CN105957024A - Blind deblurring method based on image block prior and sparse norm - Google Patents

Abstract

The invention provides a blind deblurring method based on image block prior and sparse norm, mainly for solving the problem of poor image deblurring quality of the prior art. According to the technical scheme, the blind deblurring method comprises: inputting a blurred image; initializing a fuzzy kernel, a binary mask and a candidate image; calling a pyramid model, downsampling the candidate image according to the number of pyramid layers, and upsampling the candidate image and the fuzzy kernel; updating the binary mask, updating an image block variance, and updating an image sample block; keeping parameters unchanged, and updating the fuzzy kernel and the candidate image, until the last layer of the pyramid; setting the number of iterations, keeping the fuzzy kernel and the norm of the candidate image unchanged, and performing regularity of an l1 norm which is added to the fuzzy kernel, so as to obtain a new candidate image; keeping the candidate image unchanged, and performing regularity of an l1/l2 norm which is added to the candidate image, so as to obtain a new fuzzy kernel, until iteration to the highest number of iteration. The invention enhances the effect and the robustness of blind deblurring and can be used for medical equipment, computer vision and image video processing.

Description

Blind deblurring method based on image block priori Yu sparse norm

Technical field

The invention belongs to technical field of image processing, particularly to a kind of Image Blind deblurring method, can be used for Aero-Space, doctor Treat apparatus, computer vision and image/video to process.

Background technology

Along with individual smart machine be widely popularized use, one photographing unit of staff comes true already.Intelligentized at this In the epoch, people are not coming into contacts with picture, little to the life picture of oneself shooting in life, big to divine boat seven companion Fly moonlet to Divine Land seven up to 20 minutes take pictures.The most all can relate to the relevant knowledge of image deblurring, because of This research image deblurring the most just becomes to have considerable meaningful.

Image is being formed, record, during transmission, owing to being differed by optical imaging system, and imaging diffraction, imaging non-thread Property, the impact of system noise many factors, the quality of image all can decline, the as above a series of process of image It is the degeneration of image.And image recovers, also known as image restoration, it is simply that the quality of image is as much as possible reduced or eliminated Decline, recover by the true colours of degraded image.

The image deblurring problem at initial stage can trace back to explore the outer space flourishing period last century, due to by exceedingly odious weather Or the factor impacts such as atmospheric turbulance, can cause the decline of picture quality.The decline of this picture quality has non-for scientific research The biggest impact.Therefore, the research of image deblurring is to be highly desirable to and the most challenging research contents.Exist for another example Video monitoring aspect, the acquisition of monitor video is irreversible, and the monitor video before namely having access to is found to have fuzzy scene In the presence of, monitor video can not re-shoot, and is at this moment accomplished by carrying out image deblurring and could obtain more information.This Outward, image deblurring also has critically important application, such as material science image procossing in terms of other many, public security, history, Humane photograph image restores, the aspects such as aerial reconnaissance system such as scanned document processes, spaceborne, airborne.From single width broad image Recover width image clearly and become the most basic and important studying a question already.

Whether image deblurring is process antipodal with image degradation, according to the origin cause of formation degenerated it is known that image can be gone mould Paste task is divided into image non-blind deblurring and Image Blind deblurring two kinds.The non-blind deblurring of image is i.e. to know that the degeneration of image is former Cause, thus from a width broad image, recover a width picture rich in detail, it usually needs the problem of attention is to reduce that may be present shaking Bell effect and the most quenchable noise.Image Blind deblurring is then in the case of being not aware that causes for Degradation, obscures from a width Image recovers width image clearly.The blind deblurring problems faced of image is more because the causes for Degradation of image and Picture rich in detail is all unknown.Therefore, the research for image deblurring problem is the most meaningful and is necessary.

Image blurring the most common, for deblurring process known to fuzzy core, referred to as non-blind deblurring； For the deblurring process that fuzzy core is unknown, the blindest deblurring.Each pixel of entire image is passed through identical fuzzy side Formula obscures, the fuzzyyest；Otherwise in the zones of different of image, fuzzy mode is different, the most non-all Even fuzzy.Non-homogeneous fuzzy closer in real-life blooming.

For the solution procedure of uniform blind deblurring, target is exactly to recover picture rich in detail and fuzzy core from broad image.Obviously this It is an ill-conditioning problem, i.e. organizes candidate solution more.In order to overcome the ill-conditioning problem during blind deblurring, deblurring reason in early days Opinion propose parametrization fuzzy core method, it is believed that fuzzy core be shaped as linear kernel, only length and two variablees of angle, though The problem so this method solving a part of deblurring, but the real-life fuzzy origin cause of formation is relatively complicated, so existing Sizable application limitation.Afterwards broad image is added by the work of some older generation and assume or priori.Such as, logical Often assume that it is sparse and is continuous print, usually assume that the gradient information of image obeys heavytailed distribution.After this, some New theoretical method achieves more preferable effect, it is estimated that the higher candidate image of more reliable fuzzy core, picture quality with And more preferable robustness.Then also occurring in that well application in business software, the such as Photoshop groupware provides Stabilization function.

