A scene-based multiscale non-uniformity correction method

Abstract

The application discloses a scene-based multi-scale non-uniformity correction method, which firstly establishes an image pyramid for an input image; for each layer of the image pyramid, a local linear model in guided filtering is used, i.e., an image after non-uniformity removal is taken as a guide image, in a local neighborhood window, the guide image and a filtering output image are in linear relationship, meanwhile, a low-frequency non-uniformity model in the local neighborhood window is considered as a polynomial model with coordinate axes as variables; then a weighted least square optimization method is used to constrain the non-uniformity, the original optimization problem is decomposed into two sub-problems, the alternating direction multiplier method is used to solve the two sub-problems, iteration optimization is carried out, and the non-uniformity of the image is obtained; finally, the non-uniformity obtained from each layer of the image pyramid is reconstructed, a final non-uniformity image is obtained, and a clear image is obtained by subtracting the non-uniformity from the original image.

Description

Technical Field

[0001] This invention relates to a scene-based multi-scale non-uniformity correction method, which relates to the field of digital image processing. Background Technology

[0002] Due to various factors such as detector limitations, optical lens limitations, and optical system limitations, infrared imaging systems generally suffer from non-uniformity issues. This means that while the incident light intensity and energy on the detector are uniform, the output signals from different units of the infrared detector are inconsistent. Image non-uniformity severely affects subsequent image analysis and processing.

[0003] Existing technologies cannot accurately correct non-uniformity, and are prone to leaving non-uniformity or causing image blurring. In order to accurately correct non-uniformity, this invention proposes a scene-based multi-scale non-uniformity correction method. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a scene-based multi-scale non-uniformity correction method.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] First, an image pyramid is constructed from the input image. For each layer of the pyramid, a local linear model from guided filtering is used, where the image after removing non-uniformity is used as the guide image. Within the local neighborhood window, the guide image and the filtered output image are linearly related. Simultaneously, the low-frequency non-uniformity model within the local neighborhood window is considered a polynomial model with coordinate axes as variables. Then, a weighted least squares optimization method is used to constrain the non-uniformity, decomposing the original optimization problem into two sub-problems. These are solved using the alternating direction multiplier method, iteratively optimizing to obtain the image's non-uniformity. Finally, the non-uniformity obtained from each layer of the image pyramid is reconstructed to obtain the final non-uniform image. Subtracting the non-uniformity from the original image yields the clear image.

[0007] A scene-based multi-scale non-uniformity correction method, the method comprising:

[0008] Step 1: Construct an n-layer Laplacian pyramid for the input image I: Image of the nth level of the Laplace Pyramid The resolution is wid×hei, wid≥8, hei≥8;

[0009] Step 2: First, perform non-uniform correction on the nth layer pyramid image. The input image for the nth layer pyramid non-uniform correction is... Using the local linear model in guided filtering, that is, using the image after removing non-uniformity as the guide image, within a square window w centered at pixel k with radius r. k In the middle, the guiding image and the filtered output image A linear relationship is assumed, and the window w is considered to be... k The low-frequency non-uniformity model is a polynomial model with coordinate axes as variables, and weighted least squares optimization constraints are established:

[0010]

[0011] in, The guided filter output image after removing non-uniformities from the nth pyramid layer. For a non-uniform image of the nth layer of the pyramid, w k A square window centered at pixel k with radius r, a k b k e k f k and g k It is a window w k In the linear coefficients, ε and λ are regularization coefficients; ε is used to prevent a... k λ becomes too large to prevent e from becoming too large. k and f k It became too big. and These are the index values ​​of the x and y coordinates of the nth layer pyramid image, respectively.

[0012] The original optimization problem is decomposed into two sub-problems:

[0013] 1) Determine Find a k b k :

[0014]

[0015] 2) Determine Find e k f k and g k :

[0016]

[0017] Solving subproblem 1) yields a k and b k Thus obtain Solving subproblem 2) yields e k f k and g k Thus obtain The two subproblems are solved alternately and optimized, with 8 to 12 iterations, to finally obtain the non-uniform image of the nth layer of the pyramid.

[0018] Next, non-uniform correction is performed on the pyramid images from the (n-1)th to the 1st layer, and the non-uniform correction input image for the mth (1≤m<n)th layer is... Expressed using the following formula:

[0019]

[0020] Here, pyrUp represents Gaussian upsampling, and then the same optimization method is used to obtain the non-uniform image of the m-th pyramid layer.

[0021] Step 3: Perform Gaussian pyramid reconstruction on the non-uniform images obtained from each layer of the pyramid to obtain the final non-uniform image N;

[0022] Step 4: Subtract the non-uniform image from the original input image to obtain the corrected, clear image H:

[0023] H = IN.

[0024] The beneficial effects of this invention are as follows:

[0025] 1. This invention is applicable to non-uniformity correction of images in various scenarios;

[0026] 2. This invention corrects non-uniformity without destroying the scene information of the image, and does not produce artifacts or blurred edges; Attached Figure Description

[0027] Figure 1 This is a flowchart of a scene-based multi-scale non-uniformity correction method according to the present invention;

[0028] Figure 2 The Laplacian pyramid for the image;

[0029] Figure 3 This is a flowchart illustrating the reconstruction process of the non-uniformity of each layer of the pyramid in this invention;

[0030] Figure 4 The correction results of the present invention are shown in (a)-(e), which are the original clear image, the image after adding non-uniformity, the non-uniform image added to the clear image, the image after non-uniformity correction by the present invention, and the non-uniform image obtained by the present invention, respectively. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0032] like Figure 1As shown, the present invention provides a scene-based multi-scale non-uniformity correction method, which specifically includes the following steps:

[0033] The first step is to construct an n-layer Laplacian pyramid for the input image I: Image of the nth level of the Laplace Pyramid The Laplace pyramid is defined as follows: (The resolution is wid×hei, wid≥8, hei≥8.)

