WO2016106951A1 - 一种方向自适应图像去模糊方法 - Google Patents

一种方向自适应图像去模糊方法 Download PDF

Info

Publication number
WO2016106951A1
WO2016106951A1 PCT/CN2015/072665 CN2015072665W WO2016106951A1 WO 2016106951 A1 WO2016106951 A1 WO 2016106951A1 CN 2015072665 W CN2015072665 W CN 2015072665W WO 2016106951 A1 WO2016106951 A1 WO 2016106951A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
sub
variable
operator
minimum
Prior art date
Application number
PCT/CN2015/072665
Other languages
English (en)
French (fr)
Inventor
张天序
周钢
钟奥
王亮亮
李明
鲁岑
张文
左芝勇
Original Assignee
华中科技大学
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 华中科技大学 filed Critical 华中科技大学
Priority to US15/022,872 priority Critical patent/US9582862B2/en
Publication of WO2016106951A1 publication Critical patent/WO2016106951A1/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
    • G06T5/00Image enhancement or restoration
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20004Adaptive image processing
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20004Adaptive image processing
    • G06T2207/20012Locally adaptive
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/20Special algorithmic details
    • G06T2207/20048Transform domain processing
    • G06T2207/20056Discrete and fast Fourier transform, [DFT, FFT]

Definitions

  • the invention belongs to the field of aerospace and image processing crossover technology, and more particularly to a directional adaptive image deblurring method, which is mainly suitable for deblurring a remote sensing image.
  • Space targets such as a large number of communication satellites and resource satellites launched at home and abroad can be used in applications such as network communication, aerial photography, and geodetic survey. Due to the limitation of camera spatial resolution, random noise, and the interference of atmospheric turbulence on the long-distance optical imaging system, the image acquired by the sensor is prone to ambiguity of the target, which brings huge impact on the later target positioning and target classification. The difficulty, therefore. How to effectively improve the image quality of such images has become the focus of research at home and abroad. Domestic and foreign scholars have carried out detailed research on the target deblurring algorithm under such imaging conditions, and have achieved relevant results.
  • the present invention provides a directional adaptive image correction method, which quickly and effectively solves the problem of blurring of a long-distance imaging image, and the algorithm has small calculation amount and good adaptability.
  • the present invention provides a directional adaptive image deblurring method comprising the following steps:
  • Step (1) Define the direction adaptive TV regularized image deblurring to minimize the cost function:
  • u is the restored image
  • H is the point spread function
  • f is the degraded image
  • ⁇ >0 is the regularization parameter
  • Direction vector For the gradient operator the symbol ⁇ is a vector dot product operator, The symbol ⁇ > is an inner product operator and the log is a logarithmic function; Representing the minimum value for the energy functional ⁇ 1, Hu-f log(Hu)>, and taking the u corresponding to the minimum value as the output;
  • Step (3) Introduce a penalty term to split the constrained problem in step (2) into a new minimized cost function:
  • Step (4) Convert the minimization problem in step (3) into an alternating minimum solution problem of u, d 1 , d 2 , d 3 with respect to the variable, that is, the other variables are fixedly solved for one of the variables, and the alternating minimum iterative strategy is used. Iteratively solves the above minimum solution problem and obtains the deblurred image.
  • the present invention has the following advantages:
  • the method of the present invention is capable of recovering against complex blur types or having rich texture images.
  • FIG. 1 is a flowchart of a direction adaptive image deblurring method according to the present invention
  • Figure 2 (d) is a clear image of the house in the embodiment of the present invention.
  • Figure 3 (a) is an image obtained by adding Gaussian blur and Poisson noise with a size of 15 * 15 and a standard deviation of 1.8 to Figure 2 (a), the PSNR of which is 14.97;
  • Figure 3 (b) is an image after adding a blur of 3 disc and Poisson noise degradation to Figure 2 (b), the PSNR is 21.88;
  • Figure 3 (c) is an image obtained by adding uniform blur and Poisson noise of size 7 * 7 to Figure 2 (c), the PSNR is 22.22;