At present, blind deblurring method enters the golden period of a development, mainly has three kinds of methods:

The first is method based on maximum a posteriori probability, and such method is sought most possible solution, carried to greatest extent The associating Posterior probability distribution of high fuzzy core k and candidate image x.Such method is easily understood, and shortcoming is to there may come a time when to restrain The solution being not intended to us；

The second is based on the Bayesian method of variation, and such method developed on the basis of based on maximum a posteriori probability, Owing to it considers all possible solution, thus more preferable than algorithm robustness based on maximum a posteriori probability, but shortcoming to be speed slower；

The third is method based on edge prediction, and such method thinks that fuzzy core k can be estimated from sub-fraction image border Out, and using heuristic image filter to recover sharp keen edge, such method is in fuzzy core k estimation stages speed very Hurry up, and prove effective in an experiment, but be because adding the step of heuristic filtering, so theory analysis is the most tired Difficult.

Summary of the invention

The present invention is directed to the deficiency in said method, propose a kind of blind deblurring method based on image block priori Yu sparse norm To promote the adaptability of blind deblurring, reliability and robustness.

The key problem in technology of the present invention is: in last layer of pyramid model, adds one based on best sparse of effect instantly Norm regular terms so that in iterative process, the stage sensing of ambiguous estimation core is correctly oriented, thus obtains more approaching to reality The fuzzy core of scene, implementation step includes the following:

(1) input broad image y, is set to candidate image by broad image y；

(2) take Gaussian Blur core that size is 3 × 3 as initializing fuzzy core, use k¹Represent；

(3) take the binary mask identical with image size being all 0 as initial mask, use M¹Represent, to outside sample Blocks of data integrates and learns as BSD500 standard data set, obtains the initialization external image sample block of the present invention；

(4) broad image y is initialized, obtain initial candidate image x⁰

$x^0=\arg \underset{x}{min}\underset{D_{*}}{\&Sigma;}{w}_{*}\left|\right|{\mathrm{KD}}_{*}x-D_{*}y|{|}^2+\mathrm{\&alpha;}|\left|D_{h}x\right|{|}^2+\mathrm{\&alpha;}\left|\right|D_{v}x|{|}^2,---<1>$

Wherein, K represents fuzzy core k¹Matrix form, y represent input broad image, x⁰Represent what current iteration was wanted to obtain Clear candidate image, D_*It is the matrix form of partial differential, w on different directions_*It is these marks corresponding to different directions partial differential Amount weight, D_hAnd D_vBeing respectively the matrix form of the first-order partial derivative both horizontally and vertically gone up, x is and candidate image size Identical unknown matrix,Represent the return value that object function is x during minima；

(5) gaussian pyramid model is called, according to fuzzy core k set when initializing¹Size, calculate total number of plies N of pyramid, Initial pyramid number of plies label t=1；

(6) by candidate image x⁰Carry out down-sampling according to the pyramid number of plies, obtain the candidate image x of the 1st layer of gold tower layer¹；

(7) by candidate image x^tWith fuzzy core k^tUp-sample according to the gold tower number of plies；

(8) judge whether pyramid label t is N, if it is, preserve the candidate image x of N shell^NWith fuzzy core k^NPerform rapid (9) step (14), is otherwise performed；

(9) arranging local iteration's high reps is 200, iterations label j=1, the candidate image x that will try to achieve in (8)^NWith X^jRepresent, as new candidate image, by fuzzy core k^NUse K^jRepresent, as new fuzzy core；

(10) current candidate image X is calculated^jL₂Norm；

(11) fuzzy core k is kept¹And candidate image X^jL₂Norm | | X_j||₂Keep constant, use l₁/l₂Norm sparse just Then Image Iterative direction is any limitation as, optimizes formula according to iterative shrinkage thresholding algorithm and calculate new candidate image X^j+1；

$X^{j+1}=\arg \underset{x}{min}\mathrm{\&alpha;}\left|\right|x\&CircleTimes;K^j-y|{|}_2^2+\frac{\left|\right|x|{|}_1}{\left|\right|x|{|}_2}+\mathrm{\&beta;}|\left|K^j\right|{|}_1,---<2>$

