A DR blurred image blind deconvolution restoration method and system

Abstract

The present invention relates to a blind deconvolution restoration method and system for DR blurred images, the method steps comprising: based on the grayscale distribution of DR images being concentrated and continuously constrained by using the minimum value of the image region, combining the L2 norm to characterize the sparse characteristics of the gradient, suppressing image noise by total variation, establishing a blind deconvolution mathematical model under the maximum a posteriori probability framework, solving the potential clear image and blur kernel by using the semi-quadratic splitting rule, and finally restoring the clear image by combining the R-L deconvolution method. The blind deconvolution method proposed by the present invention for the characteristics of DR blurred images can effectively remove the blur of DR images, restore the key information of the image, and achieve stable and high-quality restoration of DR blurred images. Compared with the prior art, the present invention can effectively remove the blur of DR images, restore the key information of the image, and achieve stable and high-quality restoration of DR blurred images.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a blind deconvolution restoration method and system for a DR blurred image. Background Art

[0002] In digital radiograph detection technology (DR), the object under inspection will be subject to forced vibration due to the influence of the environment, which will form relative motion with the detector, making the DR image blurred. At the same time, various degradation factors are prone to occur during the long-term storage and transmission of the image, which will also cause the DR image to contain noise and serious blur. Therefore, removing the influence of motion blur to obtain higher quality images is a prerequisite for DR image evaluation technology, which requires the use of image restoration technology.

[0003] At present, blind deconvolution methods have received extensive attention in image deblurring research. They mainly include blind deconvolution methods based on image edge estimation, deep learning methods, and methods under the framework of maximum a posteriori probability. For DR blurred images, they usually contain a large number of single background areas, and the edge information that can be used is relatively small. The image edge estimation method requires strong edge features, so it is difficult to implement; thanks to the development of neural networks in recent years, deep learning methods have also been applied to image deblurring research, but this method requires a large number of samples to train. More importantly, DR blurred images in different scenes have strong specificity, and multiple learning models are required to adapt to the blurring conditions in different scenes, which undoubtedly increases the difficulty of the method and the calculation time; the image blind deconvolution method based on the framework of maximum a posteriori probability overcomes the pathological nature of image restoration by introducing various prior information, and uses regularization terms to characterize the heavy-tailed distribution of the clear image gradient, which can effectively remove image blur; but the existing methods do not fully consider the grayscale distribution and noise characteristics of DR images, and cannot effectively improve the restoration quality. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a DR blurred image blind deconvolution restoration method and system.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A DR blurred image blind deconvolution restoration method comprises the following steps:

[0007] Step 1: Obtain the DR blurred image B and construct an image blind deconvolution model; input the DR blurred image B into the constructed image blind deconvolution model, set the initialization blur kernel K and size, and obtain the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved;

[0008] Step 2: Based on the given DR blurred image B and the size n of the initialized blur kernel K, a multi-scale hierarchical strategy is adopted to generate blurred image layers B of different scales. 1 , B 2 …B H , the corresponding input fuzzy kernel is K 1 , K 2 …K H , the blur kernel size is n 1 、n 2 …n H , set the maximum number of iterations N for each fuzzy image feature layer max ;

[0009] Step 3: Based on the blurred image B 1 , fixed fuzzy kernel variable K 1 , obtain auxiliary variables according to the objective function 1, and solve the potential clear image I through fast Fourier transform 1 ;

[0010] Step 4: Based on the blurred image B 1 , fixed latent clear image I 1 , using a fixed latent clear image I 1 The gradient information of the fuzzy kernel variable K is obtained by accelerating the conjugate gradient method. 1 Solve and get the fuzzy kernel K 1 *, K 1 * The blur kernel size is expanded to size n 2 , assigned to K 2 ;

[0011] Step 5: Blurred image feature layer B 2 To B H Repeat steps 3 and 4 to obtain the blur kernel K H *, do normalization to get K e , based on the given DR blurred image B and blur kernel K e , use RL deconvolution to restore the final clear image I e .