  • Fig. 3(d) is a diagram of adding random blur and Poisson noise degraded by 7*7 to Fig. 2(d) Like, its PSNR is 23.29;
  • each scheme has its own algorithm features.
  • the algorithm only uses some large-scale statistical prior features of the image to derive the statistical optimal solution or approximate optimal solution of the problem, without considering the small-scale geometric structure inherent in the image itself, such as the edge direction of the image.
  • the texture direction or the like is used to constrain the correction result. Therefore, for such a method that relies only on the large-scale statistical prior of the image, the correction result is usually prone to boundary blur and loss of detail at the edge of the image.
  • FIG. 1 is a flowchart of an algorithm of the present invention.
  • the present invention provides a directional adaptive image deblurring method, including the following steps:
  • Step (1) defining a new direction adaptive total variation (TV) regularized image deblurring minimization cost function
  • f is the degraded image
  • H is the linear operator, representing the point spread function that blurs the image
  • u is the clear image that is potentially recovering
  • n is the imaging noise.
  • the task of image non-blind deconvolution is to obtain a sharp image u from the known degraded image f and the point spread function H.
  • the inverse process of image restoration is ill-conditioned, and the noise is amplified during the recovery process, making the image deblurring result unstable. Since the TV regularization method has a good advantage in restoring image details, the present invention uses TV regular terms to overcome the morbidity of image restoration.
  • the local edge information of the image is incorporated into the Maximum a posteriori MAP algorithm framework, and a direction-adaptive image deblurring method is obtained, so that the edge of the restored image can be better protected.
  • the directional adaptive TV regularization cost function introduced by the present invention is defined as:
  • u is the restored image
  • H is the point spread function
  • f is the degraded image
  • ⁇ >0 is the regularization parameter
  • the direction adaptive TV regularization cost function can be expanded to:
  • Step (3) Introduce a penalty term to split the constrained problem in step (2) into a new minimized cost function:
  • Step (4) Use the alternating minimum iteration strategy to convert the minimization problem in step (3) into an alternating minimum solution problem with respect to the variables u, d 1 , d 2 , d 3 .
  • the other variables are fixed to solve one of the variables.
  • FFT represents the fast Fourier transform
  • FFT -1 represents the inverse of the fast Fourier transform
  • real represents the real part of the complex number.
  • H T represents the conjugate operator of H
  • superscript k represents the kth iteration.
  • tan -1 is the arctangent function
  • is the pi
  • w is an integer greater than 1
  • ⁇ w is the sum of the values in the neighborhood of w ⁇ w centered at the current point
  • sum(f) represents the sum of the gray values of the image
  • the method of the present invention assumes that the point spread function of the image is known, and takes the maximum number of iterations 100 times, taking the largest peak SNR (Peak Signal to Noise Ratio) PSNR image corresponding to the final image output as a clear, PSNR is calculated image u k output k-th iteration is as follows:
  • I represents a clear reference image
  • max(I) represents the grayscale maximum of the image I.
  • FIG. 2(a) is a clear circuit image in the embodiment of the present invention
  • FIG. 2(b) is a clear Cameraman image in the embodiment of the present invention
  • FIG. 2(c) is a clear CT image of the liver in the embodiment of the present invention
  • Figure 2 (d) is a clear image of the house in the embodiment of the present invention
  • Figure 3 (a) is an image obtained by adding Gaussian blur and Poisson noise with a size of 15 * 15 and a standard deviation of 1.8 to Figure 2 (a), the PSNR of which is 14.97;
  • Fig. 3(b) is a diagram showing the addition of a disc with a radius of 3 to the blur and Poisson noise degradation of Fig. 2(b) Like, its PSNR is 21.88;
  • Figure 3 (c) is an image obtained by adding uniform blur and Poisson noise of size 7 * 7 to Figure 2 (c), the PSNR is 22.22;
  • Fig. 3(d) is an image obtained by adding random blur and Poisson noise of size 7*7 to Fig. 2(d), and its PSNR is 23.29; in the present invention, the random fuzzy point spread function used is specifically:

Landscapes

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

Abstract

本发明公开了一种方向自适应图像去模糊方法,包括以下步骤:(1)定义方向自适应总变分(Total Variation)TV正则化图像去模糊最小化代价函数;(2)引入辅助变量d1Hu, d 2 =∇ x u, d 3 =∇ 3 将步骤(1)中的无约束最小化问题转换为有约束问题;(3)引入惩罚项将步骤(2)中的有约束问题转化为新的最小化代价函数;(4)使用交替最小迭代策略将步骤(3)中的最小化问题转换为关于变量的u,d1,d2,d3的交替最小求解问题。通过迭代运算最终恢复出原清晰图像u。与现有技术相比,本发明方法将局部方向信息引入最大后验概率(Maximum a posteriori)MAP算法框架,得到新的方向自适应代价函数,克服了传统TV正则项恢复图像边缘模糊的问题;且能够针对复杂模糊类型或具有丰富纹理图像进行恢复。

Description

一种方向自适应图像去模糊方法 [技术领域]
本发明属于航天与图像处理交叉技术领域,更具体地,涉及一种方向自适应图像去模糊方法,主要适用于遥感成像图像去模糊。
[背景技术]
国内外发射的大量通信卫星、资源卫星等空间目标可用于网络通信、航空摄影、大地测量等应用场所。由于相机空间分辨率限制、随机噪声、以及大气湍流对远距离光学成像系统的干扰,使得传感器获取的图像容易出现目标模糊不清的现象,这对后期的目标定位、目标分类等操作带来巨大的困难,因此。如何有效地提高这类图像的成像质量成为国内外研究的焦点。国内外学者对这类成像条件下的目标去模糊算法进行了详细的研究,并且取得了相关成果。如,何成剑,洪汉玉,张天序的“基于广义规整化的红外湍流退化图像盲复原方法”,见《红外技术》,20006年8月,第28卷,第8期。研究了一种基于广义规整化的湍流退化图像盲复原方法,对传统的规整化方法进行了扩展,提出了广义规整化的策略,并取得了较好的校正结果。但是该方法主要是针对湍流环境下红外成像,对远距离可见光成像校正结果文中未见详细报道。付长军,许东,赵剡的“湍流退化图像的最大熵盲目复原方法”,见《红外与激光工程》,2008年6月,第37卷第3期。提出了一种针对具有单一背景的湍流退化图像的最大熵盲目复原方法。复原过程中,为避免最大熵约束项的非线性带来的复杂运算,对熵的表达式进行二次近似,并利用灰度变换确保近似程度的准确性,最后采用共轭梯度法进行求解,从而大大降低了计算量。S.Setzer,G.Steidl,T.Teuber的“Deblurring Poissonian images by split Bregman techniques”,见J.Vis.Commun.Image R.21(2010)193–199.研究了一种基于分裂Bregman迭 代的图像快速去模糊方法,该文中所采用的加速技术只是提高了算法的速度,并未提高图像的校正质量。
[发明内容]
为解决现有方法运算速度慢或校正结果边缘模糊的问题,本发明提供一种方向自适应图像校正方法,快速有效地解决了远距离成像图像模糊的问题,算法计算量小,适应性好。
为实现上述目的,本发明提供了一种方向自适应图像去模糊方法,包括以下步骤:
步骤(1):定义方向自适应TV正则化图像去模糊最小化代价函数:
Figure PCTCN2015072665-appb-000001
其中,u为复原图像,H为点扩展函数,f为退化图像,λ>0为正则化参数;符号
Figure PCTCN2015072665-appb-000002
表示向量
Figure PCTCN2015072665-appb-000003
的l1-范数;
Figure PCTCN2015072665-appb-000004
为方向矢量;
Figure PCTCN2015072665-appb-000005
为梯度算子,符号·为矢量点积算子,
Figure PCTCN2015072665-appb-000006
符号<>为内积算子,log为对数函数;
Figure PCTCN2015072665-appb-000007
表示对能量泛函<1,Hu-f log(Hu)>计算最小值,并将最小值对应的u作为输出;
步骤(2):引入辅助变量d1=Hu,
Figure PCTCN2015072665-appb-000008
将步骤(1)中的无约束最小化问题转换为有约束问题;
Figure PCTCN2015072665-appb-000009
步骤(3):引入惩罚项将步骤(2)中的有约束问题分裂为新的最小化代价函数:
Figure PCTCN2015072665-appb-000010
其中,α,β,γ为大于零的惩罚参数;
步骤(4):将步骤(3)中的最小化问题转换为关于变量的u,d1,d2,d3的交替最小求解问题,即将其它变量固定求解其中一个变量,使用交替最小迭代策略迭代求解上述最小求解问题,得到去模糊后的图像。
与现有技术相比,本发明具有如下优点:
(1)将局部方向信息引入最大后验概率(Maximum a posteriori)MAP算法框架,得到新的方向自适应代价函数,克服了传统TV正则项恢复图像边缘模糊的问题;
(2)本发明方法能够针对复杂模糊类型或具有丰富纹理图像进行恢复。
[附图说明]
图1为本发明方向自适应图像去模糊方法流程图;
图2(a)为本发明实施例中清晰的电路图像;
图2(b)为本发明实施例中清晰的Cameraman图像;
图2(c)为本发明实施例中清晰的肝部CT图像;
图2(d)为本发明实施例中清晰的房子图像;
图3(a)为对图2(a)添加尺寸为15*15,标准差为1.8的高斯模糊和泊松噪声退化后的图像,其PSNR为14.97;
