WO2016183716A1 - 图像去模糊方法及系统 - Google Patents

图像去模糊方法及系统 Download PDF

Info

Publication number
WO2016183716A1
WO2016183716A1 PCT/CN2015/079039 CN2015079039W WO2016183716A1 WO 2016183716 A1 WO2016183716 A1 WO 2016183716A1 CN 2015079039 W CN2015079039 W CN 2015079039W WO 2016183716 A1 WO2016183716 A1 WO 2016183716A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
edge
model
value
kernel
Prior art date
Application number
PCT/CN2015/079039
Other languages
English (en)
French (fr)
Inventor
张欣欣
王荣刚
王振宇
高文
Original Assignee
北京大学深圳研究生院
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 北京大学深圳研究生院 filed Critical 北京大学深圳研究生院
Priority to CN201580000103.0A priority Critical patent/CN105493140B/zh
Priority to US15/571,659 priority patent/US10325358B2/en
Priority to PCT/CN2015/079039 priority patent/WO2016183716A1/zh
Publication of WO2016183716A1 publication Critical patent/WO2016183716A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/20—Image enhancement or restoration using local operators
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20112—Image segmentation details
    • G06T2207/20164—Salient point detection; Corner detection

Definitions

  • the present invention relates to the field of image enhancement, and in particular to an image deblurring method and system.
  • the model of image degradation can be expressed as the following convolution process:
  • I is the acquired blurred image
  • L is the clear image
  • k is the blur kernel (ie, the point spread function)
  • N is the noise of the image acquisition device
  • the super-Laplace constraint is widely used as a valid constraint in subsequent deblurring work; Pan et al. constrain the clear image L and the fuzzy kernel k with low rank constraint and Gaussian regular term respectively. Although the continuity of the fuzzy kernel k is guaranteed to some extent, the fuzzy kernel k is denoised by a truncation parameter. Affects the continuity of the fuzzy kernel k.
  • the selection of the significant edges of the image greatly affects the accuracy of the fuzzy kernel k estimation. It is found that the edge information larger than the blur kernel k size in the image can obtain the relatively accurate blur kernel k, and the edge smaller than the blur kernel k size makes the fuzzy kernel k estimation inaccurate. Therefore, it is very important to select the appropriate edge information for fuzzy kernel k estimation. Cho and Lee extract the sharp edges with bilateral filters and impact filters, but it is difficult to control the size of the extracted edges, which adversely affects the estimation of the fuzzy kernel k.
  • Xu and Jia proposed an edge-scale metric that effectively extracts useful significant edges. Pan et al. have improved on this basis to make the selected edges more effective. Different from the special significant edge extraction method, Xu and Pan et al. use the l 0 constraint to make the restored intermediate image L retain only the main structure and use it for fuzzy kernel k estimation, which has achieved good results.
  • an embodiment provides an image deblurring method, including the steps of:
  • Estimating the intermediate image L marking an edge region and a smooth region of the input image, respectively constraining the edge region and the smooth region, obtaining an intermediate image L, and the input image is an original blurred image;
  • Estimating the blur kernel k extracting a significant edge of the intermediate image L, the significant edge being an edge whose scale is larger than the scale of the blur kernel, and calculating the blur kernel k by using the significant edge;
  • an image deblurring system including:
  • a first estimating module configured to constrain the edge region and the smooth region according to the edge region and the smooth region of the marked input image, respectively, to obtain an intermediate image L, where the input image is an original blurred image
  • a second estimating module configured to extract a significant edge of the intermediate image L, the significant edge is an edge whose scale is larger than a scale of the fuzzy kernel, and the fuzzy kernel k is calculated by using the significant edge;
  • a recovery module for performing non-blind deconvolution based on the input image and the estimated blur kernel k to restore the input image to a sharp image.
  • the edge region and the smooth region of the input image are first marked in the process of obtaining the intermediate image L, then the edge region and the smooth region are constrained so that the obtained intermediate image can retain the edge and The noise and ringing effect of the smooth region can be effectively removed, and the fuzzy kernel is calculated by using the significant edge of the intermediate image L, so that the estimation of the fuzzy kernel is more accurate.
  • the non-blind deconvolution is performed according to the input image and the estimated fuzzy kernel k. Restore the input image to a sharp image for a good deblurring effect.
  • Figure 3 is an image restoration effect diagram
  • FIG. 4 is a schematic diagram of an image deblurring system of the present invention.
  • Figure 5 is a diagram showing an image restoration quality evaluation result
  • Figure 6 is a comparison diagram of the deblurring effect of the artificially synthesized blurred image
  • Figure 7 is a comparison of the actual image deblurring effect.
  • the deblurring of a single image mainly includes two parts: fuzzy kernel k estimation and image restoration, wherein the fuzzy kernel estimation and the intermediate image L restoration are constrained by using a robust constraint term to obtain an accurate Fuzzy kernel k and high quality restored images.
  • This example provides an image deblurring method, and the flowchart thereof is as shown in FIG. 1 , and specifically includes the following steps.
  • an image pyramid Before estimating the intermediate image L, it is necessary to first establish an image pyramid. Specifically, construct an image pyramid corresponding to the input image, and the input image is the original blurred image I; wherein the image pyramid is a structure for interpreting the image with multiple resolutions.
  • the pyramid of an image is a collection of images whose resolution is gradually reduced, and presents a pyramid shape; therefore, the image pyramid model of this example is a series of images of the blurred image I whose resolution is gradually reduced.
  • the marking process is: first, the edge region of the input image is detected by the Canny operator, the edge image is obtained, and then the edge is represented by the disk model.
  • the expansion operation is performed, the width of the fuzzy kernel is taken as the radius of the disk, and the edge region and the smooth region are respectively constrained by using the super Laplacian sparse term and the smoothing term to obtain the intermediate image L.
  • the specific process is as follows.
  • the intermediate image L estimation is performed using the degradation model of the image motion blur and the constraints on the edge region and the smooth region, wherein the calculation model of the intermediate image L is as follows:
  • a gradient representing the intermediate image L I is a blurred image
  • ⁇ 1 , ⁇ 2 are weight coefficients, where ⁇ 1 is set to 0.05, ⁇ 2 is set to 1, and ⁇ represents multiplication between matrix elements
  • J is an all-one matrix
  • M is the mark of the edge region
  • JM is the mark of the smooth region
  • the pixel value of the region is marked as a first value
  • the pixel value of the smooth region is marked as a second value different from the first value; according to the detection result, the pixel value of the image edge and its vicinity is corresponding to the element setting in M Is 1, that is, the first value is 1, the pixel value of the smooth region corresponds to the element in M is set to 0, that is, the second value is 0.
  • the first value and the second value are substituted into the calculation model of the intermediate image L, and the calculation model is subjected to a loop iterative operation to solve the minimum value of the calculation model with respect to the intermediate image L. Since the calculation model is a non-convex function, it is not possible to directly solve the intermediate image L.
  • the semi-quadratic normalization method is used to calculate the computational model. Specifically, the auxiliary variable is introduced and the auxiliary variable is substituted for the gradient of the intermediate image L.
  • the calculation model of the intermediate image L is converted into the auxiliary model; for example, by introducing the auxiliary variable u instead of the ⁇ L in the calculation model, the calculation model is converted into the following auxiliary model:
  • the auxiliary variable in the auxiliary model is taken as the unknown quantity, and the auxiliary model is reduced to the first auxiliary model, and the first auxiliary model is calculated and the auxiliary variable is calculated; for example, u in the auxiliary model is regarded as an unknown quantity, that is, in the auxiliary model, except for u
  • the external variables are all set to known quantities, and the auxiliary model is reduced to the first auxiliary model:
  • the u value can be calculated, then u is taken as the known quantity, L is the unknown quantity, that is, the intermediate image L in the auxiliary model is used as the unknown quantity, and the auxiliary model can be simplified to the second auxiliary model.
  • the L value can be calculated by minimizing the second auxiliary model.
  • the L solved by the fast Fourier transform is:
  • F and F -1 represent the fast Fourier transform and its inverse transform, respectively. Representing a conjugate operation, looping the first auxiliary model and the second auxiliary model until a suitable intermediate image L is obtained, such as a loop iteration 20 times to obtain a suitable intermediate image L.
  • Extracting the significant edge of the intermediate image L greatly affects the accuracy of the fuzzy kernel k estimation. For example, using the edge information larger than the blur kernel k size in the image can obtain a relatively accurate fuzzy kernel. k, and the edge information smaller than the k-size of the fuzzy kernel makes the estimation of the fuzzy kernel k inaccurate.
  • the significant edge of this example is the edge of the scale whose edge scale is larger than the fuzzy kernel, and the fuzzy kernel k is calculated by using the significant edge.
  • the significant edge X of the intermediate image L is extracted by using a correlation total variation (RTV) structure extraction algorithm, and the truncation parameter t is set to remove small edges and noise in the significant edge X, and an edge gradient image for performing fuzzy kernel k estimation is obtained.
  • RTV correlation total variation
  • H( ⁇ ) is a unit step function
  • the calculation model of the fuzzy kernel k is as follows:
  • the noise and block effect in Fig. 2(d) are small, and the noise of the blur kernel k is also small.
  • the non-blind deconvolution is performed according to the input image and the estimated blur kernel k, and the input image is restored to a clear image, as follows.
  • ⁇ is set to 0.001, 0.5 ⁇ ⁇ ⁇ 0.8, and the smaller ⁇ , the smoother the obtained first image L 1 .
  • the first image L 1 retains the main structure of the image, and although there is no noise and ringing effect, a lot of high frequency information is lost due to excessive smoothing, and the first image L 1 is as shown in Fig. 3(a).
  • ⁇ is set to 0.001, L 2 to obtain a second image, the second image L 2 effective to retain the high frequency information of the image, but also the presence of noise and ringing, the second image L 2 shown in FIG. 3 (b).
  • the example takes the average of the first image L 1 and the second image L 2 as the final restored clear image, that is, calculates the first image L 1 and The average value of the second image L 2 , which is a sharp image that finally restores the input image, as shown in FIG. 3( c ), the final sharp image is combined with the first image L 1 and the second image L 2 .
  • the advantages but also avoid its shortcomings, and get a high quality restored image.
  • the present example further provides an image deblurring system, and the schematic diagram thereof is as shown in FIG. 4, including:
  • the first estimating module 2 is configured to constrain the edge region and the smooth region according to the edge region and the smooth region of the marked input image to obtain the intermediate image L;
  • a second estimating module 3 configured to extract a significant edge of the intermediate image L, the significant edge is an edge whose edge scale is larger than the scale of the fuzzy kernel, and the fuzzy kernel k is calculated by using the significant edge;
  • the recovery module 4 is configured to perform non-blind deconvolution according to the input image and the estimated blur kernel k, and restore the input image to a clear image.
  • the first estimating module 2 includes: an initial unit 21, a detecting unit 22, and an arithmetic unit 23; wherein the initial unit 21 is used to initialize the blur kernel, and the detecting unit 22 is configured to perform edge detection on a series of blurred images in the image pyramid, Marking the pixel value of the edge region as the first value, marking the pixel value of the smooth region as the second value different from the first value, the operation list
  • the element 23 is configured to substitute the first value and the second value into the calculation model of the intermediate image L, and perform a loop iterative operation on the calculation model to solve the minimum value of the calculation model relative to the intermediate image L, wherein the first value and the second value respectively It is similar to the first value and the second value described above.
  • the operation unit 23 operates in the following manner: 1) introducing an auxiliary variable and replacing the gradient of the intermediate image L with the auxiliary variable, and converting the calculation model of the intermediate image L into the auxiliary model; for example, by introducing the auxiliary variable u instead of the intermediate image L ⁇ L in the calculation model, the calculation model of the intermediate image L is converted into the auxiliary model:
  • auxiliary variable in the auxiliary model is taken as the unknown quantity, the auxiliary model is reduced to the first auxiliary model, the first auxiliary model is calculated and the auxiliary variable is calculated; if u in the auxiliary model is unknown, other variables are known as the known quantity , the auxiliary model is reduced to the first auxiliary model: Similarly to the first auxiliary model described above, the arithmetic unit 23 calculates the first auxiliary model by the Newton method and calculates the u value.
  • the intermediate image L in the auxiliary model is taken as an unknown quantity, the auxiliary model is reduced to the second auxiliary model, the second auxiliary model is calculated and the intermediate image L is calculated; for example, u in the auxiliary model is used as a known quantity, L is taken as For unknowns, the auxiliary model is reduced to the second auxiliary model: Similar to the second auxiliary model described above, the arithmetic unit 23 minimizes the second auxiliary model to solve L, and the specific solution manner refers to the above-described L solving method.
  • the restoration module 4 includes a filtering unit 41, a first image calculation unit 42, a second image calculation unit 43, and a restoration unit 44.
  • the filtering unit 41 is configured to perform impact filtering on the blurred image I to obtain an image.
  • the purpose is to enhance the edge of the blurred image I.
  • the first image calculation unit 42 is configured to estimate the blurred kernel k and the image using the non-blind deconvolution algorithm of the super Laplace model. Perform deconvolution, which is to minimize the equation: A first image L 1 is obtained .
  • the second image calculation unit 43 is configured to estimate the blurred kernel k and image using the non-blind deconvolution algorithm of the TV-L 2 model. Perform deconvolution, which is to minimize the equation: A second image L 2 is obtained .
  • the restoration unit 44 is for calculating an average value of the first image L 1 and the second image L 2 , and finally restores the input image to a sharp image with the average value.
  • the image deblurring system of this example has a one-to-one correspondence with the image deblurring method described above, and the specific working modes of each module of the system are not described in detail.
  • this example uses a special non-reference de-motion fuzzy evaluation algorithm to evaluate the artificially synthesized motion blurred image set and the actual captured motion blurred image, and the existing deblurring is better.
  • the method compares, and the higher the score, the higher the quality of the restored image.
  • the scores of images A and B are -8 and -10, respectively, indicating that the quality of image A is higher than the quality of image B.
  • the image restoration quality evaluation result is shown in FIG. 5, wherein the first column in FIGS.
  • 5(a) and 5(b) is the average score of the restored image of the present invention, and it can be seen that for each data set, The average score of the restored image of the present invention is the highest, indicating that the image deblurring method of the present invention has a high quality of the restored image.
  • Figure 6 is a comparison of the deblurring results of a picture in a synthetic data set by different deblurring algorithms
  • Figure 7 is a comparison of the deblurring effects of the actually captured motion blurred pictures
  • Figures 6(a) and 7( a) are blurred images
  • Figure 6 (b) is the actual image of the blurred image of Figure 6 (a)
  • Figure 7 (b) is the enlarged image of the rectangular frame of Figure 7 (a)
  • Figure 6 (c) ⁇ 6(k) and 7(c) to 7(k) are the results of the existing 9 image deblurring methods for restoring images
  • FIG. 6(l) and FIG. 7(l) are image deblurring algorithms of the present invention. the result of. It can be intuitively seen from FIG. 6 and FIG. 7 that the image deblurring algorithm of the present invention has better detail retention characteristics and minimizes noise and ringing effects.

