WO2019085433A1 - 一种基于全变分的图像复原方法及系统 - Google Patents

一种基于全变分的图像复原方法及系统 Download PDF

Info

Publication number
WO2019085433A1
WO2019085433A1 PCT/CN2018/086406 CN2018086406W WO2019085433A1 WO 2019085433 A1 WO2019085433 A1 WO 2019085433A1 CN 2018086406 W CN2018086406 W CN 2018086406W WO 2019085433 A1 WO2019085433 A1 WO 2019085433A1
Authority
WO
WIPO (PCT)
Prior art keywords
generation
image
kth
formula
parameter
Prior art date
Application number
PCT/CN2018/086406
Other languages
English (en)
French (fr)
Inventor
白海玲
Original Assignee
上海斐讯数据通信技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海斐讯数据通信技术有限公司 filed Critical 上海斐讯数据通信技术有限公司
Publication of WO2019085433A1 publication Critical patent/WO2019085433A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00Image enhancement or restoration
    • G06T5/73Deblurring; Sharpening
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10024Color image

Definitions

  • the present invention relates to the field of image processing, and more particularly to an image restoration method and system based on total variation.
  • image restoration In the actual image acquisition process, digital imaging is often affected by degradation factors such as motion blur, optical blur, random noise, etc., and a blur-degraded image is often obtained.
  • the purpose of image restoration is to restore the original image from the obtained degraded image with maximum fidelity.
  • the energy minimization function for total variation regularization is:
  • the first term is a Total Variation (TV) regular term; the latter term is a fidelity term, which is used to measure the error of the reconstructed image, indicating the degree of fitting of the reconstructed image; ⁇ is a regular parameter. Used to balance the proportion of each item.
  • TV Total Variation
  • plays an important role.
  • the principle of choosing ⁇ is to ensure that the recovery results are in good agreement with the original data, while minimizing the noise and ringing effects.
  • is too large, and the regular term accounts for a large weight of the objective function, which is reflected in the image.
  • the ⁇ is too small, and the regularization will not start.
  • the effect of a priori constraint has little effect on eliminating ill-posed problems, resulting in a certain ringing and noise of the image. Therefore, how to adaptively and quickly estimate ⁇ while recovering a clear image is a problem to be solved.
  • the object of the present invention is to provide a method and system for image restoration based on total variation.
  • the variable separation technique is adopted to overcome the non-differentiability of TV, and combined with Morozov's
  • the deviation criterion can adaptively adjust the regular parameter ⁇ , reduce artificial blindness, improve calculation speed, and recover high quality images.
  • the regular parameter ⁇ is adaptively adjusted by iteration to reduce the artificial blindness. Increase calculation speed while recovering high quality images.
  • step S300 includes: calculating the k+1th generation restored image according to formula (1):
  • h T is the transposed matrix of the fuzzy kernel h
  • ⁇ 1 is the preset first positive parameter
  • ⁇ 2 is the preset second positive parameter
  • ( ⁇ 1 h T - ⁇ 2 ⁇ ) -1 is ( ⁇ Inversion of 1 h T - ⁇ 2 ⁇ )
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • x k is the kth generation
  • y k is the second auxiliary variable of the kth generation
  • is the Laplacian operator
  • div is the scatter operator.
  • a method of calculating f k+1 is introduced; by introducing the first auxiliary variable x and the second auxiliary variable y, only the first auxiliary variable x is related to ⁇ , and f k+1 is independent of ⁇ , thereby Make f k+1 easier to update.
  • step S400 is specifically: calculating an iteration error according to the k+1th generation restored image f k+1 and the kth generation restored image f k ; determining whether the iterative error is less than a preset iteration error threshold If it is, it is considered that the iteration stop condition is satisfied; if not, it is considered that the iteration stop condition is not satisfied.
  • the calculation formula for calculating the iterative error is formula (2), and the iteration error A is calculated:
  • A is the iteration error
  • f k+1 is the k+1th generation restored image
  • f k is the kth generation restored image
  • f k+1 is the k+1th generation restored image
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • g is the obtained blurred image
  • c is the preset third parameter
  • ⁇ 1 is the preset first positive parameter
  • a k+1 is the k+1th generation intermediate variable
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • h is a fuzzy kernel
  • f k+1 is a k+1th generation restored image
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • ⁇ 2 is the preset second positive parameter.
  • y k+1 is the kth +th generation second auxiliary variable.
  • the method further includes: when the formula (5) is established, assigning the k+1th intermediate variable a k+1 to the k+1th generation first auxiliary variable x k+1 , and the k+1th generation regularity
  • the parameter ⁇ k+1 is set to zero.
  • the parameter information updating method of the k+1th generation is given, and further iterative updating of the restored image is guaranteed.
  • the blurred images of the respective channels of the color image are respectively acquired by the three channels of RGB, and processed according to steps S100 to S700, respectively, to obtain reconstructed images corresponding to the three channels of the RBG. Then, the reconstructed images of the three channels are combined to form a reconstructed color image.
  • a method for restoring a color image is given based on a restoration method of a grayscale blurred image.
  • the present invention also provides an image restoration system based on total variation, characterized by comprising: an initialization module, configured to initialize parameter information of the 0th generation restored image f 0 and the 0th generation when the blurred image is obtained;
  • a calculation module electrically connected to the initialization module for calculating a k+1th generation restored image according to parameter information of the kth generation;
  • the calculation module is electrically connected to determine whether an iteration stop condition is satisfied according to the k+1th generation restored image f k+1 and the kth generation restored image f k ; the calculation module is further configured to satisfy The iteration stop bar, then the
  • the calculation module is further configured to calculate parameter information of the k+1th generation if the iterative stop condition is not satisfied;
  • the regular parameter ⁇ is adaptively adjusted by iteration to reduce the artificial blindness. Increase calculation speed while recovering high quality images.
  • calculation module is further configured to calculate the k+1th generation restored image according to formula (1):
  • h T is the transposed matrix of the fuzzy kernel h
  • ⁇ 1 is the preset first positive parameter
  • ⁇ 2 is the preset second positive parameter
  • ( ⁇ 1 h T - ⁇ 2 ⁇ ) -1 is ( ⁇ Inversion of 1 h T - ⁇ 2 ⁇ )
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • x k is the kth generation
  • y k is the second auxiliary variable of the kth generation
  • is the Laplacian operator
  • div is the scatter operator.
  • a calculation method is introduced; by introducing the first auxiliary variable x and the second auxiliary variable y, only the first auxiliary variable x is related and irrelevant, thereby making it easier to update.
  • the determining module is configured to determine, according to the k+1th generation restored image f k+1 and the kth generation restored image f k , whether the iterative stop condition is met, specifically:
  • the calculating module is further configured to calculate an iteration error according to the k+1th generation restored image and the kth generation restored image; and the determining module determines whether the iterative error is less than a preset iteration error threshold; , it is considered that the iteration stop condition is satisfied; if not, it is considered that the iteration stop condition is not satisfied. If not, it is considered that the iteration stop condition is not satisfied.
  • the calculation module calculates a calculation formula of the iterative error according to the k+1th generation restored image and the kth generation restored image as formula (2), and calculates an iteration error A;
  • A is the iteration error
  • f k+1 is the k+1th generation restored image
  • f k is the kth generation restored image
  • the calculation module is configured to calculate a second auxiliary variable of the k+1th generation when the iteration stop condition is not satisfied among them,
  • f k+1 is the k+1th generation restored image
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • g is the obtained blurred image
  • c is the preset third parameter
  • ⁇ 1 is the preset first positive parameter
  • a k+1 is the k+1th generation intermediate variable
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • h is a fuzzy kernel
  • f k+1 is a k+1th generation restored image
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • ⁇ 2 is the preset second positive parameter.
  • y k+1 is the kth +th generation second auxiliary variable.
  • the parameter information updating method of the k+1th generation is given, and further iterative updating of the restored image is guaranteed.
  • the calculation module is configured to assign the k+1th generation intermediate variable to the k+1th generation first auxiliary variable when the formula (5) is established, and set the k+1th generation regular parameter to 0.
  • the invention introduces two auxiliary variables x and y, uses variable separation technology to overcome the non-differentiability of TV, combines Lagrangian theory, adaptively adjusts the regular parameter ⁇ through iteration, reduces artificial blindness, and improves Calculate speed while recovering high quality images.
  • the method for restoring grayscale blurred images according to the present invention is also applicable to multi-channel images, such as restoration of color images.
  • FIG. 1 is a flow chart of an embodiment of an image restoration method based on total variation in the present invention
  • FIG. 2 is a flow chart of another embodiment of a method for image restoration based on total variation according to the present invention.
  • FIG. 3 is a flow chart of another embodiment of a method for image restoration based on total variation according to the present invention.
  • FIG. 4 is a schematic structural view of an embodiment of an image restoration system based on total variation in the present invention.
  • FIG. 5 is a recovery result diagram of an APE-ADMM method and a Wen-Chan method according to an embodiment of the all-variation-based image restoration method of the present invention
  • FIG. 6 is a PSNR-time diagram of an APE-ADMM method and a Wen-Chan method according to an embodiment of the total variation-based image restoration method of the present invention
  • FIG. 7 is a regular parameter-time diagram of the APE-ADMM method and the Wen-Chan method of an embodiment of the all-variation-based image restoration method of the present invention.
  • a method for image restoration based on total variation includes:
  • Step S100 when obtaining a blurred image, initializing parameter information of the 0th generation restored image f 0 and the 0th generation;
  • Step S300 calculates the k+1th generation restored image f k+1 according to the parameter information of the kth generation
  • Step S400 determining, according to the k+1th generation restored image f k+1 and the kth generation restored image f k , whether an iteration stop condition is satisfied;
  • Step S500 if the iteration stop condition is satisfied, the k+1th generation restored image f k+1 is a reconstructed image;
  • step S600 If the step S600 does not satisfy the iterative stop condition, the parameter information of the k+1th generation is calculated;