[0012] Preferably, the specific steps of obtaining the DR blurred image B in step 1 and constructing the image blind deconvolution model include:

[0013] The mathematical model for constructing the fuzzy model of the image is:

[0014]

[0015] Where I and B represent the clear image and blurred image respectively, K and N represent the blur kernel and noise respectively. is the convolution operator,

[0016] Get the minimum value of the image pixel area:

[0017]

[0018] i and j represent the row vector and column vector of the image matrix, respectively. R(I) represents the pixel matrix of image I. Ω(I ij ) represents the neighborhood of the pixel, R min (I ij ) represents the minimum value of this pixel area,

[0019] L 2 The norm is used as a priori condition for the gradient of the DR blurred image. At the same time, the total variation regularization is combined to suppress image noise, and an image blind deconvolution model is constructed under the maximum a posteriori probability framework:

[0020]

[0021] In the formula, represents the result of I,K when the first term is optimally solved, ‖R min (I ij )‖ 0 Count the number of extreme non-zero values ​​in the pixel area, ‖·‖ 0 is the L0 norm, ‖I‖ TV represents the second-order total variation, specifically L and W represent the length and width of the image matrix, Z 2 I is the image pixel difference obtained in the horizontal and vertical directions, Characterize the gradient of the image, For L 2 norm, represents the image gradient operator, Respectively represent the vertical and perpendicular gradient operators, α, β, δ, η, and λ are the parameters of each constraint, which are used to maintain a balance between regularization strength and fidelity. Represents the prior condition of the blur kernel.

[0022] Preferably, in step 1, the DR blurred image B is input into the constructed image blind deconvolution model, the initialization blur kernel K and size are set, and the steps of obtaining the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved include:

[0023] Based on the alternating minimization algorithm of semi-quadratic splitting, the blind deconvolution model is decomposed into two sub-problems each containing only one unknown quantity, and the objective function 1 is established:

[0024]

[0025] And establish objective function 2:

[0026]

[0027] In objective function one, the blur kernel variable K is fixed, and the clear image I is solved. After the clear image variable I is solved, the clear image I is fixed and then objective function two is optimized. Objective function one and objective function two are optimized alternately until convergence to obtain solutions I and K of the image blind deconvolution model.

[0028] Preferably, the step 2 specifically includes the following steps:

[0029] Set the number of multi-scale image layers H, and use the Gaussian filter to convolve the image and downsample it at the same time. The rule is:

[0030]

[0031] In the formula, G(m,n) represents the Gaussian filter, m and n represent the horizontal and vertical dimensions of the filter, i and j represent the row vector and column vector of the image matrix respectively, l = [1,H], and the blurred image layer from small to large is B 1 , B 2 …B H ; Corresponding input fuzzy kernel K 1 , K 2 …K H ; blur kernel size is n 1 、n 2 …n H .

[0032] Preferably, the step 3 specifically comprises the following steps:

[0033] Rewrite the image blind deconvolution model based on auxiliary variables:

[0034]

[0035] Among them, f is auxiliary variable one, q is auxiliary variable two, and p is auxiliary variable three, where p = (p h ,p v ) T Approach where p h 、p v To approximate the gradient in the horizontal and vertical directions, δ, γ, and μ are penalty coefficients.

[0036] Fix the clear image variable I to obtain the formulas for auxiliary variable 1 f, auxiliary variable 2 q, and auxiliary variable 3 p:

[0037]

[0038]

[0039]

[0040] Preferably, when solving I in step S3, f, q, and p are fixed, and through Fourier transform, the closed-form solution of I is:

[0041]

[0042] Where: and denote the forward and inverse Fourier transform, respectively. represents the conjugate Fourier transform,

[0043] Preferably, in step 3, the step of solving the clear image variable I at a certain scale is as follows:

[0044] 1) Set the maximum number of iterations of the current blurred image layer to N max , initialize the current number of iterations iter = 1, δ = set value, γ = set value, μ = set value;

[0045] 2) If δ<δ max ,δ max As the upper limit of the penalty coefficient 1, solve the auxiliary variable 1 f, and assign the penalty coefficient 1 δ: δ = 2δ;

[0046] 3) If γ<γ max ,γ max As the upper limit of the penalty coefficient 2, solve the auxiliary variable q and assign the penalty coefficient γ: γ = 2γ;

[0047] 4) If μ<μ max ,μ max For the upper limit of the penalty coefficient three, solve the auxiliary variable two p, and assign the penalty coefficient one μ: μ = 2μ;

[0048] 5) Let iter = iter + 1, and repeat steps 2) to 4) until the penalty coefficient exceeds the upper limit and the number of iterations iter>N max .