Landscapes

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

Abstract

一种图像去模糊方法，包括：估算中间图像L：标记输入图像的边缘区域和平滑区域，分别对边缘区域和平滑区域进行约束，获得中间图像L；估算模糊核k：提取中间图像L的显著边缘，显著边缘为边缘尺度大于模糊核的尺度的边缘，利用显著边缘计算模糊核k；复原输入图像：根据输入图像和估算模糊核k进行非盲反卷积，将输入图像复原成清晰图像。由于对边缘区域和平滑区域进行约束，使得获得的中间图像即能够保留边缘又能有效去除平滑区域的噪声和振铃效应，并利用中间图像L的显著边缘计算模糊核，使得模糊核的估算更加准确，最后，根据输入图像和估算的模糊核k进行非盲反卷积，将输入图像复原成清晰图像，达到了很好的去模糊效果。

Description

图像去模糊方法及系统 技术领域
本发明涉及图像增强领域，具体涉及图像去模糊方法及系统。
背景技术
由于相机和拍摄场景之间的相对运动，获取到的图像常常存在一定程度的运动模糊。图像退化的模型可以表示为如下的卷积过程：
Figure PCTCN2015079039-appb-000001
其中，I是获取到的模糊图像，L是清晰图像，k是模糊核(即点扩散函数)，N是图像获取设备的噪声，由于只有模糊图像I是已知量，所以，对模糊图像I进行复原得到一个比较清晰图像L的过程，是一个大型的病态的反问题。
由于单张图像去运动模糊具有重要的应用价值，目前，其已受到广泛关注，且有大量的去运动模糊算法解决求解清晰图像L的病态问题，如：Fergus等人将自然图像梯度的长尾分布模型表示为高斯混合模型，并用此模型对清晰图像L进行约束，用集成学习的方法得到清晰图像L；Shan等人把长尾分布表示为一个分段函数，对清晰图像L的梯度进行约束；Krishnan等人假设清晰图像L的梯度服从超拉普拉斯分布，取得了高质量的复原图像，超拉普拉斯约束项作为一个有效的约束条件被广泛应用于后来的去模糊工作中；Pan等人用低秩约束项和高斯正则项分别对清晰图像L和模糊核k进行约束，虽然在一定程度上保证了模糊核k的连续性，但是通过一个截断参数对模糊核k去噪，影响了模糊核k的连续性。
另外，图像显著边缘的选取在很大程度上影响了模糊核k估计的准确程度。研究发现利用图像中比模糊核k尺寸大的边缘信息才能够得到相对准确的模糊核k，而比模糊核k尺寸小的边缘会使得模糊核k估计不准确。所以选取合适的边缘信息进行模糊核k估计是非常重要的。Cho和Lee用双边滤波器和冲击滤波器提取锐利边缘，但是难以控制提取出的边缘的大小，对模糊核k估计造成了不利影响。Xu和Jia提出了一种边缘尺度的度量方法，能够有效提取出有用的显著边缘。Pan等人在此基础上进行了改进，使得选取的边缘更加有效。与用专门的显著边缘提 取方法不同的是，Xu和Pan等人用l0约束项使复原出的中间图像L只保留主要结构，用于模糊核k估计，取得了不错的效果。
然而，无论是模糊核k估计还是最终的图像非盲反卷积复原都存在很多问题，需要进一步提高图像去模糊的效果。
发明内容
根据第一方面，一种实施例中提供一种图像去模糊方法，包括步骤：
估算中间图像L：标记输入图像的边缘区域和平滑区域，分别对所述边缘区域和平滑区域进行约束，获得中间图像L，输入图像为原始的模糊图像；
估算模糊核k：提取所述中间图像L的显著边缘，所述显著边缘为边缘尺度大于模糊核的尺度的边缘，利用所述显著边缘计算模糊核k；
复原输入图像：根据输入图像和估算的模糊核k进行非盲反卷积，将输入图像复原成清晰图像。
根据第二方面，一种实施例中提供一种图像去模糊系统，包括：
第一估算模块，用于根据标记的输入图像的边缘区域和平滑区域，分别对所述边缘区域和平滑区域进行约束，获得中间图像L，输入图像为原始的模糊图像；
第二估算模块，用于提取所述中间图像L的显著边缘，所述显著边缘为所述边缘尺度大于模糊核的尺度的边缘，利用所述显著边缘计算模糊核k；
恢复模块，用于根据输入图像和估算的模糊核k进行非盲反卷积，将输入图像复原成清晰图像。
依据上述实施例的图像去模糊方法，由于获得中间图像L的过程中，先标记输入图像的边缘区域和平滑区域，然后对边缘区域和平滑区域进行约束，使得获得的中间图像既能够保留边缘又能有效去除平滑区域的噪声和振铃效应，并利用中间图像L的显著边缘计算模糊核，使得模糊核的估算更加准确，最后，根据输入图像和估算的模糊核k进行非盲反卷积，将输入图像复原成清晰图像，达到了很好的去模糊效果。
附图说明
图1为本发明的图像去模糊方法流程图；
图2为模糊核平滑约束项效果图；
图3为图像复原效果图；
图4为本发明的图像去模糊系统原理图；
图5为图像复原质量评价结果图；
图6为人工合成模糊图像去模糊效果对比图；
图7为实际图像去模糊效果对比图。
具体实施方式
在本发明实施例中，对单张图像去模糊，主要包括模糊核k估计和图像复原两部分，其中，利用鲁棒的约束项对模糊核k估计和中间图像L复原进行约束，以得到准确的模糊核k和高质量的复原图像。
本例提供一种图像去模糊方法，其流程图如图1所示，具体包括以下步骤。
S1：估算中间图像L。
在估算中间图像L之前，需要先建立图像金字塔，具体的，构建输入图像所对应的图像金字塔，输入图像为原始的模糊图像I；其中，图像金字塔是以多分辨率来解释图像的一种结构，一幅图像的金字塔是一系列分辨率逐步降低的图像集合，呈现金字塔形状；所以，本例的图像金字塔模型是一系列分辨率逐步降低的模糊图像I的图像集合。
建立图像金字塔的同时，还需要初始化模糊核k，将模糊核k初始化为一个3*3的矩阵。
在估算中间图像L的过程中，首先，标记出输入图像的边缘区域和平滑区域，标记过程为：首先用Canny算子检测出输入图像的边缘区域，得到边缘图，然后用圆盘模型对边缘图进行膨胀运算，模糊核的宽作为圆盘半径，并利用超拉普拉斯稀疏项和平滑项分别对该边缘区域和平滑区域进行约束，以获得中间图像L，具体过程如下。