  • the parameter information of the k+1th generation includes: a kth generation first auxiliary variable x k+1 , a k+1th second auxiliary variable y k+1 , a k+1th generation regular parameter ⁇ k+ 1.
  • the blurred image is a grayscale image.
  • g is the obtained blurred image
  • a method for image restoration based on total variation includes:
  • Step S100 when obtaining a blurred image, initializing parameter information of the 0th generation restored image f 0 and the 0th generation;
  • the parameter information of the 0th generation includes: a 0th generation first auxiliary variable x 0 , a 0th generation second auxiliary variable y 0 , a 0th generation first Lagrangian multiplier u 0 , and a 0th generation second pull Grande multiplier ⁇ 0 ;
  • Step S310 calculates the k+1th generation restored image f k+1 according to formula (1);
  • h T is the transposed matrix of the fuzzy kernel h
  • ⁇ 1 is the preset first positive parameter
  • ⁇ 2 is the preset second positive parameter
  • ( ⁇ 1 h T - ⁇ 2 ⁇ ) -1 is ( ⁇ Inversion of 1 h T - ⁇ 2 ⁇ )
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • x k is the kth generation
  • y k is the second auxiliary variable of the kth generation
  • is the Laplacian operator
  • div is the scatter operator
  • Step S410 according to the k+1th generation restored image f k+1 and the kth generation restored image f k , according to formula (2), the iteration error A is calculated;
  • Step S420 determining whether the iteration error is less than a preset iteration error threshold
  • Step S610 if no, calculate the second auxiliary variable of the k+1th generation among them,
  • Step S620 calculates the k+1th intermediate variable a k+1 according to formula (4);
  • f k+1 is the k+1th generation restored image
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • Step S630 determines whether formula (5) is established
  • Step S640 When the formula (5) is not satisfied, the k+1th regular parameter ⁇ k+1 is calculated according to the formula (6), and the k+1th first auxiliary variable x k+1 is calculated according to the formula (7);
  • g is the obtained blurred image
  • c is the preset third parameter
  • ⁇ 1 is the preset first positive parameter
  • a k+1 is the k+1th generation intermediate variable
  • Step S650 When the formula (5) is established, the k+1th intermediate variable a k+1 is assigned to the k+1th first auxiliary variable x k+1 , and the k+1th regular parameter ⁇ k is obtained. +1 is set to 0;
  • Step S660 calculates a first Lagrangian multiplier u k+1 of the k+ 1th generation according to formula (8);
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • h is a fuzzy kernel
  • f k+1 is a k+1th generation restored image
  • Step S670 calculates a second Lagrangian multiplier ⁇ k+1 of the k+ 1th generation according to formula (9);
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • ⁇ 2 is the preset second positive parameter.
  • y k+1 is the kth +th generation second auxiliary variable
  • the full variation image restoration problem is to find a suitable f that satisfies the following formula:
  • the first item is the TV regular item
  • the latter item is the fidelity item
  • is the regular parameter
  • is the support domain of f.
  • Equation 10 The augmented Lagrangian (AL) functional of equation (10) is defined as:
  • u is the first Lagrangian multiplier
  • u T is the transposed matrix of u
  • is the second Lagrangian multiplier
  • ⁇ T is the transposed matrix of ⁇
  • is y L1 norm
  • ⁇ 1 is a preset first positive parameter
  • ⁇ 2 is a preset second positive parameter
  • equation (12) involves inner iterations for each iteration, the solution (f k+1 , x k+1 , y k+1 ) is cumbersome. Therefore, an improved iterative method is used to solve this problem, that is, a solution can be obtained in one iteration (f k+1 , x k+1 , y k+1 ) in each iteration.
  • the improved iterative method is as follows:
  • the sub-problem of equation (15) is about the minimization problem of x, in which the k+1th intermediate variable a k+1 is calculated first:
  • Equation (17) is a quadratic minimization problem for x, and it has a closed form solution:
  • the iteration error A can be calculated according to formula (2);
  • the preset iterative error threshold is set according to experience, such as 10 -5 .
  • the iteration error is not less than the preset iteration error threshold, it may be further determined whether the number of iterations is greater than or equal to a preset total number of iterations; when the number of iterations reaches a preset total number of iterations, the entire iteration is terminated, and f k+1 is a reconstructed image. .
  • FIG. 5 is an original image of 300 ⁇ 300;
  • (b) of FIG. 5 is a result of a blurred image, a motion blur through the lower left corner (the blur kernel size is 53 ⁇ 53), and a Gaussian random noise with a variance of 1;
  • (c) of FIG. 5 is a result of restoration using the Wen-Chan method;
  • (d) of FIG. 5 is a result of restoration of the method (abbreviated as APE-ADMM); the condition of convergence is 10 -5 , wherein APE- The ADMM iterated 50 times and the Wen-Chan method iterated 69 times.
  • FIG. 6 shows the relationship between the peak signal to noise ratio (PSNR) of the two methods over time. It can be seen that the APE-ADMM method can achieve a higher PSNR at a faster speed;
  • Fig. 7(f) is the relationship of the regular parameters of the APE-ADMM method with time
  • (g) is the relationship of the regular parameters of the Wen-Chan method with time. From the figure, the regular parameter estimated by the APE-ADMM method is approximately It tends to be stable around 1.8s, while the Wen-Chan method is about 4.4s. This method is superior to the Wen-Chan method in terms of speed.
  • Table 1 compares the objective evaluation parameters of the two methods.
  • the objective evaluation parameters include peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). It can be seen from Table 1 that the APE-ADMM algorithm is superior to Wen-Chan. method.
  • PSNR peak signal-to-noise ratio
  • SSIM structural similarity index
  • a method for image restoration based on total variation includes:
  • Step S800 When the captured image is a color image, the blurred images of the respective channels of the color image are respectively acquired by three channels of RGB;
  • Step S810 the blurred images of the three channels are processed according to steps S100 to S700, respectively, to obtain reconstructed images corresponding to the three channels of the RBG;
  • Step S820 combines the reconstructed images of the three channels to form a reconstructed color image.
  • the blurred images of the respective channels of the color image are respectively acquired by three channels of RGB; according to the foregoing embodiment, the blurred images of the three channels are respectively restored and processed. Corresponding reconstructed images; respectively, the reconstructed images of the three channels are combined in RGB to form a reconstructed color image.
  • a full variation-based image restoration system includes:
  • the initialization module 10 is configured to initialize the 0th generation restored image f 0 and the 0th generation parameter information when the blurred image is obtained;
  • the calculation module 20 is electrically connected to the initialization module 10, and is configured to calculate a k+1th generation restored image according to the parameter information of the kth generation;
  • the determining module 30 is electrically connected to the calculating module 20, and is configured to determine, according to the k+1th generation restored image f k+1 and the kth generation restored image f k , whether an iteration stop condition is satisfied;
  • the calculating module 20 is further configured to: if the iterative stop condition is met, the k+1th generation restored image f k+1 is a reconstructed image; and if the iterative stop condition is not satisfied, calculate the k+1th generation
  • the blurred image is a grayscale image.
  • g be the obtained blurred image
  • h is the fuzzy kernel matrix
  • initialize the 0th generation restored image f 0 g, initialize the parameter information of the 0th generation: where the 0th generation first auxiliary variable x 0 is obtained according to h*f 0 , the 0th generation second auxiliary variable y 0 is based on It is calculated that the 0th generation first Lagrangian multiplier u 0 is set by the unit matrix, and the 0th generation second Lagrangian multiplier ⁇ 0 is set by the unit matrix.
  • a full variation-based image restoration system includes:
  • the initialization module 10 is configured to initialize the 0th generation restored image f 0 and the 0th generation parameter information when the blurred image is obtained;
  • the calculation module 20 is electrically connected to the initialization module 10 for calculating the k+1th generation restored image f k+1 according to formula (1):
  • h T is the transposed matrix of the fuzzy kernel h
  • ⁇ 1 is the preset first positive parameter
  • ⁇ 2 is the preset second positive parameter
  • ( ⁇ 1 h T - ⁇ 2 ⁇ ) -1 is ( ⁇ Inversion of 1 h T - ⁇ 2 ⁇ )
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • x k is the kth generation
  • y k is the second auxiliary variable of the kth generation
  • is the Laplacian operator
  • div is the scatter operator
  • the determining module 30 is electrically connected to the calculating module 20, and is configured to calculate an iteration error A according to the formula (2) according to the k+1th generation restored image f k+1 and the kth generation restored image f k ;
  • f k+1 is the k+1th generation restored image
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • g is the obtained blurred image
  • c is the preset third parameter
  • ⁇ 1 is the preset first positive parameter
  • a k+1 is the k+1th generation intermediate variable
  • u k is the first Lagrangian multiplier of the kth generation
  • ⁇ 1 is a preset first positive parameter
  • h is a fuzzy kernel
  • f k+1 is a k+1th generation restored image
  • ⁇ k is the second Lagrangian multiplier of the kth generation
  • ⁇ 2 is the preset second positive parameter.
  • y k+1 is the kth +th generation second auxiliary variable
  • the calculating module is further configured to assign the k+1th intermediate variable a k+1 to the k+1th first auxiliary variable x k+1 when the formula (5) is established, and the k+th
  • the regular parameter ⁇ k+1 of the first generation is set to zero.
  • the full variation image restoration problem is to find a suitable f that satisfies the following formula:
  • the first item is the TV regular item
  • the latter item is the fidelity item
  • is the regular parameter
  • is the support domain of f.
  • Equation 10 The augmented Lagrangian (AL) functional of equation (10) is defined as:
  • u is the first Lagrangian multiplier
  • u T is the transposed matrix of u
  • is the second Lagrangian multiplier
  • ⁇ T is the transposed matrix of ⁇
  • is y L1 norm
  • ⁇ 1 is a preset first positive parameter
  • ⁇ 2 is a preset second positive parameter
  • equation (12) involves inner iterations for each iteration, the solution (f k+1 , x k+1 , y k+1 ) is cumbersome. Therefore, an improved iterative method is used to solve this problem, that is, a solution can be obtained in one iteration (f k+1 , x k+1 , y k+1 ) in each iteration.
  • the improved iterative method is as follows:
  • the sub-problem of equation (15) is about the minimization problem of x, in which the k+1th intermediate variable a k+1 is calculated first:
  • Equation (17) is a quadratic minimization problem for x, and it has a closed form solution:
  • the iteration error A can be calculated according to formula (2);
  • the preset iterative error threshold is set according to experience, such as 10 -5 .
  • the iteration error is not less than the preset iteration error threshold, it may be further determined whether the number of iterations is greater than or equal to a preset total number of iterations; when the number of iterations reaches a preset total number of iterations, the entire iteration is terminated, and f k+1 is a reconstructed image. .