Wherein, K^jFor the fuzzy core for j iteration, x is the unknown matrix identical with candidate image size, and y is input Broad image,For two-dimensional convolution operator, the Section 1 in formula is data fidelity items, and Section 2 is the l adding x₁/l₂Model Number regular terms, last is to fuzzy core K^jThe l added₁Norm canonical, scalar weight α and β is used for representing that control is fuzzy Core K^jWith the relative intensity of image regular terms, argmin represents the value that object function is x during minima；

(12) candidate image X is kept^j+1Constant, calculate new fuzzy core K according to following formula^j+1；

$K^{j+1}=\arg \underset{k}{min}\mathrm{\&alpha;}\left|\right|X^{j+1}\&CircleTimes;k-y|{|}_2^2+\frac{\left|\right|X^{j+1}|{|}_1}{\left|\right|X^{j+1}|{|}_2},---<3>`$

Wherein, y is the broad image of input,For two-dimensional convolution operator, k is the unknown matrix identical with fuzzy core size, Arg min represents the value that object function is k during minima, and Section 1 is data fidelity items, and Section 2 is to candidate image X^j+1 The l added₁/l₂Norm regular terms, scalar weight α represents the relative intensity controlling fuzzy core, is converted into by fuzzy core Solve problems Optimization problem, uses Biconjugate gradient solution method for solving, and k value when return function minimizes, as new fuzzy core K^j+1；

(13) iterations label j adds 1, is again assigned to j, as new iterations label, it is judged that new iterations Whether label is 200, if it is, output candidate image X²⁰⁰And fuzzy core K²⁰⁰, otherwise, return step (10)；

(14) binary mask M is updated^t+1: in all image blocks, calculate the gradient information in eight directions, choose edge letter Cease the image block of stronger front 2%, by these image blocks and mask M^tRelative position puts 1, and remaining position sets to 0, as newly Binary mask M^t+1；

(15) binary mask M is kept^t+1, external image sample block vector S_iAnd candidate image x^tConstant, more new images Variance η of blockⁱ；

(16) other parameter constants are kept, in binary mask M^t+1Put all positions of 1, if study to image block be pⁱ=ηⁱSⁱ+μⁱ, ηⁱFor the variance of image block i, SⁱFor the vector form of external image sample block, μⁱFor image block i's The average of gray scale, is focused to find out and candidate image block (Q at external image sample block_ix-pⁱ)/ηⁱMost like sample block Sⁱ, Obtain new image sample block Sⁱ；

(17) keep other parameter constants, be calculated new candidate image x^t+1；

(18) keep other parameter constants, utilize equation below to solve fuzzy core k^t+1

$k^{t+1}=\underset{{\mathrm{\&delta;}}_{*}}{\&Sigma;}{\mathrm{\&omega;}}_{*}\left|\right|k^t*{\mathrm{\&delta;}}_{*}{x}^{t+1}-{\mathrm{\&delta;}}_{*}y|{|}^2+\mathrm{\&beta;}|\left|k^t\right|{|}^2---<4>$

Wherein δ_*Represent corresponding D_*Partial derivative；Y represents the broad image of input, w_*It is that these different directions partial differential institutes are right The scalar weight answered, k^tRepresent the fuzzy core of t pyramidal layer, x^t+1It is the candidate image of t+1 pyramidal layer, arranges and do not exist Mask M^t+1In gradient information δ_*x^tIt is zero；

(19) pyramid number of plies label t adds 1, is again assigned to t, as new pyramid number of plies label, returns step (7)；

The present invention compared with prior art has the advantage that

First, fuzzy core is estimated accurately

The present invention original based on external image block transcendental method on the basis of, at pyramid model last of ambiguous estimation core One layer is added the l of a fuzzy core₁/l₂Sparse norm regular terms so that in iterative process, the stage sensing of ambiguous estimation core is just True direction, it is thus possible to obtain the fuzzy core of more approaching to reality scene, improves the accuracy that fuzzy core is estimated.

Second, adaptivity is strong

The technology of more existing image deblurrings, arranges requirement higher to parameter, and parameter selects image the most easily occur The fuzzy phenomenon with over-fitting, the present invention adds l₁/l₂Sparse norm regular terms, it is not necessary to too much arrange parameter, so that estimating The process of fuzzy core has stronger adaptivity.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention；

Fig. 2 is that first group of fuzzy core estimates experimentation figure；

Fig. 3 is first group of experimental result partial enlargement comparison diagram；

Fig. 4 is that second group of fuzzy core estimates experimentation figure；

Fig. 5 is second group of experimental result partial enlargement comparison diagram；

Fig. 6 is that the 3rd group of fuzzy core estimates experimentation figure；

Fig. 7 is the 3rd group of experimental result partial enlargement comparison diagram；

Fig. 8 is that the 4th group of fuzzy core estimates experimentation figure；

Fig. 9 is the 4th group of experimental result partial enlargement comparison diagram.

