A non-uniformity correction method based on multi-scale analysis and singular value decomposition

Through multi-scale analysis and singular value decomposition methods, Curvelet transform and singular value decomposition algorithms are used to remove non-uniform noise of infrared images, which solves the problem of incomplete removal of infrared images in the prior art, and realizes no reference radiation source required efficient noise removal effect.

CN115018721BActive Publication Date: 2024-11-22西安中科立德红外科技有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210559621.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-19
Publication Date
2024-11-22
Estimated Expiration
2042-05-19

AI Technical Summary

Technical Problem

The existing infrared image non-uniformity removal method is not thorough and requires reference to the radiation source, which affects the detection performance.

Method used

Using a non-uniformity correction method based on multi-scale analysis and singular value decomposition, the multi-scale and multi-angle Curvelet coefficient matrix of infrared images is extracted and the eigenvalue correction is performed to remove non-infrared images through Curvelet transformation and singular value decomposition algorithm. Uniform noise.

Benefits of technology

Complete denoising of infrared images is achieved, reliance on reference radiation sources is avoided, image edge information is maintained, and image degradation and 'ghosting' problems are reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115018721B_ABST
    Figure CN115018721B_ABST
Patent Text Reader

Abstract

The present invention relates to an image correction method, and in particular to a non-uniformity correction method based on multi-scale analysis and singular value decomposition, which solves the technical problem that the infrared image non-uniformity removal in the prior art is not thorough and requires a reference radiation source. The present invention provides a non-uniformity correction method based on multi-scale analysis and singular value decomposition, comprising the following steps: S1, collecting an infrared image f with non-uniformity; S2, obtaining a Curvelet coefficient matrix Ct{n}; S3, establishing an equation |Ct LH {n}‑λ m E|=0, solve the equation to obtain the eigenvalues ​​λ1...λ M and the eigenvector x1{n}...x N {n}; S4, perform singular value decomposition on the Curvelet coefficient matrix Ct{n}; S5, obtain the second diagonal matrix S'{n}; S6, obtain the first coefficient matrix Ct'{n}; S7, subtract the subband coefficients in the Curvelet coefficient matrix Ct{n} from the corresponding subband coefficients in the first coefficient matrix Ct'{n}; S8, perform inverse Curvelet transform on the second coefficient matrix Ct"{n} to obtain the infrared image f with non-uniformity removed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to an image correction method, in particular to a non-uniformity correction method based on multi-scale analysis and singular value decomposition. Background Art

[0002] Infrared Focal Plane Arrays (IRFPA) are widely used in infrared detection systems due to their high sensitivity and high frame rate. However, due to the influence of its manufacturing process and readout circuit, the detection element responsiveness of IRFPA varies, resulting in the detection system producing certain non-uniform stripes and noise such as pot lids during imaging. Non-uniform stripes and noise such as pot lids are collectively referred to as image non-uniformity. This non-uniformity seriously affects the detection performance of the infrared detection system.

[0003] At present, non-uniformity removal methods are mainly divided into two categories: based on reference radiation sources and based on scenes. One-point correction, two-point correction and multi-point correction are commonly used methods based on reference radiation sources. The principle of such methods is relatively simple and easy to implement in engineering, but they need to use reference radiation sources to correct parameter drift periodically. During the correction period, the camera will temporarily stop imaging, which is inconvenient in practical applications. The correction methods based on scene adaptation mainly include time domain high-pass filtering, neural networks and other methods; although this method can achieve better correction effects, it has problems such as slow iterative convergence, image degradation and "ghosting".