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Image Processing (AREA)

Abstract

本发明公开了一种基于全变分的图像复原方法及系统,包括:步骤S100当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;步骤S200设定迭代次数k=0;步骤S300根据第k代的参数信息,计算第k+1代复原图像f k+1;步骤S400根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,判断是否满足迭代停止条件;步骤S500若是,则所述第k+1代复原图像f k+1为重构图像;步骤S600若否,则计算第k+1代的参数信息;步骤S700更新迭代次数k=k+1,并跳转到步骤S300。本发明可以在已知模糊核的情况下,能够自适应地调节正则参数λ,减少人为盲目性,提高计算速度,同时恢复出高质量的图像。

Description

一种基于全变分的图像复原方法及系统
本申请要求2017年11月06日提交的申请号为:201711077178.6、发明名称为“一种基于全变分的图像复原方法及系统”的中国专利申请的优先权,其全部内容合并在此。
技术领域
本发明涉及图像处理领域,尤指一种基于全变分的图像复原方法及系统。
背景技术
在实际的图像采集过程中,数字成像由于受运动模糊、光学模糊、随机噪声等退化因素的影响,往往获得的是一幅模糊退化的图像。图像复原的目的,就是从获得的退化图像以最大的保真度恢复原始图像。
图像复原的线性退化模型为:g=hf+n;其中,g为获得的模糊图像,f为原始图像,h为模糊核,n为随机噪声。
由于图像复原问题是一个病态求逆过程,将会导致无解或解的不唯一性。为此,较多技术采用全变分复原方法,即采用正则化方法来约束,使图像复原问题变为良态问题。
全变分正则化的能量最小化函数为:
Figure PCTCN2018086406-appb-000001
其中,第一项为全变分(Total Variation,简称TV)正则项;后一项为保真项,用来衡量重构图像的误差,表示重构图像的拟合程度;λ为正则参数,用来平衡各项的比重。
对众多基于全变分模型衍生版本的研究,发现正则参数λ,起着举足轻重的作用。选择λ的原则是既要保证复原结果与原始数据有较好的吻合,同时要尽可能地减少噪声和振铃效应。假设以图像的平滑性作为先验约束条件,λ选取的过大,正则项占目标函数的权重较大,反映在图像上会出现过平滑的现象;λ选取的过小,正则化将起不到先验约束的作用, 对消除不适定问题作用不大,从而造成图像有一定的振铃和噪声。因此,如何能够自适应地、快速地估计出λ的同时复原出清晰的图像是一个有待于解决的问题。
发明内容
本发明的目的是提供一种基于全变分的图像复原方法及系统,在已知模糊核的情况下,通过引入两个辅助变量,采用变量分离技术,克服TV的不可微性,同时结合Morozov’s偏差准则,能够自适应地调节正则参数λ,减少人为盲目性,提高计算速度,同时恢复出高质量的图像。
本发明提供的技术方案如下:
一种基于全变分的图像复原方法,包括:步骤S100当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;所述第0代的参数信息包括:第0代第一辅助变量x 0、第0代第二辅助变量y 0、第0代第一拉格朗日乘子u 0、第0代第二拉格朗日乘子ξ 0;步骤S200设定迭代次数k=0;步骤S300根据第k代的参数信息,计算第k+1代复原图像f k+1;步骤S400根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,判断是否满足迭代停止条件;步骤S500若满足迭代停止条件,则所述第k+1代复原图像f k+1为重构图像。
进一步,步骤S400之后还包括:步骤S600若不满足迭代停止条件,则计算第k+1代的参数信息;所述第k+1代的参数信息包括:第k+1代第一辅助变量x k+1、第k+1代第二辅助变量y k+1、第k+1代正则参数λ k+1、第k+1代第一拉格朗日乘子u k+1、第k+1代第二拉格朗日乘子ξ k+1;步骤S700更新迭代次数k=k+1,并跳转到步骤S300。
在上述技术方案中,引入两个辅助变量x和y,采用变量分离技术,克服TV的不可微性,结合拉格朗日理论,通过迭代,自适应地调节正则参数λ,减少人为盲目性,提高计算速度,同时恢复出高质量的图像。
进一步,所述步骤S300包括:根据公式(1)计算所述第k+1代复原图像:
f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………(1)
其中,h T为模糊核h的转置矩阵,β 1为预设的第一正参数,β 2为预设的第二正参数,(β 1h T2Δ) -1为(β 1h T2Δ)的求逆,u k是第k代的第一拉格朗日乘子,ξ k是第k代的第二拉格朗日乘子,x k为第k代的第一辅助变量,y k为第k代的第二辅助变量,Δ是拉普拉斯算符,div是散度算子。
在上述技术方案中,引入了计算f k+1的方法;通过引入第一辅助变量x和第二辅助变量y,只有第一辅助变量x与λ有关,而f k+1与λ无关,从而让f k+1更容易更新。
进一步,所述步骤S400具体为:根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,计算迭代误差;判断所述迭代误差是否小于预设迭代误差阈值;若是,则认为满足迭代停止条件;若否,则认为不满足迭代停止条件。
进一步,所述根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差的计算公式为公式(2),计算迭代误差A:
A=||f k+1-f k|| 2/||f k|| 2……………………(2);
其中,A为迭代误差,f k+1为第k+1代复原图像,f k为第k代复原图像。
在上述技术方案中,给出了迭代停止条件,当迭代停止条件得到满足时,第k+1代复原图像即为重构图像。
进一步,所述步骤S600包括:当不满足迭代停止条件时,计算第k+1代第二辅助变量
Figure PCTCN2018086406-appb-000002
其中,
Figure PCTCN2018086406-appb-000003
为y k+1的第i行、第j列的元素,i=1,2,...,M,j=1,2,...,N;
根据公式(3),计算
Figure PCTCN2018086406-appb-000004
Figure PCTCN2018086406-appb-000005
其中,
Figure PCTCN2018086406-appb-000006
是梯度算子,
Figure PCTCN2018086406-appb-000007
为第k+1代原始图像的梯度的第i行、 第j列的元素,
Figure PCTCN2018086406-appb-000008
是第k代的第二拉格朗日乘子的第i行、第j列的元素,β 2为预设的第二正参数;
根据公式(4),计算第k+1代中间变量a k+1
a k+1=hf k+1+(u k1)……………(4)
其中,f k+1为第k+1代复原图像,u k是第k代的第一拉格朗日乘子,β 1为预设的第一正参数;
判断公式(5)是否成立;
Figure PCTCN2018086406-appb-000009
其中,g为获得的模糊图像,c为预设的第三参数;
当公式(5)不成立时,根据公式(6)计算第k+1代正则参数λ k+1,及根据公式(7)计算第k+1代第一辅助变量x k+1
Figure PCTCN2018086406-appb-000010
x k+1=(λ k+1g+β 1a k+1)/(λ k+11)……………(7)
其中,g为获得的模糊图像,c为预设的第三参数,β 1为预设的第一正参数,a k+1为第k+1代中间变量;
根据公式(8)计算第k+1代的第一拉格朗日乘子u k+1
u k+1=u k1(x k+1-hf k+1)……………(8)
其中,u k为第k代的第一拉格朗日乘子,β 1为预设的第一正参数,h为模糊核,f k+1为第k+1代复原图像;
根据公式(9)计算第k+1代的第二拉格朗日乘子ξ k+1
Figure PCTCN2018086406-appb-000011
其中,ξ k为第k代的第二拉格朗日乘子,β 2为预设的第二正参数,
Figure PCTCN2018086406-appb-000012
为第k+1代原始图像的梯度,y k+1为第k+1代第二辅助变量。
进一步,还包括:当公式(5)成立时,则将第k+1代中间变量a k+1赋值给第k+1代第一辅助变量x k+1,且将第k+1代正则参数λ k+1设为0。
在上述技术方案中,给出了第k+1代的参数信息更新方法,保证了复原图像的进一步迭代更新。
进一步,当拍摄的图像为彩色图像时,按RGB三个通道分别获取所 述彩色图像的各个通道的模糊图像,并分别按照步骤S100至步骤S700进行处理,得到RBG三个通道对应的重构图像,再将三个通道各自的重构图像组合形成重构的彩色图像。
在上述技术方案中,基于灰度模糊图像的复原方法,给出了彩色图像的复原方法。
本发明还提供一种基于全变分的图像复原系统,其特征在于,包括:初始化模块,用于当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;所述第0代的参数信息包括:第0代第一辅助变量x 0、第0代第二辅助变量y 0、第0代第一拉格朗日乘子u 0、第0代第二拉格朗日乘子ξ 0;以及设定迭代次数k=0;计算模块,与所述初始化模块电连接,用于根据第k代的参数信息,计算第k+1代复原图像;判断模块,与所述计算模块电连接,用于根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,判断是否满足迭代停止条件;所述计算模块,进一步用于若满足迭代停止条,则所述第k+1代复原图像f k+1为重构图像。
进一步,计算模块,进一步用于若不满足迭代停止条件则计算第k+1代的参数信息;所述第k+1代的参数信息包括:第k+1代第一辅助变量x k+1、第k+1代第二辅助变量y k+1、第k+1代正则参数λ k+1、第k+1代第一拉格朗日乘子u k+1、第k+1代第二拉格朗日乘子ξ k+1;以及更新迭代次数k=k+1,并重新计算第k+1代复原图像。
在上述技术方案中,引入两个辅助变量x和y,采用变量分离技术,克服TV的不可微性,结合拉格朗日理论,通过迭代,自适应地调节正则参数λ,减少人为盲目性,提高计算速度,同时恢复出高质量的图像。
进一步,所述计算模块进一步用于根据公式(1)计算所述第k+1代复原图像:
f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………(1)
