A hyperspectral anomaly detection method based on multi-level tensor prior constraints
Through multi-level tensor prior constraints and ADMM algorithm optimization, combined with low rank, sparse and segmented smooth constraints, the problems of insufficient detection accuracy of hyperspectral anomaly and high false alarm rate are solved, and more efficient abnormal object detection is achieved.
Patent Information
- Application Number
- CN202111525268.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-14
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-12-14
AI Technical Summary
The existing hyperspectral anomaly target detection method ignores the structural characteristics of the hyperspectral image when utilizing the background address characteristics and the target sparse characteristics, resulting in insufficient detection accuracy and high false alarm rate.
Using multi-level tensor prior constraint method, the hyperspectral image is decomposed into background tensors and abnormal tensors, and regularized representation is applied along the spectral dimension and spatial dimension to construct a unified Lagrangian function, optimized using ADMM algorithm, and fused low Rank, sparse, and segmented smooth prior constraints.
The accuracy of hyperspectral abnormal object detection is improved, the false alarm rate is reduced, and the global structural features are effectively extracted through tensor decomposition, which improves detection performance.
Smart Images

Figure CN114331976B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of hyperspectral data application, and in particular to a hyperspectral anomaly detection method based on multi-level tensor prior constraints. Background Art
[0002] Hyperspectral abnormal target detection technology is a hot research topic in the field of hyperspectral data application. Its purpose is to accurately detect abnormal targets in hyperspectral images without any prior information about the target. In order to achieve this goal, many abnormality detection methods have been proposed, including Reed-Xiaoli (RX) detection method, detection method based on collaborative representation and detection method based on low-rank sparse characteristics. Among them, the detection method based on low-rank sparse characteristics can extract the sparse features of abnormal targets from the global background low-rank features, and has received widespread attention and research. However, since only the address characteristics of the background and the sparse characteristics of the target are used and the structural characteristics of the hyperspectral image itself are ignored, the accuracy of hyperspectral abnormal target detection needs to be improved. Summary of the invention
[0003] The technical problem to be solved by the present invention is to provide a hyperspectral anomaly detection method based on multi-level tensor prior constraints, which can improve the anomaly detection accuracy and reduce the corresponding false alarm rate.
[0004] In order to solve the above technical problems, the present invention provides a hyperspectral anomaly detection method based on multi-level tensor prior constraints, comprising the following steps:
[0005] Step 1: Input the original hyperspectral image , decompose it into background tensor and abnormal tensor, and then initialize the background tensor , the abnormal target tensor
[0006] Step 2, expand the background tensor along the spectral dimension to apply regularization to characterize the piecewise smooth prior;
[0007] Step 3, expand the background tensor along the spatial dimension to apply regularization to represent the low-rank prior;
[0008] Step 4: Expand the anomaly tensor along the spectral dimension to apply a regularized representation sparse prior;
[0009] Step 5: Based on the prior representations of Step 2, Step 3 and Step 4, a new Lagrangian function is constructed using a unified framework;
[0010] Step 6: Use the ADMM algorithm to optimize the function in step 5;
[0011] Step 7: Calculate the abnormal target tensor S * =L spe (S), through Get the anomaly detection graph , where L spe It means that the matrix expanded along the spectral dimension is inverted into a tensor, M means there are M spectral bands, S * (i,j,l) represents the element of the i-th row, j-th column, and l-th band of the anomaly tensor.
[0012] Preferably, in step 2, the background tensor is expanded along the spectral dimension to apply a regularized representation of the piecewise smooth prior. Specifically, the background tensor is expanded along the spectral dimension to apply a regularized representation of the piecewise smooth prior. Expanded into a two-dimensional matrix along the spectral dimension , create l 0 -l 1 The mixed total variation regularization term is where x and y are the two spatial dimensions of the hyperspectral image, D represents the discrete difference operator of the periodic boundary, ξ represents the diagonal matrix with binary elements 0 and 1, ζ represents the index of the diagonal matrix, and D x represents the discrete difference operator in the horizontal direction, D y represents the discrete difference operator in the vertical direction, D d is a one-dimensional finite difference operator, the function Designed to enhance image edges.
[0013] Preferably, in step 3, the background tensor is expanded along the spatial dimension to apply regularization to characterize the low-rank prior. Specifically, two matrices are given: and , meet PP T =QQ T =I r×r , create a truncated nuclear norm regularization term for the background matrix X Given, where r represents the number of the largest singular values, represents the nuclear norm of X, ω m represents the mth largest singular value of X, and Tr(·) represents the trace of the matrix.
[0014] Preferably, in step 4, the anomaly tensor is expanded along the spectral dimension to apply a regularized representation sparse prior, specifically: the anomaly target tensor Expanded along the spectral dimension into a matrix S spe , create S spe l 2,1 The norm regularization term is Where d represents the length of the spectral dimension and represents the value of S in the i-th row and j-th column.
