CN103093436A - Blurring kernel multi-scale iteration estimation method using directional derivative of image local structure - Google Patents

Abstract

The invention discloses a blurring kernel multi-scale iteration estimation method using a directional derivative of an image local structure. The blurring kernel multi-scale iteration estimation method comprises inter-scale updating, intra-scale iteration estimation and inter-scale iteration termination judgment. A blurring kernel and a recovering image are estimated from coarse scale iteration to fine scale iteration. The intra-scale iteration estimation comprises the following steps: calculating an enhanced type direction gradient filed of a current scale; estimating a blurring kernel of the current scale rapidly; calculating a directional derivative approximate spectrum of a current image; and recovering an image of the current scale. Gradient information of image jumping edges is utilized, with a quick Fourier transform technology based on, the blurring kernel is quickly estimated in a small time complexity, and the blurring kernel multi-scale iteration estimation method can be used for conducting blind deblurring on various actual blurring images.

Description

Utilize the multiple dimensioned iterative estimate method of fuzzy core of Local Structure of Image directional derivative

Technical field

The invention belongs to the computer digital image process field, particularly a kind of multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative.

Background technology

Camera shake is a problem that is always perplexing the cameraman, and unsettled camera can cause the fuzzy of photo.Particularly present popular HIGH RESOLUTION digital camera, make them be difficult in use stably hold because fuselage weight is light.What many photographs were caught is some the of short duration precious moments that can't reappear, if at this moment wait, camera shake has occured, and these precious moments just lose fully so.Therefore, be devoted to seek the method for removing camera shake from the single width photograph and just seem particularly important.

Image deblurring mainly can be divided into two large classes: the non-blind deblurring of image and Image Blind deblurring.The non-blind deblurring of image refers to ask under the known fuzzy core condition that causes image degradation the process of original picture rich in detail, and the method for problems is more skillful; Iff known degraded image, and in the situation of fuzzy core the unknown, this class image deblurring problem is called the Image Blind deblurring.The estimation technique of fuzzy core is in fact vital part in Image Blind deblurring technology, namely how to estimate the fuzzy core that causes original image to degenerate from only degraded image, obtain just can utilizing the algorithm of the non-blind deblurring of some maturations to recover original image after fuzzy core.Because utilizable priori is fewer, make the Image Blind deblurring very difficult, therefore blind deblurring is more challenging, also more has practicality simultaneously.

Traditional fuzzy core method of estimation supposes that normally fuzzy core has certain special parametric form, and the shake path of camera Parameter Expression simply in reality.Therefore, fuzzy core for curve shape, the people such as Rob Fergus propose the variational Bayesian method based on the distribution of natural image gradient, referring to article " Removing Camera Shake from a Single Photograph ", 2006, ACM Transactions on Graphics, vol.25, pp.787-794.This methods analyst the gradient of image statistically obey heavy-tailed distribution, and mix with Gauss's yardstick match is carried out in this distribution, first estimate fuzzy core, the classical Lucy-Richardson method of recycling is carried out the non-blind deblurring of image, and the method has been introduced the framework of multiple dimensioned iterative estimate fuzzy core simultaneously.Yet this method is also not accurate enough to the estimation of fuzzy core, causes final deblurring result to have a lot of ringing effects.Afterwards, the people such as Sunghyun Cho had proposed the method based on the fuzzy edge sharpening, referring to article " Fast Motion Deblurring ", and 2009, ACM Transactions on Graphics, vol.28, pp.145:1-145:8.This method is utilized the marginal information of blurred picture, fuzzy edge is carried out impact filtering, more alternately iterative estimate fuzzy core and restored image, the method can be fast and ambiguous estimation core preferably, if but for the more image of texture, this method is unsatisfactory to the recovery effects of image.

Summary of the invention

The object of the invention is to provide a kind of multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative, with ambiguous estimation core more accurately, remove ringing effect, and algorithm takes full advantage of fast Fourier transform techniques, make deblurring can access quick realization.