Specific implementation method

Referring to the drawings technical scheme and effect are described in further detail.

With reference to Fig. 1, the present invention to realize step as follows:

Step 1: input broad image y, and broad image y is set to candidate image.

This example chooses 4 different natural images, as shown in accompanying drawing 2 (a), 4 (a), 6 (a), 8 (a), its name Being respectively as follows: brige, Boats, Beverage and tower, their picture size size is respectively as follows: Boats and tower Being 256 × 256, the size of image brige is 419 × 566, and the size of image Beverage is 520 × 395；Wherein image Boats is gray level image, and tower, brige and Beverage are color RGB image.It is carried out at artificial fuzzy's mixing Reason, obtains broad image y as shown in accompanying drawing 2 (b), 4 (b), 6 (b), 8 (b).

Step 2: initialize fuzzy core

With the fspecial function of matlab generate size be the Gaussian Blur core of 3 × 3 as initial outermost layer fuzzy core, Use k⁰Represent；

Step 3: initialize binary mask, initializes external image sample block.

Take that size is identical with image array, numerical value is all the matrix of 1 as initializing binary mask, uses M¹Represent；

Outside sample blocks of data collection BSD500 standard data set is learnt as follows:

First, to each dimensions of 500 figures all in this data set with oversampling ratio for the downward down-sampling of 1:2, with initially Binary mask M¹Centered by, from 500 images, extract 220K the image block that size is 5 × 5；

Then, being normalized these image blocks, finally arranging clusters number is 2560, uses K mean algorithm pair 220K image block clusters, and forms 2560 clustering cluster, extracts the cluster centre of these 2560 clustering cluster as needs External image sample block.

Step 4: input broad image y is carried out following formula such as and processes, obtains initializing candidate image x⁰

$x^0=\arg \underset{x}{min}\underset{D_{*}}{\&Sigma;}{w}_{*}\left|\right|{\mathrm{KD}}_{*}x-D_{*}y|{|}^2+\mathrm{\&alpha;}|\left|D_{h}x\right|{|}^2+\mathrm{\&alpha;}\left|\right|D_{v}x|{|}^2,---<1>$

Wherein, K represents fuzzy core k¹Matrix form, y represent input broad image, x⁰Represent what current iteration was wanted to obtain Clear candidate image, D_*It is the matrix form of partial differential, w on different directions_*It is these marks corresponding to different directions partial differential Amount weight, D_hAnd D_vBeing respectively the matrix form of the first-order partial derivative both horizontally and vertically gone up, x is and candidate image size Identical unknown matrix,Represent the return value that object function is x during minima | | | |²Representing matrix one norm square, ∑ is summation symbol.

Step 5: call gaussian pyramid model, calculating total number of plies N of pyramid:

Wherein, N is the total number of plies of gold tower,For downward floor operation, log represents the log operations with 2 as the end, and b is according to mould Stick with paste the customer parameter that core size determines, initial pyramid number of plies label t=1.

Step 6: to initializing candidate image x⁰Carry out down-sampling with the pyrDown function of MATLAB, down-sampling result is made Candidate image x for the 1st layer of gold tower layer¹。

Step 7: the candidate image x to gold tower t layer^tUp-sample with the pyrDown function of MATLAB, and by upper Sampled result is assigned to x again^t。

PyrDown function k to fuzzy core MATLAB of gold tower t layer^tUp-sample, and will up-sampling result weight Newly it is assigned to k^t。

Step 8: judge whether pyramid label t is pyramid number of plies N, if it is, preserve candidate's figure of pyramid n-th layer As x^NWith fuzzy core k^NPerform rapid 9, otherwise perform step 14；

Step 9: arranging local iteration's high reps is 200, iterations label j=1, the candidate image x that will try to achieve in (8)^N Use X^jRepresent, as new candidate image, fuzzy core k^NUse K^jRepresent, as new fuzzy core.

Step 10: calculate the candidate image X of iteration j^jL₂Norm | | X_j||₂；

Step 11: keep fuzzy core K^jAnd candidate image X^jL₂Norm | | X_j||₂Keep constant, the l of employing₁/l₂Norm Image Iterative direction is any limitation as by sparse canonical, optimizes formula according to iterative shrinkage thresholding algorithm and calculates new candidate image X^j+1；

$X^{j+1}=\arg \underset{x}{min}\mathrm{\&alpha;}\left|\right|x\&CircleTimes;K^j-y|{|}_2^2+\frac{\left|\right|x|{|}_1}{\left|\right|x|{|}_2}+\mathrm{\&beta;}|\left|K^j\right|{|}_1,---<2>$

