CN102915527A - Face image super-resolution reconstruction method based on morphological component analysis - Google Patents

Abstract

The invention discloses a face image super-resolution reconstruction method based on morphological component analysis. The method comprises the following steps: upsampling a low-resolution input image to acquire an interpolated image; acquiring a whole high-resolution face image from the interpolated image by a morphological component analysis method; downsampling the whole high-resolution face image and subtracting the whole high-resolution face image from the low-resolution input image to acquire a low-resolution residual image, and performing face detail information compensation by performing neighboring construction on an image block at each face position to acquire a high-resolution residual image; and combining the high-resolution residual image with the whole high-resolution face image to acquire the face image super-resolution result and finish resolution reconstruction. The invention also provides a method for performing face image super-resolution reconstruction and expression normalization simultaneously on the basis of morphological component analysis. By the method, detailed face detail information can be acquired, the ringing phenomenon is eliminated and high image quality is achieved.

Description

Face image super-resolution reconstruction method based on the morphology constituent analysis

Technical field

The present invention relates to image processing field, particularly a kind of face image super-resolution reconstruction method based on the morphology constituent analysis.

Background technology

The task of image super-resolution is to infer high-definition picture from one or more low resolution input picture.Be widely used in real life, particularly the Super-resolution reconstruction of facial image is built the aspects such as long-distance video monitoring or Video processing in important application.

Traditional face image super-resolution method combines the peculiar classical priori of facial image, and a lot of face image super-resolution methods are appearring closely during the decade, as make up based on differential information radio-frequency component, based on principal component analysis (PCA) (PCA) set up block mold reconstruct high-resolution human face image, in conjunction with based on the block mold of PCA and the methods such as " whole faces+residual error face " that makes up based on the partial model of belief propagation.Although existing face image super-resolution reconstruction method has been obtained good effect, but still have greatly improved the space, a part that wherein needs to promote is whole face feature, existing algorithm often causes people's face of reconstruct and real people's face unlike same person, and generation ringing effect, this is because existing estimation model only to global face rather than all the components modeling of input picture, has therefore reduced the degree of accuracy of estimating.

Summary of the invention

Fundamental purpose of the present invention is to overcome the shortcoming of prior art with not enough, face image super-resolution reconstruction method based on the morphology constituent analysis is provided, people's face that the method is rebuild is more as real people's face, can obtain more detailed facial detail information, eliminate ringing, obtain better image quality.In addition, the present invention can also be extended to and solve simultaneously face image super-resolution and the normalized problem of expression.

Purpose of the present invention realizes by following technical scheme: the face image super-resolution reconstruction method based on the morphology constituent analysis may further comprise the steps:

(1) by interpolation method low resolution input picture up-sampling, obtain the image after the interpolation;

(2) image after using the morphology component analyzing method to the resulting interpolation of step (1) carries out picture breakdown, resolves into whole high-resolution human face image and anti-sharpening mask;

(3) the whole high-resolution human face image that step (2) is obtained carries out down-sampling, then reduced by the low resolution input picture, obtain the low-resolution residual error image, then come image is carried out the facial detail information compensation by the image block on the face location of low-resolution residual error image everywhere is carried out neighbour's reconstruct, obtain the high-resolution residual error image;

(4) the whole high-resolution human face image that the high-resolution residual error image in the step (3) and step (2) is obtained merges, and obtains final face image super-resolution result, finishes super-resolution rebuilding.

Up-sampling Interpolation Process formula in the described step (1) is: x _Zoom=s ↑ x _l=s ↑ s ↓ (h*x _h) ≈ h*x _h, s ↑ be up-sampling interpolation operation symbol wherein, s ↓ be the down-sampling operational character, x _lThe low resolution input picture, x _ZoomThe image after the interpolation, x _hBe true high-definition picture, h is the some propagator of camera, and * is the convolution operation symbol; Image x after the interpolation _ZoomWith equation expression be: x _Zoom=x _h+ x _s, x wherein _sBe comprise on the occasion of with the radio-frequency component of negative value, be called anti-sharpening mask, the image x after the interpolation _ZoomX _hVersion after fuzzy.

