CN103208109A - Local restriction iteration neighborhood embedding-based face hallucination method - Google Patents

Abstract

The invention relates to a local restriction iteration neighborhood embedding-based face hallucination method. The method comprises the following steps of: establishing high-resolution and low-resolution image block sets to be used as high-resolution and low-resolution image block dictionaries; sampling on inputted image blocks of a low-resolution face image to obtain estimation high-resolution image blocks, seeking K nearest image blocks at the corresponding position in the high-resolution image block dictionary, and expressing the inputted low-resolution image blocks by using the corresponding K low-resolution image blocks to acquire a weight coefficient; reconstructing K neighbor high-resolution image blocks by utilizing the weight coefficient to form new estimation high-resolution image blocks, and performing the operation repeatedly until the most satisfied estimation high-resolution image blocks are obtained; and integrating into a high-resolution image according to the positional relations of the low-resolution image blocks. According to the method, two manifold structures are considered simultaneously on the basis of position apriority and local manifold restriction, and K neighbor points and reconstruction weights are updated continuously in an iteration form on the basis of a result of last reconstruction to achieve a high-quality reconstruction effect which is close to the real condition.

Description

A kind of unreal structure method of people's face that embeds based on local restriction iteration neighborhood

Technical field

The present invention relates to the image super-resolution field, be specifically related to a kind of unreal structure method of people's face that embeds based on local restriction iteration neighborhood.

Background technology

In in the past 20 years, face recognition technology has obtained development rapidly.Simultaneously because video monitoring system has the network bandwidth limited, restrictions such as server stores cause the face-image resolution that photographs low, make that people's face information that can provide is very limited, and this becomes in the biological identification technology one of challenging problem of tool.Recently, super-resolution technique has been used to handle low resolution (Low-Resolution, LR) image, it can provide more high resolving power (High-Resolution, HR) image of multiaspect portion details for follow-up identification process from a sequence of low resolution pictures or energy of single frames low-resolution image generation.

Baker in 2000 and Kanade are at document 1(S.Baker and T.Kanade.Hallucinating faces.In FG, Grenoble, France, Mar.2000,83-88.) in the method for the unreal structure of a kind of people's face (face hallucination) has been proposed, utilize the prior imformation of facial image in the training set, obtain the high-resolution human face image of low resolution facial image correspondence by the method for study.This is the initiative work of the unreal structure technical field of people's face.After this, many diverse ways and model are introduced into, wherein the most representative technology is that people such as Chang are at document 2(H.Chang, D.Yeung, and Y.Xiong.Super-resolution through neighbor embedding[A] .In Proc.IEEE CVPR ' 04[C] .Washington, 2004.275 – 282.) the middle a kind of method based on manifold learning that proposes, they suppose that the high-definition picture piece has identical local geometry with the low-resolution image piece, and then (locally linear embedding LLE) learns the corresponding relation that high low-resolution image piece flows shape to utilize local linear the embedding.Human face analysis studies show that with the correlation theory of synthesizing, the positional information of facial image human face analysis with synthetic in extremely important, inspired by this, people such as Ma are at document 3(X.Ma, J.P Zhang, and C.Qi.Hallucinating face by position-patch.Pattern Recognition, 43 (6): 3178 – 3194,2010.) and document 4(X.Ma, J.Zhang, and C.Qi, " Position-based face hallucination method, " in Proc.IEEE Conf.on Multimedia and Expo (ICME), 2009, pp.290 – 293.) proposed a kind of unreal structure method of people's face of position-based piece in, it is synthetic that the image block of the given low resolution facial image position image block by all same positions in the training set facial image is carried out linearity.Yet because this method adopts least square method to find the solution, when the number of image in the training sample was bigger than the dimension of image block, the expression coefficient of image block was not unique.For this reason, people such as Jung were at document 5(C.Jung in 2011, L.Jiao, B.Liu, and M.Gong, " Position-Patch Based Face Hallucination Using Convex Optimization; " IEEE Signal Process.Lett., vol.18, no.6, pp.367 – 370,2011.) utilize sparse regularization method to obtain the optimal reconstruction weight of the unreal structure of people's face in.Recently, patent 1(Hu Ruimin, Jiang Junjun, Wang Bing, Han Zhen, Huang Kebin, Lu Tao, Wang Yimin, a kind of human face super-resolution method for reconstructing of representing based on local restriction, number of patent application: the unreal structure method of people's face of further having improved the position-based piece 201110421452.3), image block being carried out retrain the sparse constraint of having replaced in the document 5 with the local geometric that flows shape in the process of reconstruction, make reconstructed results have sparse property and locality simultaneously.Up to the present, the unreal structure method of people's face represented of this local restriction is effect the best way.