[0004] Curvelet transform is a multi-scale geometric transform that can effectively represent edges and curves at different angles of an image and has very good sparse expression capabilities. Researchers often apply Curvelet transform to image denoising, but it is rarely used to remove non-uniformity problems in infrared images. Summary of the invention

[0005] The purpose of the present invention is to provide a non-uniformity correction method based on multi-scale analysis and singular value decomposition to achieve image denoising, in view of the technical problems that the infrared image non-uniformity removal in the prior art is not thorough and requires a reference radiation source.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A non-uniformity correction method based on multi-scale analysis and singular value decomposition, which is special in that it includes the following steps:

[0008] S1, collecting infrared images f with non-uniformity;

[0009] S2. For the infrared image f, use the second generation fast Curvelet transform to obtain the Curvelet coefficient matrix Ct{n}; where n represents the scale of the Curvelet coefficient matrix Ct{n}, and n∈[1,6];

[0010] S3. Establish equation |Ct LH {n}-λ m E|=0, solve the equation to obtain the eigenvalue λ of the Curvelet coefficient matrix Ct{n} 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n};

[0011] In the formula, Ct LH {n} represents the product matrix of Curvelet coefficient matrix Ct{n} and its transpose;

[0012] λ m represents the mth eigenvalue of the Curvelet coefficient matrix Ct{n}, where m is 1...M, and M is a positive integer;

[0013] E represents the identity matrix;

[0014] N represents the number of eigenvectors, and N=M;

[0015] S4. Perform singular value decomposition on the Curvelet coefficient matrix Ct{n}, and obtain Ct{n}=U{n}*S{n}*V{n}, and obtain the first diagonal matrix S{n}, the first orthogonal matrix U{n} and the second orthogonal matrix V{n};

[0016] S5, performing linear correction on the eigenvalues ​​of the first diagonal matrix S{n} after taking square root, to obtain a second diagonal matrix S'{n};

[0017] S6. Multiply the second diagonal matrix S'{n} by the first orthogonal matrix U{n} on the left and by the second orthogonal matrix V{n} on the right to obtain a first coefficient matrix Ct'{n};

[0018] S7, subtract the subband coefficients in the Curvelet coefficient matrix Ct{n} from the corresponding subband coefficients in the first coefficient matrix Ct'{n} to obtain a second coefficient matrix Ct"{n};

[0019] S8. Perform an inverse Curvelet transform on the second coefficient matrix Ct″{n} to obtain an infrared image f with non-uniformity removed.

[0020] Furthermore, S3 is specifically:

[0021] S3.1. Find the transpose of the Curvelet coefficient matrix Ct{n} and get Ct T {n};

[0022] S3.2. Compare the Curvelet coefficient matrix Ct{n} with Ct in S3.1 T {n} multiply to get Ct LH {n};

[0023] S3.3. Solve the equation |Ct LH {n}-λ m E|=0, the eigenvalue of the Curvelet coefficient matrix Ct{n} is obtained as λ 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n}.

[0024] Furthermore, S4 is specifically:

[0025] S4.1. For eigenvalue λ 1 ... M The square root is and will Assign values ​​to the diagonal elements of the first diagonal matrix S{n} in sequence, and take the element in the first row and first column of the first diagonal matrix S{n} as the starting point to obtain the first diagonal matrix S{n};

[0026] S4.2. Ct LH The eigenvector x of {n} 1 {n}...x N {n} performs normalization operation to obtain normalized feature vector

[0027] S4.3, let the second orthogonal matrix According to Ct{n}=U{n}*S{n}*V{n}, we can get the first orthogonal matrix U{n}=Ct{n}*V -1 {n}*S -1 {n};

[0028] In the formula, S -1 {n} represents the inverse matrix of the first diagonal matrix S{n}, V -1 {n} represents the inverse matrix of the second orthogonal matrix V{n}.

[0029] Furthermore, S5 is specifically:

[0030] S5.1. Yes Perform linear corrections respectively to obtain the corrected eigenvalue λ 1 '...λK ', the subscript K represents the number of corrected eigenvalues;

[0031] S5.2. Correct the eigenvalue λ 1 '...λ K 'Assign values ​​to the diagonal elements of the second diagonal matrix S'{n} in sequence, and take the element on the first row and first column of the second diagonal matrix S'{n} as the starting point to obtain the second diagonal matrix S'{n}.

[0032] Furthermore, in S4.1, the first diagonal matrix S{n} and the second diagonal matrix S'{n} in S5.2 are of the same size.