图3(b)为对图2(b)添加半径为3的圆盘模糊和泊松噪声退化后的图像,其PSNR为21.88;
图3(c)为对图2(c)添加尺寸为7*7的均匀模糊和泊松噪声退化后的图像,其PSNR为22.22;
图3(d)为对图2(d)添加尺寸为7*7的随机模糊和泊松噪声退化后的图 像,其PSNR为23.29;
图4(a)为对图3(a)使用本发明算法进行去模糊的效果,其PSNR为19.56,正则化参数λ=5.5666×10-7
图4(b)为对图3(b)使用本发明算法进行去模糊的效果,其PSNR为23.93,正则化参数λ=6.2986×10-7
图4(c)为对图3(c)使用本发明算法进行去模糊的效果,其PSNR为24.71,正则化参数λ=3.3193×10-6
图4(d)为对图3(d)使用本发明算法进行去模糊的效果,其PSNR为28.32,正则化参数λ=2.7896×10-7
[具体实施方式]
为了使本发明的目的、技术方案及优点更加清楚明白,以下结合附图及实施例,对本发明进行进一步详细说明。应当理解,此处所描述的具体实施例仅仅用以解释本发明,并不用于限定本发明。此外,下面所描述的本发明各个实施方式中所涉及到的技术特征只要彼此之间未构成冲突就可以相互组合。
由于现有的研究成果都对该类图像去模糊化问题的特殊情况提出了各自的解决方案,各个方案具有自己的算法特点。但是算法都只利用了图像的某种大尺度统计先验特性来推导出问题的统计最优解或近似最优解,并未考虑图像本身内在的小尺度几何结构特性,例如图像的边缘方向、纹理方向等来约束校正结果。因此,对于此类仅仅依赖图像大尺度统计先验的方法,其校正结果在图像边缘处通常容易出现边界模糊、细节丢失等现象。针对上述问题,我们将图像小尺度几何特性(图像的局部边缘方向)融合到最大后验概率(Maximum a posteriori)MAP算法框架中,提出了一种基于方向自适应的图像去模糊方法,采用交替最小迭代和快速傅里叶技术实现问题的加速求解。试验结果显示提出的方法能较好地恢复出清晰的图像 边缘和细节信息,且具有较快的运算速度。
如图1所示为本发明算法流程图,本发明提供了一种方向自适应图像去模糊方法,包括以下步骤:
步骤(1):定义一种新的方向自适应总变分(Total Variation)TV正则化图像去模糊最小化代价函数;
在图像去模糊中,大多数图像的退化可以看成线性过程,可以用以下式子表达:f=Hu+n。
其中f为退化图像,H为线性算子,代表使图像模糊的点扩散函数,u表示为潜在需要恢复的清晰图像,n为成像噪声。图像非盲反卷积的任务就是根据已知的退化图像f和点扩展函数H得到清晰图像u。图像恢复这一逆过程存在着病态性,在恢复过程中会使噪声放大,使得图像去模糊结果很不稳定。由于TV正则化方法对恢复图像细节有很好的优势,因此本发明使用TV正则项来克服图像恢复的病态性。同时将图像的局部边缘信息融入到最大后验概率(Maximum a posteriori)MAP算法框架中,得到一种方向自适应的图像去模糊方法,使得恢复图像的边缘能够得到更好的保护。对于泊松噪声污染图像,本发明引入的方向自适应TV正则化代价函数定义为:
Figure PCTCN2015072665-appb-000011
其中,u为复原图像,H为点扩展函数,f为退化图像,λ>0为正则化参数,符号||x||1表示向量x=(x1,x2,...xn)的l1-范数;其定义为:
Figure PCTCN2015072665-appb-000012
|xi|表示对变量xi取绝对值,i为下标指示符号,表示取向量x的第i个分量。
Figure PCTCN2015072665-appb-000013
表示对|xi|求和。
Figure PCTCN2015072665-appb-000014
为方向矢量;
Figure PCTCN2015072665-appb-000015
为梯度算子,符号·为矢量点积算子,
Figure PCTCN2015072665-appb-000016
符号<>为内积算子,log为对数函数,
Figure PCTCN2015072665-appb-000017
表示对能量泛函g(u)计算最小值,并将最小值对应的u作为输出。
方向自适应TV正则化代价函数可展开为:
Figure PCTCN2015072665-appb-000018
步骤(2):引入辅助变量d1=Hu,
Figure PCTCN2015072665-appb-000019
将步骤(1)中的无约束最小化问题转换为有约束问题;
Figure PCTCN2015072665-appb-000020
步骤(3):引入惩罚项将步骤(2)中的有约束问题分裂为新的最小化代价函数:
Figure PCTCN2015072665-appb-000021
其中,α,β,γ为大于零的惩罚参数。
步骤(4):使用交替最小迭代策略将步骤(3)中的最小化问题转换为关于变量的u,d1,d2,d3的交替最小求解问题。即将其它变量固定求解其中一个变量,这几个子问题分别为:
(4.1)关于变量u的子问题:
Figure PCTCN2015072665-appb-000022
(4.2)关于变量d1的子问题:
Figure PCTCN2015072665-appb-000023
(4.3)关于变量d2的子问题:
Figure PCTCN2015072665-appb-000024
(4.4)关于变量d3的子问题:
Figure PCTCN2015072665-appb-000025
在求解(4.1),(4.2)时,由于函数是可微的,本发明方法中采用求偏微分与快速傅里叶变换(FFT)的方法直接求解:
Figure PCTCN2015072665-appb-000026