The technical solution that realizes the object of the invention is: a kind of multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative, comprise iteration system initialization procedure and multiple dimensioned iterative estimate process, and step is as follows:

1.1 described iteration system initialization procedure is:

(1) the blurred picture y of input one width M to be recovered * N size, the initialization fuzzy core k of thick yardstick ₀, treat the support set size h of ambiguous estimation core ₁* h ₂, blurred picture y down-sampling is obtained multiple dimensioned blurred picture sequence { y ₁, y ₂, L, y _S, calculate the out to out number

Wherein: M, N is line number and the columns of presentation video respectively, h ₁, h ₂The line number and the columns that represent respectively fuzzy core support set, max (h ₁, h ₂) expression h ₁And h ₂In maximal value, Int () is rounding operation, log ₁₀The expression denary logarithm;

(2) need initialized iteration system parameter to comprise: maximum iteration time N in yardstick _maxConstant compensating parameter β, difference operator contribution parameters γ, image difference operator contribution parameters α;

(3) calculated level direction difference operator D ₁Fourier transform F (the D of=[1 ,-1] ₁) and vertical direction difference operator D ₂=[1 ,-1] ^TFourier transform F (D ₂);

1.2 described multiple dimensioned iterative process comprise between yardstick upgrade, iteration stops judgement between the iterative estimate in yardstick and yardstick:

1.2.1 upgrade between yardstick, yardstick is s, 1≤s≤S:

(1) utilize blind restored image x under the s-1 yardstick _s-1With the fuzzy core k that estimates _s-1, they are carried out respectively up-sampling, the initial pictures under current yardstick s is updated to Initial fuzzy core under current yardstick s is updated to

During for yardstick s=1, the primary iteration solution of image is

The primary iteration solution of fuzzy core is

Wherein: the operation of symbol " ↑ " expression up-sampling;

(2) calculate y _sStructure tensor, and then obtain position (i, j) perpendicular to the normalization proper vector of partial structurtes direction

Calculate simultaneously y _sNormal gradients vector field (u _s, v _s), node-by-node algorithm y thus _sDirection gradient vector, its normal component

And tangential component ${\left({\▿}_a{y}_s\right)}_{i,j}={(-{\mathrm{\υ}}_s,{\mathrm{\μ}}_s)}_{i,j}\·{(u_s,v_s)}_{i,j}^T;$

(3) calculate blurred picture y under current yardstick s _sFrequency spectrum F (y _s), and y _sThe frequency spectrum of direction gradient vector field

1.2.2 the estimation of yardstick inner iteration, n is iterations, 1≤n≤N _max:

Step 1 generates the enhancement mode direction gradient vector field of iterative image, the image that the n-1 time iteration obtained

Utilize Meyer cartoon-texture decomposition method to obtain the cartoon composition, and carry out impact filtering, obtain edge enhanced images

By calculating The structure tensor node-by-node algorithm

Normal component

And tangential component The direction gradient vector field that obtains the n time iterative image is

Step 2, iterative image direction gradient vector field is calculated in the quick estimation of current yardstick fuzzy core the n time Frequency spectrum

Then obtain the fuzzy core of the n time iteration in frequency domain

Step 3, the approximate frequency spectrum of calculated direction derivative, according to

Calculate the method directional derivative of the n time iteration and cut the approximate frequency spectrum of directional derivative

Step 4, the Fast Restoration of current scalogram picture, fuzzy core in calculation procedure 2

Frequency spectrum

Then try to achieve a rough estimate of the n time iteration restored image

Step 5, the yardstick inner iteration stops judgement: if n 〉=N _max, go to iteration termination judgement between yardstick, and will As the final fuzzy core of estimating of current yardstick,

As the blind restored image of current yardstick, inner iteration count n is set to 1; Otherwise upgrade inner iteration count n=n+1, and go to step 1;

1.2.3 between yardstick, iteration stops judgement: if s 〉=S, between yardstick, iteration stops, and will

As final fuzzy core,

As final restored image; Otherwise upgrade yardstick iteration count s=s+1, go between yardstick and upgrade.