Wherein, K^jFor fuzzy core, x is the unknown matrix identical with candidate image size, and y is the broad image of input, For two-dimensional convolution operator.Section 1 in formula is data fidelity items, and Section 2 is the l adding x₁/l₂Norm regular terms, Latter is to fuzzy core K^jThe l added₁Norm canonical, α and β is scalar weight, is used for representing control fuzzy core K^jWith The relative intensity of image regular terms, argmin represents the value that object function is x during minima, | | | |₁Representing matrix one norm, | | |₂ Representing matrix two norm,Representing matrix two norm square；

Step 12: keep candidate image X^j+1Constant, fuzzy core K of iteration j it is calculated according to following formula^j+1；

$K^{j+1}=\arg \underset{k}{min}\mathrm{\&alpha;}\left|\right|X^{j+1}\&CircleTimes;k-y|{|}_2^2+\frac{\left|\right|X^{j+1}|{|}_1}{\left|\right|X^{j+1}|{|}_2},---<3>`$

Wherein, X^j+1For candidate image, y is the broad image of input,For two-dimensional convolution operator, k is and fuzzy core The unknown matrix that size is identical, arg min represents the value that object function is k during minima；Section 1 is data fidelity items, Section 2 is to candidate image X^j+1The l added₁/l₂Norm regular terms, scalar weight α represents the relative intensity controlling fuzzy core, ||·||₁Representing matrix one norm, | | | |₂Representing matrix two norm,Representing matrix two norm square, fuzzy core Solve problems is turned Turning to optimization problem, use Biconjugate gradient solution method for solving, k value when return function minimizes, as new fuzzy core K^j+1；

Step 13: iterations label j adds 1, is again assigned to j, as new iterations label, it is judged that new iteration Whether number of times label is 200, if it is, output candidate image X²⁰⁰And fuzzy core K²⁰⁰, as the termination of this example Really, as shown in accompanying drawing 2 (c), 4 (c), 6 (c), 8 (c), otherwise, step 10 is returned.

Step 14: update binary mask M^t+1。

In all image blocks, calculate eight directions gradient information, choose marginal information stronger front 2% image block, by this A little image blocks and mask M^tRelative position puts 1, and remaining position sets to 0, as new binary mask M^t+1。

Step 15: keep binary mask M^t+1, external image sample block vector S_iAnd candidate image x^tConstant, enter with two Matrix processed extracts operator Q_iIn extraction candidate image x, at the i of position, size is the image block of 5 × 5 pixels, and updates image block Variance ηⁱ；

(15a) v is made_i=Q_ix^t-pⁱ, calculate weight coefficient ω_i:

${\mathrm{\&omega;}}_i={(2{\mathrm{\&epsiv;}}^2+v_i^T{v}_i)}^{-1}---<5>$

Wherein, Q_iBeing that binary matrix extracts operator, ε is the gray threshold set in advance, pⁱ=ηⁱSⁱ+μⁱ, SⁱFor position Put the vector form of external image sample block corresponding at i, ηⁱFor the variance yields of original image block, μⁱFor the gray scale of image block i, x^tIt is the candidate image of pyramid t layer,It is intermediate variable v_iMatrix transpose；

(15b) equation below is utilized to be calculated the variance of new image block

${{\mathrm{\&eta;}}_{*}}^i=\frac{\frac{{\mathrm{\&omega;}}_i\mathrm{\&beta;}}{\left|M^{t+1}\right|}{S}^{iT}(Q_i{x}^t-{\mathrm{\&mu;}}^i)-F_{ref}^{-1}\left(F\right({\mathrm{\&eta;}}^i\left)\right)}{\frac{{\mathrm{\&omega;}}_i\mathrm{\&beta;}}{\left|M^{t+1}\right|}{S}^{iT}{S}^i}---<6>$

Wherein, M^t+1It is binary mask, ω_iFor weight coefficient, β is regularized image intensity, SⁱFor correspondence at the i of position The vector form of external image sample block, S^iTRepresent SⁱMatrix transpose, Q_iIt is that binary matrix extracts operator, μⁱFor figure As the gray value of block i, | | representing and calculate determinant of a matrix, β is regularized image intensity, x^tIt it is pyramid t layer Candidate image, F is current candidate image variance ηⁱEmpirical cumulative distribution, ηⁱFor the variance yields of original image block, F_refIt is The local contrast of external image sample block is with reference to cumulative distribution.