利用图像运动模糊的退化模型，以及对边缘区域和平滑区域的约束进行中间图像L估计，其中，中间图像L的计算模型如下所示：
计算模型：
Figure PCTCN2015079039-appb-000002
其中，
Figure PCTCN2015079039-appb-000003
表示中间图像L的梯度，I是模糊图像，λ1，λ2是权重系数，其中，λ1设置为0.05，λ2设置为1，ο表示矩阵元素之间相乘，
Figure PCTCN2015079039-appb-000004
表示卷积运算，J是全1矩阵，M是边缘区域的标记，J-M是平滑区域的标记；并利用Canny算子对图像金字塔中的一系列模糊图像 进行边缘检测，并进行膨胀运算，将边缘区域的像素值标记为第一值，将平滑区域的像素值标记为不同于第一值的第二值；根据检测结果，将图像边缘及其附近区域的像素值相对应于M中的元素设为1，即第一值为1，平滑区域的像素值相对应于M中的元素设为0，即第二值为0。
将第一值和第二值代入中间图像L的计算模型，对计算模型进行循环迭代运算，求解计算模型相对于中间图像L的最小值。由于计算模型是一个非凸函数，不能够直接进行中间图像L的求解运算，本例采用半二次规整化方法来运算计算模型，具体的，引入辅助变量并将辅助变量代替中间图像L的梯度，将中间图像L的计算模型转化为辅助模型；如，引入辅助变量u代替计算模型中的▽L，则将计算模型转化为如下的辅助模型：
辅助模型：
Figure PCTCN2015079039-appb-000005
当权重系数λ3趋近于∞时，辅助模型的解就收敛于计算模型，然后通过循环迭代的方法求解出辅助模型中的u和L即可，其求解过程如下：
将辅助模型中的辅助变量作为未知量，辅助模型简化为第一辅助模型，运算第一辅助模型并计算出辅助变量；如，将辅助模型中的u作为未知量，即辅助模型中除u之外的变量全部设为已知量，则辅助模型简化为第一辅助模型：
第一辅助模型：
Figure PCTCN2015079039-appb-000006
利用牛顿法运算第一辅助模型即可计算出u值，然后把u作为已知量，L作为未知量，即将辅助模型中的中间图像L作为未知量，辅助模型可以简化为第二辅助模型，运算第二辅助模型并计算出中间图像L：
第二辅助模型：
Figure PCTCN2015079039-appb-000007
通过将第二辅助模型最小化即可计算出L值，本例基于帕斯瓦尔定理，利用快速傅里叶变换求解出的L为：
Figure PCTCN2015079039-appb-000008
其中，F和F-1分别表示快速傅里叶变换及其逆变换运算，
Figure PCTCN2015079039-appb-000009
表示共轭运算，循环迭代第一辅助模型和第二辅助模型，直至获取合适的中 间图像L，如循环迭代20次就可以得到一个合适的中间图像L。
S2：估算模糊核k。
提取中间图像L的显著边缘，由于显著边缘的选取在很大程度上影响了模糊核k估计的准确程度，如，利用图像中比模糊核k尺寸大的边缘信息才能够得到相对准确的模糊核k，而比模糊核k尺寸小的边缘信息会使得模糊核k估计不准确，本例的显著边缘为边缘尺度大于模糊核的尺度的边缘，并利用显著边缘计算模糊核k。
具体的，利用相关总变差(RTV)结构提取算法提取中间图像L的显著边缘X，设置截断参数t去除显著边缘X中的细小边缘和噪声，获得用于进行模糊核k估计的边缘梯度图像▽S；
Figure PCTCN2015079039-appb-000010
式中，H(·)为单位阶跃函数；
利用边缘梯度图像▽S，同时，加入稀疏约束和平滑约束计算出模糊核k，模糊核k的计算模型如下：
Figure PCTCN2015079039-appb-000011
其中，
Figure PCTCN2015079039-appb-000012
是高斯正则项，用来保证模糊核k的稀疏性，
Figure PCTCN2015079039-appb-000013
是梯度稀疏项，可以在保证模糊核k的连续性的前提下去除模糊核k中的噪声，权值γ1和γ2分别控制模糊核k的稀疏性和平滑性，其中，γ1设置为2，通过快速傅里叶变换运算上述能量方程，计算出的模糊核k为：
Figure PCTCN2015079039-appb-000014
本例结合稀疏约束和平滑约束估计出的模糊核k同时保证了稀疏性和连续性，并且减小了噪声，如图2所示，其中，图2(a)是模糊图像I，图2(b)和图2(c)是当γ2＝0时复原出来的图像及模糊核k，图2(d)和图2(e)是当γ2＝50时复原出来的图像及模糊核k，通过比较可以看出图2(d)中的噪声和块效应较小，且模糊核k的噪声也比较小。
S3：复原输入图像。
根据输入图像和估算的模糊核k进行非盲反卷积，将输入图像复原成清晰图像，具体过程如下。
对模糊图像I进行冲击滤波，获得图像
Figure PCTCN2015079039-appb-000015
目的是增强模糊图像I的边缘。
然后，利用超拉普拉斯模型的非盲反卷积算法对估算的模糊核k和图像
Figure PCTCN2015079039-appb-000016
进行反卷积，即最小化方程：
Figure PCTCN2015079039-appb-000017
ρ设置为0.001，0.5<α<0.8，α越小，得到的第一图像L1越平滑。第一图像L1保留了图像的主要结构，虽然没有噪声和振铃效应，但是由于过度平滑也丢失了很多高频信息，第一图像L1如图3(a)所示。
再利用TV-L2模型的非盲反卷积算法对估算的模糊核k和图像
Figure PCTCN2015079039-appb-000018
进行反卷积，即最小化方程：
Figure PCTCN2015079039-appb-000019
μ设置为0.001，