[0033] Furthermore, in S2, the value of n is 1 to 5.

[0034] Furthermore, in S3, the values ​​of M and N are both 10.

[0035] Furthermore, in S5.1, the modified characteristic value λ 1 The gain coefficient a is greater than the modified characteristic value λ 2 '...λ 10 'Gain coefficient b.

[0036] Furthermore, the modified characteristic value λ 1 'The gain coefficient a is 0.64;

[0037] Corrected eigenvalue λ 2 '...λ 10 'The gain coefficient b is 0.52.

[0038] Furthermore, S1 specifically includes: using an infrared thermal imager to collect an infrared image f with non-uniformity.

[0039] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0040] 1. The present invention uses Curvelet transform to extract the multi-scale and multi-angle Curvelet coefficient matrix Ct{n} of the infrared image f, and uses the singular value decomposition (SVD) algorithm to obtain the first diagonal matrix S{n} corresponding to the Curvelet coefficient matrix Ct{n} at each scale; according to the Curvelet transform theory, a larger Curvelet coefficient matrix corresponds to the high-frequency information of the stronger infrared image, and a smaller Curvelet coefficient matrix corresponds to the low-frequency noise. Therefore, after the Curvelet coefficient matrix Ct{n} is decomposed by SVD, the larger eigenvalue in the first diagonal matrix S{n} obtained corresponds to the high-frequency information of the infrared image f. Since the non-uniformity of the infrared image f includes both high-frequency non-uniformity and partial low-frequency non-uniformity, linear correction is performed on some eigenvalues ​​to obtain the non-uniform noise extracted from the infrared image.

[0041] 2. The present invention provides a non-uniformity correction method based on multi-scale analysis and singular value decomposition. It does not require a reference radiation source and applies Curvelet transform to remove the non-uniformity of infrared images. It can completely remove the non-uniformity in the infrared image f and achieve image denoising.

[0042] 3. The Curvelet transform used in the present invention is a multi-scale, multi-directional geometric analysis method developed based on the wavelet transform. Compared with the traditional wavelet transform, the Curvelet transform has a better expression of information such as straight lines and curves in infrared images, and is more suitable for analyzing the curve or straight line edge features in infrared images. With the help of the Curvelet transform, a more ideal effect of removing the non-uniformity of infrared images can be achieved compared with the wavelet transform.

[0043] 4. Combining Curvelet transform with singular value decomposition algorithm can be unaffected by image scene information and does not require reference radiation source. It can be widely used in the non-uniform correction needs of infrared images in different scenes and environments, and can better maintain the edge information of infrared images. It can suppress the degradation and "ghosting" problems of infrared images to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The present invention is a flowchart of an embodiment of a non-uniformity correction method based on multi-scale analysis and singular value decomposition.

[0045] Figure 2 It is a schematic diagram of an infrared image with infrared non-uniformity collected by the infrared thermal imager in an embodiment of the present invention.

[0046] Figure 3 It is a schematic diagram of an infrared image after removing non-uniformity in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the technical solution in the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0048] Curvelet transform is a multi-scale geometric transform that can effectively represent the edges and curves of different angles of an image, and has very good sparse expression capabilities. Researchers often apply Curvelet transform to image denoising, and it is rarely used to remove the non-uniformity of infrared images. Singular value decomposition (SVD) can convert the original matrix into the product of two orthogonal matrices and a diagonal matrix, where the values ​​of the diagonal positions in the diagonal matrix are the eigenvalues, and the eigenvalues ​​represent different degrees of high-frequency and low-frequency information of the image from large to small.