No matter be to use rarefaction representation or local restriction, above mentioned method all is by exploring rational priori, seeking the most representative image block and obtain optimal weights and carry out the unreal structure of people's face.Therefore, how to seek rational K neighbour's sample image piece and to obtain optimal weights be two links of most critical in the unreal structure technology of people's face.Above-mentioned all methods have all only been considered low-resolution image piece stream shape, and have ignored the geometry information of high-definition picture piece, make reconstructed results lack reliability and identification.

Summary of the invention

The object of the invention is to provide a kind of unreal structure method of people's face that embeds based on local restriction iteration neighborhood.Utilize whole training sample image piece as dictionary, considered the local geometry feature of low resolution manifold structure and high resolving power manifold structure simultaneously, remedied the deficiency of in the past only considering a certain manifold structure, through iterative step repeatedly, make and rebuild effect further near the original high resolution image simultaneously.

For achieving the above object, the technical solution used in the present invention is a kind of unreal structure method of people's face that embeds based on local restriction iteration neighborhood, comprises the steps:

Step 1, low resolution facial image to input carries out up-sampling and obtains estimating the high-resolution human face image, to the low resolution facial image of input, estimate all low resolution people face sample images in high-resolution human face image, the low resolution training set and all the high-resolution human face sample images in the high resolving power training set are divided overlapped image block respectively;

Step 2 for each image block in the low resolution facial image of input, is carried out following steps and is obtained reconstruct high-definition picture piece,

Step 2.1 is got the image block of each low resolution people face sample image relevant position in the low resolution training set as sample point, sets up low resolution people face sample block space; Get the image block of each high-resolution human face sample image relevant position in the high resolving power training set as sample point, set up high-resolution human face sample block space; Get the image block of estimating high-resolution human face image relevant position, obtain estimating the high-definition picture piece;

Step 2.2, calculate the current K of reconstruct high-definition picture piece on high-resolution human face sample block space of last iteration gained nearest image block, seek K the image block of this K image block in low resolution people face sample block space, adopt during execution in step 2.2 step 2.1 gained to estimate the high-definition picture piece as current reconstruct high-definition picture piece first; And utilize K image block in the low resolution people face sample block space that the low-resolution image piece of importing is carried out linear reconstruction, obtain optimum weights coefficient; Utilize K image block on optimum weights coefficient and the high-resolution human face sample block space, linear reconstruction obtains the reconstruct high-definition picture piece of this iteration; K is default value;

Step 2.3, the reconstruct high-definition picture piece current according to step 2.2 gained returns repeating step 2.2, reaches the value that sets in advance up to iterations;

Step 3 divides other corresponding reconstruct high-definition picture piece to superpose according to the position all images piece in the low resolution facial image of input, divided by the overlapping number of times of each location of pixels, reconstructs the high-resolution human face image then.