Preferably, the step of carrying out picture breakdown of the use morphology component analyzing method in the described step (2) specifically refers to:

(2-1) choose at random different people's face high-definition pictures, will obtain people's face low-resolution image behind its down-sampling, obtain image after the interpolation by interpolation method again, the image after the interpolation and corresponding high-definition picture are subtracted each other, obtain anti-sharpening mask; With high-definition picture and anti-sharpening mask as training sample, by the average of principal component analysis (PCA) (Principal Component Analysis, PCA) methodology acquistion to true high-definition picture

Average with anti-sharpening mask

Obtain all N eigenwert of corresponding true high-definition picture

All N eigenwert with anti-sharpening mask

And the eigenmatrix P of corresponding true high-definition picture _hEigenmatrix P with anti-sharpening mask _s, x then _hAnd x _sBe expressed as respectively:

α wherein _hAnd α _sProjection coefficient, P _hAnd P _sBeing used as dictionary describes respectively

With

(2-2) based on the morphology component analyzing method, its core is to find the solution an optimized process, and the formula of finding the solution is $\underset{{\mathrm{\α}}_h,{\mathrm{\α}}_s}{\min }\left|\right|{\mathrm{\α}}_h|{|}_0+{\left|\right|{\mathrm{\α}}_s\left|\right|}_0,$ Constraint condition is:

Wherein, ε is the fault tolerant rate, and (m, n) is pixel coordinate, and S is enlargement factor, x _lIt is the low resolution input picture; Finally, by optimum solution

Produce the subsignal of separation

With

Above-mentioned solution formula is a NP Solve problems, and its sparse solution can approach with the solution of L1 Norm minimum, can find the solution by using alternately the method for iteration threshold, and each iteration all will be the iteration result super ellipsoid that stretches back

Within the scope, and guarantee Reconstruction Constraints, namely input low-resolution image x _lWith high-definition picture x _hRelation need to satisfy And threshold value and iterations can obtain suitable value by experiment.Solution procedure is specific as follows:

(2-2-1) initialization: input maximum iteration time I _MaxInput threshold value maxT _h, maxT _s, minTh and minT _s, these threshold values are: ${\max T}_h={\left|\right|P_h^T(x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s})\left|\right|}_{\∞},$

MinT _hAnd minT _sInput value is the square root of t eigenvalue of maximum corresponding to the corresponding true high-definition picture that uses the PCA method and obtain and t major component of anti-sharpening mask; Input back projection operational symbol p; Set x _h, x _sInitial value is 0, T _h=maxT _h, T _s=maxT _sSet iteration variable k, allow k get I from l _Max, step (2-2-2) is the process of iteration;

(2-2-2) keep x _sConstant calculating x _hThe process of new value:

Calculate first residual error $r=x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s}-x_h-x_s,$

Then calculate x _h+ r is to P _hProjection coefficient:

Passing threshold T _hTo parameter alpha _hCarrying out the hard-threshold processing obtains Convergent-divergent Obtain to satisfy constraint condition

α _sGet in the last iterative process and upgrade x _sObtain

Calculate, constraint condition is as follows:

${\left|\right|x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s}-P_h{\mathrm{\α}}_h-P_s{\mathrm{\α}}_s\left|\right|}_2\≤\mathrm{\ϵ},$

i=1…N,

$x_l(m,n)=\frac{1}{S^2}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{x}_h(\mathrm{Sm}+i,\mathrm{Sn}+j).;$

Reconstruct $x_h=\stackrel{\‾}{x_h}+P_h{\mathrm{\α}}_h^2;$

Back projection correction: with x _hReplace with

Then, keep x _hConstant calculating x _sNew value process is as follows:

Calculate first residual error

Then calculate x _s+ r is to P _sProjection coefficient:

Passing threshold T _sTo parameter alpha _sCarrying out the hard-threshold processing obtains

Convergent-divergent

Obtain to satisfy constraint condition

α _hCalculate with front portion Constraint condition is as follows:

${\left|\right|x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s}-P_h{\mathrm{\α}}_h-P_s{\mathrm{\α}}_s\left|\right|}_2\≤\mathrm{\ϵ},$

i=1…N,

$x_l(m,n)=\frac{1}{S^2}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{x}_h(\mathrm{Sm}+i,\mathrm{Sn}+j).;$

Reconstruct $x_s=\stackrel{\‾}{x_s}+P_s{\mathrm{\α}}_s^2;$

Upgrade threshold value $T_h=T_h-\frac{{\max T}_h-{\min T}_h}{I_{\max }-1},$ $T_s=T_s-\frac{{\max T}_s-{\min T}_s}{I_{\max }-1};$

When iterations then withdraws from iterative program and obtains optimum solution greater than the maximum iteration time that arranges

Produce the subsignal of separation

With

Concrete, in the described step (3), the concrete steps that obtain the low-resolution residual error image are: the whole high-resolution human face image that is obtained by step (2)

With low resolution input picture x _l, definition low-resolution residual error image is:

By based on neighbour's reconstruct of the image block on the face location of everywhere from

Set up the high-resolution residual error image

Estimation.

Concrete, described step (3) is to neighbour's reconstructing method of the image block on the face location of everywhere, and is specific as follows:

Obtain people's face low-resolution image as training sample behind the people's face high-definition picture that (3-1) will choose at random and its down-sampling.Low-resolution image applying step (2) is obtained whole high-resolution human face image, then by the subduction of people's face full resolution pricture, obtain high-resolution residual error image training sample.Obtained low-resolution residual error image training sample by the subduction of people's face low-resolution image behind the whole high-resolution human face image down sampling.

(3-2) with the low-resolution residual error image of inputting

Be divided into image block, make between the image block to overlap.

(3-3) for the low-resolution residual error image block of each input, k arest neighbors low-resolution residual error image block training sample by same face location carries out portfolio restructuring, then the weights that keeping combination synthesize corresponding high-resolution residual error image block by substituting low-resolution residual error image block training sample for corresponding high-resolution residual error image block training sample.

(3-4) obtain complete high-resolution residual error image by the integrated high-resolution residual error image block of averaged overlapping portions

As another kind of preferred version, based on the Multichannel Decomposition ability of morphology component analyzing method, to the expression shape change modeling in the morphology constituent analysis, expansion algorithm is to carrying out simultaneously face image super-resolution and expression normalization by the design dictionary.Face image super-resolution reconstruction method based on the morphology constituent analysis may further comprise the steps:

(1) study obtains the mutual incoherent expression dictionary of several correspondences to some classes expression by principal component analytical method;

(2) for the low resolution input picture of input, judge the expression that it is affiliated.Then, on the one hand carry out up-sampling, obtain the image of the espressiove after the interpolation; Use on the other hand the morphology component analyzing method, based on its corresponding expression dictionary of expressing one's feelings, the low resolution input picture is carried out picture breakdown, resolve into the amimia input picture of whole low resolution and low resolution espressiove mask;

(3) image of the espressiove after using the morphology component analyzing method to the resulting interpolation of step (2) carries out picture breakdown, based on its corresponding expression dictionary of expressing one's feelings, resolves into the amimia image of whole high resolving power, expression mask and anti-sharpening mask;

(4) the amimia image of the resulting whole high resolving power of step (3) is carried out down-sampling, then and the amimia input picture of whole low resolution that obtains of step (2) subtract each other, obtain the low-resolution residual error image; To this low-resolution residual error image, come image is carried out the facial detail information compensation by the image block on the face location of everywhere being carried out neighbour's reconstruct, obtain the high-resolution residual error image;

(5) the amimia image of whole high resolving power that the high-resolution residual error image in the step (4) and step (3) is obtained merges, and obtains final face image super-resolution result, finishes super-resolution rebuilding.

Further, in the described step (2), by adopting LBP(Local Binary Patterns, local binary patterns) sorter judges the expression in the low resolution input picture.