其中,h T为模糊核h的转置矩阵,β 1为预设的第一正参数,β 2为预设的第二正参数,(β 1h T2Δ) -1为(β 1h T2Δ)的求逆,u k是第k代的第一拉格朗日乘子,ξ k是第k代的第二拉格朗日乘子,x k为第k代的第一辅助变量, y k为第k代的第二辅助变量,Δ是拉普拉斯算符,div是散度算子。
在上述技术方案中,引入了计算的方法;通过引入第一辅助变量x和第二辅助变量y,只有第一辅助变量x与有关,而与无关,从而让更容易更新。
进一步,所述判断模块,用于根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,判断是否满足迭代停止条件具体为:
所述计算模块,进一步用于根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差;所述判断模块,判断所述迭代误差是否小于预设迭代误差阈值;若是,则认为满足迭代停止条件;若否,则认为不满足迭代停止条件。若否,则认为不满足迭代停止条件。
进一步,所述计算模块根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差的计算公式为公式(2),计算迭代误差A;
A=||f k+1-f k|| 2/||f k|| 2……………………(2);
其中,A为迭代误差,f k+1为第k+1代复原图像,f k为第k代复原图像。
在上述技术方案中,给出了迭代停止条件,当迭代停止条件得到满足时,第k+1代复原图像即为重构图像。
进一步,所述计算模块,用于当不满足迭代停止条件时,计算第k+1代第二辅助变量
Figure PCTCN2018086406-appb-000013
其中,
Figure PCTCN2018086406-appb-000014
为y k+1的第i行、第j列的元素,i=1,2,...,M,j=1,2,...,N;
以及,根据公式(3),计算
Figure PCTCN2018086406-appb-000015
Figure PCTCN2018086406-appb-000016
其中,
Figure PCTCN2018086406-appb-000017
是梯度算子,
Figure PCTCN2018086406-appb-000018
为第k+1代原始图像的梯度的第i行、第j列的元素,
Figure PCTCN2018086406-appb-000019
是第k代的第二拉格朗日乘子的第i行、第j列的元素,β 2为预设的第二正参数;
以及,根据公式(4),计算第k+1代中间变量a k+1
a k+1=hf k+1+(u k1)……………(4)
其中,f k+1为第k+1代复原图像,u k是第k代的第一拉格朗日乘子,β 1为预设的第一正参数;
以及,判断公式(5)是否成立;
Figure PCTCN2018086406-appb-000020
其中,g为获得的模糊图像,c为预设的第三参数;
以及,当公式(5)不成立时,根据公式(6)计算第k+1代正则参数λ k+1,及根据公式(7)计算第k+1代第一辅助变量x k+1
Figure PCTCN2018086406-appb-000021
x k+1=(λ k+1g+β 1a k+1)/(λ k+11)……………(7)
其中,g为获得的模糊图像,c为预设的第三参数,β 1为预设的第一正参数,a k+1为第k+1代中间变量;
以及,根据公式(8)计算第k+1代的第一拉格朗日乘子u k+1
u k+1=u k1(x k+1-hf k+1)……………(8)
其中,u k为第k代的第一拉格朗日乘子,β 1为预设的第一正参数,h为模糊核,f k+1为第k+1代复原图像;
以及,根据公式(9)计算第k+1代的第二拉格朗日乘子ξ k+1
Figure PCTCN2018086406-appb-000022
其中,ξ k为第k代的第二拉格朗日乘子,β 2为预设的第二正参数,
Figure PCTCN2018086406-appb-000023
为第k+1代原始图像的梯度,y k+1为第k+1代第二辅助变量。
在上述技术方案中,给出了第k+1代的参数信息更新方法,保证了复原图像的进一步迭代更新。
进一步,所述计算模块,用于当公式(5)成立时,则将第k+1代中间变量赋值给第k+1代第一辅助变量,且将第k+1代的正则参数设为0。
通过本发明提供的一种基于全变分的图像复原方法及系统,能够带来以下至少一种有益效果:
1、本发明通过引入两个辅助变量x和y,采用变量分离技术,克服 TV的不可微性,结合拉格朗日理论,通过迭代,自适应地调节正则参数λ,减少人为盲目性,提高计算速度,同时恢复出高质量的图像。
2、本发明基于灰度模糊图像的复原方法,同样适用于多通道图像,如彩色图像的复原。
附图说明
下面将以明确易懂的方式,结合附图说明优选实施方式,对一种基于全变分的图像复原方法及系统的上述特性、技术特征、优点及其实现方式予以进一步说明。
图1是本发明基于全变分的图像复原方法一个实施例的流程图;
图2是本发明基于全变分的图像复原方法另一个实施例的流程图;
图3是本发明基于全变分的图像复原方法另一个实施例的流程图;
图4是本发明基于全变分的图像复原系统一个实施例的结构示意图;
图5是本发明基于全变分的图像复原方法一个实施例的APE-ADMM方法与Wen-Chan方法的复原结果图;
图6是本发明基于全变分的图像复原方法一个实施例的APE-ADMM方法与Wen-Chan方法的PSNR-时间图;
图7是本发明基于全变分的图像复原方法一个实施例的APE-ADMM方法与Wen-Chan方法的正则参数-时间图。
附图标号说明:
10.初始化模块,20.计算模块,30.判断模块。
具体实施方式
为了更清楚地说明本发明实施例或现有技术中的技术方案,下面将对照附图说明本发明的具体实施方式。显而易见地,下面描述中的附图仅仅是本发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图,并获得其他的实施方式。
为使图面简洁,各图中只示意性地表示出了与本发明相关的部分,它 们并不代表其作为产品的实际结构。另外,以使图面简洁便于理解,在有些图中具有相同结构或功能的部件,仅示意性地绘示了其中的一个,或仅标出了其中的一个。在本文中,“一个”不仅表示“仅此一个”,也可以表示“多于一个”的情形。
在本发明的一个实施例中,如图1所示,一种基于全变分的图像复原方法,包括:
步骤S100当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;
所述第0代的参数信息包括:第0代第一辅助变量x 0、第0代第二辅助变量y 0、第0代第一拉格朗日乘子u 0、第0代第二拉格朗日乘子ξ 0
步骤S200设定迭代次数k=0;
步骤S300根据第k代的参数信息,计算第k+1代复原图像f k+1
步骤S400根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,判断是否满足迭代停止条件;
步骤S500若满足迭代停止条件,则所述第k+1代复原图像f k+1为重构图像;
步骤S600若不满足迭代停止条件,则计算第k+1代的参数信息;
所述第k+1代的参数信息包括:第k+1代第一辅助变量x k+1、第k+1代第二辅助变量y k+1、第k+1代正则参数λ k+1、第k+1代第一拉格朗日乘子u k+1、第k+1代第二拉格朗日乘子ξ k+1
步骤S700更新迭代次数k=k+1,并跳转到步骤S300。
具体的,模糊图像为灰度图像。假设g为所获得的模糊图像,h为模糊核矩阵;初始化第0代复原图像f 0=g,初始化第0代的参数信息:比如,第0代第一辅助变量x 0根据h*f 0得到,第0代第二辅助变量y 0根据
Figure PCTCN2018086406-appb-000024
计算得到,第0代第一拉格朗日乘子u 0按单位矩阵设置,第0代第二拉格朗日乘子ξ 0按单位矩阵设置。
设定迭代次数k=0;f 0、x 0、y 0、u 0、ξ 0,即为第0代的参数信息;根据第0代的参数信息,计算第1代复原图像f 1,根据f 1和f 0,判断是否 满足迭代停止条件;若是,则f 1为重构图像,整个迭代结束;若否,则计算第1代的参数信息:第1代第一辅助变量x 1、第1代第二辅助变量y 1、第1代正则参数λ 1、第1代第一拉格朗日乘子u 1、第1代第二拉格朗日乘子ξ 1;更新迭代次数k=1,计算第2代复原图像f 2,根据f 2和f 1,判断是否满足迭代停止条件;若是,则f 2为重构图像,整个迭代结束;若否,则计算第2代的参数信息;如此迭代循环,直至满足迭代停止条件,所得到的第k+1代复原图像f k+1即为重构图像。
在本发明的另一个实施例中,如图2所示,一种基于全变分的图像复原方法,包括:
步骤S100当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;
所述第0代的参数信息包括:第0代第一辅助变量x 0、第0代第二辅助变量y 0、第0代第一拉格朗日乘子u 0、第0代第二拉格朗日乘子ξ 0
步骤S200设定迭代次数k=0;
步骤S310根据公式(1)计算所述第k+1代复原图像f k+1
f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………(1)
其中,h T为模糊核h的转置矩阵,β 1为预设的第一正参数,β 2为预设的第二正参数,(β 1h T2Δ) -1为(β 1h T2Δ)的求逆,u k是第k代的第一拉格朗日乘子,ξ k是第k代的第二拉格朗日乘子,x k为第k代的第一辅助变量,y k为第k代的第二辅助变量,Δ是拉普拉斯算符,div是散度算子;
步骤S410根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,根据公式(2),计算迭代误差A;
A=||f k+1-f k|| 2/||f k|| 2……………………(2)
步骤S420判断所述迭代误差是否小于预设迭代误差阈值;
步骤S500若是,则所述第k+1代复原图像f k+1为重构图像;
步骤S610若否,则计算第k+1代第二辅助变量
Figure PCTCN2018086406-appb-000025
其中,
Figure PCTCN2018086406-appb-000026
为y k+1的第i行、第j列的元素,i=1,2,...,M,j=1,2,...,N;
根据公式(3),计算
Figure PCTCN2018086406-appb-000027
Figure PCTCN2018086406-appb-000028
其中,
Figure PCTCN2018086406-appb-000029
是梯度算子,
Figure PCTCN2018086406-appb-000030
为第k+1代原始图像的梯度的第i行、第j列的元素,
Figure PCTCN2018086406-appb-000031
是第k代的第二拉格朗日乘子的第i行、第j列的元素,β 2为预设的第二正参数;
步骤S620根据公式(4),计算第k+1代中间变量a k+1
a k+1=hf k+1+(u k1)……………(4)