Figure PCTCN2015072665-appb-000027
式中,FFT表示快速傅里叶变换,FFT-1表示快速傅里叶变换逆变换,real表示取复数的实部,
Figure PCTCN2015072665-appb-000028
表示
Figure PCTCN2015072665-appb-000029
的转置算子,HT表示H的共轭算子,上标k表示第k次迭代。
问题(4.3),(4.4)可直接采用二维shrink算子求解:
Figure PCTCN2015072665-appb-000030
Figure PCTCN2015072665-appb-000031
本发明中,我们取α=β=γ=1,方向矢量
Figure PCTCN2015072665-appb-000032
其中,cos,sin分别为余弦,正弦函数,max(X,Y)表示取X,Y中的最大值,角度θ计算公式如下:
Figure PCTCN2015072665-appb-000033
式中,tan-1为反正切函数,π为圆周率,w为大于1的整数,∑w表示 对以当前点为中心的w×w的邻域内的值进行求和,本发明中我们取w=5。
因此,一种方向自适应图像去模糊方法算法过程如下:
Figure PCTCN2015072665-appb-000034
式中,sum(f)表示对图像的灰度值求和;在去模糊过程中,本发明方法假设图像的点扩展函数是已知的,且取最大迭代次数为100次,取最大的峰值信噪比(Peak Signal to Noise Ratio)PSNR对应的图像作为最终的清晰图像输出,第k次迭代输出的图像uk的PSNR计算公式如下:
Figure PCTCN2015072665-appb-000035
式中,I表示清晰的参考图像,max(I)表示图像I的灰度最大值。
为了让本领域技术人员更好地理解该算法,我们对不同图像进行了仿真验证。如图2(a)为本发明实施例中清晰的电路图像;图2(b)为本发明实施例中清晰的Cameraman图像;图2(c)为本发明实施例中清晰的肝部CT图像;图2(d)为本发明实施例中清晰的房子图像;
图3(a)为对图2(a)添加尺寸为15*15,标准差为1.8的高斯模糊和泊松噪声退化后的图像,其PSNR为14.97;
图3(b)为对图2(b)添加半径为3的圆盘模糊和泊松噪声退化后的图 像,其PSNR为21.88;
图3(c)为对图2(c)添加尺寸为7*7的均匀模糊和泊松噪声退化后的图像,其PSNR为22.22;
图3(d)为对图2(d)添加尺寸为7*7的随机模糊和泊松噪声退化后的图像,其PSNR为23.29;本发明中,使用的随机模糊点扩展函数具体为:
Figure PCTCN2015072665-appb-000036
图4(a)为对图3(a)使用本发明算法进行去模糊的效果,其PSNR为19.56,正则化参数λ=5.5666×10-7,算法在k=63时取得PSNR最大值;
图4(b)为对图3(b)使用本发明算法进行去模糊的效果,其PSNR为23.93,正则化参数λ=6.2986×10-7,算法在k=50时取得PSNR最大值;
图4(c)为对图3(c)使用本发明算法进行去模糊的效果,其PSNR为24.71,正则化参数λ=3.3193×10-6,算法在k=36时取得PSNR最大值;
图4(d)为对图3(d)使用本发明算法进行去模糊的效果,其PSNR为28.32,正则化参数λ=2.7896×10-7,算法在k=42时取得PSNR最大值;
本领域的技术人员容易理解,以上所述仅为本发明的较佳实施例而已,并不用以限制本发明,凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等,均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种方向自适应图像去模糊方法,其特征在于,所述方法包括如下步骤:
    步骤(1):定义方向自适应总变分(Total Variation)TV正则化图像去模糊最小化代价函数:
    Figure PCTCN2015072665-appb-100001
    其中,u为复原图像,H为点扩展函数,f为退化图像,λ>0为正则化参数;符号
    Figure PCTCN2015072665-appb-100002
    表示向量
    Figure PCTCN2015072665-appb-100003
    的l1-范数;
    Figure PCTCN2015072665-appb-100004
    为方向矢量;
    Figure PCTCN2015072665-appb-100005
    为梯度算子,符号·为矢量点积算子,
    Figure PCTCN2015072665-appb-100006
    符号
    Figure PCTCN2015072665-appb-100007
    为内积算子,log为对数函数;
    Figure PCTCN2015072665-appb-100008
    表示对能量泛函<1,Hu-f log(Hu)>计算最小值,并将最小值对应的u作为输出;
    步骤(2):引入辅助变量d1=Hu,
    Figure PCTCN2015072665-appb-100009
    Figure PCTCN2015072665-appb-100010
    将步骤(1)中的无约束最小化问题转换为有约束问题;
    Figure PCTCN2015072665-appb-100011
    步骤(3):引入惩罚项将步骤(2)中的有约束问题分裂为新的最小化代价函数:
    Figure PCTCN2015072665-appb-100012
    其中,α,β,γ为大于零的惩罚参数;
    步骤(4):将步骤(3)中的最小化问题转换为关于变量的u,d1,d2,d3的交替最小求解问题,即将其它变量固定求解其中一个变量,使用交替最 小迭代策略迭代求解上述最小求解问题,得到去模糊后的图像。
  2. 如权利要求1所述的方法,其特征在于,所述步骤(4)中的最小求解问题具体为:
    (4.1)关于变量u的子问题:
    Figure PCTCN2015072665-appb-100013
    (4.2)关于变量d1的子问题:
    Figure PCTCN2015072665-appb-100014
    (4.3)关于变量d2的子问题:
    Figure PCTCN2015072665-appb-100015
    (4.4)关于变量d3的子问题:
    Figure PCTCN2015072665-appb-100016
  3. 如权利要求2所述的方法,其特征在于,所述步骤上(4)中使用交替最小迭代策略迭代求解上述最小求解问题,得到去模糊后的图像,具体包括:
    初始化待恢复图像
    Figure PCTCN2015072665-appb-100017
    d1 0=d2 0=d3 0=0;
    初始化最大迭代次数kMax,并初始化迭代次数k=0;
    判断迭代次数k是否小于最大迭代次数kMax,如果不小于则终止迭代;如果小于则继续进行下述迭代操作:
    更新迭代次数k=k+1;
    求解步骤(4)中的子问题(4.1)以更新复原图像uk
    求解步骤(4)中的子问题(4.2)以更新辅助变量d1 k
    求解步骤(4)中的子问题(4.2)以更新复原图像d2 k
    求解步骤(4)中的子问题(4.3)以更新复原图像d3 k
    计算更新后的复原图像uk的PSNR;
    迭代终止后,取最大的PSNR对应的恢复图像作为最终的清晰图像输出。
  4. 如权利要求2或3所述的方法,其特征在于,所述子问题(4.1)中的变量u采用求偏微分与快速傅里叶变换的方法求解:
    Figure PCTCN2015072665-appb-100018
    式中,FFT表示快速傅里叶变换,FFT-1表示快速傅里叶变换逆变换,real表示取复数的实部,
    Figure PCTCN2015072665-appb-100019
    表示
    Figure PCTCN2015072665-appb-100020
    的转置算子,HT表示H的共轭算子,上标k表示第k次迭代。
  5. 如权利要求2或3所述的方法,其特征在于,所述子问题(4.2)中的变量d1采用求偏微分与快速傅里叶变换的方法求解:
    Figure PCTCN2015072665-appb-100021
    式中,FFT表示快速傅里叶变换,FFT-1表示快速傅里叶变换逆变换,real表示取复数的实部,
    Figure PCTCN2015072665-appb-100022
    表示
    Figure PCTCN2015072665-appb-100023
    的转置算子,HT表示H的共轭算子,上标k表示第k次迭代。
  6. 如权利要求2或3所述的方法,其特征在于,所述子问题(4.3)中的变量d2采用二维shrink算子求解:
    Figure PCTCN2015072665-appb-100024
    式中,方向矢量
    Figure PCTCN2015072665-appb-100025
    其中,cos,sin分别为余弦,正弦函数,max(X,Y)表示取X,Y中的最大值,角度θ计算公式如下:
    Figure PCTCN2015072665-appb-100026
    式中,tan-1为反正切函数,π为圆周率,w为大于1的整数,∑w表示对以当前点为中心的w×w的邻域内的值进行求和。
  7. 如权利要求2或3所述的方法,其特征在于,所述子问题(4.4)中的变量d3采用二维shrink算子求解:
    Figure PCTCN2015072665-appb-100027
    式中,方向矢量
    Figure PCTCN2015072665-appb-100028
    其中,cos,sin分别为余弦,正弦函数,max(X,Y)表示取X,Y中的最大值,角度θ计算公式如下:
    Figure PCTCN2015072665-appb-100029
    式中,tan-1为反正切函数,π为圆周率,w为大于1的整数,∑w表示对以当前点为中心的w×w的邻域内的值进行求和。
  8. 如权利要求3至7任一项所述的方法,其特征在于,所述图像uk的PSNR计算公式如下:
    Figure PCTCN2015072665-appb-100030
    式中,I表示清晰的参考图像,max(I)表示图像I的灰度最大值。
  9. 如权利要求6或7所述的方法,其特征在于,α=β=γ=1。
  10. 如权利要求6或7所述的方法,其特征在于,所述w=5。
PCT/CN2015/072665 2014-12-30 2015-02-10 一种方向自适应图像去模糊方法 WO2016106951A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/022,872 US9582862B2 (en) 2014-12-30 2015-02-10 Direction-adaptive image deblurring method

