 is minimized; The definition of performing total variation constraint on the background component and the target component is as follows: ； In the formula, denotes Difference operator in horizontal direction and vertical direction, denotes Difference operator in horizontal direction and third dimension; i, j, k respectively represent the position of pixels in three dimensions, and the B and T are the background tensor and target tensor to be decomposed by the model.

7. The spatio-temporal tensor decomposition model based on sparse target sensing and total variation constraint according to claim 6, characterized in that, In the step S400, the expression of the new optimization model is as follows: ； In the formula, R psR×psC×45 , respectively, represent the background tensor, the target tensor and the noise tensor, and the balance between the components is controlled by the coefficient .

8. The spatio-temporal tensor decomposition model based on sparse target sensing and total variation constraint according to claim 7, characterized in that, In the step S500, the optimization model in the step S400 is solved by using an alternating direction multiplier method (ADMM), and the step specifically comprises the following steps: Step S501: Introducing auxiliary variables Rewriting the optimization model in step S400; Step S502: representing the rewritten model by using an augmented Lagrange equation; Step S503: decomposing the augmented Lagrange equation into sub-problems for solving, fixing other variables when updating one variable, and alternately updating all variables; Step S504: updating the Lagrange multiplier and the penalty factor.

9. The spatio-temporal tensor decomposition model based on sparse target sensing and total variation constraint according to claim 8, wherein: In the step S501, the rewritten optimization model is: ； In the step S502, the augmented Lagrange equation is expressed as: ； In the formula, to denotes the Lagrange multiplier, is a penalty factor; In the step S503, assuming that k is the iteration number, the sub-problem of the augmented Lagrange equation in the step S502 and the solving method can be described as follows: The sub-problem is: ； Using partial singular value thresholding operator to solve the above equation, assuming X ∈ R n1×n2 , Y ∈ R n1×n2 , l = min(n1, n2), N ≤ l, Y can be expressed as the sum of two matrices, that is, Y = Y1 + Y2, and Y1 = (U Y1 )D Y1 (V Y1 ) T , Y2 = (U Y2 )D Y2 (V Y2 ) T , wherein U Y1 and V Y1 correspond to the first N singular values of Y, U Y2 and V Y2 correspond to the remaining singular values, and a minimization problem is defined as follows: ； In the formula, denotes the sum of the singular values, and the singular value threshold operator is defined as: ； The optimal solution of the above formula is: ； In the formula, , , ; The sub-problem is: ； The equivalent linear equation set of the above formula is: ； In the formula, denotes the matrix transpose, , , , ; The closed-form solution of the above equivalent linear equation set is: ； In the formula, respectively represent complex conjugate, fast Fourier transform, and inverse Fourier transform. The sub-problem is: ； The above formula is solved by a soft threshold operator, and the specific solution is: ； Assume that τ > 0, X ∈ R n1×n2 , Y ∈ R n1×n2 , define a norm minimization problem as follows: ； The closed-form solution S exists for the above equation τ ( Y ), where S τ is a soft threshold operator defined as follows: ； The sub-problem is: ； The equivalent linear equation set of the above formula is: ； In the formula, , ; The closed-form solution of the above equivalent linear equation set is: ； The sub-problem is: ； The above formula is solved by a soft threshold operator, and the specific solution is: ； The sub-problem is: ； The solving expression of the above formula is: ； In the step S504, the Lagrange multiplier is updated to The expression is: ； The updating expression of the penalty factor is: ； In the formula, represents a penalty factor The growth rate of the penalty factor, the stop condition is set to: ； In the formula, is a constant representing the minimum error acceptable at the time of stopping the iteration.

Images