As preferably, face image super-resolution reconstruction method based on the morphology constituent analysis, all low resolution input pictures are all judged whether espressiove of its people's face by sorter before processing, if do not have, then adopt first method to process, if espressiove then adopts second method to process.

The present invention compared with prior art has following advantage and beneficial effect:

1, the present invention is interpreted as the picture breakdown problem to the face image super-resolution problem, and the framework that uses the morphology component analyzing method to set up one three step has been realized the decomposition of image effectively, experiment shows, method of the present invention is more effective than existing method, can obtain more detailed facial detail information, eliminated ringing, obtained better image quality, and under the scene of laboratory environment and reality, can both realize satisfactory results.

2, the invention allows for a kind of method, be extended to and solve simultaneously face image super-resolution and the normalized problem of expression, this is first successful method in the existing method.Thereby can further improve the accuracy of facial image identification.

Description of drawings

Fig. 1 is the algorithm flow chart of the embodiment of the invention 1;

Fig. 2 is the algorithm flow chart of the embodiment of the invention 2.

Embodiment

The present invention is described in further detail below in conjunction with embodiment and accompanying drawing, but embodiments of the present invention are not limited to this.

Embodiment 1

As shown in Figure 1, the face image super-resolution reconstruction method based on the morphology constituent analysis may further comprise the steps:

(1) by interpolation method low resolution input picture up-sampling, obtain the image after the interpolation;

(2) image after using the morphology component analyzing method to the resulting interpolation of step (1) carries out picture breakdown, resolves into whole high-resolution human face image and anti-sharpening mask;

(3) the whole high-resolution human face image that step (2) is obtained carries out down-sampling, then reduced by the low resolution input picture, obtain the low-resolution residual error image, then come image is carried out the facial detail information compensation by the image block on the face location of low-resolution residual error image everywhere is carried out neighbour's reconstruct, obtain the high-resolution residual error image;

(4) the whole high-resolution human face image that the high-resolution residual error image in the step (3) and step (2) is obtained merges, and obtains final face image super-resolution result, finishes super-resolution rebuilding.

Up-sampling Interpolation Process formula in the described step (1) is: x _Zoom=s ↑ x _l=s ↑ s ↓ (h*x _h) ≈ h*x _h, s ↑ be up-sampling interpolation operation symbol wherein, s ↓ be the down-sampling operational character, x _lThe low resolution input picture, x _ZoomThe image after the interpolation, x _hBe true high-definition picture, h is the some propagator of camera, and * is the convolution operation symbol; Image x after the interpolation _ZoomWith equation expression be: x _Zoom=x _h+ x _s, x wherein _sBe comprise on the occasion of with the radio-frequency component of negative value, be called anti-sharpening mask, the image x after the interpolation _ZoomX _hVersion after fuzzy.

The step that use morphology component analyzing method in the described step (2) carries out picture breakdown specifically refers to:

(2-1) choose at random different people's face high-definition pictures, will obtain people's face low-resolution image behind its down-sampling, obtain image after the interpolation by interpolation method again, the image after the interpolation and corresponding high-definition picture are subtracted each other, obtain anti-sharpening mask; High-definition picture and anti-sharpening mask as training sample, are obtained the average of true high-definition picture by principal component analytical method study

Average with anti-sharpening mask Obtain all N eigenwert of corresponding true high-definition picture

All N eigenwert with anti-sharpening mask

And the eigenmatrix P of true high-definition picture _hEigenmatrix P with anti-sharpening mask _s, x then _hAnd x _sBe expressed as respectively:

α wherein _hAnd α _sProjection coefficient, P _hAnd P _sJust being used as dictionary describes respectively

With

(2-2) based on the morphology component analyzing method, solution formula is:

$\underset{{\mathrm{\α}}_h,{\mathrm{\α}}_s}{\min }{\left|\right|{\mathrm{\α}}_h\left|\right|}_0+{\left|\right|{\mathrm{\α}}_s\left|\right|}_0,$

Constraint condition is as follows:

${\left|\right|x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s}-P_h{\mathrm{\α}}_h-P_s{\mathrm{\α}}_s\left|\right|}_2\≤\mathrm{\ϵ},$

i=1…N,

$x_l(m,n)=\frac{1}{S^2}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{x}_h(\mathrm{Sm}+i,\mathrm{Sn}+j).$

Wherein, ε is the fault tolerant rate, and (m, n) is pixel coordinate, and S is enlargement factor, x _lIt is the low resolution input picture; Finally, by optimum solution

Produce the subsignal of separation

With

Above-mentioned solution formula is a NP Solve problems, and its sparse solution can approach with the solution of L1 Norm minimum, can find the solution by using alternately the method for iteration threshold, and each iteration all will be the iteration result super ellipsoid that stretches back

Within the scope, and guarantee Reconstruction Constraints, namely input low-resolution image x _lWith high-definition picture x _hRelation need to satisfy

And threshold value and iterations can obtain suitable value by experiment.

In the described step (3), the concrete steps that obtain the low-resolution residual error image are: the whole high-resolution human face image that is obtained by step (2)

With low resolution input picture x _l, definition low-resolution residual error image is:

By based on neighbour's reconstruct of the image block on the face location of everywhere from

Set up the high-resolution residual error image

Estimation.

Described step (3) is to neighbour's reconstructing method of the image block on the face location of everywhere, and is specific as follows:

Obtain people's face low-resolution image as training sample behind the people's face high-definition picture that (3-1) will choose at random and its down-sampling.Low-resolution image applying step (2) is obtained whole high-resolution human face image, then by the subduction of people's face full resolution pricture, obtain high-resolution residual error image training sample.Obtained low-resolution residual error image training sample by the subduction of people's face low-resolution image behind the whole high-resolution human face image down sampling.

(3-2) with the low-resolution residual error image of inputting

Be divided into image block, make between the image block to overlap.

(3-3) for the low-resolution residual error image block of each input, k arest neighbors low-resolution residual error image block training sample by same face location carries out portfolio restructuring, then the weights that keeping combination synthesize corresponding high-resolution residual error image block by substituting low-resolution residual error image block training sample for corresponding high-resolution residual error image block training sample.

(3-4) obtain complete high-resolution residual error image by the integrated high-resolution residual error image block of averaged overlapping portions

Embodiment 2

Present embodiment except following characteristics other structures with embodiment 1: based on the Multichannel Decomposition ability of morphology component analyzing method, to the expression shape change modeling in the morphology constituent analysis, expansion algorithm is to carrying out simultaneously face image super-resolution and expression normalization by the design dictionary.Face image super-resolution reconstruction method based on the morphology constituent analysis may further comprise the steps:

(1) set five kinds of basic human face expressions: close one's eyes, frown, smile, surprised, open one's mouth.Respectively five kinds of expression study are obtained mutual incoherent expression dictionary by principal component analytical method, specific as follows:

(1-1) to each expression, choose at random different people face high-definition picture and the corresponding amimia high-definition picture of this expression.People's face low-resolution image and the corresponding amimia low-resolution image of this expression will be obtained behind its down-sampling.Obtain image after the interpolation by interpolation method again.With this facial expression image after the interpolation and corresponding amimia image subtraction, obtain high resolving power expression mask.With this facial expression image of low resolution and corresponding amimia image subtraction, obtain low resolution expression mask.

(1-2) with the expression mask of high-resolution and low-resolution as training sample, learn respectively to obtain expression average, expressive features value and the corresponding expression dictionary of high-resolution and low-resolution by principal component analytical method.