其中,f k+1为第k+1代复原图像,u k是第k代的第一拉格朗日乘子,β 1为预设的第一正参数;
步骤S630判断公式(5)是否成立;
Figure PCTCN2018086406-appb-000032
其中,g为获得的模糊图像,c为预设的第三参数;
步骤S640当公式(5)不成立时,根据公式(6)计算第k+1代正则参数λ k+1,及根据公式(7)计算第k+1代第一辅助变量x k+1
Figure PCTCN2018086406-appb-000033
x k+1=(λ k+1g+β 1a k+1)/(λ k+11)……………(7)
其中,g为获得的模糊图像,c为预设的第三参数,β 1为预设的第一正参数,a k+1为第k+1代中间变量;
步骤S650当公式(5)成立时,则将第k+1代中间变量a k+1赋值给第k+1代第一辅助变量x k+1,且将第k+1代正则参数λ k+1设为0;
步骤S660根据公式(8)计算第k+1代的第一拉格朗日乘子u k+1
u k+1=u k1(x k+1-hf k+1)……………(8)
其中,u k为第k代的第一拉格朗日乘子,β 1为预设的第一正参数,h为模糊核,f k+1为第k+1代复原图像;
步骤S670根据公式(9)计算第k+1代的第二拉格朗日乘子ξ k+1
Figure PCTCN2018086406-appb-000034
其中,ξ k为第k代的第二拉格朗日乘子,β 2为预设的第二正参数,
Figure PCTCN2018086406-appb-000035
为第k+1代原始图像的梯度,y k+1为第k+1代第二辅助变量;
步骤S710更新迭代次数k=k+1,并跳转到步骤S310。
具体的,全变分图像复原问题是要找到合适的f,满足下式:
Figure PCTCN2018086406-appb-000036
其中,第一项为TV正则项,后一项为保真项,λ为正则参数。
图像f的全变分计算公式为:
Figure PCTCN2018086406-appb-000037
其中,Ω是f的支持域。
根据Morozov’s偏差准则,上述问题等价于下面的约束最优问题:
Figure PCTCN2018086406-appb-000038
Figure PCTCN2018086406-appb-000039
其中,c为预设的第三参数,根据c=τmnσ 2计算,σ 2为噪声方差,τ是一个依赖于噪声的参数,比如,设为1;m、n为原始图像的维度。
采用变量分离方法,引入两个辅助变量,分别是第二辅助变量y,y∈Q,等于
Figure PCTCN2018086406-appb-000040
和第一辅助变量x,x∈V,等于hf,其中,V表示欧式空间R mm,Q=V×V,从而得到等价约束形式为:
Figure PCTCN2018086406-appb-000041
公式(10)的增广拉格朗日(Augmented Lagrangian,简称AL)泛函定义为:
Figure PCTCN2018086406-appb-000042
其中,u是第一拉格朗日乘子,u T是u的转置矩阵,ξ是第二拉格朗日乘子,ξ T是ξ的转置矩阵,||y||是y的L1范数,β 1为预设的第一正参数,β 2为预设的第二正参数;
将公式(10)的最小化问题转为求解公式(11)增广拉格朗日泛函的鞍点,具体迭代框架如下所示:
Figure PCTCN2018086406-appb-000043
由于公式(12)每次迭代都涉及到内迭代,所以解(f k+1,x k+1,y k+1)是比较繁琐的。因此采用改进的迭代方法来解决这个问题,即在每次迭代中只需一步运算就可以求得解(f k+1,x k+1,y k+1)。
改进的迭代方法如下所示:
Figure PCTCN2018086406-appb-000044
Figure PCTCN2018086406-appb-000045
Figure PCTCN2018086406-appb-000046
μ k+1=μ k1(x k+1-hf k+1)………………(8)
Figure PCTCN2018086406-appb-000047
公式(13)子问题,是关于f的最小化问题,是一个二次方程,形式如(16)所示;根据公式(16)得到公式(1):
Figure PCTCN2018086406-appb-000048
f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………………(1)
其中,
Figure PCTCN2018086406-appb-000049
是拉普拉斯算符;对于一个m×n大小的图像,经过两次FFT和一次逆FFT,即可求得式(1)的解,其算法复杂度为O(mn log(mn))。
公式(14)子问题,是关于y的最小化问题:
Figure PCTCN2018086406-appb-000050
此式是近端最小化(Proximal Minimization,PM)问题,利用二维收缩求得其解为:
Figure PCTCN2018086406-appb-000051
公式(15)子问题,是关于x的最小化问题,其中,先计算第k+1代中间变量a k+1
a k+1=hf k+1+(u k1)…………………(4)
通过求解公式(15)得公式(17):
Figure PCTCN2018086406-appb-000052
公式(17)是一个关于x的二次最小化问题,并且它有一个闭合形式的解:
x k+1=(λ k+1g+β 1a k+1)/(λ k+11)…………………(7)
依据a k+1的范围,每次迭代中λ的解存在两种情况:一种情况是
Figure PCTCN2018086406-appb-000053
此时λ k+1=0同时x k+1=a k+1,很明显x k+1满足Morozov’s偏差准则;另一种情况是
Figure PCTCN2018086406-appb-000054
根据Morozov’s偏差准则,应求解如下的方程:
Figure PCTCN2018086406-appb-000055
将公式(7)中的x k+1代入公式(18),得到:
Figure PCTCN2018086406-appb-000056
如此,通过引入辅助变量x,hf从严格的Morozov’s偏差准则中释放了出来,得益于此,在每次迭代中无需附加条件即可获得一个闭合形式的λ解。
在根据公式(1)计算出第k+1代复原图像f k+1后,可以根据公式(2),计算迭代误差A;
A=||f k+1-f k|| 2/||f k|| 2……………………(2)
判断是否需要继续迭代;
当迭代误差小于预设迭代误差阈值时,则第k+1代复原图像f k+1为重构图像,整个迭代结束;预设迭代误差阈值,根据经验设置,比如10 -5
当迭代误差不小于预设迭代误差阈值时,可以进一步判断迭代次数是否大于等于预设总迭代次数;当迭代次数达到预设总迭代次数时,则终止 整个迭代,f k+1为重构图像。
当迭代误差不小于预设迭代误差阈值,且迭代次数未达到预设总迭代次数时,需要更新迭代次数,进一步迭代。
以上是针对已知模糊核情况下的处理,对于真实拍摄的模糊图像,需要利用目前现有的方法如刃边法,先进行模糊核的估计。
对本方法与Wen-Chan方法(基于原始-对偶问题的TV模型,首先求解鞍点,然后利用原始-对偶临近点算法求解最优解)进行了测试比较:
图5的(a)是300×300的原始图像;图5的(b)是模糊图像,经过左下角的运动模糊(模糊核大小为53×53)及方差为1的高斯随机噪声的结果;图5的(c)是利用Wen-Chan方法复原的结果;图5的(d)是本文方法(简称为APE-ADMM)复原的结果;两者收敛的条件是10 -5,其中,APE-ADMM迭代了50次,Wen-Chan方法迭代了69次。
从左下角的细节放大图可以明显地看出图5的(d)图细节更清楚,同原图比较接近,而图5的(c)图略显平滑。
图6是两种方法的峰值信噪比PSNR(Peak Signal to Noise Ratio)随时间的变化关系,可见APE-ADMM方法可以以较快的速度达到较高的PSNR;
图7的(f)是APE-ADMM方法的正则参数随时间的变化关系,(g)是Wen-Chan方法的正则参数随时间的变化关系,从图可知,APE-ADMM方法估计的正则参数大约在1.8s附近趋于稳定,而Wen-Chan方法差不多在4.4s左右,可见本方法在速度方面优于Wen-Chan方法。
表1是两种方法的客观评价参数的对比,客观评价参数包括峰值信噪比PSNR、结构相似性SSIM(structural similarity index measurement);从表1可以看出,APE-ADMM算法优于Wen-Chan方法。
表1
Figure PCTCN2018086406-appb-000057
Figure PCTCN2018086406-appb-000058
在本发明的另一个实施例中,如图3所示,一种基于全变分的图像复原方法,除与上述相同的之外,还包括:
步骤S800当拍摄的图像为彩色图像时,按RGB三个通道分别获取所述彩色图像的各个通道的模糊图像;
步骤S810将所述三个通道的模糊图像分别按照步骤S100至步骤S700进行处理,得到RBG三个通道对应的重构图像;
步骤S820将所述三个通道各自的重构图像组合形成重构的彩色图像。
具体的,当拍摄的图像为彩色图像时,按RGB三个通道分别获取所述彩色图像的各个通道的模糊图像;按照前述实施例,分别对所述三个通道的模糊图像进行复原处理,得到各自对应的重构图像;再将三个通道各自的重构图像按照RGB组合形成重构的彩色图像。
在本发明的另一个实施例中,如图4所示,一种基于全变分的图像复原系统,包括:
初始化模块10,用于当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;所述第0代的参数信息包括:第0代第一辅助变量x 0、第0代第二辅助变量y 0、第0代第一拉格朗日乘子u 0、第0代第二拉格朗日乘子ξ 0;以及设定迭代次数k=0;
计算模块20,与所述初始化模块10电连接,用于根据第k代的参数信息,计算第k+1代复原图像;
判断模块30,与所述计算模块20电连接,用于根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,判断是否满足迭代停止条件;
所述计算模块20,进一步用于若满足迭代停止条件,则所述第k+1代复原图像f k+1为重构图像;以及,若不满足迭代停止条件,则计算第k+1代的参数信息;所述第k+1代的参数信息包括:第k+1代第一辅助变量x k+1、第k+1代第二辅助变量y k+1、第k+1代正则参数λ k+1、第k+1代第一拉格朗日乘子u k+1、第k+1代第二拉格朗日乘子ξ k+1;以及更新迭代次数k=k+1,并重新计算第k+1代复原图像。
具体的,模糊图像为灰度图像。假设g为所获得的模糊图像,h为模糊 核矩阵;初始化第0代复原图像f 0=g,初始化第0代的参数信息:其中第0代第一辅助变量x 0根据h*f 0得到,第0代第二辅助变量y 0根据