The present invention compared with prior art, its remarkable advantage: (1) can realize the accurate estimation of fuzzy core.Compare with solution in the past, the inventive method is by utilizing gradient fields information accurately, can effectively suppress the noise in estimated fuzzy core, make fuzzy core more accurate, on vision, the continuity of fuzzy core curve is stronger, thus the ringing effect that the error that minimizing is estimated because of fuzzy core causes.(2) utilize fast Fourier transform techniques, thereby made computation complexity greatly reduce.The inventive method is at road vehicles monitoring, and digital photograph is processed, and the aspects such as recovery of medical science and astronomic graph picture all have wide practical use, and also provides new thinking for image super-resolution simultaneously.

Below in conjunction with accompanying drawing, the present invention is described in further detail.

Description of drawings

Fig. 1 is the process flow diagram of inventive method Global Iterative Schemes structure.

Fig. 2 is iterative estimate unit step 1 data flow diagram in yardstick.

Fig. 3 is iterative estimate unit step 2 data flow diagram in yardstick.

Fig. 4 is the iterative estimate unit step 3 and step 4 data flow diagram in yardstick.

Fig. 5 is Woman picture rich in detail, the first fuzzy core and the blurred picture thereof that emulation experiment of the present invention is used.

Fig. 6 is Vase picture rich in detail, the second fuzzy core and the blurred picture thereof that emulation experiment of the present invention is used.

Fig. 7 is that the present invention is to the blind deblurring result of Fig. 5 (c) and the fuzzy core of estimation.

Fig. 8 is that the present invention is to the blind deblurring result of Fig. 6 (c) and the fuzzy core of estimation.

Fig. 9 is the actual blurred picture fishes that the present invention uses.

Figure 10 is that the present invention is to the deblurring result of Fig. 9 and the fuzzy core of estimation.

Figure 11 is that existing method based on the fuzzy edge sharpening is to the deblurring result of Fig. 9 and the fuzzy core of estimation.

Figure 12 is that the existing method that distributes based on the natural image gradient is to the deblurring result of Fig. 9 and the fuzzy core of estimation.

Embodiment

The present invention utilizes the multiple dimensioned iterative estimate method of the fuzzy core of Local Structure of Image directional derivative, take picture breakdown, partial structurtes directional derivative as the basis, by multiple dimensioned direction gradient vector field, mode predictive fuzzy core and the restored image to replace of finding the solution iteratively, reach the purpose that fuzzy core is estimated.The present invention utilizes picture breakdown that the blind restored image in iterative process is decomposed into cartoon part and texture part, obtain the direction gradient field of cartoon part after impact filtering, the gradient fields that utilization obtains is detailed predicting fuzzy core and restored image alternately, and iteration obtains final fuzzy core after stopping.

As shown in Figure 1, the present invention utilizes the multiple dimensioned iterative estimate method of the fuzzy core of Local Structure of Image directional derivative, comprises iteration system initialization procedure and multiple dimensioned iterative estimate process.

1, described iteration system initialization procedure comprises 3 parts:

(1) the blurred picture y of input one width M to be recovered * N size, the initialization fuzzy core k of thick yardstick ₀, treat the support set size h of ambiguous estimation core ₁* h ₂, blurred picture y down-sampling is obtained multiple dimensioned blurred picture sequence { y ₁, y ₂, L, y _S, calculate the out to out number

Wherein: M, N is line number and the columns of presentation video respectively, h ₁, h ₂The line number and the columns that represent respectively fuzzy core support set, max (h ₁, h ₂) expression h ₁And h ₂In maximal value, Int () is rounding operation, log ₁₀The expression denary logarithm.