(2) for the low resolution input picture of input, judge the expression that it is affiliated.Then, on the one hand carry out up-sampling, obtain the image of the espressiove after the interpolation; Use on the other hand the morphology component analyzing method, based on its express one's feelings corresponding low resolution expression average, expressive features value and expression dictionary, the low resolution input picture is carried out picture breakdown, resolve into the amimia input picture of whole low resolution With low resolution espressiove mask x _Le

(3) because the facial expression image after the interpolation

Can be expressed as the amimia image x after the interpolation _ZoomWith high resolving power expression mask x _eAnd:

And the amimia image after the interpolation can further be expressed as x _Zoom=x _h+ x _s, then Therefore, the image of the espressiove after using the morphology component analyzing method to the resulting interpolation of step (2) carries out picture breakdown, based on its express one's feelings corresponding high resolving power expression average, expressive features value and expression dictionary, resolve into the amimia image of whole high resolving power The expression mask

With anti-sharpening mask

(4) the amimia image of the resulting whole high resolving power of step (3) is carried out down-sampling, then and the amimia input picture of whole low resolution that obtains of step (2) subtract each other, obtain the low-resolution residual error image

To this low-resolution residual error image, come image is carried out the facial detail information compensation by the image block on the face location of everywhere being carried out neighbour's reconstruct, obtain the high-resolution residual error image

(5) the amimia image of whole high resolving power that the high-resolution residual error image in the step (4) and step (3) is obtained merges, and obtains final face image super-resolution result

Finish super-resolution rebuilding.

In the described step (2), judge expression in the low resolution input picture by adopting the LBP sorter.

The same on the process of using the morphology component analyzing method in the described step (1), step (3) and the Method And Principle of embodiment 1, just embodiment 1 becomes 2 parts to picture breakdown, and embodiment 2 has added the expression mask, therefore need to become 3 parts to picture breakdown.

Embodiment 3

Present embodiment except following characteristics other structures with embodiment 1: in the present embodiment, face image super-resolution reconstruction method based on the morphology constituent analysis, all low resolution input pictures are all judged whether espressiove of its people's face by sorter before processing, if do not have, then adopt embodiment 1 described method to process, if espressiove then adopts embodiment 2 described methods to process.

The present invention describes effect of the present invention by following experiment: experimental selection CES-PEAL-R1 face data storehouse, this database comprises 99594 images of 1040 people, all image scalings on the storehouse are done simple calibration to the 128*96 pixel and according to the position of eyes and face, as shown in table 1, the subset of 5 positive face images is used in two experiments, and existing certain methods also has been implemented with as a comparison:

The content of form 1.CAS-PEAL-R1 database 5 subsets

First experiment is the visual quality evaluation experimental.The common collection of CAS-PEAL-R1 is used in this experiment, common concentrated each personage comprises 1040 images, with amplification factor 4 (or 8) original high-definition picture is carried out the level and smooth and down-sampled low-resolution image that obtains corresponding 32*24 (perhaps 16*12), and they are called high resolving power-low-resolution image pair.For each amplification factor, select 1000 high resolving power-low-resolution images to as training set, remaining 40 images are to as test set, in particular for the framework that uses " whole face+residual error face ", with 700 images to training block mold, with 300 images to training the residual compensation model.For morphology constituent analysis parameter, maximum threshold value is Lookup protocol as follows: ${\max T}_h={\left|\right|P_h^T(x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s})\left|\right|}_{\∞}$ With

Threshold value minT _hAnd minT _sBe the square root that uses t eigenvalue of maximum corresponding to corresponding true high-definition picture that the PCA method obtains and t major component of anti-sharpening mask, algorithm can obtain stable output when the observation number of times that can obtain working as iteration was above 50 times by experiment.For the quantized image vision quality, calculate the difference that 2 indexs are measured image and the true high-definition picture of reconstructing human face super resolution, one of them is square error (MSE), formula is

$\mathrm{MSE}={\mathrm{\Σ}}_{i=1}^N{\left|\right|I_{\mathrm{test},i}^h-J_{\mathrm{test},i}^h\left|\right|}_2^2/\left(\mathrm{WHN}\right)$ Wherein

True high-definition picture,

Be the facial image of reconstruct, it is picture traverse and height that W and H differentiate, and N is the quantity of test pattern, and another one is structural similarity index (SSIM).Present embodiment institute employing method is that the MSE value that obtained in 4 o'clock is that 85, SSIM value is 0.9 in amplification factor, is that the MSE value that obtained in 8 o'clock is that 260, SSIM value is 0.79 in amplification factor, has good image quality assessment compared to more existing algorithms.