[0015] Preferably, in step 5, based on the prior representations of steps 2, 3 and 4, a new Lagrangian function is constructed using a unified framework, specifically: all spatial-spectral regularization terms are integrated to create a new Lagrangian function Among them A 1 ,A 2 ,A 3 and A 4 represents auxiliary variables, Φ,B 1 and B 2 represents the Lagrange multiplier, α and τ are two positive constants used to balance the contribution of each term, μ and σ represent non-negative penalty parameters, G = DX, E = D x XD d 、F=D y XD d , Y spe , X spe They represent the hyperspectral matrix and background matrix expanded along the spectral dimension, respectively.
[0016] Preferably, in step 6, using the ADMM algorithm to optimize the function in step 5 specifically includes the following steps:
[0017] Step 6.1: Fix other variables and use E=S α / 2 [D x XD d +A 1 ] Update the variable E, where the symbolic operator The final update solution is obtained by applying the operator S ε [x] = sgn(x)*max(|x|-ε,0);
[0018] Step 6.2: Fix other variables and use F=S α / 2 [D y XD d +A 2 ]Update variable F;
[0019] Step 6.3: Fix other variables and use G = G′ + (I-ξ)(DX + A 3 ) Update the variable G, where I represents the identity matrix, G′=ξ(DX+A 3 ), at the kth iteration, G′ (k) is in descending order;
[0020] Step 6.4 Written as Fix other variables and update variable A through the above formula 3 ;
[0021] Step 6.5: Fix other variables, use the singular value decomposition SVD method, select the r largest singular values, and solve Update P and Q; where Σ represents the singular value matrix and SVDs represents the singular value decomposition function;
[0022] Step 6.6, fix other variables and update The sub-optimization problem of is modeled in vector form as Use the least squares method to solve and update the variables ,in Expressed as Kronecker product; where s, a 1 、a 2 , g, a 3 They are the vector forms corresponding to the matrices.
[0023] Step 6.7: Fix other variables by Update variables and ,in and Respectively represent the expansion of the tensor along the spatial dimension and the spectral dimension, expressed by the formula X 1 =U 1 (X),X 2 =U 2 (X) and X 3 =U 3 (X);
[0024] Step 6.8: Fix other variables and solve Update variable S spe ;
[0025] Step 6.9: Fix other variables by Update the Lagrange multipliers Φ, A respectively n and B n .
[0026] The beneficial effects of the present invention are as follows: (1) The present invention creates a new hyperspectral anomaly detection model, including a tensor low-rank prior, a tensor sparse prior and a global spatial piecewise smooth prior, which can fully utilize the spatial-spectral structural information of the hyperspectral image; (2) The present invention uses the ADMM algorithm to solve the new hyperspectral anomaly detection model, thereby improving the accuracy of hyperspectral anomaly target detection and reducing the false alarm rate under the same conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a schematic diagram of the method flow of the present invention.
[0028] FIG. 2( a ) is a pseudo-camouflage image of the San Diego dataset used by the detection method MSSR of the present invention.
[0029] FIG2( b ) is a ground truth image of the San Diego dataset used by the detection method MSSR of the present invention.
[0030] Figure 3 This is a ROC curve diagram of the San Diego data set used in the detection method MSSR of the present invention compared with other methods.
[0031] Figure 4 This is a graph showing the detection results of the San Diego dataset used by the detection method MSSR of the present invention compared with other methods. DETAILED DESCRIPTION
[0032] like Figure 1 As shown, a hyperspectral anomaly detection method based on multi-level tensor prior constraints includes the following steps:
[0033] Step 1: Input the original hyperspectral image , decompose it into background tensor and abnormal tensor, and then initialize the background tensor , the abnormal target tensor
[0034] Step 2: The background tensor Expanded into a two-dimensional matrix along the spectral dimension , create l 0 -l 1 The mixed total variation regularization term is where x and y are the two spatial dimensions of the hyperspectral image, D represents the discrete difference operator of the periodic boundary, ξ represents the diagonal matrix with binary elements 0 and 1, ζ represents the index of the diagonal matrix, and D x represents the discrete difference operator in the horizontal direction, D y represents the discrete difference operator in the vertical direction, D d is a one-dimensional finite difference operator, the function Designed to enhance image edges;
[0035] Step 3: Given two matrices and Meet PP T =QQ T =I r×r , create a truncated nuclear norm regularization term for the background matrix X Given, where r represents the number of the largest singular values, represents the nuclear norm of X, ω m represents the mth largest singular value of X, Tr(·) represents the trace of the matrix;
[0036] Step 4: Abnormal target tensor Expanded along the spectral dimension into a matrix S spe , create S spe l 2,1 The norm regularization term is Where d represents the length of the spectral dimension, which represents the value of S in the i-th row and j-th column;
[0037] Step 5: Fusion of all spatial-spectral regularization terms to create a new Lagrangian function Among them A 1 ,A 2 ,A 3 and A 4 represents auxiliary variables, Φ,B 1 and B 2 represents the Lagrange multiplier, α and τ are two positive constants used to balance the contribution of each term, μ and σ represent non-negative penalty parameters, G = DX, E = D x XD d 、F=D y XD d , Y spe , X spe They represent the hyperspectral matrix and background matrix expanded along the spectral dimension, respectively.