(2) iteration system parameter initialization needs initialized iteration system parameter to comprise: maximum iteration time N in yardstick _maxConstant compensating parameter β, difference operator contribution parameters γ, image difference operator contribution parameters α.Following initial parameter value is set: maximum iteration time N in yardstick _max∈ [6, L, 10]; Constant compensating parameter β ∈ [5,10], difference operator contribution parameters γ=2 β; Image difference operator contribution parameters α ∈ [0.1,0.3];

(3) calculated level direction difference operator D ₁Fourier transform F (the D of=[1 ,-1] ₁) and vertical direction difference operator D ₂=[1 ,-1] ^TFourier transform F (D ₂).

2, described multiple dimensioned iterative process comprise between yardstick upgrade, iteration stops judgement between the iterative estimate in yardstick and yardstick:

2.1 upgrade between yardstick (yardstick is s, 1≤s≤S):

(1) utilize blind restored image x under the s-1 yardstick _s-1With the fuzzy core k that estimates _s-1, they are carried out respectively up-sampling, the initial pictures under current yardstick s is updated to

Initial fuzzy core under current yardstick s is updated to

(the primary iteration solution of image is during for yardstick s=1

The primary iteration solution of fuzzy core is

Wherein: the operation of symbol " ↑ " expression up-sampling.

(2) calculate y _sStructure tensor, and then obtain position (i, j) perpendicular to the normalization proper vector of partial structurtes direction

Calculate simultaneously y _sNormal gradients vector field (u _s, v _s), direction gradient vector that thus can node-by-node algorithm ys, its normal component

And tangential component ${\left({\▿}_a{y}_s\right)}_{i,j}={(-{\mathrm{\υ}}_s,{\mathrm{\μ}}_s)}_{i,j}\·{(u_s,v_s)}_{i,j}^T.$

In the present invention the computing method of structure tensor are referring to the people's such as Stefan Roth paper " Steerable Random Fields ", 2007, ICCV, and parameter arranges same the method.

(3) calculate blurred picture y under current yardstick s _sFrequency spectrum F (y _s), and y _sThe frequency spectrum of direction gradient vector field

2.2 the yardstick inner iteration estimates that (the n time iteration, n is iterations, 1≤n≤N _max):

Step 1: the enhancement mode direction gradient vector field that generates iterative image.The detailed process of this step is as follows as shown in Figure 2, the image that the n-1 time iteration obtained

Utilize Meyer cartoon-texture decomposition method to obtain the cartoon composition, and carry out impact filtering (shock filter), obtain edge enhanced images

By calculating

The structure tensor node-by-node algorithm Normal component

And tangential component

The direction gradient vector field that obtains the n time iterative image is

What use in this step of the present invention is a kind of rapid image decomposition method, " the Fast Cartoon+Texture Image Filters " that delivers referring to people such as Antoni Buades, 2010, IEEE Transactions on Image Processing, 19 (8): 1978-1986, parameter arranges same the method.

Step 2: the quick estimation of current yardstick fuzzy core.The detailed process of this step is as follows as shown in Figure 3, calculates iterative image direction gradient vector field the n time

Frequency spectrum

Then obtain the fuzzy core of the n time iteration in frequency domain

Namely

Calculate iterative image direction gradient vector field the n time

Frequency spectrum

Then ambiguous estimation core in frequency domain, can be found the solution by following relational expression and obtain:

$k_s^{\left(n\right)}=\mathrm{Shrink}\left(F^{-1}\left(\frac{F^{*}\left({\▿}_o{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_o{y}_s\right)+F^{*}\left({\▿}_a{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_a{y}_s\right)}{F^{*}\left({\▿}_o{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_o{S}_s^{\left(n\right)}\right)+F^{*}\left({\▿}_a{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_a{S}_s^{\left(n\right)}\right)+\mathrm{\γ}(F^{*}\left(D_1\right)\mathrm{oF}\left(D_1\right)+F^{*}\left(D_2\right)\mathrm{oF}\left(D_2\right))+\mathrm{\β}}\right)\right)---\left(1\right)$

Wherein: β is that constant compensating parameter, γ are the difference operator contribution parameters, F ^*The complex conjugate of () expression F (), F ^-1() expression inverse fourier transform, the pointwise of " o " representing matrix element is multiplied each other, and fraction represents that between two matrixes of molecule denominator, the element pointwise is divided by, shrinkage operation operator Shrink ()=max (, 0.01), s is yardstick.