Step 16: keep other parameter constants, in binary mask M^t+1Put all positions of 1, if the image block that study is arrived For pⁱ=ηⁱSⁱ+μⁱ, ηⁱFor the variance of image block i, SⁱFor the vector form of external image sample block, μⁱFor image block i The average of gray scale, be focused to find out and candidate image block (Q at external image sample block_ix-pⁱ)/ηⁱMost like sample block Sⁱ, Obtain new image sample block Sⁱ；

Step 17: keep other parameter constants, is calculated the candidate image x of gold tower layer t+1 layer^t+1:

$x^{t+1}=F^{-1}\left(B\right)+\frac{\mathrm{\&beta;}}{\left|M^{t+1}\right|}\underset{i\&Element;M^{t+1}}{\&Sigma;}\frac{2}{2{\mathrm{\&epsiv;}}^2+v_i^T{v}_i}{Q}_i^T({\mathrm{\&eta;}}^i{S}^i+{\mathrm{\&mu;}}^i)---<7>$

B is intermediate variable, and it is expressed as follows:

In formula, x^t+1For the candidate image of t+1 pyramidal layer, F^-1Representing Fourier inversion, β is regularized image intensity, M^t+1It is binary mask, | | representing and calculate determinant of a matrix, ε is the gray threshold set, vⁱ=Q_ix^t-pⁱ, x^t It is the candidate image of pyramid t layer, Q_iIt is that a binary matrix extracts operator, Q_i ^TRepresent that binary matrix extracts to calculate The matrix transpose of son, ηⁱFor the variance of image block, μ at the i of positionⁱFor the gray average of image block, SⁱFor corresponding at the i of position The vector form of external image sample block；In intermediate variable B,Complex conjugate operation, δ are asked in representative_*Represent corresponding micro- The partial derivative ⊙ of sub matrix is Element-Level multiplication operator, k^tBeing the fuzzy core of pyramid t layer, y is the mould of initial input Stick with paste image.

Step 18: keep other parameter constants, utilizes equation below to solve fuzzy core k of t+1 pyramidal layer^t+1

$k^{t+1}=\underset{{\mathrm{\&delta;}}_{*}}{\&Sigma;}{\mathrm{\&omega;}}_{*}\left|\right|k^t*{\mathrm{\&delta;}}_{*}{x}^{t+1}-{\mathrm{\&delta;}}_{*}y|{|}^2+\mathrm{\&beta;}|\left|k^t\right|{|}^2---<4>$

Wherein δ_*Represent corresponding D_*Partial derivative, y represent input broad image, ω_*It is that these different directions partial differential institutes are right The scalar weight answered, k^tRepresent the fuzzy core of pyramid t layer, x^t+1It is the candidate image of pyramid t+1 layer gained, if Put not at mask M^t+1In gradient information δ * x^t+1Being zero, β represents control fuzzy core k^tThe relative intensity of regular terms, ∑ is Summation symbol, | | | | representing matrix one norm, | | | |₂Representing matrix two norm,Representing matrix two norm square.

Step 19: add 1 to pyramid number of plies label t, be again assigned to t, as new pyramid number of plies label, returns Step 7.

The effect of the present invention can be further illustrated by following experiment:

1, simulated conditions:

The software environment of the present invention is 64 systems of Windows 7 Ultimate, MATLAB 2014a.Hardware environment is Intel The CPU of Core2Duo 3.2GHz, and in save as and run in the environment of DDR34GB.

2. emulation content:

Emulation 1, concentrates the nonlinear smearing verification picture rich in detail figure choosing 19 × 19 to carry out from existing nonlinear smearing Nuclear Data Fuzzy hybrid processes, and obtains the broad image figure of synthesis, the broad image figure of synthesis is carried out blind deblurring process, obtains mould Image after paste and the fuzzy core figure estimated, as shown in Figure 2.Wherein

Fig. 2 (a) is the original picture rich in detail of brige, and the image lower right corner is the fuzzy fuzzy core used of synthesis；

Fig. 2 (b) is the broad image of synthesis；

Fig. 2 (c) is the image after deblurring, and the lower right corner of image is the fuzzy core estimated；

By the Detail contrast of image, result such as Fig. 3 after the broad image shown in Fig. 2 (b) and Fig. 2 (c) deblurring.Can from Fig. 3 Seeing, the fuzzy core that the present invention estimates approaches real fuzzy core, and after deblurring, image is close to picture rich in detail.

Emulation 2, concentrates the nonlinear smearing verification picture rich in detail figure choosing 17 × 17 to carry out from existing nonlinear smearing Nuclear Data Fuzzy hybrid processes, and obtains the broad image figure of synthesis, the broad image figure of synthesis is carried out blind deblurring process, obtains mould Image after paste and the fuzzy core figure estimated, as shown in Figure 4.Wherein

Fig. 4 (a) is the original picture rich in detail of boats, and the image lower right corner is the fuzzy fuzzy core used of synthesis；

Fig. 4 (b) is the broad image of synthesis；

Fig. 4 (c) is the image after deblurring, and the lower right corner of image is the fuzzy core estimated；

By the Detail contrast of image, result such as Fig. 5 after the broad image shown in Fig. 4 (b) and Fig. 4 (c) deblurring.Can from Fig. 5 See, the present invention in this group figure after deblurring image clear, the fuzzy core estimated is almost identical with real fuzzy core.