[0049] Preferably, in step 4, the following steps are specifically included:

[0050] For the blur kernel variable K, using the gradient information of the clear image I, the objective function 2 is rewritten as:

[0051]

[0052] And further rewritten as:

[0053]

[0054] Among them, v is auxiliary variable four, is the penalty factor four;

[0055] The blur kernel K is quickly solved by Fourier transform, and its closed-form solution calculation formula is:

[0056]

[0057] Solve for the auxiliary variable four v:

[0058]

[0059] A DR blurred image blind deconvolution restoration system, preferably, the system comprises an objective function acquisition module, a blurred image layer acquisition module, a potential clear image acquisition module, a blur kernel acquisition module, and a clear image acquisition module.

[0060] The objective function acquisition module is used to acquire the DR blurred image B and construct an image blind deconvolution model; the DR blurred image B is input into the constructed image blind deconvolution model, the initialization blur kernel K and the size are set, and the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved are obtained;

[0061] The fuzzy image layer acquisition module is used to generate fuzzy image layers B of different scales based on a given DR fuzzy image B and the size n of the initialized fuzzy kernel K, using a multi-scale hierarchical strategy. 1 , B 2 …B H , the corresponding input blur kernel is K 1 , K 2 …K H , the blur kernel size is n 1 、n 2 …n H , set the maximum number of iterations N for each fuzzy image feature layer max ;

[0062] The potential clear image acquisition module is used to obtain the blurred image B 1 , fixed fuzzy kernel variable K 1 , obtain auxiliary variables according to the objective function 1, and solve the potential clear image I through fast Fourier transform 1 ;

[0063] The blur kernel acquisition module is used to obtain the blur kernel based on the blurred image B 1 , fixed latent clear image I 1 , using a fixed latent clear image I 1 The gradient information of the fuzzy kernel variable K is obtained by accelerating the conjugate gradient method. 1 Solve and get the fuzzy kernel K 1*, K 1 * The blur kernel size is expanded to size n 2 , assigned to K 2 ;

[0064] The clear image acquisition module is used to obtain the blurred image feature layer B 2 To B H , use the potential clear image acquisition module and the blur kernel acquisition module to repeat the process and obtain the blur kernel K H *, do normalization to get K e , based on the given DR blurred image B and blur kernel K e , use RL deconvolution to restore the final clear image I e .

[0065] Preferably, the specific steps of constructing the image blind deconvolution model by the objective function acquisition module include:

[0066] The mathematical model for constructing the fuzzy model of the image is:

[0067]

[0068] Where I and B represent the clear image and blurred image respectively, K and N represent the blur kernel and noise respectively. is the convolution operator,

[0069] Get the minimum value of the image pixel area:

[0070]

[0071] i and j represent the row vector and column vector of the image matrix, respectively. R(I) represents the pixel matrix of image I. Ω(I ij ) represents the neighborhood of the pixel, R min (I ij ) represents the minimum value of this pixel area,

[0072] L 2 The norm is used as a priori condition for the gradient of the DR blurred image. At the same time, the total variation regularization is combined to suppress image noise, and an image blind deconvolution model is constructed under the maximum a posteriori probability framework:

[0073]

[0074] In the formula, represents the result of I,K when the first term is optimally solved, ‖R min (I ij )‖ 0 Count the number of extreme non-zero values ​​in the pixel area, ‖·‖ 0 is the L0 norm, ‖I‖ TV represents the second-order total variation, specifically L and W represent the length and width of the image matrix, Z 1 I. Z 2 I is the image pixel difference obtained in the horizontal and vertical directions, Characterize the gradient of the image, For L 2 norm, represents the image gradient operator, Respectively represent the vertical and perpendicular gradient operators, α, β, δ, η, and λ are the parameters of each constraint, which are used to maintain a balance between regularization strength and fidelity. Represents the prior condition of the blur kernel.