Figure PCTCN2018086406-appb-000059
计算得到,第0代第一拉格朗日乘子u 0按单位矩阵设置,第0代第二拉格朗日乘子ξ 0按单位矩阵设置。
设定迭代次数k=0;f 0、x 0、y 0、u 0、ξ 0,即为第0代的参数信息;根据第0代的参数信息,计算第1代复原图像f 1,根据f 1和f 0,判断是否满足迭代停止条件;若是,则f 1为重构图像,整个迭代结束;若否,则计算第1代的参数信息:第1代第一辅助变量x 1、第1代第二辅助变量y 1、第1代正则参数λ 1、第1代第一拉格朗日乘子u 1、第1代第二拉格朗日乘子ξ 1;更新迭代次数k=1,计算第2代复原图像f 2,根据f 2和f 1,判断是否满足迭代停止条件;若是,则f 2为重构图像,整个迭代结束;若否,则计算第2代的参数信息;如此迭代循环,直至满足迭代停止条件,所得到的第k+1代复原图像f k+1即为重构图像。
在本发明的另一个实施例中,如图4所示,一种基于全变分的图像复原系统,包括:
初始化模块10,用于当获得模糊图像时,初始化第0代复原图像f 0、第0代的参数信息;所述第0代的参数信息包括:第0代第一辅助变量x 0、第0代第二辅助变量y 0、第0代第一拉格朗日乘子u 0、第0代第二拉格朗日乘子ξ 0;以及设定迭代次数k=0;以及设定迭代次数k=0;
计算模块20,与所述初始化模块10电连接,用于根据公式(1)计算所述第k+1代复原图像f k+1
f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………(1)
其中,h T为模糊核h的转置矩阵,β 1为预设的第一正参数,β 2为预设的第二正参数,(β 1h T2Δ) -1为(β 1h T2Δ)的求逆,u k是第k代的第一拉格朗日乘子,ξ k是第k代的第二拉格朗日乘子,x k为第k代的第一辅助变量,y k为第k代的第二辅助变量,Δ是拉普拉斯算符,div是散度算子;
判断模块30,与所述计算模块20电连接,用于根据所述第k+1代复原图像f k+1和所述第k代复原图像f k,根据公式(2),计算迭代误差A;
A=||f k+1-f k|| 2/||f k|| 2……………………(2)
以及,判断所述迭代误差是否小于预设迭代误差阈值;若是,则认为满足迭代停止条件;若否,则认为不满足迭代停止条件;
所述计算模块,进一步用于当满足迭代停止条件时,则所述第k+1代复原图像f k+1为重构图像;以及,当不满足迭代停止条件时,则计算第k+1代第二辅助变量
Figure PCTCN2018086406-appb-000060
其中,
Figure PCTCN2018086406-appb-000061
为y k+1的第i行、第j列的元素,i=1,2,...,M,j=1,2,...,N;
根据公式(3),计算
Figure PCTCN2018086406-appb-000062
Figure PCTCN2018086406-appb-000063
其中,
Figure PCTCN2018086406-appb-000064
是梯度算子,
Figure PCTCN2018086406-appb-000065
为第k+1代原始图像的梯度的第i行、第j列的元素,
Figure PCTCN2018086406-appb-000066
是第k代的第二拉格朗日乘子的第i行、第j列的元素,β 2为预设的第二正参数;
以及,根据公式(4),计算第k+1代中间变量a k+1
a k+1=hf k+1+(u k1)……………(4)
其中,f k+1为第k+1代复原图像,u k是第k代的第一拉格朗日乘子,β 1为预设的第一正参数;
以及,判断公式(5)是否成立;
Figure PCTCN2018086406-appb-000067
其中,g为获得的模糊图像,c为预设的第三参数;
以及,当公式(5)不成立时,根据公式(6)计算第k+1代正则参数λ k+1,及根据公式(7)计算第k+1代第一辅助变量x k+1
Figure PCTCN2018086406-appb-000068
x k+1=(λ k+1g+β 1a k+1)/(λ k+11)……………(7)
其中,g为获得的模糊图像,c为预设的第三参数,β 1为预设的第一正参数,a k+1为第k+1代中间变量;
以及,根据公式(8)计算第k+1代的第一拉格朗日乘子u k+1
u k+1=u k1(x k+1-hf k+1)……………(8)
其中,u k为第k代的第一拉格朗日乘子,β 1为预设的第一正参数,h为模糊核,f k+1为第k+1代复原图像;
以及,根据公式(9)计算第k+1代的第二拉格朗日乘子ξ k+1
Figure PCTCN2018086406-appb-000069
其中,ξ k为第k代的第二拉格朗日乘子,β 2为预设的第二正参数,
Figure PCTCN2018086406-appb-000070
为第k+1代原始图像的梯度,y k+1为第k+1代第二辅助变量;
以及,更新迭代次数k=k+1,并重新计算第k+1代复原图像;
所述计算模块,进一步用于当公式(5)成立时,则将第k+1代中间变量a k+1赋值给第k+1代第一辅助变量x k+1,且将第k+1代的正则参数λ k+1设为0。
具体的,全变分图像复原问题是要找到合适的f,满足下式:
Figure PCTCN2018086406-appb-000071
其中,第一项为TV正则项,后一项为保真项,λ为正则参数。
图像f的全变分计算公式为:
Figure PCTCN2018086406-appb-000072
其中,Ω是f的支持域。
根据Morozov’s偏差准则,上述问题等价于下面的约束最优问题:
Figure PCTCN2018086406-appb-000073
Figure PCTCN2018086406-appb-000074
其中,c为预设的第三参数,根据c=τmnσ 2计算,σ 2为噪声方差,τ是一个依赖于噪声的参数,比如,设为1;m、n为原始图像的维度。
采用变量分离方法,引入两个辅助变量,分别是第二辅助变量y,y∈Q,等于
Figure PCTCN2018086406-appb-000075
和第一辅助变量x,x∈V,等于hf,其中,V表示欧式空间R mm,Q=V×V,从而得到等价约束形式为:
Figure PCTCN2018086406-appb-000076
公式(10)的增广拉格朗日(Augmented Lagrangian,简称AL)泛函定义为:
Figure PCTCN2018086406-appb-000077
其中,u是第一拉格朗日乘子,u T是u的转置矩阵,ξ是第二拉格朗日乘子,ξ T是ξ的转置矩阵,||y||是y的L1范数,β 1为预设的第一正参数,β 2为预设的第二正参数;
将公式(10)的最小化问题转为求解公式(11)增广拉格朗日泛函的鞍点,具体迭代框架如下所示:
Figure PCTCN2018086406-appb-000078
由于公式(12)每次迭代都涉及到内迭代,所以解(f k+1,x k+1,y k+1)是比较繁琐的。因此采用改进的迭代方法来解决这个问题,即在每次迭代中只需一步运算就可以求得解(f k+1,x k+1,y k+1)。
改进的迭代方法如下所示:
Figure PCTCN2018086406-appb-000079
Figure PCTCN2018086406-appb-000080
Figure PCTCN2018086406-appb-000081
μ k+1=μ k1(x k+1-hf k+1)………………(8)
Figure PCTCN2018086406-appb-000082
公式(13)子问题,是关于f的最小化问题,是一个二次方程,形式如(16)所示;根据公式(16)得到公式(1):
Figure PCTCN2018086406-appb-000083
f k+1=(β 1h T2Δf k) -1[h T1x k-u k)-div(β 2y kk)]………………(1)
其中,
Figure PCTCN2018086406-appb-000084
是拉普拉斯算符;对于一个m×n大小的图像,经过两次FFT和一次逆FFT,即可求得式(1)的解,其算法复杂度为O(mn log(mn))。
公式(14)子问题,是关于y的最小化问题:
Figure PCTCN2018086406-appb-000085
此式是近端最小化(Proximal Minimization,PM)问题,利用二维收缩求得其解为:
Figure PCTCN2018086406-appb-000086
公式(15)子问题,是关于x的最小化问题,其中,先计算第k+1代中间变量a k+1
a k+1=hf k+1+(u k1)…………………(4)
通过求解公式(15)得公式(17):
Figure PCTCN2018086406-appb-000087
公式(17)是一个关于x的二次最小化问题,并且它有一个闭合形式的解:
x k+1=(λ k+1g+β 1a k+1)/(λ k+11)…………………(7)
依据a k+1的范围,每次迭代中λ的解存在两种情况:一种情况是
Figure PCTCN2018086406-appb-000088
此时λ k+1=0同时x k+1=a k+1,很明显x k+1满足Morozov’s偏差准则;另一种情况是
Figure PCTCN2018086406-appb-000089
根据Morozov’s偏差准则,应求解如下的方程:
Figure PCTCN2018086406-appb-000090
将公式(7)中的x k+1代入公式(18),得到:
Figure PCTCN2018086406-appb-000091
如此,通过引入辅助变量x,hf从严格的Morozov’s偏差准则中释放了出来,得益于此,在每次迭代中无需附加条件即可获得一个闭合形式的λ解。
在根据公式(1)计算出第k+1代复原图像f k+1后,可以根据公式(2),计算迭代误差A;
A=||f k+1-f k|| 2/||f k|| 2……………………(2)
判断是否需要继续迭代;
当迭代误差小于预设迭代误差阈值时,则第k+1代复原图像f k+1为重构图像,整个迭代结束;预设迭代误差阈值,根据经验设置,比如10 -5
当迭代误差不小于预设迭代误差阈值时,可以进一步判断迭代次数是否大于等于预设总迭代次数;当迭代次数达到预设总迭代次数时,则终止整个迭代,f k+1为重构图像。
当迭代误差不小于预设迭代误差阈值,且迭代次数未达到预设总迭代次数时,需要更新迭代次数,进一步迭代。
以上是针对已知模糊核情况下的处理,对于真实拍摄的模糊图像,需要利用目前现有的方法如刃边法,先进行模糊核的估计。
应当说明的是,上述实施例均可根据需要自由组合。以上所述仅是本发明的优选实施方式,应当指出,对于本技术领域的普通技术人员来说,在不脱离本发明原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本发明的保护范围。