[0038] Step 6: Use the ADMM algorithm to optimize the function in step 5; specifically, the following steps are included:
[0039] Step 6.1: Fix other variables and use E=S α / 2 [D x XD d +A 1 ] Update the variable E, where the symbolic operator The final update solution is obtained by applying the operator S ε [x] = sgn(x)*max(|x|-ε,0);
[0040] Step 6.2: Fix other variables and use F=S α / 2 [D y XD d +A 2 ]Update variable F;
[0041] Step 6.3: Fix other variables and use G = G′ + (I-ξ)(DX + A 3 ) Update the variable G, where I represents the identity matrix, G′=ξ(DX+A 3 ), at the kth iteration, G′ (k) is in descending order;
[0042] Step 6.4 Written as Fix other variables and update variable A through the above formula 3 ;
[0043] Step 6.5: Fix other variables, use the singular value decomposition SVD method, select the r largest singular values, and solve Update P and Q; where Σ represents the singular value matrix and SVDs represents the singular value decomposition function.
[0044] Step 6.6, fix other variables and update The sub-optimization problem of is modeled in vector form as Use the least squares method to solve and update the variables ,in Expressed as Kronecker product; where s, a 1 、a 2 , g, a 3 They are the vector forms corresponding to the matrices.
[0045] Step 6.7: Fix other variables by Update variables and ,in and Respectively represent the expansion of the tensor along the spatial dimension and the spectral dimension, expressed by the formula X 1 =U 1 (X),X 2 =U 2 (X) and X 3 =U 3 (X);
[0046] Step 6.8: Fix other variables and solve Update variable S spe ;
[0047] Step 6.9: Fix other variables by Update the Lagrange multipliers Φ, A respectively n and B n ;
[0048] Step 7: Calculate the abnormal target tensor S * =L spe (S), through Get the anomaly detection graph . Where L spe It means that the matrix expanded along the spectral dimension is inverted into a tensor, M means there are M spectral bands, S * (i,j,l) represents the element of the i-th row, j-th column, and l-th band of the anomaly tensor.
[0049] In order to better demonstrate the advantages of the multi-spectral space representation (MSSR) based on tensor decomposition of the present invention, a specific example is combined below to compare the detection method of the present invention with several existing advanced detection algorithms.
[0050] Example:
[0051] The comparison method is as follows: abnormal target detection is performed on the real hyperspectral image San Diego, and the detection accuracy that can be achieved by each method is compared. The detection accuracy is measured using the receiver operating characteristic (ROC) and the area under the curve (AUC), where the ROC curve represents the corresponding relationship between the false alarm rate and the detection rate of each threshold segmentation result, and the AUC value is obtained by calculating the area under the ROC curve of the anomaly detector. Under the same false alarm rate, the higher the AUC value, the higher the detection performance. The hyperspectral image used contains 100×100 spatial pixels, and the background mainly contains asphalt roads, roads, roofs and shadows. Three aircraft occupying 58 pixels of the image are considered to be abnormal targets. The pseudo-camouflage image of the hyperspectral image used and the ground truth image are shown in Figure 2(a) and Figure 2(b), and their detailed information is shown in Table 1.
[0052] Table 1 Hyperspectral image parameters
[0053]
[0054] Table 2 shows the AUC comparison of the anomaly detection results of each method. Figure 3 It is the ROC curve comparison of the anomaly detection results of each method. Figure 4 The label diagrams of the anomaly detection results of each method are compared. From the results, it can be seen that the method proposed in the present invention has better detection performance than other methods.