Figure PCTCN2015079039-appb-000020
得到第二图像L2，第二图像L2有效的保留了图像的高频信息，但是同时也存在噪声和振铃效应，第二图像L2如图3(b)所示。
为了保留清晰图像中高频信息的同时去除噪声和振铃效应，本例采取用第一图像L1和第二图像L2的平均值作为最终复原出来的清晰图像，即计算第一图像L1和第二图像L2的平均值，该平均值即为将输入图像最终复原出来的清晰图像，如图3(c)所示，最终清晰图像结合了第一图像L1和第二图像L2的优点，又避免了其缺点，得到了高质量的复原图像。
依据上述的图像去模糊方法，本例还提供一种图像去模糊系统，其原理图如图4所示，包括：
创建模块1，用于构建输入图像所对应的图像金字塔，输入图像为原始的模糊图像；
第一估算模块2，用于根据标记的输入图像的边缘区域和平滑区域，分别对所述边缘区域和平滑区域进行约束，获得中间图像L；
第二估算模块3，用于提取所述中间图像L的显著边缘，显著边缘为边缘尺度大于模糊核的尺度的边缘，利用所述显著边缘计算模糊核k；
恢复模块4，用于根据输入图像和估算的模糊核k进行非盲反卷积，将输入图像复原成清晰图像。
进一步，第一估算模块2包括：初始单元21、检测单元22和运算单元23；其中，初始单元21用于初始化模糊核，检测单元22用于对图像金字塔中的一系列模糊图像进行边缘检测，将边缘区域的像素值标记为第一值，将平滑区域的像素值标记为不同于第一值的第二值，运算单 元23用于将第一值和第二值代入中间图像L的计算模型，对计算模型进行循环迭代运算，求解计算模型相对于中间图像L的最小值，其中，第一值和第二值分别与上述的第一值和第二值类同。
具体的，运算单元23的工作方式为：1)引入辅助变量并将辅助变量代替中间图像L的梯度，将中间图像L的计算模型转化为辅助模型；如，通过引入辅助变量u代替中间图像L的计算模型中的▽L，将中间图像L的计算模型转化为辅助模型：
计算模型：
Figure PCTCN2015079039-appb-000021
与上述的中间图像L的计算模型类同，具体参数不作赘述；
辅助模型：
Figure PCTCN2015079039-appb-000022
与上述的辅助模型类同，具体参数不作赘述。
2)将辅助模型中的辅助变量作为未知量，辅助模型简化为第一辅助模型，运算第一辅助模型并计算出辅助变量；如将辅助模型中的u作为未知量，其他变量作为已知量，则辅助模型简化为第一辅助模型：
Figure PCTCN2015079039-appb-000023
与上述的第一辅助模型类同，运算单元23利用牛顿法运算第一辅助模型并计算出u值。
3)将辅助模型中的中间图像L作为未知量，辅助模型简化为第二辅助模型，运算第二辅助模型并计算出中间图像L；如，将辅助模型中的u作为已知量，L作为未知量，则辅助模型简化为第二辅助模型：
Figure PCTCN2015079039-appb-000024
与上述的第二辅助模型类同，运算单元23将第二辅助模型最小化以求解L，具体求解方式参照上述L的求解方法。
4)循环迭代第一辅助模型和第二辅助模型，直至获取合适的中间图像L，迭代次数为20次。
恢复模块4包括滤波单元41、第一图像计算单元42、第二图像计算单元43和复原单元44。
其中，滤波单元41用于对模糊图像I进行冲击滤波，获得图像
Figure PCTCN2015079039-appb-000025
目的是增强模糊图像I的边缘。
第一图像计算单元42用于利用超拉普拉斯模型的非盲反卷积算法对估算的模糊核k和图像
Figure PCTCN2015079039-appb-000026
进行反卷积，即最小化方程：
Figure PCTCN2015079039-appb-000027
得到第一图像L1。
第二图像计算单元43用于利用TV-L2模型的非盲反卷积算法对估算的模糊核k和图像
Figure PCTCN2015079039-appb-000028
进行反卷积，即最小化方程：
Figure PCTCN2015079039-appb-000029
得到第二图像L2。
复原单元44用于计算第一图像L1和第二图像L2的平均值，并以该平均值将输入图像最终复原出清晰图像。
本例的图像去模糊系统与上述的图像去模糊方法是一一对应的，系统的各个模块的具体工作方式不作详细赘述。
为了检验本发明的有效性，本例用专门的无参考去运动模糊评价算法在人工合成的运动模糊图片集和实际拍摄的运动模糊图像上进行评价，与现有的去模糊较好的9种方法进行对比，得分越高则说明复原图像的质量越高。例如，图像A和B的得分分别为-8和-10，说明图像A的质量高于图像B的质量。图像复原质量评价结果如图5所示，其中，图5(a)和图5(b)中的第一列均为本发明的复原图像的平均得分，可以看到，对于每一个数据集，本发明的复原图像的平均得分都是最高的，说明本发明的图像去模糊方法具有较高的复原图像的质量。
图6是不同去模糊算法对人工合成数据集中的一张图片的去模糊结果的对比，图7是对实际拍摄的运动模糊图片的去模糊效果的对比图，其中图6(a)和7(a)均为模糊图像，图6(b)为图6(a)模糊图像的实际图像，图7(b)为图7(a)中矩形框中的图像放大图，图6(c)～图6(k)及图7(c)～图7(k)分别是现有9种图像去模糊方法复原图像的结果，图6(l)和图7(l)为本发明图像去模糊算法的结果。从图6和图7可以直观的看出本发明的图像去模糊算法具有较好的细节保持特性，并且噪声和振铃效应最少。
本领域技术人员可以理解，上述实施方式中各种方法的全部或部分步骤可以通过程序来指令相关硬件完成，该程序可以存储于一计算机可读存储介质中，存储介质可以包括：只读存储器、随机存储器、磁盘或光盘等。
以上应用了具体个例对本发明进行阐述，只是用于帮助理解本发明，并不用以限制本发明。对于本领域的一般技术人员，依据本发明的思想，可以对上述具体实施方式进行变化。