Claims (15)

  1. 一种基于全变分的图像复原方法,其特征在于,包括:
    步骤S100当获得模糊图像时,初始化第0代复原图像、第0代的参数信息;
    所述第0代的参数信息包括:第0代第一辅助变量、第0代第二辅助变量、第0代第一拉格朗日乘子、第0代第二拉格朗日乘子;
    步骤S200设定迭代次数k=0;
    步骤S300根据第k代的参数信息,计算第k+1代复原图像;
    步骤S400根据所述第k+1代复原图像和所述第k代复原图像,判断是否满足迭代停止条件;
    步骤S500若满足迭代停止条件,则所述第k+1代复原图像f k+1为重构图像。
  2. 根据权利要求1所述的基于全变分的图像复原方法,其特征在于,所述步骤S400之后还包括:
    步骤S600若不满足迭代停止条件,则计算第k+1代的参数信息;
    所述第k+1代的参数信息包括:第k+1代第一辅助变量、第k+1代的第二辅助变量、第k+1代正则参数、第k+1代第一拉格朗日乘子、第k+1代第二拉格朗日乘子;
    步骤S700更新迭代次数k=k+1,并跳转到步骤S300。
  3. 根据权利要求1所述的基于全变分的图像复原方法,其特征在于,所述步骤S300包括:
    根据公式(1)计算所述第k+1代复原图像f k+1
    f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………(1)
    其中,h T为模糊核h的转置矩阵,β 1为预设的第一正参数,β 2为预设的第二正参数,(β 1h T2Δ) -1为(β 1h T2Δ)的求逆,u k是第k代的第一拉格朗日乘子,ξ k是第k代的第二拉格朗日乘子,x k为第k代的第一辅助变量, y k为第k代的第二辅助变量,Δ是拉普拉斯算符,div是散度算子。
  4. 根据权利要求1所述的基于全变分的图像复原方法,其特征在于,所述步骤S400具体为:
    根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差;
    判断所述迭代误差是否小于预设迭代误差阈值;
    若是,则认为满足迭代停止条件;
    若否,则认为不满足迭代停止条件。
  5. 根据权利要求4所述的基于全变分的图像复原方法,其特征在于,所述根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差的计算公式为公式(2):
    A=||f k+1-f k|| 2/||f k|| 2……………………(2);
    其中,A为迭代误差,f k+1为第k+1代复原图像,f k为第k代复原图像。
  6. 根据权利要求2所述的基于全变分的图像复原方法,其特征在于,所述步骤S600包括:
    当不满足迭代停止条件时,计算第k+1代第二辅助变量
    Figure PCTCN2018086406-appb-100001
    其中,
    Figure PCTCN2018086406-appb-100002
    为y k+1的第i行、第j列的元素,i=1,2,...,M,j=1,2,...,N;
    根据公式(3),计算
    Figure PCTCN2018086406-appb-100003
    Figure PCTCN2018086406-appb-100004
    其中,
    Figure PCTCN2018086406-appb-100005
    是梯度算子,
    Figure PCTCN2018086406-appb-100006
    为第k+1代原始图像的梯度的第i行、第j列的元素,
    Figure PCTCN2018086406-appb-100007
    是第k代的第二拉格朗日乘子的第i行、第j列的元素,β 2为预设的第二正参数;
    根据公式(4),计算第k+1代中间变量a k+1
    a k+1=hf k+1+(u k1)……………(4)
    其中,f k+1为第k+1代复原图像,u k是第k代的第一拉格朗日乘子,β 1为预设的第一正参数;
    判断公式(5)是否成立;
    Figure PCTCN2018086406-appb-100008
    其中,g为获得的模糊图像,c为预设的第三参数;
    当公式(5)不成立时,根据公式(6)计算第k+1代正则参数λ k+1,及根据公式(7)计算第k+1代第一辅助变量x k+1
    Figure PCTCN2018086406-appb-100009
    x k+1=(λ k+1g+β 1a k+1)/(λ k+11)……………(7)
    其中,g为获得的模糊图像,c为预设的第三参数,β 1为预设的第一正参数,a k+1为第k+1代中间变量;
    根据公式(8)计算第k+1代的第一拉格朗日乘子u k+1
    u k+1=u k1(x k+1-hf k+1)……………(8)
    其中,u k为第k代的第一拉格朗日乘子,β 1为预设的第一正参数,h为模糊核,f k+1为第k+1代复原图像;
    根据公式(9)计算第k+1代的第二拉格朗日乘子ξ k+1
    Figure PCTCN2018086406-appb-100010
    其中,ξ k为第k代的第二拉格朗日乘子,β 2为预设的第二正参数,
    Figure PCTCN2018086406-appb-100011
    为第k+1代原始图像的梯度,y k+1为第k+1代第二辅助变量。
  7. 根据权利要求6所述的基于全变分的图像复原方法,其特征在于:
    当公式(5)成立时,则将第k+1代中间变量赋值给第k+1代第一辅助变量,且将第k+1代正则参数设为0。
  8. 根据权利要求2所述的基于全变分的图像复原方法,其特征在于:
    当拍摄的图像为彩色图像时,按RGB三个通道分别获取所述彩色图像的各个通道的模糊图像,并分别按照步骤S100至步骤S700进行处理,得到RBG三个通道对应的重构图像,再将三个通道各自的重构图像组合形成重构的彩色图像。
  9. 一种基于全变分的图像复原系统,其特征在于,包括:
    初始化模块,用于当获得模糊图像时,初始化第0代复原图像、第0代的参数信息;所述第0代的参数信息包括:第0代第一辅助变量、第0代第二辅助变量、第0代第一拉格朗日乘子、第0代第二拉格朗日乘子;以及设定迭代次数k=0;
    计算模块,与所述初始化模块电连接,用于根据第k代的参数信息,计算第k+1代复原图像;
    判断模块,与所述计算模块电连接,用于根据所述第k+1代复原图像和所述第k代复原图像,判断是否满足迭代停止条件;若满足迭代停止条件,则所述第k+1代复原图像为重构图像。
  10. 根据权利要求9所述的基于全变分的图像复原系统,其特征在于:
    所述计算模块,进一步用于若不满足迭代停止条件,则计算第k+1代的参数信息;所述第k+1代的参数信息包括:第k+1代第一辅助变量、第k+1代第二辅助变量、第k+1代正则参数、第k+1代第一拉格朗日乘子、第k+1代第二拉格朗日乘子;以及更新迭代次数k=k+1,并重新计算第k+1代复原图像。
  11. 根据权利要求9所述的基于全变分的图像复原系统,其特征在于:
    所述计算模块用于根据公式(1)计算所述第k+1代复原图像f k+1
    f k+1=(β 1h T2Δ) -1[h T1x k-u k)-div(β 2y kk)]………(1)
    其中,h T为模糊核h的转置矩阵,β 1为预设的第一正参数,β 2为预设 的第二正参数,(β 1h T2Δ) -1为(β 1h T2Δ)的求逆,u k是第k代的第一拉格朗日乘子,ξ k是第k代的第二拉格朗日乘子,x k为第k代的第一辅助变量,y k为第k代的第二辅助变量,Δ是拉普拉斯算符,div是散度算子。
  12. 根据权利要求9所述的基于全变分的图像复原系统,其特征在于,所述判断模块,用于根据所述第k+1代复原图像和所述第k代复原图像,判断是否满足迭代停止条件具体为:
    所述计算模块,进一步用于根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差;
    所述判断模块,判断所述迭代误差是否小于预设迭代误差阈值;若是,则认为满足迭代停止条件;若否,则认为不满足迭代停止条件。
  13. 根据权利要求12所述的基于全变分的图像复原系统,其特征在于,所述计算模块根据所述第k+1代复原图像和所述第k代复原图像,计算迭代误差的计算公式为公式(2):
    A=||f k+1-f k|| 2/||f k|| 2……………………(2);
    其中,A为迭代误差,f k+1为第k+1代复原图像,f k为第k代复原图像。
  14. 根据权利要求10所述的基于全变分的图像复原系统,其特征在于:
    所述计算模块,用于当不满足迭代停止条件时,计算第k+1代第二辅助变量
    Figure PCTCN2018086406-appb-100012
    其中,
    Figure PCTCN2018086406-appb-100013
    为y k+1的第i行、第j列的元素,i=1,2,...,M,j=1,2,...,N;
    以及,根据公式(3),计算
    Figure PCTCN2018086406-appb-100014
    Figure PCTCN2018086406-appb-100015
    其中,
    Figure PCTCN2018086406-appb-100016
    是梯度算子,
    Figure PCTCN2018086406-appb-100017
    为第k+1代原始图像的梯度的第i行、第j列的元素,
    Figure PCTCN2018086406-appb-100018
    是第k代的第二拉格朗日乘子的第i行、第j列的元素,β 2为预设的第二正参数;
    以及,根据公式(4),计算第k+1代中间变量a k+1
    a k+1=hf k+1+(u k1)……………(4)
    其中,f k+1为第k+1代复原图像,u k是第k代的第一拉格朗日乘子,β 1为预设的第一正参数;
    以及,判断公式(5)是否成立;
    Figure PCTCN2018086406-appb-100019
    其中,g为获得的模糊图像,c为预设的第三参数;
    以及,当公式(5)不成立时,根据公式(6)计算第k+1代正则参数λ k+1,及根据公式(7)计算第k+1代第一辅助变量x k+1
    Figure PCTCN2018086406-appb-100020
    x k+1=(λ k+1g+β 1a k+1)/(λ k+11)……………(7)
    其中,g为获得的模糊图像,c为预设的第三参数,β 1为预设的第一正参数,a k+1为第k+1代中间变量;
    以及,根据公式(8)计算第k+1代的第一拉格朗日乘子u k+1
    u k+1=u k1(x k+1-hf k+1)……………(8)
    其中,u k为第k代的第一拉格朗日乘子,β 1为预设的第一正参数,h为模糊核,f k+1为第k+1代复原图像;
    以及,根据公式(9)计算第k+1代的第二拉格朗日乘子ξ k+1
    Figure PCTCN2018086406-appb-100021
    其中,ξ k为第k代的第二拉格朗日乘子,β 2为预设的第二正参数,
    Figure PCTCN2018086406-appb-100022
    为第k+1代原始图像的梯度,y k+1为第k+1代第二辅助变量。
  15. 根据权利要求14所述的基于全变分的图像复原系统,其特征在于:
    所述计算模块,用于当公式(5)成立时,则将第k+1代中间变量赋值给第k+1代第一辅助变量,且将第k+1代的正则参数设为0。
PCT/CN2018/086406 2017-11-06 2018-05-10 一种基于全变分的图像复原方法及系统 WO2019085433A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711077178.6 2017-11-06
CN201711077178.6A CN107945121A (zh) 2017-11-06 2017-11-06 一种基于全变分的图像复原方法及系统