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201410844605.9A CN104537620B (zh) 2014-12-30 2014-12-30 一种方向自适应图像去模糊方法
CN2014108446059 2014-12-30

Publications (1)

Publication Number Publication Date
WO2016106951A1 true WO2016106951A1 (zh) 2016-07-07

Family

ID=52853139

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/072665 WO2016106951A1 (zh) 2014-12-30 2015-02-10 一种方向自适应图像去模糊方法

Country Status (3)

Country Link
US (1) US9582862B2 (zh)
CN (1) CN104537620B (zh)
WO (1) WO2016106951A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108198149A (zh) * 2018-01-24 2018-06-22 闽南师范大学 一种图像去模糊方法
CN108615224A (zh) * 2016-12-13 2018-10-02 三星电机株式会社 图像校正设备和方法以及非易失性计算机可读存储介质
CN112184567A (zh) * 2020-08-28 2021-01-05 江苏海洋大学 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
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 (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106408519B (zh) * 2015-11-10 2019-06-21 青岛大学 一种基于全变分的非局部图像复原方法
CN106204485B (zh) * 2016-07-11 2019-03-22 西安理工大学 基于正弦积分的图像复原边界振铃效应抑制方法
CN106504209A (zh) * 2016-10-27 2017-03-15 西安电子科技大学 无源毫米波雷达图像的迭代重加权盲反卷积方法
CN107133923B (zh) * 2017-03-02 2020-10-27 杭州电子科技大学 一种基于自适应梯度稀疏模型的模糊图像非盲去模糊方法
CN107038685B (zh) * 2017-04-12 2019-07-12 重庆大学 一种基于多幅离焦图像的超分辨率图像重构方法
CN107590781B (zh) * 2017-08-17 2020-11-27 天津大学 基于原始对偶算法的自适应加权tgv图像去模糊方法
CN108305220B (zh) * 2017-12-29 2020-06-02 华中科技大学 一种机载红外退化图像校正方法
CN109949244B (zh) * 2019-03-21 2023-01-24 青岛大学 一种基于曲率项的水下图像盲复原变分方法
US11575865B2 (en) 2019-07-26 2023-02-07 Samsung Electronics Co., Ltd. Processing images captured by a camera behind a display
CN110400280B (zh) * 2019-08-02 2023-02-03 电子科技大学 一种基于人造信标和相位屏的大气湍流退化图像复原方法
CN110599429B (zh) * 2019-09-26 2022-09-13 河海大学常州校区 一种高能x射线图像非盲去模糊方法
CN110796616B (zh) * 2019-10-23 2022-05-10 武汉工程大学 基于范数约束和自适应加权梯度的湍流退化图像恢复方法
CN110930324A (zh) * 2019-11-12 2020-03-27 上海航天控制技术研究所 一种模糊星图复原方法
CN111047544B (zh) * 2020-01-08 2022-09-23 华中科技大学 一种基于非线性退化模型的饱和图像去模糊方法
CN111583149B (zh) * 2020-05-08 2023-02-17 西安电子科技大学 一种基于l1-l0范数最小化的气动热辐射图像自动校正方法
CN112116541B (zh) * 2020-09-24 2024-05-14 南京航空航天大学 基于梯度l0范数和总变分正则化约束的模糊图像复原方法
US11721001B2 (en) 2021-02-16 2023-08-08 Samsung Electronics Co., Ltd. Multiple point spread function based image reconstruction for a camera behind a display
US11722796B2 (en) 2021-02-26 2023-08-08 Samsung Electronics Co., Ltd. Self-regularizing inverse filter for image deblurring
CN113139920B (zh) * 2021-05-12 2023-05-12 闽南师范大学 一种古籍图像修复方法、终端设备及存储介质
CN113822813A (zh) * 2021-09-17 2021-12-21 南京信息工程大学 基于四方向交叠组合稀疏全变分的图像复原方法及系统

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101540043A (zh) * 2009-04-30 2009-09-23 南京理工大学 单幅图像复原的解析迭代快速频谱外推方法
CN102147915A (zh) * 2011-05-06 2011-08-10 重庆大学 一种权重的稀疏边缘正则化图像复原方法
CN102184533A (zh) * 2011-06-10 2011-09-14 西安电子科技大学 基于非局部约束的全变分图像去模糊方法
CN102222320A (zh) * 2011-05-24 2011-10-19 西安电子科技大学 基于全变分迭代反向投影的单帧图像空间分辨率增强方法
CN102354395A (zh) * 2011-09-22 2012-02-15 西北工业大学 基于稀疏表示的模糊图像盲复原方法
CN102682437A (zh) * 2012-05-17 2012-09-19 浙江大学 一种基于总变分正则约束的图像解卷积方法
WO2013148139A1 (en) * 2012-03-29 2013-10-03 Nikon Corporation Algorithm for minimizing latent sharp image and point spread function cost functions with spatial mask fidelity
CN103914818A (zh) * 2014-03-06 2014-07-09 中国人民解放军国防科学技术大学 一种基于全向全变分的全向图像稀疏重构方法
CN104036473A (zh) * 2014-05-30 2014-09-10 南京邮电大学 基于分裂Bregman迭代的快速鲁棒图像运动去模糊方法
CN104134196A (zh) * 2014-08-08 2014-11-05 重庆大学 基于非凸高阶全变差模型的Split Bregman权值迭代图像盲复原方法