Claims (9)

  1. 一种图像去模糊方法，其特征在于，包括步骤：
    估算中间图像L：标记输入图像的边缘区域和平滑区域，分别对所述边缘区域和平滑区域进行约束，获得中间图像L，输入图像为原始的模糊图像；
    估算模糊核k：提取所述中间图像L的显著边缘，所述显著边缘为边缘尺度大于模糊核的尺度的边缘，利用所述显著边缘计算模糊核k；
    复原输入图像：根据所述输入图像和估算的模糊核k进行非盲反卷积，将所述输入图像复原成清晰图像。
  2. 如权利要求1所述的方法，其特征在于，所述估算中间图像L之前，还包括构建所述输入图像所对应的图像金字塔模型，估算中间图像L包括：
    初始化模糊核；
    对图像金字塔中的一系列模糊图像进行边缘检测，将边缘区域的像素值标记为第一值，将平滑区域的像素值标记为不同于第一值的第二值；
    将第一值和第二值代入中间图像L的计算模型，对计算模型进行循环迭代运算，求解计算模型相对于中间图像L的最小值。
  3. 如权利要求2所述的方法，其特征在于，所述对计算模型进行循环迭代运算步骤为：
    引入辅助变量并将所述辅助变量代替所述中间图像L的梯度，将所述中间图像L的计算模型转化为辅助模型；
    将所述辅助模型中的辅助变量作为未知量，所述辅助模型简化为第一辅助模型，运算所述第一辅助模型并计算出所述辅助变量；
    将所述辅助模型中的中间图像L作为未知量，所述辅助模型简化为第二辅助模型，运算所述第二辅助模型并计算出所述中间图像L；
    循环迭代所述第一辅助模型和第二辅助模型，直至计算出合适的中间图像L。
  4. 如权利要求3所述的方法，其特征在于，所述估算模糊核k的具体步骤为：
    利用相关总变差结构提取算法提取所述中间图像L的显著边缘X，设置截断参数t去除显著边缘X中的细小边缘和噪声，获得用于进行模糊核k估计的边缘梯度图像
    Figure PCTCN2015079039-appb-100001
    利用所述边缘梯度图像
    Figure PCTCN2015079039-appb-100002
    同时，加入稀疏约束和平滑约束计算出模糊核k。
  5. 如权利要求4所述的方法，其特征在于，所述复原输入图像的具体步骤为：
    对模糊图像I进行冲击滤波，获得图像
    Figure PCTCN2015079039-appb-100003
    利用超拉普拉斯模型的非盲反卷积算法对估算的模糊核k和图像
    Figure PCTCN2015079039-appb-100004
    进行反卷积，得到第一图像L1；
    利用TV-L2模型的非盲反卷积算法对估算的模糊核k和图像
    Figure PCTCN2015079039-appb-100005
    进行反卷积，得到第二图像L2；
    计算第一图像L1和第二图像L2的平均值，所述平均值为将输入图像最终复原出来的清晰图像。
  6. 一种图像去模糊系统，其特征在于，包括：
    第一估算模块，用于根据标记的输入图像的边缘区域和平滑区域，分别对所述边缘区域和平滑区域进行约束，获得中间图像L，输入图像为原始的模糊图像；
    第二估算模块，用于提取所述中间图像L的显著边缘，所述显著边缘为所述边缘区域中大于给定边缘尺度的区域，利用所述显著边缘计算模糊核k；
    恢复模块，用于根据所述输入图像和估算的模糊核k进行非盲反卷积，将所述输入图像复原成清晰图像。
  7. 如权利要求6所述的系统，其特征在于，还包括创建模块，用于构建输入图像所对应的图像金字塔。
  8. 如权利要求7所述的系统，其特征在于，所述第一估算模块包括：
    初始单元，用于初始化模糊核；
    检测单元，用于对图像金字塔中的一系列模糊图像进行边缘检测，将边缘区域的像素值标记为第一值，将平滑区域的像素值标记为不同于第一值的第二值；
    运算单元，用于将第一值和第二值代入中间图像L的计算模型，对 计算模型进行循环迭代运算，求解计算模型相对于中间图像L的最小值。
  9. 如权利要求8所述的系统，其特征在于，所述恢复模块包括：
    滤波单元，用于对模糊图像I进行冲击滤波，获得图像
    Figure PCTCN2015079039-appb-100006
    第一图像计算单元，用于利用超拉普拉斯模型的非盲反卷积算法对估算的模糊核k和图像
    Figure PCTCN2015079039-appb-100007
    进行反卷积，得到第一图像L1；
    第二图像计算单元，用于利用TV-L2模型的非盲反卷积算法对估算的模糊核k和图像
    Figure PCTCN2015079039-appb-100008
    进行反卷积，得到第二图像L2；
    复原单元，用于计算第一图像L1和第二图像L2的平均值，并以所述平均值将输入图像最终复原出清晰图像。
PCT/CN2015/079039 2015-05-15 2015-05-15 图像去模糊方法及系统 WO2016183716A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CN201580000103.0A CN105493140B (zh) 2015-05-15 2015-05-15 图像去模糊方法及系统
US15/571,659 US10325358B2 (en) 2015-05-15 2015-05-15 Method and system for image de-blurring
PCT/CN2015/079039 WO2016183716A1 (zh) 2015-05-15 2015-05-15 图像去模糊方法及系统

