CN104200443A - Alpha channel and image segmentation combined local motion blur restoration algorithm - Google Patents

Abstract

The invention relates to the field of images, and discloses an alpha channel and image segmentation combined local motion blur restoration algorithm. The algorithm includes: firstly, estimating blur parameters according to information of a blurred image, estimating a point spread function (PSF) through relation between an alpha channel and the PSF, and restoring the blurred image; secondly, accurately positioning the restored image in a background image by the aid of knowledge of image segmentation; finally acquiring the restored image according to provided image synthesis method. The algorithm has the advantages that simulation results show that the provided restoration algorithm is high in estimation accuracy when used for performing blur parameter estimation, restoration effect of motion blur caused by rectilinear motion is good, the provided image synthesis method is effective, and the problem about restoration of local motion blur is solved.

Description

The local motion fuzzy restoration algorithm of cutting apart in conjunction with alpha passage and image

Technical field

The present invention relates to image processing field, particularly local motion blurred picture restores field.

Background technology

The important carrier that image has become the mankind carries out acquisition of information and information interchange has obtained application widely in science and technology, military affairs, safety, industrial or agricultural and daily life.As the carrier of information interchange, whether image can be by clearly discrimination of human eye, has determined that the mankind carry out the success or not of acquisition of information.Picture pick-up device self and outside environment, all may cause picture quality decline in various degree, and such as object ratio distortion, details are unintelligible or have a ghost image etc.

The mankind are carried out to acquisition of information in the decline of picture quality and information interchange has caused sizable difficulty, such as target identification, graphical analysis etc.In order effectively to extract the interested information of the mankind from degraded image, digital image processing techniques fast development is got up.The problem causing in order to solve image quality decrease, Chinese scholars has been studied the method for various image restorations.Twentieth century sixties, Harris carries out model analysis to space is image blurring, has proposed restoration algorithm.Mcglamery adopts the mode of experiment on the basis of Harris theoretical research, utilizes point spread function to deconvolute to atmospheric disturbance.From then on, deconvolute and just become a kind of image restoration of standard technology.But the relative merits of this method all clearly, in the time there is no noise, can realize accurately the recovery of blurred picture, and while there is noise, it is very poor that the effect of image restoration can become.

Scholar afterwards has also done a large amount of research work on this, proposes the Image Restoration Algorithm of many blind signals and non-blind signal.

Summary of the invention

The present invention, in order to solve the problem of prior art, restores for local motion blurred picture, discloses the local motion fuzzy restoration algorithm that a kind of combination alpha passage and image are cut apart, and this algorithm comprises the following steps: (1): obtain blurred picture.Use background subtraction method to extract the fuzzy region in local motion blurred picture.

(2): obtain fuzzy parameter.Utilize the blurred picture that (1) step obtains to become a series of similar blurred pictures through spatial variations, utilize acquisition of information blurred length and the blur direction of blurred picture.

(3): solution point spread function.Solve the alpha passage that blurred picture is corresponding, recycling point spread function is estimated and the relation of alpha passage, is solved point spread function corresponding to fuzzy region.

(4): utilize the point spread function obtaining to carry out restoration disposal to fuzzy region, obtain restored image.

(5): the restored image that (4) step is obtained navigates in background, complete the recovery of entire image.

Beneficial effect

The invention has the beneficial effects as follows, the restoration algorithm of proposition estimated accuracy in the time carrying out fuzzy parameter estimation is high, and good to the motion blur recovery effect being caused by rectilinear motion, the image combining method of proposition is effective, has solved the fuzzy recovery problem of local motion.

Brief description of the drawings

Fig. 1 is that the present invention proposes local motion fuzzy restoration algorithm FB(flow block).

Fig. 2 is that the point spread function that the present invention proposes is estimated process flow diagram.

Fig. 3 is image synthetic zone schematic diagram

Embodiment

Below in conjunction with accompanying drawing, technical scheme of the present invention is explained.