[0075] Compared with the prior art, the present invention has the following advantages:

[0076] The blind deconvolution restoration method of DR blurred image of the present invention, based on the continuity of grayscale distribution of DR blurred image, gradient obeying heavy-tailed distribution and noise characteristics, uses the minimum value of image region to constrain grayscale distribution, combines L2 norm to characterize gradient sparse characteristics, and suppresses image noise by total variation. A blind deconvolution model is established under the framework of maximum a posteriori probability, and the semi-quadratic splitting rule is introduced to solve the blur kernel and potential clear image. At the same time, the image multi-scale strategy is used to estimate the blur kernel layer by layer from the smallest layer to the largest layer, effectively improving the accuracy of the blur kernel. This method can better remove DR image blur, effectively restore key areas such as image weld defects and marking information, and provide effective help for subsequent quality evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 A flowchart of a blind deconvolution restoration method for a DR blurred image according to the present invention;

[0078] Figure 2 The blur kernels of images of various scales are estimated based on a blind deconvolution restoration method for DR blurred images according to an embodiment of the present invention;

[0079] Figure 3 An initial blurred image based on a DR blurred image blind deconvolution restoration method according to an embodiment of the present invention;

[0080] Figure 4 A potential clear image based on a DR blurred image blind deconvolution restoration method according to an embodiment of the present invention;

[0081] Figure 5 This is a clear image based on a DR blurred image blind deconvolution restoration method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0082] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. Note that the following embodiments are merely illustrative in nature, and the present invention is not intended to limit its applicable objects or uses, and the present invention is not limited to the following embodiments.

[0083] Example

[0084] A blind deconvolution restoration method for DR blurred images, such as Figure 1 As shown, the following steps are included:

[0085] Step 1: Obtain the DR blurred image B and construct an image blind deconvolution model; input the DR blurred image B into the constructed image blind deconvolution model, set the initialization blur kernel K and size, and obtain the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved;

[0086] Step 1.1 The mathematical model of the fuzzy model of the image is expressed as:

[0087]

[0088] Where I and B represent the clear image and blurred image respectively, K and N represent the blur kernel and noise respectively. is the convolution operator; the minimum value of the image pixel area is defined as:

[0089] R min (I ij )=min i,j∈R(I) Ω(I ij ) (2)

[0090] i and j represent the row vector and column vector of the image matrix, respectively. R(I) represents the pixel matrix of image I. Ω(I ij ) represents the neighborhood of the pixel, R min (I ij ) represents the minimum value of this pixel area. 2 The norm is used as a priori condition for the gradient of the DR blurred image, and the total variation regularization is combined to suppress image noise. The following image blind deconvolution model is constructed under the maximum a posteriori probability framework:

[0091]

[0092] In the formula, represents the result of I,K when the first term obtains the optimal solution; ‖R min (I ij )‖ 0 The number of non-zero extreme values ​​in the pixel region, ‖·‖ 0 is the L0 norm; ‖I‖ TV represents the second-order total variation, specifically ||I|| TV = L and W represent the length and width of the image matrix, Z 1 I. Z 2 I is the image pixel difference obtained in the horizontal and vertical directions respectively; Characterize the gradient of the image, For L 2 norm, represents the image gradient operator, Represent the vertical and perpendicular gradient operators respectively; α, β, δ, η, and λ are the parameters of each constraint, maintaining a balance between regularization strength and fidelity. Represents the prior condition of the blur kernel.

[0093] Step 1.2 Based on the alternating minimization algorithm of semi-quadratic splitting, the blind deconvolution model is decomposed into two sub-problems each containing only one unknown quantity, and objective function 1 (4) and objective function 2 (5) are established:

[0094]

[0095]

[0096] In objective function 1 (3), the blur kernel variable K is fixed and the clear image I is solved. After the clear image variable I is solved, the clear image I is fixed and objective function 2 (5) is optimized again. Objective function 1 (4) and objective function 2 (5) are optimized alternately until convergence to obtain solutions I and K of the image blind deconvolution model.

[0097] Step 2: Based on the given DR blurred image B and the size n of the initialized blur kernel K, a multi-scale hierarchical strategy is adopted to generate blurred image layers B of different scales. 1 , B 2 …B H , the corresponding input blur kernel is K 1 , K 2 …K H , the blur kernel size is n 1 、n 2 …n H , set the maximum number of iterations N for each fuzzy image feature layer max

[0098] Set the number of multi-scale image layers H, use the Gaussian filter to convolve the image, and downsample it at the same time. Rule 1 is shown in formula (6)

[0099]

[0100] In the formula, G(m,n) represents the Gaussian filter, m and n represent the horizontal and vertical dimensions of the filter, i and j represent the row vector and column vector of the image matrix respectively, l = [1,H], and the blurred image layer from small to large is B 1 , B 2 …B H ; Corresponding input fuzzy kernel K 1 , K 2 …K H ; blur kernel size is n 1 、n 2 …n H .