Emulation 3, concentrates the nonlinear smearing verification picture rich in detail figure choosing 15 × 15 to carry out from existing nonlinear smearing Nuclear Data Fuzzy hybrid processes, and obtains the broad image figure of synthesis, the broad image figure of synthesis is carried out blind deblurring process, obtains mould Image after paste and the fuzzy core figure estimated, as shown in Figure 6.Wherein

Fig. 6 (a) is the original picture rich in detail of Beverage, and the image lower right corner is the fuzzy fuzzy core used of synthesis；

Fig. 6 (b) is the broad image of synthesis；

Fig. 6 (c) is the image after deblurring, and the lower right corner of image is the fuzzy core estimated；

By the Detail contrast of image, result such as Fig. 7 after the broad image shown in Fig. 6 (b) and Fig. 6 (c) deblurring.Can from Fig. 7 Seeing, after deblurring of the present invention, image is clear, and the fuzzy core estimated is almost identical with real fuzzy core.

Emulation 4, concentrates the nonlinear smearing verification picture rich in detail figure choosing 13 × 13 to carry out from existing nonlinear smearing Nuclear Data Fuzzy hybrid processes, and obtains the broad image figure of synthesis, the broad image figure of synthesis is carried out blind deblurring process, obtains mould Image after paste and the fuzzy core figure estimated, as shown in Figure 8.Wherein

Fig. 8 (a) is the original picture rich in detail of Beverage, and the image lower right corner is the fuzzy fuzzy core used of synthesis；

Fig. 8 (b) is the broad image of synthesis；

Fig. 8 (c) is the image after deblurring, and the lower right corner of image is the fuzzy core estimated；

By the Detail contrast of image, result such as Fig. 9 after the broad image shown in Fig. 8 (b) and Fig. 8 (c) deblurring.Can from Fig. 9 Seeing, after deblurring of the present invention, image is clear, and the fuzzy core estimated is almost identical with real fuzzy core.

3, interpretation:

Comprehensive four groups of experimental results it can be seen that after deblurring of the present invention image clear, the fuzzy core estimated approaches true mould The fuzzy core stuck with paste, and overcome to a certain extent in fuzzy core and there is the phenomenon that class is noise-like, embody the present invention to fuzzy core Estimation there is good accuracy.The fuzzy core estimated result performance of four groups of experiments is stable, embodies the present invention and has well Adaptability and stability.

Claims (3)

1. a blind deblurring method based on image block priori Yu sparse norm, including:

(1) input broad image y, is set to candidate image by broad image y；

(2) take Gaussian Blur core that size is 3 × 3 as initializing fuzzy core, use k¹Represent；

(3) take the binary mask identical with image size being all 0 as initial mask, use M¹Represent, outside sample blocks of data is integrated and learns as BSD500 standard data set, obtain the initialization external image sample block of the present invention；

(4) broad image y is initialized, obtain initial candidate image x⁰

Wherein, k¹Represent fuzzy core k¹Matrix form, y represent input broad image, x⁰Represent the clear candidate image that current iteration is wanted to obtain, D_*It is the matrix form of partial differential, w on different directions_*It is these scalar weight corresponding to different directions partial differential, D_hAnd D_vBeing respectively the matrix form of the first-order partial derivative both horizontally and vertically gone up, x is the unknown matrix identical with candidate image size,Represent the return value that object function is x during minima；

(5) gaussian pyramid model is called, according to fuzzy core k set when initializing¹Size, calculate total number of plies N of pyramid, initial pyramid number of plies label t=1；

(6) by candidate image x⁰Carry out down-sampling according to the pyramid number of plies, obtain the candidate image x of the 1st layer of gold tower layer¹；

(7) by candidate image x^tWith fuzzy core k^tUp-sample according to the gold tower number of plies；

(8) judge whether pyramid label t is N, if it is, preserve the candidate image x of N shell^NWith fuzzy core k^NPerform rapid (9), otherwise perform step (14)；

(9) arranging local iteration's high reps is 200, iterations label j=1, the candidate image x that will try to achieve in (8)^NUse X^jRepresent, as new candidate image, by fuzzy core k^NUse K^jRepresent, as new fuzzy core；

(10) current candidate image X is calculated^jL₂Norm；

(11) fuzzy core k is kept¹And candidate image X^jL₂Norm | | X_j||₂Keep constant, use l₁/l₂Image Iterative direction is any limitation as by the sparse canonical of norm, optimizes formula according to iterative shrinkage thresholding algorithm and calculates new candidate image X^j+1；