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2015/079039 WO2016183716A1 (zh) 2015-05-15 2015-05-15 图像去模糊方法及系统

Publications (1)

Publication Number Publication Date
WO2016183716A1 true WO2016183716A1 (zh) 2016-11-24

Family

ID=55678496

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/079039 WO2016183716A1 (zh) 2015-05-15 2015-05-15 图像去模糊方法及系统

Country Status (3)

Country Link
US (1) US10325358B2 (zh)
CN (1) CN105493140B (zh)
WO (1) WO2016183716A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105741243A (zh) * 2016-01-27 2016-07-06 北京航空航天大学 一种模糊图像复原方法
CN108399620A (zh) * 2018-03-19 2018-08-14 嘉兴学院 一种基于低秩稀疏矩阵分解的图像质量评价方法
CN112561819A (zh) * 2020-12-17 2021-03-26 温州大学 一种基于自支持模型的自滤波图像去雾算法
CN113222853A (zh) * 2021-05-26 2021-08-06 武汉博宇光电系统有限责任公司 一种基于噪声估计的渐进式红外图像降噪方法
CN113344800A (zh) * 2020-03-02 2021-09-03 通用汽车环球科技运作有限责任公司 用于训练非盲图像去模糊模块的系统和方法

Families Citing this family (26)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106056556B (zh) * 2016-06-07 2018-12-07 浙江大学 一种块效应抑制的降质模糊图像非盲复原方法
CN106339996B (zh) * 2016-09-09 2018-11-30 江南大学 一种基于超拉普拉斯先验的图像盲去模糊方法
CN107292842B (zh) * 2017-06-15 2020-08-07 北京大学深圳研究生院 基于先验约束和离群值抑制的图像去模糊方法
CN107274370A (zh) * 2017-06-19 2017-10-20 南京末梢信息技术有限公司 一种基于交替求极值算法的快速图像去模糊方法
CN107403414B (zh) * 2017-07-14 2018-11-02 华中科技大学 一种利于模糊核估计的图像区域选择方法和系统
CN108257096A (zh) * 2017-12-28 2018-07-06 三维通信股份有限公司 一种基于多先验协同的盲去模糊方法
CN108230276B (zh) * 2018-02-06 2020-06-02 江西理工大学 一种融合无参图像质量评价的自然场景图像去模糊方法
CN108550125B (zh) * 2018-04-17 2021-07-30 南京大学 一种基于深度学习的光学畸变修正方法
CN109934775B (zh) * 2018-11-20 2022-12-27 腾讯科技（深圳）有限公司 一种图像处理、模型训练、方法、装置和存储介质
CN109495759B (zh) * 2018-11-27 2021-04-27 深圳大学 图像处理方法及装置
CN109685750A (zh) * 2018-12-14 2019-04-26 厦门美图之家科技有限公司 图像增强方法及计算设备
CN109934778B (zh) * 2019-01-30 2024-02-23 长视科技股份有限公司 一种家用监控视频截图的盲去模糊方法
CN110503611A (zh) * 2019-07-23 2019-11-26 华为技术有限公司 图像处理的方法和装置
CN110766641B (zh) * 2019-11-08 2023-05-09 上海健康医学院 结合分数阶总变分和非凸总变分的模糊核估计方法及应用
CN110852973B (zh) * 2019-11-12 2022-09-23 华中科技大学 一种脉冲噪声模糊图像非线性复原方法及系统
CN113450264A (zh) * 2020-03-27 2021-09-28 上海途擎微电子有限公司 去振铃效应的方法及其相关装置
CN111815537B (zh) * 2020-07-16 2022-04-29 西北工业大学 一种新型图像盲解去模糊方法
CN112085674B (zh) * 2020-08-27 2022-08-26 河海大学 基于神经网络的航拍图像去模糊算法
CN112634163A (zh) * 2020-12-29 2021-04-09 南京大学 基于改进型循环生成对抗网络去图像运动模糊方法
CN113592728B (zh) * 2021-07-01 2024-04-05 温州理工学院 一种图像复原方法、系统、处理终端及计算机介质
TWI783721B (zh) * 2021-10-08 2022-11-11 瑞昱半導體股份有限公司 決定去模糊化濾波器係數的方法
CN113822823B (zh) * 2021-11-17 2022-03-15 武汉工程大学 气动光学效应图像空变模糊核的点近邻复原方法及系统
CN113935925B (zh) * 2021-12-17 2022-03-15 武汉工程大学 气动光学效应空变模糊图像复原方法及系统
CN114708166A (zh) * 2022-04-08 2022-07-05 Oppo广东移动通信有限公司 图像处理方法、装置、存储介质以及终端
CN116309190B (zh) * 2023-05-17 2023-08-15 武汉工程大学 一种基于最佳区域中值先验的湍流退化图像恢复方法
CN117173058B (zh) * 2023-11-03 2024-02-02 武汉工程大学 空变模糊图像统一复原方法及系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7433086B2 (en) * 2004-09-27 2008-10-07 General Electric Company Edge detection and correcting system and method
CN102231204A (zh) * 2011-06-20 2011-11-02 南京航空航天大学 一种序列图像自适应正则超分辨率重建方法
CN103413277A (zh) * 2013-08-19 2013-11-27 南京邮电大学 基于l0稀疏先验的盲相机抖动去模糊方法