And, establish the low resolution facial image X to input ^tDividing gained image block collection is

Estimate high-resolution human face image Y ^t(0) dividing gained image block collection is

All high-resolution human face sample images in the high resolving power training set are divided respectively obtain high-definition picture piece collection Y={y _Ij| 1≤i≤N, 1≤j≤M} divides respectively all low resolution people face sample images in the low resolution training set and to obtain low-resolution image piece collection X={x _Ij| 1≤i≤N, 1≤j≤M}; Wherein, sign i represents the sequence number of low resolution people face sample image in the sequence number of high resolving power training set middle high-resolution people face sample image and the low resolution training set, piece position number on the sign j presentation video;

Be to estimate high-resolution human face image Y ^t(1) image block at position j place,

Be the low resolution facial image X of input ^tThe image block at j place, position; y _IjBe that i opens the image block at j place, picture position in the high resolving power training set, x _IjBe that i opens the image block at j place, picture position in the low resolution training set; The number of the number of low resolution people face sample image and high resolving power training set middle high-resolution people face sample image all is designated as N in the low resolution training set, and M is the piece number of every width of cloth image partitioned image piece;

In the step 2, to arbitrary image block in the low resolution facial image of input

Carry out following substep,

Step 2.1, the initial value that makes iterations p is 1; Image block to the arbitrary position in the low resolution facial image of input

Set up low resolution people face sample block space X _j={ x _Ij| 1≤i≤N} sets up high-resolution human face sample block space Y as low-resolution image piece dictionary _j={ y _Ij| 1≤i≤N} seeks the relevant position and estimates the high-definition picture piece as high-definition picture piece dictionary

Wherein ${\{y}_j^t\left(1\right)=\mathrm{Bicubic}\left(x_j^t\right)|1\≤j\≤M\};$

Step 2.2, the reconstruct high-definition picture piece that execution in step 2.2 obtains when utilizing last iteration calculates the reconstruct high-definition picture piece of this iteration, comprises following substep:

Step 2.2.1 is according to current reconstruct high-definition picture piece

Calculate the high-definition picture piece dictionary Y with the relevant position _j={ y _Ij| the distance of each image block among 1≤i≤N}, and it is as follows to seek K minimum image block of distance,

$\mathrm{dist}\left(p\right)={\left|\right|y_j^t\left(p\right)-y_{\mathrm{ij}}\left|\right|}_2,i=1,\·\·\·,N$

$N_K\left(y_j^t\left(p\right)\right)=\mathrm{support}\left({\mathrm{dist}|}_K\right)$

Wherein, dist (p) ∈ R ^N, dist (p) expression reconstruct high-definition picture piece

With each image block y in the high-definition picture piece dictionary _IjDistance, R ^NExpression N dimension real number space; Dist| _KK value of minimum among the expression dist (p),

It is current high-definition picture piece Index set with K image block of high-definition picture piece dictionary middle distance minimum;

For the first time during execution in step 2.2.1, current reconstruct high-definition picture piece

The high-definition picture piece is estimated in employing

Afterwards during execution in step 2.2.1, current reconstruct high-definition picture piece

The reconstruct high-definition picture piece that execution in step 2.2.3 obtains during for last iteration;

Step 2.2.2 seeks the set of step 2.2.1 gained

Middle K image block y _kCorrespondence image piece x in low-resolution image piece dictionary respectively _k, and to image block

Carry out linear reconstruction, obtain optimum weights coefficient

As shown in the formula,

Wherein,

Return the value about function w (p) when obtaining minimum value of weights coefficient w (p), w _k(p) corresponding x among the expression weights coefficient w (p) _kComponent, τ is the regularization parameter that balance is rebuild the sum of errors local restriction;

Step 2.2.3 rebuilds the reconstruct high-definition picture piece of this iteration by following formula,

$y_j^t(p+1)=\underset{y_k\∈N_K\left(y_j^t\left(p\right)\right)}{\mathrm{\Σ}}{\hat{w}}_k\left(p\right)y_k$

Wherein,

It is the optimum weights coefficient of step 2.2.2 gained

Middle correspondence image piece y _kComponent;

Step 2.3 is judged whether iterations reaches the value that sets in advance, otherwise is made p=p+1, carries out next iteration according to returning repeating step 2.2 at step 2.2.3 gained reconstruct high-definition picture piece in this iteration; Be with in this iteration in the reconstruct high-definition picture piece of step 2.2.3 gained reconstruct high-definition picture piece as final gained, finishing iteration.