Wherein, K^jFor the fuzzy core for j iteration, x is the unknown matrix identical with candidate image size, and y is the broad image of input,For two-dimensional convolution operator, the Section 1 in formula is data fidelity items, and Section 2 is the l adding x₁/l₂Norm regular terms, last is to fuzzy core K^jThe l added₁Norm canonical, scalar weight α and β is used for representing control fuzzy core K^jWith the relative intensity of image regular terms, argmin represents the value that object function is x during minima；

(12) candidate image X is kept^j+1Constant, calculate new fuzzy core K according to following formula^j+1；

Wherein, y is the broad image of input,For two-dimensional convolution operator, k is the unknown matrix identical with fuzzy core size, and arg min represents the value that object function is k during minima, and Section 1 is data fidelity items, and Section 2 is to candidate image X^j+1The l added₁/l₂Norm regular terms, scalar weight α represents the relative intensity controlling fuzzy core, and fuzzy core Solve problems is converted into optimization problem, uses Biconjugate gradient solution method for solving, and k value when return function minimizes, as new fuzzy core K^j+1；

(13) iterations label j adds 1, is again assigned to j, as new iterations label, it is judged that whether new iterations label is 200, if it is, output candidate image X²⁰⁰And fuzzy core K²⁰⁰, otherwise, return step (10)；

(14) binary mask M is updated^t+1: in all image blocks, calculate the gradient information in eight directions, choose marginal information stronger front 2% image block, by these image blocks and mask M^tRelative position puts 1, and remaining position sets to 0, as new binary mask M^t+1；

(15) binary mask M is kept^t+1, external image sample block vector S_iAnd candidate image x^tConstant, update variance η of image blockⁱ；

(16) other parameter constants are kept, in binary mask M^t+1Put all positions of 1, if study to image block be pⁱ=ηⁱSⁱ+μⁱ, ηⁱFor the variance of image block i, SⁱFor the vector form of external image sample block, μⁱFor the average of the gray scale of image block i, it is focused to find out and candidate image block (Q at external image sample block_ix-pⁱ)/ηⁱMost like sample block Sⁱ, obtain new image sample block Sⁱ；

(17) keep other parameter constants, be calculated new candidate image x^t+1；

(18) keep other parameter constants, utilize equation below to solve fuzzy core k^t+1

Wherein δ_*Represent corresponding D_*Partial derivative；Y represents the broad image of input, w_*It is these scalar weight corresponding to different directions partial differential, k^tRepresent the fuzzy core of t pyramidal layer, x^t+1It is the candidate image of t+1 pyramidal layer, arranges not at mask M^t+1In gradient information δ_*x^tIt is zero；

(19) pyramid number of plies label t adds 1, is again assigned to t, as new pyramid number of plies label, returns step (7).

Blind deblurring method based on image block priori Yu sparse norm the most according to claim 1, is wherein calculated the variance of new images block in step (15)Carry out as follows:

(15a) v is madeⁱ=Q_ix^t-pⁱ, calculate weight coefficient ω_i:

Wherein, Q_iBeing that binary matrix extracts operator, ε is the gray threshold set in advance, pⁱ=ηⁱSⁱ-μⁱ, SⁱFor the vector form of external image sample block corresponding at the i of position, ηⁱFor the variance yields of original image block, μⁱGray value for image block i；

(15b) equation below is utilized to be calculated variance η of new image block_* ⁱ:

Wherein, M^t+1It is binary mask, ηⁱFor the variance yields of original image block, Q_iIt is that binary matrix extracts operator, SⁱFor the vector form of external image sample block corresponding at the i of position, μⁱFor the gray value of image block i, β is regularized image intensity, x^tBeing the candidate image of pyramid t layer, F is current candidate image variance ηⁱEmpirical cumulative distribution, F_refIt is that the local contrast of external image sample block is with reference to cumulative distribution.

Blind deblurring method based on image block priori Yu sparse norm the most according to claim 1, wherein step (17) calculates new candidate image x^t+1, carried out by equation below:

Wherein, x^t+1For the candidate image of t+1 pyramidal layer, F^-1Representing Fourier inversion, β is regularized image intensity, M^t+1Being binary mask, ε is the gray threshold set, vⁱ=Q_ix^t-pⁱ, x^tIt is the candidate image of pyramid t layer, Q_iIt is that a binary matrix extracts operator, Q_i ^TRepresent that binary matrix extracts the matrix transpose of operator, ηⁱFor the variance of image block, μ at the i of position_iFor the gray average of image block, SⁱVector form for external image sample block corresponding at the i of position；In intermediate variable B,Complex conjugate operation, δ are asked in representative_*Representing the partial derivative of corresponding differential matrix, ⊙ is Element-Level multiplication operator, k^tBeing the fuzzy core of pyramid t layer, y is the broad image of initial input.