[0101] Step 3: Based on the blurred image feature layer B1, fix the blur kernel variable K1, introduce auxiliary variables according to the objective function 1, and solve the potential clear image I1 through fast Fourier transform:

[0102] Introduce auxiliary variables one f, two q, and three p, where p = (p h ,p v ) T Approach where p h 、p v In order to approximate the gradients in the horizontal and vertical directions, weight variables δ, γ, and μ are introduced at the same time, so formula (3) can be rewritten as formula (7):

[0103]

[0104] In the formula, δ, γ, and μ are penalty coefficients. The four variables I, f, q, and p in formula (6) can be solved independently by fixing other variables. Therefore, when solving I, f, q, and p are fixed. Through Fourier transform, the closed-form solution of I is:

[0105]

[0106] Where: and denote the forward and inverse Fourier transform, respectively. represents the conjugate Fourier transform,

[0107] Fix the clear image variable I, and rewrite equation (7) as follows with respect to auxiliary variable 1 f, auxiliary variable 2 q, and auxiliary variable 3 p:

[0108]

[0109]

[0110]

[0111] The steps to solve the clear image variable I(l) at a certain image scale are as follows:

[0112] 1) Set the maximum number of iterations of the current blurred image layer to N max , initialize the current number of iterations iter = 1, δ = γ = set value; such as 0.0002, μ = set value, such as 0.0004, δ max =γ max =μ max = set value, such as 50000;

[0113] 2) If δ<δ max ,δ max is the upper limit of the penalty coefficient 1, and the auxiliary variable 1f is solved by formula (9), and the penalty coefficient 1δ is assigned: δ = 2δ;

[0114] 3) If γ<γ max ,γ max is the upper limit of the penalty coefficient 2, and the auxiliary variable q is solved by formula (10), and the penalty coefficient γ is assigned: γ = 2γ;

[0115] If μ<μ max ,μ max is the upper limit of the penalty coefficient three. The auxiliary variable two p is solved by formula (11), and the penalty coefficient one μ is assigned: μ = 2μ;

[0116] Let iter = iter + 1, and repeat steps 2) to 4) until the penalty coefficient exceeds the upper limit and the number of iterations iter>N max .

[0117] Step 4: For the fuzzy kernel variable K, using the gradient information of the clear image I, rewrite the objective function 2 (5) as follows:

[0118]

[0119] Introduce auxiliary variable 4v, penalty coefficient 4 Rewrite equation (12) as

[0120]

[0121] The blur kernel K is quickly solved by Fourier transform, and its closed-form solution is calculated as shown in formula (14):

[0122]

[0123] Solve for the auxiliary variable four v:

[0124]

[0125] A DR blurred image blind deconvolution restoration method, based on the above process, the specific implementation steps of this embodiment are:

[0126] 1) Input blurred image B, initialize blur kernel K, and set the maximum number of iterations to N maI , δ=γ=set value; such as 0.0002, μ=set value, such as 0.0004,

[0127] 2) Multi-scale layering of blurred image B, number of layers H, current feature layer l = 1;

[0128] 3) Initialize the number of iterations iter = 1;

[0129] 4) Solve the potential clear image I(l) of the current blurred image feature layer through equation (8);

[0130] 5) Solve the blur kernel K(l) of the current blurred image feature layer through equation (12);

[0131] 6) Let iter = iter + 1, and repeat steps 4) and 5) until the number of iterations iter>NmaI;

[0132] 7) After the iteration is completed, set l = l + 1 and jump to step 3) until l>H, and output K(l);

[0133] 8) Normalize K(l) to get the blur kernel Ke; then use the blur kernel Ke and the blurred image B to perform RL deconvolution to get the clear image I e .

[0134] In this embodiment, the estimated blur kernels of each scale are obtained. Figure 2 The initial blurred image, potential clear image, and clear image are obtained in sequence. Figures 3 to 5 Compared with other blind deconvolution methods, the present invention can estimate the blur kernel more accurately and restore the weld and marking information area better according to the characteristics of DR blurred images. The overall image visual effect is significantly improved, which provides effective help for subsequent weld quality evaluation.

[0135] A DR blurred image blind deconvolution restoration system, the system comprises an objective function acquisition module, a blurred image layer acquisition module, a potential clear image acquisition module, a blur kernel acquisition module, and a clear image acquisition module.