[0055] Table 2 AUC comparison of anomaly detection results of each method
[0056]
[0057] In summary, this paper proposes a new hyperspectral anomaly detection method, MSSR. It cleverly combines the prior properties (low rank, sparsity, and segmentation smoothness) with the tensor decomposition of hyperspectral data. Different regularization methods are used to embed these priors for tensors of different dimensions. Among them, the low-rank prior along the background spectral dimension is represented by the truncated nuclear norm, and the piecewise smoothness of the background spatial dimension is represented by l 0 -l 1 Mixed total variation regularization representation, sparse prior anomaly component is represented by l 2,1 norm regularized representation. In addition, the tensor decomposition representation can effectively extract global structural features, thereby better separating anomalies from the background. All regularization constraints are fused into a convex optimization function and iteratively optimized using the ADMM algorithm. Finally, when the iteration converges, the detection map is obtained. In experiments, the performance of our proposed MSSR is proven to be robust and outperforms several advanced anomaly detection methods.
Claims
1. A hyperspectral anomaly detection method based on multi-level tensor prior constraints, It is characterized in that The steps include: Step 1: Input the original hyperspectral image Decompose it into background tensor and abnormal tensor, and then initialize the background tensor Abnormal target tensor S = 0; Step 2: Expand the background tensor along the spectral dimension to apply regularization to characterize the piecewise smooth prior; Expanded into a two-dimensional matrix along the spectral dimension Create 0 -l 1 The mixed total variation regularization term is where x and y are the two spatial dimensions of the hyperspectral image, D represents the discrete difference operator of the periodic boundary, ξ represents the diagonal matrix with binary elements 0 and 1, ζ represents the index of the diagonal matrix, and D x represents the discrete difference operator in the horizontal direction, D y represents the discrete difference operator in the vertical direction, D d is a one-dimensional finite difference operator, the function Designed to enhance image edges; Step 3: Expand the background tensor along the spatial dimension to apply regularization to represent the low-rank prior; give two matrices and Meet PP T =QQ T =I r×r , create a truncated nuclear norm regularization term for the background matrix X Given, where r represents the number of the largest singular values, represents the nuclear norm of X, ω m represents the mth largest singular value of X, Tr(·) represents the trace of the matrix; Step 4: Expand the anomaly tensor along the spectral dimension to apply a regularized representation sparse prior; Expand the abnormal target tensor S along the spectral dimension into a matrix S spe , create S spe l 2,1 The norm regularization term is Where d represents the length of the spectral dimension, which represents the value of S in the i-th row and j-th column; Step 5: Based on the prior representations of steps 2, 3, and 4, a new Lagrangian function is constructed using a unified framework; all spatial-spectral regularization terms are integrated to create a new Lagrangian function. Among them A 1 ,A 2 ,A 3 and A 4 represents auxiliary variables, Φ,B 1 and B 2 represents the Lagrange multiplier, α and τ are two positive constants used to balance the contribution of each term, μ and σ represent non-negative penalty parameters, G = DX, E = D x XD d 、F=D y XD d , Y spe , X spe They represent the hyperspectral matrix and background matrix expanded along the spectral dimension respectively; Step 6: Use the ADMM algorithm to optimize the function in step 5; Step 7: Calculate the abnormal target tensor S * =L spe (S), through Get the anomaly detection graph Where L spe It means that the matrix expanded along the spectral dimension is inverted into a tensor, M means there are M spectral bands, S * (i,j,l) represents the element of the i-th row, j-th column, and l-th band of the anomaly tensor.
2. The hyperspectral anomaly detection method based on multi-level tensor prior constraints as claimed in claim 1, It is characterized in that In step 6, using the ADMM algorithm to optimize the function in step 5 specifically includes the following steps: Step 6.1: Fix other variables and use E=S α2 [D x XD d +A 1 ] Update the variable E, where the symbolic operator The final update solution is obtained by applying the operator S ε [x] = sgn(x)*max(x|-ε,0); Step 6.2: Fix other variables and use F=S α2 [D y XD d +A 2 ]Update variable F; Step 6.3: Fix other variables and use G = G′ + (I-ξ)(DX + A 3 ) Update the variable G, where I represents the identity matrix, G′=ξ(DX+A 3 ), at the kth iteration, G′ (k) is in descending order; Step 6.4 Written as Fix other variables and update variable A through the above formula 3 ; Step 6.5: Fix other variables, use the singular value decomposition SVD method, select the r largest singular values, and solve Update P and Q; where Σ represents the singular value matrix and SVDs represents the singular value decomposition function; Step 6.6, fix other variables and update The sub-optimization problem of is modeled in vector form as Use the least squares method to solve and update the variable X, where Expressed as Kronecker product; where s, a 1 、a 2 , g, a 3 They are the vector forms corresponding to the matrices; Step 6.7: Fix other variables by Update variables and in and Respectively represent the expansion of the tensor along the spatial dimension and the spectral dimension, expressed by the formula X 1 =U 1 (X),X 2 =U 2 (X) and X 3 =U 3 (X); Step 6.8: Fix other variables and solve Update variable S spe ; Step 6.9: Fix other variables by Update the Lagrange multipliers Φ, A respectively n and B n .