[0049] like Figure 1 As shown, the present invention provides a non-uniformity correction method based on multi-scale analysis and singular value decomposition, comprising the following steps:

[0050] S1, collecting infrared images f with non-uniformity;

[0051] S2. For the infrared image f, use the second generation fast Curvelet transform to obtain the Curvelet coefficient matrix Ct{n}; where n represents the scale of the Curvelet coefficient matrix Ct{n}, and n∈[1,6];

[0052] S3. Establish equation |Ct LH {n}-λ m E|=0, solve the equation to obtain the eigenvalue λ of the Curvelet coefficient matrix Ct{n} 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n};

[0053] In the formula, Ct LH {n} represents the product matrix of Curvelet coefficient matrix Ct{n} and its transpose;

[0054] λ m represents the mth eigenvalue of the Curvelet coefficient matrix Ct{n}, where m is 1...M, and M is a positive integer;

[0055] E represents the identity matrix;

[0056] N represents the number of eigenvectors, and N=M;

[0057] S3.1. Find the transpose of the Curvelet coefficient matrix Ct{n} and get Ct T {n};

[0058] S3.2. Compare the Curvelet coefficient matrix Ct{n} with Ct in S3.1 T {n} multiply to get Ct LH {n};

[0059] S3.3. Solve the equation |Ct LH {n}-λ m E|=0, the eigenvalue of the Curvelet coefficient matrix Ct{n} is obtained as λ 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n}.

[0060] S4. Perform singular value decomposition on the Curvelet coefficient matrix Ct{n}, and obtain Ct{n}=U{n}*S{n}*V{n}, and obtain the first diagonal matrix S{n}, the first orthogonal matrix U{n} and the second orthogonal matrix V{n};

[0061] S4.1. For eigenvalue λ 1 ... M The square root is and will Assign values ​​to the diagonal elements of the first diagonal matrix S{n} in sequence, and take the element in the first row and first column of the first diagonal matrix S{n} as the starting point to obtain the first diagonal matrix S{n};

[0062] S4.2. Ct LH The eigenvector x of {n} 1 {n}...x N {n} performs normalization operation to obtain normalized feature vector

[0063] S4.3, let the second orthogonal matrix According to Ct{n}=U{n}*S{n}*V{n}, we can get the first orthogonal matrix U{n}=Ct{n}*V -1 {n}*S -1 {n};

[0064] In the formula, S -1 {n} represents the inverse matrix of the first diagonal matrix S{n}, V -1 {n} represents the inverse matrix of the second orthogonal matrix V{n}.

[0065] S5. Perform linear correction on the eigenvalues ​​of the first diagonal matrix S{n} to obtain a second diagonal matrix S'{n};

[0066] S5.1. Yes Perform linear corrections respectively to obtain the corrected eigenvalue λ 1 '...λ K ', the subscript K represents the number of corrected eigenvalues;

[0067] S5.2. Correct the eigenvalue λ 1 '...λ K 'Assign values ​​to the diagonal elements of the second diagonal matrix S'{n} in sequence, and take the element on the first row and first column of the second diagonal matrix S'{n} as the starting point to obtain the second diagonal matrix S'{n}.

[0068] S6. Multiply the second diagonal matrix S'{n} by the first orthogonal matrix U{n} on the left and by the second orthogonal matrix V{n} on the right to obtain a first coefficient matrix Ct'{n};

[0069] S7, subtract the subband coefficients in the Curvelet coefficient matrix Ct{n} from the corresponding subband coefficients in the first coefficient matrix Ct'{n} to obtain a second coefficient matrix Ct"{n};

[0070] S8. Perform an inverse Curvelet transform on the second coefficient matrix Ct″{n} to obtain an infrared image f with non-uniformity removed.

[0071] The following is a specific embodiment of the present invention, which is used to illustrate the above method.

[0072] Step 1: If Figure 2 As shown, an infrared image f containing non-uniformity is collected using an infrared thermal imager, wherein the infrared image f represents the image data actually collected;

[0073] Step 2: Use the second-generation fast Curvelet transform on the infrared image f to obtain the Curvelet coefficient matrix Ct{n}, where n represents the scale of the Curvelet coefficient matrix Ct{n}, and n∈[1,6];

[0074] The larger the n, the larger the scale of the Curvelet coefficient matrix Ct{n}. When the scale is large, it represents more high-frequency information, especially when n = 6, it represents the high-frequency information of the infrared image f. After many experiments, when n = 6, it basically does not contain the non-uniformity of the image, so the Curvelet coefficient matrix Ct{n} when n = 6 is not processed.