Step 3: the approximate frequency spectrum of calculated direction derivative.The detailed process of this step is as follows as shown in the part of Fig. 4, according to

Calculate the method directional derivative of the n time iteration and cut the approximate frequency spectrum of directional derivative

Namely

According to

The computation structure tensor, and then obtain the normalization proper vector that position (i, j) is orthogonal to the partial structurtes direction

Thus can be according to relational expression (2) and (3) the current method directional derivative of node-by-node algorithm and cut the approximate frequency spectrum of directional derivative respectively,

$F\left({\mathrm{Do}}_s^{\left(n\right)}\right)={\widehat{\mathrm{\μ}}}_s^{(n-1)}\mathrm{oF}\left(D_1\right)+{\hat{v}}_s^{(n-1)}\mathrm{oF}\left(D_2\right)---\left(2\right)$

$F\left({\mathrm{Da}}_s^{\left(n\right)}\right)={-\hat{v}}_s^{(n-1)}\mathrm{oF}\left(D_1\right)+{\widehat{\mathrm{\μ}}}_s^{(n-1)}\mathrm{oF}\left(D_2\right)---\left(3\right)$

Wherein: the pointwise of " o " representing matrix element is multiplied each other.

Step 4: the Fast Restoration of current scalogram picture.The detailed process of this step is as follows as shown in the part of Fig. 4, fuzzy core in calculation procedure 2

Frequency spectrum

Then try to achieve a rough estimate of the n time iteration restored image

Namely

Fuzzy core in calculation procedure 2

Frequency spectrum

Then tried to achieve a rough estimate of restored image by following relational expression:

$x_s^{\left(n\right)}=F^{-1}\left(\frac{F^{*}\left(k_s^{\left(n\right)}\right)\mathrm{oF}\left(y_s\right)o(1+0.5F^{*}\left({\mathrm{Do}}_s^{\left(n\right)}\right)+0.5F^{*}\left({\mathrm{Da}}_s^{\left(n\right)}\right)\mathrm{oF}\left({\mathrm{Da}}_s^{\left(n\right)}\right))}{F^{*}\left(k_s^{\left(n\right)}\right)\mathrm{oF}\left(k_s^{\left(n\right)}\right)o(1+0.5F^{*}\left({\mathrm{Do}}_s^{\left(n\right)}\right)\mathrm{oF}\left({\mathrm{Do}}_s^{\left(n\right)}\right)+0.5F^{*}\left({\mathrm{Da}}_s^{\left(n\right)}\right)\mathrm{oF}\left({\mathrm{Da}}_s^{\left(n\right)}\right))+\mathrm{\α}(F^{*}\left(D_1\right)\mathrm{oF}\left(D_1\right)+F^{*}\left(D_2\right)\mathrm{oF}\left(D_2\right))}\right)---\left(4\right)$

Wherein: α is image difference operator contribution parameters, F ^*The complex conjugate of () expression F (), F ^-1() expression inverse fourier transform, the pointwise of " o " representing matrix element is multiplied each other, and fraction represents that between two matrixes of molecule denominator, the element pointwise is divided by, and s is yardstick.

Step 5: the yardstick inner iteration stops judgement.If n 〉=N _max, go to iteration termination judgement between yardstick, and will

As the final fuzzy core of estimating of current yardstick,

As the blind restored image of current yardstick, inner iteration count n is set to 1; Otherwise upgrade inner iteration count n=n+1, and go to step 1.

2.3 iteration termination judgement between yardstick (yardstick s, 1≤s≤S):

If s 〉=S, between yardstick, iteration stops, and will

As final fuzzy core,

As final restored image; Otherwise upgrade yardstick iteration count s=s+1, go between yardstick and upgrade.

To shown in Figure 12, come specifically technique effect and the practicality of bright the inventive method by two experiments below in conjunction with Fig. 5.