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013116709A1 (en) * 2012-02-01 2013-08-08 The Research Foundation of States University of New York Computerized image reconstruction method and apparatus

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101540043A (zh) * 2009-04-30 2009-09-23 南京理工大学 单幅图像复原的解析迭代快速频谱外推方法
CN102147915A (zh) * 2011-05-06 2011-08-10 重庆大学 一种权重的稀疏边缘正则化图像复原方法
CN102222320A (zh) * 2011-05-24 2011-10-19 西安电子科技大学 基于全变分迭代反向投影的单帧图像空间分辨率增强方法
CN102184533A (zh) * 2011-06-10 2011-09-14 西安电子科技大学 基于非局部约束的全变分图像去模糊方法
CN102354395A (zh) * 2011-09-22 2012-02-15 西北工业大学 基于稀疏表示的模糊图像盲复原方法
WO2013148139A1 (en) * 2012-03-29 2013-10-03 Nikon Corporation Algorithm for minimizing latent sharp image and point spread function cost functions with spatial mask fidelity
CN102682437A (zh) * 2012-05-17 2012-09-19 浙江大学 一种基于总变分正则约束的图像解卷积方法
CN103914818A (zh) * 2014-03-06 2014-07-09 中国人民解放军国防科学技术大学 一种基于全向全变分的全向图像稀疏重构方法
CN104036473A (zh) * 2014-05-30 2014-09-10 南京邮电大学 基于分裂Bregman迭代的快速鲁棒图像运动去模糊方法
CN104134196A (zh) * 2014-08-08 2014-11-05 重庆大学 基于非凸高阶全变差模型的Split Bregman权值迭代图像盲复原方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
WANG, JING ET AL.: "Total Variant Image Deblurring Based on Split Bregman Method", CHINESE JOURNAL OF ELECTRONICS, vol. 40, no. 8, 31 August 2012 (2012-08-31), pages 1503 - 1508 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108615224A (zh) * 2016-12-13 2018-10-02 三星电机株式会社 图像校正设备和方法以及非易失性计算机可读存储介质
CN108198149A (zh) * 2018-01-24 2018-06-22 闽南师范大学 一种图像去模糊方法
CN112184567A (zh) * 2020-08-28 2021-01-05 江苏海洋大学 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
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
US20160321788A1 (en) 2016-11-03
US9582862B2 (en) 2017-02-28
CN104537620A (zh) 2015-04-22
CN104537620B (zh) 2017-04-12

Similar Documents

Publication Publication Date Title
WO2016106951A1 (zh) 一种方向自适应图像去模糊方法
CN108492274B (zh) 一种长波红外偏振特征提取与融合的图像增强方法
CN101441764B (zh) 一种mtfc遥感图像复原方法
CN102147915B (zh) 一种权重的稀疏边缘正则化图像复原方法
CN110675347A (zh) 一种基于组稀疏表示的图像盲复原方法
CN101540043B (zh) 单幅图像复原的解析迭代快速频谱外推方法
Dharejo et al. A color enhancement scene estimation approach for single image haze removal
Shu et al. Alternating minimization algorithm for hybrid regularized variational image dehazing
Fan et al. Noise suppression and details enhancement for infrared image via novel prior
Hong et al. Multi-frame real image restoration based on double loops with alternative maximum likelihood estimation
Nieuwenhuizen et al. Deep learning for software-based turbulence mitigation in long-range imaging
Hong et al. Blind restoration of real turbulence-degraded image with complicated backgrounds using anisotropic regularization
Yang et al. Fractional‐order tensor regularisation for image inpainting
CN111899196A (zh) 一种基于经典复原算法的叶片缺陷运动模糊图像复原方法
Halder et al. Simple algorithm for correction of geometrically warped underwater images
Karnaukhov et al. Analysis of linear distortion characteristics in problems of restoration of multispectral images
Shi et al. Accurate estimation of motion blur parameters in noisy remote sensing image
Zhu et al. A new method for superresolution image reconstruction based on surveying adjustment
Zhu et al. Passive millimeter wave image denoising based on adaptive manifolds
Halder et al. A centroid algorithm for stabilization of turbulence-degraded underwater videos
Su et al. Richardson-Lucy deblurring for the star scene under a thinning motion path
Setia et al. Image Deblurring using Wiener Filtering and Siamese Neural Network
CN112330739B (zh) 一种基于光学概率图模型的卫星探测方法
Xu et al. An object detection method for heavy fog scenes based on image defogging and sample enhancement
van Eekeren et al. Evaluation of turbulence mitigation methods

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 15022872

Country of ref document: US

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

Ref document number: 15874644

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: 15874644

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 15874644

Country of ref document: EP

Kind code of ref document: A1