[0075] Step 3: Establish the equation |CtLH {n}-λ m E|=0, solve the equation to obtain the eigenvalue λ of the Curvelet coefficient matrix Ct{n} 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n};

[0076] In the formula, Ct LH {n} represents the product matrix of Curvelet coefficient matrix Ct{n} and its transpose;

[0077] λ m represents the eigenvalue of the Curvelet coefficient matrix Ct{n};

[0078] E represents the identity matrix;

[0079] The subscript m represents the number of eigenvalues, and the value of m is 1...M, where M is a positive integer;

[0080] The subscript N indicates the number of eigenvectors, and N = M = 10;

[0081] S3.1. Find the transpose of the Curvelet coefficient matrix Ct{n} and get Ct T {n};

[0082] S3.2. Compare the Curvelet coefficient matrix Ct{n} with Ct in S3.1 T {n} multiply to get Ct LH {n};

[0083] S3.3. Solve the equation in S3 and obtain the eigenvalue of the Curvelet coefficient matrix Ct{n} as λ 1 ... 10 With Ct LH The eigenvector x of {n} 1 {n}...x 10 {n};

[0084] Step 4: When n takes the values ​​of 1, 2, 3, 4, and 5, they represent the first five scales respectively. Perform SVD decomposition on the Curvelet coefficient matrix Ct{n} at the first five scales to obtain the first diagonal matrix S{n}, that is:

[0085] Ct{n}=U{n}*S{n}*V{n}

[0086] Where S{n} represents the first diagonal matrix of the Curvelet coefficient matrix Ct{n};

[0087] U{n} and V{n} represent the first orthogonal matrix and the second orthogonal matrix of the Curvelet coefficient matrix Ct{n} respectively;

[0088] S4.1. For the eigenvalue λ in S 3.3 1 ... 10 The square root is and will Assign values ​​to the diagonal elements of the first diagonal matrix S{n} in sequence, and take the element in the first row and first column of the first diagonal matrix S{n} as the starting point to obtain the first diagonal matrix S{n};

[0089] S4.2. Ct LH The eigenvector x of {n} 1 {n}...x 10 {n} performs normalization operation to obtain normalized feature vector

[0090] S4.3, let the second orthogonal matrix According to S4, we can get the first orthogonal matrix U{n}=Ct{n}*V -1 {n}*S -1 {n};

[0091] In the formula, S -1 {n} represents the inverse matrix of the first diagonal matrix S{n}, V -1 {n} represents the inverse matrix of the second orthogonal matrix V{n}.

[0092] In this embodiment, U{n}, V{n} and S{n} are matrices that satisfy Ct{n}=U{n}*S{n}*V{n}. At this time, the process of decomposing the Curvelet coefficient matrix Ct{n} into two orthogonal matrices and a diagonal matrix is ​​completed, that is, the process of performing SVD decomposition of the Curvelet coefficient matrix Ct{n} is completed.

[0093] Step 5: Perform linear transformation respectively to obtain the second diagonal matrix S'{n};

[0094] S5.1. Yes Perform linear transformation respectively to obtain the modified eigenvalue λ 1 '...λ 10 ', the subscript K represents the number of modified eigenvalues, K = M = 10;

[0095] S5.2. Correct the eigenvalue λ 1 '...λ 10'Assign values ​​to the diagonal elements of the second diagonal matrix S'{n} in sequence, and take the element on the first row and first column of the second diagonal matrix S'{n} as the starting point to obtain the second diagonal matrix S'{n}.