[0034]

[0035] in, This represents the m-th level in Laplace's pyramid. Let represent the m-th level in the Gaussian pyramid, and pyrUp represent Gaussian upsampling. For example... Figure 2 As shown, this is a four-tiered Laplace pyramid.

[0036] The second step is to first perform non-uniform correction on the nth layer pyramid image. The input image for the nth layer pyramid non-uniform correction is... Using the local linear model in guided filtering, that is, using the image after removing non-uniformity as the guide image, within a square window w centered at pixel k with radius r. k In the middle, the guiding image and the filtered output image A linear relationship is assumed, and the window w is considered to be... k The low-frequency non-uniformity model is a polynomial model with coordinate axes as variables, and weighted least squares optimization constraints are established:

[0037]

[0038] in, The guided filter output image after removing non-uniformities from the nth pyramid layer. For a non-uniform image of the nth layer of the pyramid, w k A square window centered at pixel k with radius r, a k b k e k f k and g k It is a window w k In the linear coefficients, ε and λ are regularization coefficients; ε is used to prevent a... k λ becomes too large to prevent e from becoming too large. k and f k It became too big. and These are the index values ​​of the x and y coordinates of the nth layer of the pyramid image, respectively.

[0039] The original optimization problem is decomposed into two sub-problems:

[0040] 3) Determine Find a k b k :

[0041]

[0042] 4) Determine Find e k f k and g k :

[0043]

[0044] Solving subproblem 1) yields a k and b k Thus obtain Solving subproblem 2) yields e k f k and g k Thus obtain The two subproblems are solved alternately and optimized, with 8 to 12 iterations, to finally obtain the non-uniform image of the nth layer of the pyramid.

[0045] Next, non-uniform correction is performed on the pyramid images from the (n-1)th to the 1st layer, and the non-uniform correction input image for the mth (1≤m<n)th layer is... Expressed using the following formula:

[0046]

[0047] Then, using the same optimization method, the non-uniform image of the m-th pyramid is obtained.

[0048] The third step is to reconstruct the Gaussian pyramid based on the non-uniformity obtained from each layer of the pyramid:

[0049]

[0050] in, This represents the non-uniform image after reconstruction at the m-th layer, and the final non-uniform image. like Figure 3 As shown.

[0051] The fourth step is to subtract the non-uniform image from the original input image to obtain the corrected, clear image H:

[0052] H = IN.

[0053] Figure 4The correction results of the present invention are shown in (a)-(e), which are the original clear image, the image after adding non-uniformity, the non-uniform image added to the clear image, the image after non-uniformity correction by the present invention, and the non-uniform image obtained by the present invention, respectively.

[0054] The parts of this invention not described in detail are well-known in the art. The embodiments described above are merely preferred embodiments of the present invention, and do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Various modifications and improvements to the technical solutions of this invention made by those skilled in the art without departing from the spirit of the invention should fall within the protection scope defined by the claims of this invention.

Claims

1. A scene-based multiscale non-uniformity correction method, characterized in that, The method comprises the following steps: Step one, establish n-layer Laplacian pyramid for input image I: the n-layer Laplacian pyramid image resolution is wid x hei, wid ≥ 8, hei ≥ 8; Step two, first, the n-th layer pyramid image is subjected to non-uniform correction, the n-th layer pyramid non-uniform correction input image The local linear model in guided filtering is utilized, that is, the image after non-uniformity removal is taken as a guided image, in a square window w centered at pixel k with a radius of r k , the guided image and the filtering output image are in a linear relationship, and it is considered that the low-frequency non-uniformity model in the window w k is a polynomial model with coordinate axes as variables, and a weighted least square optimization constraint is established: wherein, is the guided filter output image of the n-th layer pyramid after removing non-uniformity, is the non-uniform image of the n-th layer pyramid, w k is a square window centered at pixel k with radius r, a k , b k , e k , f k , and g k are linear coefficients in the window w k , ε and λ are regularization coefficients, ε is to prevent a k from becoming too large, and λ is to prevent e k and f k from becoming too large, and are index values of the horizontal and vertical coordinates of the n-th layer pyramid image, respectively. The original optimization problem is decomposed into two sub-problems: 1) determine find a k , b k : 2) determine find e k , f k , and g k : Solving sub-problem 1) gives a k and b k , so that Solving sub-problem 2) gives e k , f k and g k , so that The two sub-problems are solved alternately and optimized, and after 8-12 iterations, the non-uniform image of the nth layer of the pyramid is obtained Next, the non-uniform correction is performed on the n-1th to 1st layer pyramid images, respectively, and the non-uniform correction input image of the mth (1≤m This is expressed by the following equation: where pyrUp denotes a Gaussian upsampling followed by the same optimization method to obtain the non-uniform image of the m-th pyramid level Step three, the non-uniformity image obtained by each layer of the pyramid is reconstructed by Gaussian pyramid to obtain a final non-uniform image N; Step four, the original input image is subtracted from the non-uniform image to obtain a corrected clear image H: H = I - N.

2. The method of claim 1, wherein, In step one, the definition form of the Laplace pyramid is: wherein, represents the m-th layer in the Laplacian pyramid, represents the m-th layer in the Gaussian pyramid, pyrUp represents Gaussian up-sampling.

3. The method of claim 2, wherein, In step three, the non-uniformity obtained by each layer of the pyramid is reconstructed by Gaussian pyramid: wherein, denotes the reconstructed non-uniformity image of the m-th layer, the final non-uniformity image

Images