Thinking of the present invention is that elder generation goes out fuzzy parameter according to the information estimator of blurred picture, estimates point spread function by the relation between alpha passage and point spread function (PSF), realizes the recovery of blurred picture; Then the knowledge of utilizing image to cut apart, navigates to restored image in background image exactly; Finally, according to the image combining method proposing, obtain restored image.Below theory derivation and the concrete methods of realizing of this algorithm are described in detail.

In order to obtain the fuzzy part in image, adopt the method for digital matting.In digital matting, think that the pixel value of image any point is by the linear combination of foreground elements color and background element color, be expressed as I _g=α F _g+ (1-α) B _g.Wherein, I _grepresentative image pixel value; F _gand B _grepresent respectively the pixel value of foreground elements and background element; The span of α is 0 to 1, and it can represent foreground elements shared proportion in image pixel, and in like manner 1-α represents background element shared proportion in image pixel value.

The method that obtains fuzzy region is called foreground extraction, and conventional foreground extracting method has three kinds: optical flow method, frame difference method and background subtraction method.Optical flow method arithmetic speed is slower, and it is larger that frame difference method is extracted error, and background subtraction method is the method that uses light method during current video image is processed.The present invention also extracts fuzzy region by employing background subtraction method.

Suppose that local motion blurred picture is C (x, y), the fuzzy part of prospect is F (x, y), and the clear part of background is B (x, y).By setting certain threshold value T, blurred picture C (x, y) and background image B (x, y) carry out reducing and just can obtain the approximate evaluation of foreground image F (x, y),

Extract after fuzzy region, utilize the alpha value of blurred picture to obtain picture point spread function: first to analyze the relation between blurred picture alpha channel value and point spread function, then utilize alpha channel value to realize the estimation of point spread function.

Suppose that object is along horizontal motion, corresponding PSF is designated as h, utilizes matrix knowledge, h can be changed into h _{(2n+1) × 1}=[h _-n, h _-(n-1), L.h ₀, Lh _n-1, h _n] wherein the line number of matrix determined by blurred length, the head and the tail element value h of matrix _-nand h _nall non-vanishing.If L _irepresent arbitrary the movement locus that object carries out tangential movement, vector represent that object is at path L _ialpha passage corresponding to On Local Fuzzy image that upper motion causes, is expressed as ${\mathrm{\α}}_G^{L_i}=[{\mathrm{\α}}_G^{L_i}\left(1\right),{\mathrm{\α}}_G^{L_i}\left(2\right),L{\mathrm{\α}}_G^{L_i}\left(m\right)]$ Wherein the number m of middle element has represented path L _ithe non-zero α that upper blurred picture is corresponding _gsum, vector in head and the tail element with all non-vanishing.The alpha passage that picture rich in detail is corresponding can become ${\mathrm{\α}}_L^{L_i}=[\mathrm{1,1},L\mathrm{1,1,0,0},L\mathrm{0,0}]$

In entry of a matrix element, there is m-2n 1,2n individual 0.By distortion obtains ${\left({\mathrm{\α}}_G^{L_i}\right)}^T={({\mathrm{\α}}_L^{L_i}\⊗h)}^T=H{\left({\mathrm{\α}}_L^{L_i}\right)}^T$

According to multiplication of matrices rule, m × 1 matrix, also be m × 1 matrix, therefore need degeneration system matrix H to be extended to m × m matrix.Expansion process is as follows

$H_{m\×m}={\left[\begin{array}{ccccc}{h}_{-n}& 0& L& h_{-(n-2)}& h_{-(n-1)}\\ h_{-(n-1)}& h_{-n}& L& h_{-(n-3)}& h_{-(n-2)}\\ h_{-(n-2)}& h_{-(n-1)}& L& h_{-(n-4)}& h_{-(n-3)}\\ M& M& O& M& M\\ h_n& h_{n-1}& L& 0& 0\\ 0& h_n& L& 0& 0\\ M& M& O& M& M\\ 0& 0& L& 0& 0\\ 0& 0& L& h_{-n}& 0\\ 0& 0& L& h_{-(n-1)}& h_{-n}\end{array}\right]}_{m\×m}$