[0096] In this embodiment, the coordinates of the first row and first column elements of the first diagonal matrix S{n} are defined as (1,1), and the coordinates of the first row and first column elements of the first diagonal matrix S{n} are (1,1) as the starting point, the first ten elements of the diagonal position of the first diagonal matrix S{n} from the upper left corner to the lower right corner are kept unchanged, and the other elements are set to 0; and the elements of the first row and first column pixel position are set to Perform a linear transformation to obtain the modified eigenvalue λ 1 ',Right now

[0097]

[0098] Among them, a represents Gain coefficient, in this embodiment, a=0.64;

[0099] The elements corresponding to the pixel positions with coordinates (2, 2), (3, 3), (4, 4), (5, 5), (6, 6), (7, 7), (8, 8), (9, 9), (10, 10) in the first diagonal matrix S{n} are Perform linear transformation respectively to obtain the modified eigenvalue λ 2 ',λ 3 ',λ 4 ',λ 5 ',λ 6 ',λ 7 ',λ 8 ',λ 9 ',λ 10 ',Right now:

[0100] Where b represents Gain coefficient, in this embodiment, b = 0.52;

[0101] The corrected eigenvalue λ 1 ' as the diagonal element at the first row and first column of the second diagonal matrix S'{n}, and λ 2 ',λ 3 ',λ 4 ',λ 5 ',λ 6 ',λ 7 ',λ 8 ',λ 9 ',λ 10' are respectively used as the diagonal elements corresponding to the pixel positions with coordinates of the second diagonal matrix S'{n} (2, 2), (3, 3), (4, 4), (5, 5), (6, 6), (7, 7), (8, 8), (9, 9), and (10, 10), and the other elements are set to 0 to obtain the second diagonal matrix S'{n}.

[0102] Step 6: Multiply the second diagonal matrix S'{n} by the first orthogonal matrix U{n} on the left and by the second orthogonal matrix V{n} on the right to obtain the first coefficient matrix Ct'{n};

[0103] Step 7: Subtract the subband coefficients in the Curvelet coefficient matrix Ct{n} from the corresponding subband coefficients in the first coefficient matrix Ct'{n} to obtain a second coefficient matrix Ct"{n};

[0104] Step 8: Perform inverse Curvelet transform on the second coefficient matrix Ct"{n} to obtain the final infrared image f data with non-uniformity removed, see Figure 3 .

[0105] In this embodiment, the second diagonal matrix S'{n} is the same size as the first diagonal matrix S'{n}; the columns where the first ten elements on the diagonal of the first diagonal matrix S{n} are located are used as the eigenvectors of the Curvelet coefficient matrix Ct{n}. The eigenvectors closer to the upper left corner can better represent the low-frequency information of the Curvelet coefficient matrix Ct{n}, and the eigenvectors closer to the lower right corner can better represent the high-frequency information of the Curvelet coefficient matrix Ct{n}.

[0106] The present invention utilizes Curvelet transform to extract multi-scale and multi-angle coefficient matrices of infrared images, and utilizes singular value decomposition (SVD) algorithm to obtain the first diagonal matrix corresponding to the Curvelet coefficient matrix at each scale. According to Curvelet transform theory, larger elements in the Curvelet coefficient matrix correspond to high-frequency information of a stronger infrared image, and smaller elements correspond to low-frequency noise. After SVD decomposition is performed on the Curvelet coefficient matrix, the larger the diagonal elements in the first diagonal matrix obtained contain, the more image information. Since the non-uniformity of the infrared image includes both high-frequency non-uniformity and part of the low-frequency non-uniformity, the image non-uniformity is contained in the larger diagonal elements. Therefore, linear correction is performed on part of the eigenvalues ​​to obtain the non-uniform noise extracted from the infrared image.