[0136] The objective function acquisition module is used to acquire the DR blurred image B and construct an image blind deconvolution model; the DR blurred image B is input into the constructed image blind deconvolution model, the initialization blur kernel K and the size are set, and the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved are obtained;

[0137] The fuzzy image layer acquisition module is used to generate fuzzy image layers B of different scales based on a given DR fuzzy image B and the size n of the initialized fuzzy kernel K, using a multi-scale hierarchical strategy. 1 , B 2 …B H , the corresponding input fuzzy kernel is K 1 , K 2 …K H , the blur kernel size is n 1 、n 2 …n H , set the maximum number of iterations N for each fuzzy image feature layer max ;

[0138] The potential clear image acquisition module is used to obtain the blurred image B 1 , fixed fuzzy kernel variable K 1 , obtain auxiliary variables according to the objective function 1, and solve the potential clear image I through fast Fourier transform 1 ;

[0139] The blur kernel acquisition module is used to obtain the blur kernel based on the blurred image B 1 , fixed latent clear image I 1 , using a fixed latent clear image I 1 The gradient information of the fuzzy kernel variable K is obtained by accelerating the conjugate gradient method. 1 Solve and get the fuzzy kernel K 1 *, K 1 * The blur kernel size is expanded to size n 2 , assigned to K 2 ;

[0140] The clear image acquisition module is used to obtain the blurred image feature layer B 2 To B H , use the potential clear image acquisition module and the blur kernel acquisition module to repeat the process and obtain the blur kernel K H *, do normalization to get K e , based on the given DR blurred image B and blur kernel K e , use RL deconvolution to restore the final clear image I e .

[0141] The specific steps of constructing the image blind deconvolution model by the objective function acquisition module include:

[0142] The mathematical model for constructing the fuzzy model of the image is:

[0143]

[0144] Where I and B represent the clear image and blurred image respectively, K and N represent the blur kernel and noise respectively. is the convolution operator,

[0145] Get the minimum value of the image pixel area:

[0146]

[0147] i and j represent the row vector and column vector of the image matrix, respectively. R(I) represents the pixel matrix of image I. Ω(I ij ) represents the neighborhood of the pixel, R min (I ij ) represents the minimum value of this pixel area,

[0148] The L2 norm is used as the prior condition of the DR blurred image gradient, and the total variation regularization is combined to suppress image noise. The image blind deconvolution model is constructed under the maximum a posteriori probability framework:

[0149]

[0150] In the formula, represents the result of I,K when the first term is optimally solved, ‖R min (I ij )‖ 0 Count the number of extreme non-zero values ​​in the pixel area, ‖·‖ 0 is the L0 norm, ‖I‖ TV represents the second-order total variation, specifically L and W represent the length and width of the image matrix, Z 1 I. Z 2 I is the image pixel difference obtained in the horizontal and vertical directions, Characterize the gradient of the image, is the L2 norm, represents the image gradient operator, Respectively represent the vertical and perpendicular gradient operators, α, β, δ, η, and λ are the parameters of each constraint, which are used to maintain a balance between regularization strength and fidelity. Represents the prior condition of the blur kernel.

[0151] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0152] The above embodiments are merely examples and do not limit the scope of the present invention. These embodiments can also be implemented in various other ways, and various omissions, substitutions, and changes can be made without departing from the technical concept of the present invention.