1, experiment condition:

Testing computing environment used is Intel i3-2100CPU 3.1GHz, in save as 4GB1333MHz, programming platform is Matlab R2011a.Test image used and derive from Berkeley Segmentation Benchmark, network address is http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/, the image size is 321 * 481, and natural image fishes, and size is 858 * 558.

2, experiment content:

This experiment specifically is divided into two son experiments: the corresponding respectively simulated experiment that provides in emulation fuzzy core situation, and the True Data experiment that only provides the actual blurred picture that obtains.

Experiment one:

1. to Women image shown in Fig. 5 (a), use the first fuzzy core shown in Fig. 5 (b) to simulate, the fuzzy core size is 19 * 19, obtain the blurred picture as shown in Fig. 5 (c), with the inventive method, blurred picture is carried out core and estimate and deblurring, the fuzzy core that obtains estimating and restoration result such as Fig. 7.

2. to Vase image shown in Fig. 6 (a), use the first fuzzy core shown in Fig. 6 (b) to simulate, the fuzzy core size is 19 * 19, obtain the blurred picture as shown in Fig. 6 (c), with the inventive method, blurred picture is carried out core and estimate and deblurring, the fuzzy core that obtains estimating and restoration result such as Fig. 8.

Experiment two:

To fishes image shown in Figure 9 respectively with the inventive method, existing based on the method for fuzzy edge sharpening with carry out blind deblurring based on the method that the natural image gradient distributes, result such as Figure 10-Figure 12 wherein estimate and deblurring result such as Figure 10 with the image blurring core of fishes that the present invention obtains; Estimate and deblurring result such as Figure 11 with the existing image blurring core of fishes that obtains based on the method for fuzzy edge sharpening; Estimate and deblurring result such as Figure 12 with the image blurring core of fishes that the existing method that distributes based on the natural image gradient obtains.

Can see that from Fig. 7 and Fig. 8 core estimated result that the present invention obtains can keep the continuity of fuzzy core effectively, quite approaching with original real fuzzy core contrast in Fig. 5 (b) and Fig. 6 (b), the result of deblurring is clear undistorted, ringing effect is also less, and fine edge and the details that has kept image proved absolutely validity of the present invention.

For the true blurred picture that directly obtains from reality, from Figure 10 (a) as seen, deblurring result of the present invention meets vision requirement more, and the details of image is comparatively clear, and the edge has also obtained keeping preferably; From Figure 10 (b) as seen, fuzzy core estimated result of the present invention is also more reasonable, and its continuity is fine, and almost noiseless exists.

From Figure 11 (a) as seen, although based on the effective marginal information of Recovery image of the result of the method deblurring of fuzzy edge sharpening, image detail is too level and smooth, and certain distortion phenomenon is arranged; From Figure 11 (b) as seen, the estimated result of the method fuzzy core has more noise, and continuity is also good not.

From Figure 12 (a) as seen, the result of the method deblurring that distributes based on the natural image gradient is clear not, and it is good not that details keeps; From Figure 12 (b) as seen, there is certain noise in the estimated result of the method fuzzy core, and continuity is equally also good not.

In sum, no matter synthesize simulated experiment or True Data experiment, the inventive method has all obtained effect preferably, the result that core is estimated is very accurate, the edge as a result of deblurring and details keep significantly having good application prospect and value, and practicality is also very strong.

Claims (5)

1. the multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative, is characterized in that comprising iteration system initialization procedure and multiple dimensioned iterative estimate process, and step is as follows:

1.1 described iteration system initialization procedure is:

(1) the blurred picture y of input one width M to be recovered * N size, the initialization fuzzy core k of thick yardstick ₀, treat the support set size h of ambiguous estimation core ₁* h ₂, blurred picture y down-sampling is obtained multiple dimensioned blurred picture sequence { y ₁, y ₂, L, y _S, calculate the out to out number