Publications (1)

Publication Number Publication Date
WO2019085433A1 true WO2019085433A1 (zh) 2019-05-09

Family

ID=61934311

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/086406 WO2019085433A1 (zh) 2017-11-06 2018-05-10 一种基于全变分的图像复原方法及系统

Country Status (2)

Country Link
CN (1) CN107945121A (zh)
WO (1) WO2019085433A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113570522A (zh) * 2021-07-28 2021-10-29 南京航空航天大学 一种基于混合约束条件的雾霾图像复原方法及终端
CN114022387A (zh) * 2021-11-11 2022-02-08 上海联影医疗科技股份有限公司 基于全变分最小化的图像优化方法、装置和计算机设备
CN114052701A (zh) * 2021-11-01 2022-02-18 河南师范大学 一种电容耦合电阻层析成像图像重建方法
CN114282348A (zh) * 2021-11-17 2022-04-05 南京邮电大学 基于隐式正则与算法bdca的相位恢复方法
CN115082333A (zh) * 2022-05-16 2022-09-20 西北核技术研究所 基于归一化加权总变分法的图像去模糊方法、计算机程序产品

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107945121A (zh) * 2017-11-06 2018-04-20 上海斐讯数据通信技术有限公司 一种基于全变分的图像复原方法及系统
CN111986123B (zh) * 2020-09-24 2024-03-12 南京航空航天大学 基于kl散度和l0范数约束的模糊图像复原方法
CN112215779B (zh) * 2020-10-28 2023-10-03 青岛大学 一种图像处理方法、装置、设备及计算机可读存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004233124A (ja) * 2003-01-29 2004-08-19 Mitsubishi Electric Corp レーダ画像処理装置
WO2013076613A1 (en) * 2011-11-23 2013-05-30 Koninklijke Philips Electronics N.V. Image domain de-noising
CN104751420A (zh) * 2015-03-06 2015-07-01 湖南大学 一种基于稀疏表示和多目标优化的盲复原方法
CN106780645A (zh) * 2016-11-28 2017-05-31 四川大学 动态mri图像重建方法及装置
CN106934778A (zh) * 2017-03-10 2017-07-07 北京工业大学 一种基于小波域结构和非局部分组稀疏的mr图像重建方法
CN107945121A (zh) * 2017-11-06 2018-04-20 上海斐讯数据通信技术有限公司 一种基于全变分的图像复原方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2004233124A (ja) * 2003-01-29 2004-08-19 Mitsubishi Electric Corp レーダ画像処理装置
WO2013076613A1 (en) * 2011-11-23 2013-05-30 Koninklijke Philips Electronics N.V. Image domain de-noising
CN104751420A (zh) * 2015-03-06 2015-07-01 湖南大学 一种基于稀疏表示和多目标优化的盲复原方法
CN106780645A (zh) * 2016-11-28 2017-05-31 四川大学 动态mri图像重建方法及装置
CN106934778A (zh) * 2017-03-10 2017-07-07 北京工业大学 一种基于小波域结构和非局部分组稀疏的mr图像重建方法
CN107945121A (zh) * 2017-11-06 2018-04-20 上海斐讯数据通信技术有限公司 一种基于全变分的图像复原方法及系统

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113570522A (zh) * 2021-07-28 2021-10-29 南京航空航天大学 一种基于混合约束条件的雾霾图像复原方法及终端
CN113570522B (zh) * 2021-07-28 2024-04-16 南京航空航天大学 一种基于混合约束条件的雾霾图像复原方法及终端
CN114052701A (zh) * 2021-11-01 2022-02-18 河南师范大学 一种电容耦合电阻层析成像图像重建方法
CN114022387A (zh) * 2021-11-11 2022-02-08 上海联影医疗科技股份有限公司 基于全变分最小化的图像优化方法、装置和计算机设备
CN114282348A (zh) * 2021-11-17 2022-04-05 南京邮电大学 基于隐式正则与算法bdca的相位恢复方法
CN115082333A (zh) * 2022-05-16 2022-09-20 西北核技术研究所 基于归一化加权总变分法的图像去模糊方法、计算机程序产品

Also Published As

Publication number Publication date
CN107945121A (zh) 2018-04-20

Similar Documents

Publication Publication Date Title
WO2019085433A1 (zh) 一种基于全变分的图像复原方法及系统
US8755628B2 (en) Image de-hazing by solving transmission value
WO2020015330A1 (zh) 基于增强的神经网络的图像复原方法、存储介质及系统
CN106803235B (zh) 基于各向异性时空域全变分非均匀性校正的方法
CN105160687B (zh) 一种气动光学热辐射噪声校正方法与系统
US9406108B2 (en) Deblurring of an image from a sequence of images
WO2014036950A1 (zh) 基于非局部均值的视频去噪方法和装置
CN108460783B (zh) 一种脑部核磁共振图像组织分割方法
KR101698314B1 (ko) 영상 통계정보에 기반한 정지장면 분할장치 및 그 방법
CN110147816B (zh) 一种彩色深度图像的获取方法、设备、计算机存储介质
CN114418883A (zh) 一种基于深度先验的盲图像去模糊方法
CN104008533B (zh) 基于分块自适应特征跟踪的多传感器图像融合方法
CN110033414B (zh) 一种基于均值化处理的红外图像非均匀性校正方法及系统
US10803557B2 (en) Non-uniformity correction method for infrared image based on guided filtering and high-pass filtering
CN104809706B (zh) 一种基于图像颜色平缓变化先验的单透镜计算成像方法
CN110517197A (zh) 基于参数模型的单幅图像去模糊方法
CN110796616A (zh) 基于分数阶微分算子的l0范数约束和自适应加权梯度的湍流退化图像恢复方法
CN109636747A (zh) 应用于单幅红外图像非均匀校正的深度高频网络校正算法
CN106210448A (zh) 一种视频图像抖动消除处理方法
Dong et al. A multi-scale vectorial L τ-TV framework for color image restoration
Marais et al. Proximal-gradient methods for poisson image reconstruction with bm3d-based regularization
CN112184567A (zh) 一种基于交替最小化的多通道盲识别自适应光学图像复原方法
CN111179333B (zh) 一种基于双目立体视觉的散焦模糊核估计方法
CN107295217A (zh) 一种基于主成分分析的视频噪声估计方法
Chen et al. Fast linearized alternating direction minimization algorithm with adaptive parameter selection for multiplicative noise removal

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18872096

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18872096

Country of ref document: EP

Kind code of ref document: A1