Claims

1. A blind deconvolution restoration method for DR blurred images, characterized in that: The following steps are involved: Step 1: Obtain the DR blurred image B and construct an image blind deconvolution model; input the DR blurred image B into the constructed image blind deconvolution model, set the initial blur kernel and size, and obtain the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved; Step 2: Based on the given DR blurred image B and the size of the initialized blur kernel n, a multi-scale hierarchical strategy is adopted to generate blurred image layers B of different scales. 1 , B 2 …B H , the corresponding input fuzzy kernel variable is K 1 , K 2 …K H , the blur kernel size is n 1 、n 2 …n H , set the maximum number of iterations N for each blurred image layer max ; Step 3: Blurred image layer B 1 , fixed fuzzy kernel variable K 1 , obtain auxiliary variables according to the objective function 1, and solve the potential clear image I through fast Fourier transform 1 ; Step 4: Blurred Image Layer B 1 , fixed latent clear image I 1 , using a fixed latent clear image I 1 The gradient information of the fuzzy kernel variable K is obtained by accelerating the conjugate gradient method. 1 Solve and get the fuzzy kernel K 1 *, K 1 * The blur kernel size is expanded to size n 2 , assigned to K 2 ; Step 5: Blurred image layer B 2 To B H Repeat steps 3 and 4 to obtain the blur kernel K H *, do normalization to get K e , based on the given DR blurred image B and blur kernel K e , use RL deconvolution to restore the final clear image I e ; The specific steps of obtaining the DR blurred image B in step 1 and constructing the image blind deconvolution model include: The mathematical model for constructing the fuzzy model of the image is: In the formula, I and B represent the clear image and blurred image respectively, K and N represent the fuzzy kernel variable and noise respectively. is the convolution operator, Get the minimum value of the image pixel area: i and j represent the row vector and column vector of the image matrix, respectively. R(I) represents the pixel matrix of image I. Ω(I ij ) represents the neighborhood of the pixel, R min (I ij ) represents the minimum value of this pixel area, The L2 norm is used as the prior condition of the DR blurred image gradient, and the total variation regularization is combined to suppress image noise. The image blind deconvolution model is constructed under the maximum a posteriori probability framework: In the formula, represents the result of I,K when the first term is optimally solved, ‖R min (I ij )‖0 counts the number of extreme non-zero values ​​in the pixel area, ‖·‖0 is the L0 norm, ‖I‖ TV represents the second-order total variation, specifically L and W represent the length and width of the image matrix, Z1I and Z2I are the image pixel differences obtained in the horizontal and vertical directions respectively. Characterize the gradient of the image, is the L2 norm, represents the image gradient operator, Respectively represent the vertical and perpendicular gradient operators, α, β, δ, η, and λ are the parameters of each constraint, which are used to maintain a balance between regularization strength and fidelity. Represents the prior condition of the blur kernel.

2. The method for blind deconvolution restoration of DR blurred images according to claim 1, characterized in that: In step 1, the DR blurred image B is input into the constructed image blind deconvolution model, the initial blur kernel and size are set, and the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved are obtained. The steps include: Based on the alternating minimization algorithm of semi-quadratic splitting, the blind deconvolution model is decomposed into two sub-problems each containing only one unknown quantity, and the objective function 1 is established: And establish objective function 2: In objective function one, the blur kernel variable K is fixed, and the clear image I is solved. After the clear image variable I is solved, the clear image I is fixed and then objective function two is optimized. Objective function one and objective function two are optimized alternately until convergence to obtain solutions I and K of the image blind deconvolution model.

3. The method for blind deconvolution restoration of DR blurred images according to claim 1, characterized in that: The step 2 specifically comprises the following steps: Set the number of multi-scale image layers H, and use the Gaussian filter to convolve the image and downsample it at the same time. The rule is: In the formula, G(m,n) represents the Gaussian filter, m and n represent the horizontal and vertical dimensions of the filter, i and j represent the row vector and column vector of the image matrix respectively, l = [1,H], and the blurred image layer from small to large is B 1 , B 2 …B H ; Corresponding input fuzzy kernel variable K 1 , K 2 …K H ; blur kernel size is n 1 、n 2 …n H .

4. The method for blind deconvolution restoration of DR blurred images according to claim 2, characterized in that: The step 3 specifically comprises the following steps: Rewrite the image blind deconvolution model based on auxiliary variables: Among them, f is auxiliary variable one, q is auxiliary variable two, and p is auxiliary variable three, where p = (p h ,p v ) T Approach where p h 、p v To approximate the gradient in the horizontal and vertical directions, δ, γ, and μ are penalty coefficients. Fix the clear image variable I to obtain the formulas for auxiliary variable 1 f, auxiliary variable 2 q, and auxiliary variable 3 p:

5. The method for blind deconvolution restoration of DR blurred images according to claim 4, characterized in that: When solving I in step S3, f, q, and p are fixed, and the closed-form solution of I is obtained through Fourier transform: Where: and denote the forward and inverse Fourier transform, respectively. represents the conjugate Fourier transform, 6. The method for blind deconvolution restoration of DR blurred images according to claim 5, characterized in that: In step 3, the steps of solving the clear image variable I at a certain scale are as follows: 1) Set the maximum number of iterations of the current blurred image layer to N max , initialize the current number of iterations iter = 1, δ = set value, γ = set value, μ = set value; 2) If δ<δ max ,δ max As the upper limit of the penalty coefficient 1, solve the auxiliary variable 1 f, and assign the penalty coefficient 1 δ: δ = 2δ; 3) If γ<γ max ,γ max For the upper limit of the penalty coefficient 2, solve the auxiliary variable q, and assign the penalty coefficient γ: γ = 2γ; 4) If μ<μ max ,μ max For the upper limit of the penalty coefficient three, solve the auxiliary variable two p, and assign the penalty coefficient one μ: μ = 2μ; 5) Let iter = iter + 1, and repeat steps 2) to 4) until the penalty coefficient exceeds the upper limit and the number of iterations iter>N max .