Wherein: M, N is line number and the columns of presentation video respectively, h ₁, h ₂The line number and the columns that represent respectively fuzzy core support set, max (h ₁, h ₂) expression h ₁And h ₂In maximal value, Int () is rounding operation, log ₁₀The expression denary logarithm;

(2) need initialized iteration system parameter to comprise: maximum iteration time N in yardstick _maxConstant compensating parameter β, difference operator contribution parameters γ, image difference operator contribution parameters α;

(3) calculated level direction difference operator D ₁Fourier transform F (the D of=[1 ,-1] ₁) and vertical direction difference operator D ₂=[1 ,-1] ^TFourier transform F (D ₂);

1.2 described multiple dimensioned iterative process comprise between yardstick upgrade, iteration stops judgement between the iterative estimate in yardstick and yardstick:

1.2.1 upgrade between yardstick, yardstick is s, 1≤s≤S:

(1) utilize blind restored image x under the s-1 yardstick _s-1With the fuzzy core k that estimates _s-1, they are carried out respectively up-sampling, the initial pictures under current yardstick s is updated to Initial fuzzy core under current yardstick s is updated to

During for yardstick s=1, the primary iteration solution of image is

The primary iteration solution of fuzzy core is

Wherein: the operation of symbol " ↑ " expression up-sampling;

(2) calculate y _sStructure tensor, and then obtain position (i, j) perpendicular to the normalization proper vector of partial structurtes direction

Calculate simultaneously y _sNormal gradients vector field (u _s, v _s), node-by-node algorithm y thus _sDirection gradient vector, its normal component

And tangential component ${\left({\▿}_a{y}_s\right)}_{i,j}={(-{\mathrm{\υ}}_s,{\mathrm{\μ}}_s)}_{i,j}\·{(u_s,v_s)}_{i,j}^T;$

(3) calculate blurred picture y under current yardstick s _sFrequency spectrum F (y _s), and y _sThe frequency spectrum of direction gradient vector field

1.2.2 the estimation of yardstick inner iteration, n is iterations, 1≤n≤N _max:

Step 1 generates the enhancement mode direction gradient vector field of iterative image, the image that the n-1 time iteration obtained

Utilize Meyer cartoon-texture decomposition method to obtain the cartoon composition, and carry out impact filtering, obtain edge enhanced images

By calculating

The structure tensor node-by-node algorithm

Normal component And tangential component

The direction gradient vector field that obtains the n time iterative image is

Step 2, iterative image direction gradient vector field is calculated in the quick estimation of current yardstick fuzzy core the n time

Frequency spectrum

Then obtain the fuzzy core of the n time iteration in frequency domain

Step 3, the approximate frequency spectrum of calculated direction derivative, according to

Calculate the method directional derivative of the n time iteration and cut the approximate frequency spectrum of directional derivative

Step 4, the Fast Restoration of current scalogram picture, fuzzy core in calculation procedure 2

Frequency spectrum

Then try to achieve a rough estimate of the n time iteration restored image

Step 5, the yardstick inner iteration stops judgement: if n 〉=N _max, go to iteration termination judgement between yardstick, and will

As the final fuzzy core of estimating of current yardstick,

As the blind restored image of current yardstick, inner iteration count n is set to 1; Otherwise upgrade inner iteration count n=n+1, and go to step 1;

1.2.3 between yardstick, iteration stops judgement: if s 〉=S, between yardstick, iteration stops, and will

As final fuzzy core,

As final restored image; Otherwise upgrade yardstick iteration count s=s+1, go between yardstick and upgrade.

2. the multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative according to claim 1, it is characterized in that iteration system parameter initialization in the iteration system initialization procedure, initial parameter value arranges as follows: maximum iteration time N in yardstick _max∈ [6, L, 10]; Constant compensating parameter β ∈ [5,10], difference operator contribution parameters γ=2 β; Image difference operator contribution parameters α ∈ [0.1,0.3].