Can be calculated

${\mathrm{\α}}_G^{L_i}\left(j\right)=\left\{\begin{array}{cc}\underset{k=-n}{\overset{j-(n+1)}{\mathrm{\Σ}}}{h}_k& 1\≤j\≤2n+1\\ 1& 2n+1\≤j\≤m-2n\\ \underset{k=n+j-m}{\overset{n}{\mathrm{\Σ}}}{h}_k& m-2n\≤j\≤m\end{array}\right.$

By calculating can draw h _kexpression formula

$h_k=\left\{\begin{array}{cc}{\mathrm{\α}}_G^{L_i}(k+n+1)-{\mathrm{\α}}_G^{L_i}(k+n)& -n\≤k\≤n\\ {\mathrm{\α}}_G^{L_i}\left(1\right)& k=n\end{array}\right.$

By to h _kwe find the derivation of expression formula, only need to process the fuzzy part of local motion blur image, try to achieve corresponding alpha channel value, just can estimate the point spread function of motion blur process.

In the one-dimensional space, provide a kind of method of solution point spread function above.The problem that only two dimensional motion of image need to be transformed into motion in one dimension, just can obtain the estimation of point spread function.Estimate to solve motion blur angle θ by fuzzy parameter, then by image rotation θ angle, just the blur direction of image can be become to horizontal direction, i.e. fuzzy angle vanishing, problem just becomes above-mentioned model solution.

The precision of estimating in order to improve point spread function, the present invention has done following processing.Suppose arbitrary line of motion L in blurred picture _icorresponding alpha passage, can utilize in value realize solving of point spread function.By alpha passage in for the new vector of value composition of solution point spread function be expressed as ${\mathrm{\α}}_{\mathrm{head}}^{L_i}=[{\mathrm{\α}}_G^{L_i}\left(1\right),{\mathrm{\α}}_G^{L_i}\left(2\right),L,{\mathrm{\α}}_G^{L_i}(2n+1)]$

In same fuzzy region, select arbitrarily two different motion path L _iand L _j, its correspondence for PSF estimate alpha channel value vector should equate, correspondingly, the related coefficient between the two is

$r=\frac{E\left\{\right[{\mathrm{\α}}_{\mathrm{head}}^{L_i}-E\left({\mathrm{\α}}_{\mathrm{head}}^{L_i}\right)\left]\right[{\mathrm{\α}}_{\mathrm{head}}^{L_j}-E\left({\mathrm{\α}}_{\mathrm{head}}^{L_j}\right)\left]\right\}}{\sqrt{E\left\{{[{\mathrm{\α}}_{\mathrm{head}}^{L_i}-E\left({\mathrm{\α}}_{\mathrm{head}}^{L_i}\right)]}^2\right\}}g\sqrt{E\left\{{[{\mathrm{\α}}_{\mathrm{head}}^{L_j}-E\left({\mathrm{\α}}_{\mathrm{head}}^{L_j}\right)]}^2\right\}}}=1.$

But in practice, the alpha value that same fuzzy region obtains along different motion line is also different, and then by passage element the difference vector of composition between related coefficient be not 1 substantially.Screen so need to do in the time that the alpha value of utilizing fuzzy region realizes the estimation of fuzzy region point spread function responsively.Due to vectors different under perfect condition between related coefficient should be r=1, the vector that is the alpha value composition that related coefficient is higher is more accurate in the time estimating for point spread function, carry out the estimation for point spread function therefore can select the alpha value vector that the degree of correlation is high by setting threshold, correspondingly improved the degree of accuracy that PSF estimates.

Estimate toward the method for upper strata successive iteration on the blurred length l of fuzzy part and the basis of blur direction θ from lower floor in use, with reference to Fig. 2, point spread function is estimated, mainly contain seven steps, as follows respectively:

The first step: by the rotation of image travel direction, make its blur direction become horizontal direction (being θ=0) according to the blur direction θ trying to achieve, obtain new blurred picture g.