7. The method for blind deconvolution restoration of DR blurred images according to claim 2, characterized in that: In step 4, the following steps are specifically included: For the blur kernel variable K, using the gradient information of the clear image I, the objective function 2 is rewritten as: And further rewritten as: Among them, v is auxiliary variable four, is the penalty factor four; The fuzzy kernel variable K is quickly solved by Fourier transform, and its closed-form solution calculation formula is: Solve for the auxiliary variable four v:

8. A DR blurred image blind deconvolution restoration system, characterized in that: The system comprises an objective function acquisition module, a fuzzy image layer acquisition module, a potential clear image acquisition module, a fuzzy kernel acquisition module, and a clear image acquisition module. The objective function acquisition module is used to acquire the DR blurred image B and construct an image blind deconvolution model; the DR blurred image B is input into the constructed image blind deconvolution model, the initialization blur kernel and size are set, and the objective function 1 about the clear image variable I to be restored and the objective function 2 about the blur kernel variable K to be solved are obtained; The fuzzy image layer acquisition module is used to generate fuzzy image layers B of different scales based on a given DR fuzzy image B and the size n of the initialized fuzzy kernel, using a multi-scale hierarchical strategy. 1 , B 2 …B H , the corresponding input fuzzy kernel variable is K 1 , K 2 …K H , the blur kernel size is n 1 、n 2 …n H , set the maximum number of iterations N for each blurred image layer max ; The potential clear image acquisition module is used to obtain the image based on the blurred image layer B 1 , fixed fuzzy kernel variable K 1 , obtain auxiliary variables according to the objective function 1, and solve the potential clear image I through fast Fourier transform 1 ; The blur kernel acquisition module is used to obtain the blur kernel based on the blurred image layer B. 1 , fixed latent clear image I 1 , using a fixed latent clear image I 1 The gradient information of the fuzzy kernel variable K is obtained by accelerating the conjugate gradient method. 1 Solve and get the fuzzy kernel K 1 *, K 1 * The blur kernel size is expanded to size n 2 , assigned to K 2 ; The clear image acquisition module is used to obtain the blurred image layer B 2 To B H , use the potential clear image acquisition module and the blur kernel acquisition module to repeat the process and obtain the blur kernel K H *, do normalization to get K e , based on the given DR blurred image B and blur kernel K e , use RL deconvolution to restore the final clear image I e ; The specific steps of constructing the image blind deconvolution model by the objective function acquisition module include: The mathematical model for constructing the fuzzy model of the image is: In the formula, I and B represent the clear image and blurred image respectively, K and N represent the fuzzy kernel variable and noise respectively. is the convolution operator, Get the minimum value of the image pixel area: i and j represent the row vector and column vector of the image matrix, respectively. R(I) represents the pixel matrix of image I. Ω(I ij ) represents the neighborhood of the pixel, R min (I ij ) represents the minimum value of this pixel area, The L2 norm is used as the prior condition of the DR blurred image gradient, and the total variation regularization is combined to suppress image noise. The image blind deconvolution model is constructed under the maximum a posteriori probability framework: In the formula, represents the result of I,K when the first term is optimally solved, ‖R min (I ij )‖0 counts the number of extreme non-zero values ​​in the pixel area, ‖·‖0 is the L0 norm, ‖I‖ TV represents the second-order total variation, specifically L and W represent the length and width of the image matrix, Z1I and Z2I are the image pixel differences obtained in the horizontal and vertical directions respectively. Characterize the gradient of the image, is the L2 norm, represents the image gradient operator, Respectively represent the vertical and perpendicular gradient operators, α, β, δ, η, and λ are the parameters of each constraint, which are used to maintain a balance between regularization strength and fidelity. Represents the prior condition of the blur kernel.

Images