3. the multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative according to claim 1, is characterized in that in the step 2 of yardstick inner iteration estimation the fuzzy core of the n time iteration

Solution procedure is to obtain fuzzy core in frequency domain, is found the solution by following relational expression to obtain:

$k_s^{\left(n\right)}=\mathrm{Shrink}\left(F^{-1}\left(\frac{F^{*}\left({\▿}_o{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_o{y}_s\right)+F^{*}\left({\▿}_a{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_a{y}_s\right)}{F^{*}\left({\▿}_o{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_o{S}_s^{\left(n\right)}\right)+F^{*}\left({\▿}_a{S}_s^{\left(n\right)}\right)\mathrm{oF}\left({\▿}_a{S}_s^{\left(n\right)}\right)+\mathrm{\γ}(F^{*}\left(D_1\right)\mathrm{oF}\left(D_1\right)+F^{*}\left(D_2\right)\mathrm{oF}\left(D_2\right))+\mathrm{\β}}\right)\right)$

Wherein: β is that constant compensating parameter, γ are the difference operator contribution parameters, F ^*The complex conjugate of () expression F (), F ^-1() expression inverse fourier transform, the pointwise of " o " representing matrix element is multiplied each other, and fraction represents that between two matrixes of molecule denominator, the element pointwise is divided by, shrinkage operation operator Shrink ()=max (, 0.01).

4. the multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative according to claim 1 is characterized in that in step 3 that the yardstick inner iteration estimates, the method directional derivative of the n time iteration with cut the approximate frequency spectrum of directional derivative

Computing formula be:

$F\left({\mathrm{Do}}_s^{\left(n\right)}\right)={\widehat{\mathrm{\μ}}}_s^{(n-1)}\mathrm{oF}\left(D_1\right)+{\hat{v}}_s^{(n-1)}\mathrm{oF}\left(D_2\right)---\left(2\right)$

$F\left({\mathrm{Da}}_s^{\left(n\right)}\right)={-\hat{v}}_s^{(n-1)}\mathrm{oF}\left(D_1\right)+{\widehat{\mathrm{\μ}}}_s^{(n-1)}\mathrm{oF}\left(D_2\right)---\left(3\right)$

Wherein:

With

Represent respectively the n-1 time iteration restored image of current yardstick

The corresponding normalization proper vector that is orthogonal to partial structurtes of structure tensor, the pointwise of " o " representing matrix element is multiplied each other.

5. the multiple dimensioned iterative estimate method of fuzzy core of utilizing the Local Structure of Image directional derivative according to claim 1, is characterized in that in the step 4 of yardstick inner iteration estimation the restored image rough estimate of the n time iteration

To find the solution be to be tried to achieve by following relational expression

$x_s^{\left(n\right)}=F^{-1}\left(\frac{F^{*}\left(k_s^{\left(n\right)}\right)\mathrm{oF}\left(y_s\right)o(1+0.5F^{*}\left({\mathrm{Do}}_s^{\left(n\right)}\right)+0.5F^{*}\left({\mathrm{Da}}_s^{\left(n\right)}\right)\mathrm{oF}\left({\mathrm{Da}}_s^{\left(n\right)}\right))}{F^{*}\left(k_s^{\left(n\right)}\right)\mathrm{oF}\left(k_s^{\left(n\right)}\right)o(1+0.5F^{*}\left({\mathrm{Do}}_s^{\left(n\right)}\right)\mathrm{oF}\left({\mathrm{Do}}_s^{\left(n\right)}\right)+0.5F^{*}\left({\mathrm{Da}}_s^{\left(n\right)}\right)\mathrm{oF}\left({\mathrm{Da}}_s^{\left(n\right)}\right))+\mathrm{\α}(F^{*}\left(D_1\right)\mathrm{oF}\left(D_1\right)+F^{*}\left(D_2\right)\mathrm{oF}\left(D_2\right))}\right)$

Wherein: α is image difference operator contribution parameters, F ^*The complex conjugate of () expression F (), F ^-1() expression inverse fourier transform, the pointwise of " o " representing matrix element is multiplied each other, and fraction represents that between two matrixes of molecule denominator, the element pointwise is divided by.

Images