Second experiment is super-resolution reconstruction and expression normalization experiment: select 240 images as test set from the expression subset, remaining image is used as training set in the subset, and other parameters of experiment arrange consistent with experiment before.The result shows that the method for present embodiment can obtain expressionless high-definition picture from the low resolution input picture of multiple expression, and the recognition of face rate is higher by 15% than the method for bilinear interpolation, and is higher by 5% than the method for using original high-definition picture with multiple expression.

Above-described embodiment is the better embodiment of the present invention; but embodiments of the present invention are not restricted to the described embodiments; other any do not deviate from change, the modification done under Spirit Essence of the present invention and the principle, substitutes, combination, simplify; all should be the substitute mode of equivalence, be included within protection scope of the present invention.

Claims (7)

1. based on the face image super-resolution reconstruction method of morphology constituent analysis, it is characterized in that, may further comprise the steps:

(1) by interpolation method low resolution input picture up-sampling, obtain the image after the interpolation;

(2) image after using the morphology component analyzing method to the resulting interpolation of step (1) carries out picture breakdown, resolves into whole high-resolution human face image and anti-sharpening mask;

(3) the whole high-resolution human face image that step (2) is obtained carries out down-sampling, then reduced by the low resolution input picture, obtain the low-resolution residual error image, then come image is carried out the facial detail information compensation by the image block on the face location of low-resolution residual error image everywhere is carried out neighbour's reconstruct, obtain the high-resolution residual error image;

(4) the whole high-resolution human face image that the high-resolution residual error image in the step (3) and step (2) is obtained merges, and obtains final face image super-resolution result, finishes super-resolution rebuilding.

2. the face image super-resolution reconstruction method based on the morphology constituent analysis according to claim 1 is characterized in that, the up-sampling Interpolation Process formula in the described step (1) is: x _Zoom=s ↑ x _i=s ↑ s ↓ (h*x _h) ≈ h*x _h, s ↑ be up-sampling interpolation operation symbol wherein, s ↓ be the down-sampling operational character, x _lThe low resolution input picture, x _ZoomThe image after the interpolation, x _hBe true high-definition picture, h is the some propagator of camera, and * is the convolution operation symbol; Image x after the interpolation _ZoomWith equation expression be: x _Zoom=x _h+ x _s, x wherein _sBe comprise on the occasion of with the radio-frequency component of negative value, be called anti-sharpening mask, the image x after the interpolation _ZoomX _hVersion after fuzzy.

3. the face image super-resolution reconstruction method based on the morphology constituent analysis according to claim 1 is characterized in that, the step that the use morphology component analyzing method in the described step (2) carries out picture breakdown specifically refers to:

(2-1) choose at random different people's face high-definition pictures, will obtain people's face low-resolution image behind its down-sampling, obtain image after the interpolation by interpolation method again, the image after the interpolation and corresponding high-definition picture are subtracted each other, obtain anti-sharpening mask; High-definition picture and anti-sharpening mask as training sample, are obtained the average of true high-definition picture by principal component analytical method study Average with anti-sharpening mask

Obtain all N eigenwert of corresponding true high-definition picture

All N eigenwert with anti-sharpening mask

And the eigenmatrix P of corresponding true high-definition picture _hEigenmatrix P with anti-sharpening mask _s, x then _hAnd x _sBe expressed as respectively:

α wherein _hAnd α _sProjection coefficient, P _hAnd P _sBeing used as dictionary describes respectively

With

(2-2) based on the morphology component analyzing method, solution formula is:

$\underset{{\mathrm{\α}}_h,{\mathrm{\α}}_s}{\min }{\left|\right|{\mathrm{\α}}_h\left|\right|}_0+{\left|\right|{\mathrm{\α}}_s\left|\right|}_0,$

Constraint condition is as follows:

${\left|\right|x_{\mathrm{zoom}}-\stackrel{\‾}{x_h}-\stackrel{\‾}{x_s}-P_h{\mathrm{\α}}_h-P_s{\mathrm{\α}}_s\left|\right|}_2\≤\mathrm{\ϵ},$

i=1…N,

$x_l(m,n)=\frac{1}{S^2}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{x}_h(\mathrm{Sm}+i,\mathrm{Sn}+j);$

Wherein, ε is the fault tolerant rate, and (m, n) is pixel coordinate, and S is enlargement factor, x _lIt is the low resolution input picture; Finally, produce the subsignal of separation

With

Be whole high-resolution human face image.

4. the face image super-resolution reconstruction method based on the morphology constituent analysis according to claim 1, it is characterized in that, in the described step (3), the concrete steps that obtain the low-resolution residual error image are: the whole high-resolution human face image that is obtained by step (2)

With low resolution input picture x _l, definition low-resolution residual error image is: $x_l^l=x_l-\frac{1}{S^2}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\underset{j=1}{\overset{S}{\mathrm{\Σ}}}{x}_h^g(\mathrm{Sm}+i,\mathrm{Sn}+j),$ By based on neighbour's reconstruct of the image block on the face location of everywhere from

Set up the high-resolution residual error image

Estimation.

5. the face image super-resolution reconstruction method based on the morphology constituent analysis according to claim 4 is characterized in that, described step (3) is to neighbour's reconstructing method of the image block on the face location of everywhere, and is specific as follows:

Obtain people's face low-resolution image as training sample behind the people's face high-definition picture that (3-1) will choose at random and its down-sampling; Low-resolution image applying step (2) is obtained whole high-resolution human face image, then by the subduction of people's face full resolution pricture, obtain high-resolution residual error image training sample; Obtained low-resolution residual error image training sample by the subduction of people's face low-resolution image behind the whole high-resolution human face image down sampling;

(3-2) with the low-resolution residual error image of inputting

Be divided into image block, make between the image block to overlap;

(3-3) for the low-resolution residual error image block of each input, k arest neighbors low-resolution residual error image block training sample by same face location carries out portfolio restructuring, then the weights that keeping combination synthesize corresponding high-resolution residual error image block by substituting low-resolution residual error image block training sample for corresponding high-resolution residual error image block training sample;

(3-4) obtain complete high-resolution residual error image by the integrated high-resolution residual error image block of averaged overlapping portions

6. based on the face image super-resolution reconstruction method of morphology constituent analysis, it is characterized in that, may further comprise the steps:

(1) study obtains the mutual incoherent expression dictionary of several correspondences to some classes expression by principal component analytical method;

(2) for the low resolution input picture of input, judge the expression under it, then, carry out up-sampling on the one hand, obtain the image of the espressiove after the interpolation; Use on the other hand the morphology component analyzing method, based on its corresponding expression dictionary of expressing one's feelings, the low resolution input picture is carried out picture breakdown, resolve into the amimia input picture of whole low resolution and low resolution espressiove mask;

(3) image of the espressiove after using the morphology component analyzing method to the resulting interpolation of step (2) carries out picture breakdown, based on its corresponding expression dictionary of expressing one's feelings, resolves into the amimia image of whole high resolving power, expression mask and anti-sharpening mask;

(4) the amimia image of the resulting whole high resolving power of step (3) is carried out down-sampling, then and the amimia input picture of whole low resolution that obtains of step (2) subtract each other, obtain the low-resolution residual error image; To this low-resolution residual error image, come image is carried out the facial detail information compensation by the image block on the face location of everywhere being carried out neighbour's reconstruct, obtain the high-resolution residual error image;

(5) the amimia image of whole high resolving power that the high-resolution residual error image in the step (4) and step (3) is obtained merges, and obtains final face image super-resolution result, finishes super-resolution rebuilding.

7. the face image super-resolution reconstruction method based on the morphology constituent analysis according to claim 6 is characterized in that, in the described step (2), judges expression in the low resolution input picture by adopting the LBP sorter.

Images