Claims

1. A non-uniformity correction method based on multi-scale analysis and singular value decomposition, It is characterized in that The following steps are involved: S1, collecting infrared images f with non-uniformity; S2. For the infrared image f, use the second generation fast Curvelet transform to obtain the Curvelet coefficient matrix Ct{n}; where n represents the scale of the Curvelet coefficient matrix Ct{n}, and n∈[1,6]; S3. Establish equation |Ct LH {n}-λ m E|=0, solve the equation to obtain the eigenvalue λ of the Curvelet coefficient matrix Ct{n} 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n}; Where, Ct LH {n} represents the product matrix of Curvelet coefficient matrix Ct{n} and its transpose; λ m represents the mth eigenvalue of the Curvelet coefficient matrix Ct{n}, where m is 1...M, and M is a positive integer; E represents the identity matrix; N represents the number of eigenvectors, and N=M; S4. Perform singular value decomposition on the Curvelet coefficient matrix Ct{n}, and obtain Ct{n}=U{n}*S{n}*V{n}, and obtain the first diagonal matrix S{n}, the first orthogonal matrix U{n} and the second orthogonal matrix V{n}; S5, performing linear correction on the eigenvalues ​​of the first diagonal matrix S{n} after taking square root, to obtain a second diagonal matrix S'{n}; S6. Multiply the second diagonal matrix S'{n} by the first orthogonal matrix U{n} on the left and by the second orthogonal matrix V{n} on the right to obtain a first coefficient matrix Ct'{n}; S7, subtract the subband coefficients in the Curvelet coefficient matrix Ct{n} from the corresponding subband coefficients in the first coefficient matrix Ct'{n} to obtain a second coefficient matrix Ct"{n}; S8. Perform an inverse Curvelet transform on the second coefficient matrix Ct″{n} to obtain an infrared image f with non-uniformity removed.

2. According to the non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 1, It is characterized in that S3 is specifically: S3.

1. Find the transpose of the Curvelet coefficient matrix Ct{n} and get Ct T {n}; S3.

2. Compare the Curvelet coefficient matrix Ct{n} with Ct in S3.1 T {n} multiply to get Ct LH {n}; S3.

3. Solve the equation |Ct LH {n}-λ m E|=0, the eigenvalue of the Curvelet coefficient matrix Ct{n} is obtained as λ 1 ... M With Ct LH The eigenvector x of {n} 1 {n}...x N {n}.

3. According to the non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 2, It is characterized in that S4 is specifically: S4.

1. For eigenvalue λ 1 ... M The square root is and will Assign values ​​to the diagonal elements of the first diagonal matrix S{n} in sequence, and take the element in the first row and first column of the first diagonal matrix S{n} as the starting point to obtain the first diagonal matrix S{n}; S4.

2. Ct LH The eigenvector x of {n} 1 {n}...x N {n} performs normalization operation to obtain normalized feature vector S4.3, let the second orthogonal matrix According to Ct{n}=U{n}*S{n}*V{n}, we can get the first orthogonal matrix U{n}=Ct{n}*V -1 {n}*S -1 {n}; In the formula, S -1 {n} represents the inverse matrix of the first diagonal matrix S{n}, V -1 {n} represents the inverse matrix of the second orthogonal matrix V{n}.

4. According to the non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 3, It is characterized in that S5 is as follows: S5.

1. Yes Perform linear corrections respectively to obtain the corrected eigenvalue λ 1 '...λ K ', the subscript K represents the number of corrected eigenvalues; S5.

2. Correct the eigenvalue λ 1 '...λ K 'Assign values ​​to the diagonal elements of the second diagonal matrix S'{n} in sequence, and take the element on the first row and first column of the second diagonal matrix S'{n} as the starting point to obtain the second diagonal matrix S'{n}.

5. According to the non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 4, Features: In S4.1, the first diagonal matrix S{n} and the second diagonal matrix S'{n} in S5.2 have the same size.

6. The non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 5, Features: In S2, the value of n is 1 to 5.

7. The non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 6, Features: In S3, the values ​​of M and N are both 10.

8. The non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 7, Features: In S5.1, the modified characteristic value λ 1 The gain coefficient a is greater than the modified characteristic value λ 2 '...λ 10 'Gain coefficient b.

9. The non-uniformity correction method based on multi-scale analysis and singular value decomposition according to claim 8, Features: The modified characteristic value λ 1 'The gain coefficient a is 0.64; The modified characteristic value λ 2 '...λ 10 'The gain coefficient b is 0.

52.

10. A non-uniformity correction method based on multi-scale analysis and singular value decomposition according to any one of claims 1 to 9, It is characterized in that S1 specifically includes: using an infrared thermal imager to collect an infrared image f with non-uniformity.