Calculate its corresponding alpha channel value α _g, composition matrix [α _g] _{m × N}.

The 3rd step: to matrix [α _g] _{m × N}in the element of arbitrary row screen, the element screening forms new vector ${\mathrm{\α}}_{\mathrm{head}}^i=[{\mathrm{\α}}_g^i\left(j\right),{\mathrm{\α}}_g^i(j+1),L,{\mathrm{\α}}_g^i(j+l-1)],$ Wherein represent the matrix that the capable element of electing of alpha channel value matrix i forms, 1≤i≤M, 1≤j≤N-l+1.When screening, must meet condition below: (1) simultaneously (2) (3) t _fspan is 0.97～0.99.

The 4th step: to [α _g] _{m × N}the operation of the 3rd step is repeated in each provisional capital of matrix, obtains altogether M matrix; Then by the M obtaining be configured to new matrix A _head, formation rule is it is matrix A _headeach row are matrix [α _g] _{m × N}in corresponding row, screen inversion.

Compute matrix A _headthe related coefficient of every two row, deposits the result obtaining in matrix of coefficients C _headin.Wherein Matrix C _headelements C _head(ij) representing matrix A _headin the related coefficient of i column vector and j column vector.

The setting of passing threshold, utilizes correlation matrix C _headselection is used for carrying out the vector of point spread function estimation by setting global threshold T _c(T _cbe worth greatlyr, correlativity is higher, and general value is 0.95～0.98), select matrix of coefficients C _headin be greater than threshold value T _celement value, form new Matrix C ' _head.Statistical matrix C' _headthe number of nonzero element in each row, and be deposited into new matrix D _headin, D _headin i element value representing matrix C' _headin the number of i row nonzero element, by threshold value T is set _d, select D _headin be greater than threshold value T _delement value form new matrix D ' _head.Finally by D' _headmiddle nonzero element value is corresponding vector is as being used for carrying out the input of point spread function estimation, and the alpha channel value vector that all selections are obtained deposit point spread function estimated matrix F in _headin.Wherein threshold value T _dsetting rule be: wherein, it is matrix D _headin greatest member value.

The 7th step: carry out the calculating of PSF.For the matrix F of solution point spread function _headin each column matrix can be used for carrying out the estimation of point spread function.If obtain S through conditional filtering above matrix, so point spread function estimated matrix $F_{\mathrm{head}}=\left[\right[{\left({\mathrm{\α}}_{\mathrm{head}}^1\right)}^T,{\left({\mathrm{\α}}_{\mathrm{head}}^2\right)}^T,L,{\left({\mathrm{\α}}_{\mathrm{head}}^S\right)}^T],$ In order further to eliminate due to the error of introducing in matrixing process, by matrix F _headin column vector between obtain α after corresponding element arithmetic mean _head, utilize h _kexpression formula solves point spread function h _head.Finally by h _headbe rotated counterclockwise the point spread function estimation h that fuzzy angle θ just can obtain blurred picture _g.

After having calculated point spread function, just can adopt L-R algorithm to carry out restoration disposal to image.But for the fuzzy image of local motion, fuzzy part mainly concentrates in foreground area, most of background area is clearly.If being carried out to the recovery of L-R algorithm, integral image is certain to allow originally background clearly thicken.For this situation, the recovery work of entire image can be divided into two and walk greatly.The first step: to extracting blurred picture (the foreground image F obtaining _g) carry out L-R and restore and obtain corresponding restored image F _g'; Second step: recovery is obtained to F _g' and background image B clearly _gcombine by certain mode, obtain the restored image (array mode as shown in Figure 3) of full width image.In this process, there are 2 needs to note especially.The first point: image is synthetic is to use background image information B _gwith foreground image information F _g' superposeing obtains image, i.e. I _sum=B _g+ F _g', how by foreground image F _g' correctly navigate to background image B _gmiddlely will be the committed step that affects image restoration.Second point: foreground image F _g' and background image B _gcross-shaped portion branch there is significantly splicing vestige, how making the image after combination can natural transition be also the key that affects image restoration quality in the pixel in this region.