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100990791B1 (ko) * 2008-12-31 2010-10-29 포항공과대학교 산학협력단 영상의 블러 제거 방법 및 블러 제거 방법이 기록된 기록매체
US9076205B2 (en) * 2012-11-19 2015-07-07 Adobe Systems Incorporated Edge direction and curve based image de-blurring

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7433086B2 (en) * 2004-09-27 2008-10-07 General Electric Company Edge detection and correcting system and method
CN102231204A (zh) * 2011-06-20 2011-11-02 南京航空航天大学 一种序列图像自适应正则超分辨率重建方法
CN103413277A (zh) * 2013-08-19 2013-11-27 南京邮电大学 基于l0稀疏先验的盲相机抖动去模糊方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
LI, XIN'GANG;: "Blind Deblurring from Single Motion Blurred Image", CHINA MASTER'S THESES FULL-TEXT DATABASE, 16 July 2013 (2013-07-16) *
LI, XU ET AL.: "Two-Phase Kernel Estimation for Robust Motion Deblurring", COMPUTER VISION-ECCV 2010, 5 September 2010 (2010-09-05), XP019150505 *

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105741243A (zh) * 2016-01-27 2016-07-06 北京航空航天大学 一种模糊图像复原方法
CN105741243B (zh) * 2016-01-27 2018-05-11 北京航空航天大学 一种模糊图像复原方法
CN108399620A (zh) * 2018-03-19 2018-08-14 嘉兴学院 一种基于低秩稀疏矩阵分解的图像质量评价方法
CN108399620B (zh) * 2018-03-19 2022-04-19 嘉兴学院 一种基于低秩稀疏矩阵分解的图像质量评价方法
CN113344800A (zh) * 2020-03-02 2021-09-03 通用汽车环球科技运作有限责任公司 用于训练非盲图像去模糊模块的系统和方法
CN113344800B (zh) * 2020-03-02 2023-09-29 通用汽车环球科技运作有限责任公司 用于训练非盲图像去模糊模块的系统和方法
CN112561819A (zh) * 2020-12-17 2021-03-26 温州大学 一种基于自支持模型的自滤波图像去雾算法
CN113222853A (zh) * 2021-05-26 2021-08-06 武汉博宇光电系统有限责任公司 一种基于噪声估计的渐进式红外图像降噪方法

Also Published As

Publication number Publication date
CN105493140A (zh) 2016-04-13
US10325358B2 (en) 2019-06-18
US20180137606A1 (en) 2018-05-17
CN105493140B (zh) 2018-02-02

Similar Documents

Publication Publication Date Title
WO2016183716A1 (zh) 图像去模糊方法及系统
Pan et al. Kernel estimation from salient structure for robust motion deblurring
Michaeli et al. Blind deblurring using internal patch recurrence
Zhong et al. Handling noise in single image deblurring using directional filters
US8380000B2 (en) Methods of deblurring image and recording mediums having the same recorded thereon
Wang et al. Recent progress in image deblurring
US9692939B2 (en) Device, system, and method of blind deblurring and blind super-resolution utilizing internal patch recurrence
Cho et al. Fast motion deblurring
Sun et al. Gradient profile prior and its applications in image super-resolution and enhancement
Mao et al. Non rigid geometric distortions correction-application to atmospheric turbulence stabilization
US9589328B2 (en) Globally dominant point spread function estimation
US20140355901A1 (en) Algorithm for minimizing latent sharp image cost function and point spread function cost function with a spatial mask in a regularization term
Li et al. A maximum a posteriori estimation framework for robust high dynamic range video synthesis
Zhang et al. Image deblurring using robust sparsity priors
US9008453B2 (en) Blur-kernel estimation from spectral irregularities
Lee et al. Recent advances in image deblurring
Dong et al. Blind image deconvolution using the Fields of Experts prior
KR101152525B1 (ko) 모션 블러 제거 방법 및 이러한 방법을 수행하는 장치
Han et al. A nonblind deconvolution method by bias correction for inaccurate blur kernel estimation in image deblurring
Chen et al. Blind restoration for nonuniform aerial images using nonlocal Retinex model and shearlet-based higher-order regularization
CN110121016B (zh) 一种基于双重曝光先验的视频去模糊方法及装置
Li et al. Joint motion deblurring with blurred/noisy image pair
Zhang Image motion deblurring based on salient structure selection and L0-2 norm kernel estimation
Feng et al. A restoration method for the turbulent degraded images based on the salient edge selection and the L0 norm constraint
Xie et al. Distortion-driven turbulence effect removal using variational model

Legal Events

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

Ref document number: 201580000103.0

Country of ref document: CN

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

Ref document number: 15892097

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 15571659

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 04/05/2018)

122 Ep: pct application non-entry in european phase

Ref document number: 15892097

Country of ref document: EP

Kind code of ref document: A1