For by restore foreground image F _g' correct navigate to clear background image B _gin, utilize background subtraction method to extract the alpha passage staying after foreground image and can realize accurate location.In order to solve the obviously problem of splicing vestige that exists, can be at foreground image F _g' and background image B _gintersection specify a piece region, the pixel in this region is reset.Piece zone definitions is ambiguous prospect image F _gmiddle alpha value is the part (as Fig. 3 region 3) between 0 to 1, and this area pixel value is by I ₃=w*I _f'+ (1-w) I _bdetermine.Wherein w represents weight, I _f'and I _bbe respectively F _g' and B _gpixel value.W is defined as follows: the scope of supposing overlapping region projection on x axle is x _min: x _max, when w during to 0 variation, just completes image pixel by F by 1 _g' to B _gtransition.

Above content is in conjunction with concrete preferred implementation further description made for the present invention, can not assert that specific embodiment of the invention is confined to these explanations.For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, can also make some simple deduction or replace, all should be considered as belonging to protection scope of the present invention.

Claims (8)

1. a local motion fuzzy restoration algorithm of cutting apart in conjunction with alpha passage and image, is characterized in that: can effectively from the On Local Fuzzy image being caused by motion, restore Useful Information and can carry out preferably image synthetic, comprise the steps:

(1) use background subtraction method to extract the fuzzy region in local motion blurred picture, obtain blurred picture;

(2) utilize the blurred picture that the first step obtains to become a series of similar blurred pictures through spatial variations, utilize acquisition of information blurred length and the blur direction of blurred picture;

(3) solve the alpha passage that blurred picture is corresponding, recycling point spread function is estimated and the relation of alpha passage, is solved point spread function corresponding to fuzzy region;

(4) utilize the point spread function obtaining to carry out restoration disposal to fuzzy region, obtain restored image;

(5) restored image step (4) being obtained navigates in background, completes the recovery of entire image.

2. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, is characterized in that, the described local motion blurred picture of step (1) refers to the image that the moving object in shooting static background obtains.

3. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, is characterized in that, in step (1), the fuzzy region of definition is wherein C (x, y) is motion blur image, and F (x, y) is the fuzzy part of prospect, and B (x, y) is the clear part of background.

4. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, it is characterized in that, in step (2), adopt blurred length and the blur direction of the acquisition of the method toward the upper strata successive iteration blurred picture from lower floor.

5. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, is characterized in that, the relational expression that the alpha passage that in step (3), derivation obtains and point spread function are estimated is $h_k=\left\{\begin{array}{cc}{\mathrm{\&alpha;}}_G^{L_i}(k+n+1)-{\mathrm{\&alpha;}}_G^{L_i}(k+n)& -n\&le;k<n\\ {\mathrm{\&alpha;}}_G^{L_i}\left(1\right)& k=n\end{array}\right..$

6. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, is characterized in that, the estimation of the point spread function in step (3) has comprised following steps:

(1), by image rotation, obtain new image;

(2) extract fuzzy part and calculate corresponding matrix;

(3) screening element obtains candidate matrices

(4) repeating step (3), obtains every a line form new matrix $A_{\mathrm{head}}=[{\left({\mathrm{\&alpha;}}_{\mathrm{head}}^1\right)}^T,{\left({\mathrm{\&alpha;}}_{\mathrm{head}}^2\right)}^T,L,{\left({\mathrm{\&alpha;}}_{\mathrm{head}}^M\right)}^T];$

(5) solve correlation matrix;

(6) setting threshold;

(7) point spread function is estimated.

7. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, is characterized in that, the restoration disposal in step (4) is the point spread function obtaining according to estimating, substitution L-R algorithm carries out image restoration.

8. the local motion fuzzy restoration algorithm that combination alpha passage according to claim 1 and image are cut apart, it is characterized in that, image in step (5) is synthetic is that the alpha passage that utilizes background subtraction method to extract to stay after foreground image is realized accurate location.