A kind of unreal structure method of people's face that embeds based on local restriction iteration neighborhood that the present invention proposes, only considered the method for the manifold structure of low-resolution image before being different from, increased the constraint of high-definition picture manifold structure to reconstructed coefficients, when disclosing high low resolution people's face stream shape space immanent structure similarity, also utilized the essential difference in two kinds of stream shape spaces, made reconstructed results have stronger identification; Method by iteration is constantly upgraded rebuilding weight and neighbour point, selects more can represent neighbour's point and the reconstructed coefficients of target high-definition picture piece, finally obtains the high-definition picture more approaching with truth.This method has been improved traditional neighborhood embedding grammar, makes the expression coefficient of input picture piece more accurate, finally obtains higher-quality high-resolution human face image.

Description of drawings

Fig. 1 is the process flow diagram of the embodiment of the invention.

Embodiment

Technical solution of the present invention can adopt software engineering to realize the automatic flow operation.Below in conjunction with drawings and Examples technical solution of the present invention is further described.Referring to Fig. 1, embodiment of the invention concrete steps are as follows:

Step 1, low resolution facial image to input carries out initialization, be that the Bicubic up-sampling obtains estimating the high-resolution human face image, to the low resolution facial image of input, estimate all low resolution people face sample images in high-resolution human face image, the low resolution training set and all the high-resolution human face sample images in the high resolving power training set are divided overlapped image block in the same way;

Low resolution training set and high resolving power training set provide predefined training sample right, comprise low resolution people face sample image in the low resolution training set, comprise the high-resolution human face sample image in the high resolving power training set.Among the embodiment, all high-resolution human face sample images are the facial image (manual alignment eyes or face position in advance) through the alignment registration, and pixel size is 112 * 100.Each low resolution people face sample image is obtained with 4 * 4 smothing filterings and 4 times of down-samplings by a high-resolution human face sample image in the high resolving power training set in the low resolution training set, and the low-resolution image pixel size is 28 * 25.Accordingly, the low resolution facial image pixel size of input is 28 * 25, also can be in advance and people's face aligned in position registration of high-resolution human face sample image, and estimating the high-resolution human face image pixel size is 112 * 100.Overlapping division belongs to this area common technology, and when specifically implementing, those skilled in the art can specify the overlaid pixel size.The embodiment unification is decided to be 12 * 12 with all high-resolution human face sample images and the high-definition picture block size of estimating the high-resolution human face image, the overlaid pixel value is made as 4, the low-resolution image block size of the low resolution facial image of all low resolution people face sample images and input is 3 * 3, and the overlaid pixel value is 1.

Among the embodiment, to the low resolution facial image X of input ^tDividing the image block collection that constitutes behind the overlapped image block is

M is the image block number.Input low resolution facial image is carried out the Bicubic up-sampling, obtain estimating high-resolution human face image Y ^t(1)=Bicubic (X ^t), it is divided with identical method obtain behind the overlapped image block estimating high-resolution human face image block collection with low-resolution image piece set pair is answered

Obtain high-definition picture piece collection Y={y after people's face sample image in the high-resolution and low-resolution training set divided overlapped image block respectively with identical method _Ij| 1≤i≤N, 1≤j≤M} and low-resolution image piece collection X={x _Ij| 1≤i≤N, 1≤j≤M}, wherein, sign i represents the sequence number of low resolution people face sample image in the sequence number of high resolving power training set middle high-resolution people face sample image and the low resolution training set, sign j represents the piece position number on every image. Be to estimate high-resolution human face image Y ^t(1) image block at position j place,

Be the low resolution facial image X of input ^tThe image block at j place, position, y _Ij, x _IjBe respectively that i opens the image block at j place, picture position in the high-resolution and low-resolution training set, the number of the number of low resolution people face sample image and high resolving power training set middle high-resolution people face sample image all is designated as N in the low resolution training set, and M is the piece number of every width of cloth image partitioned image piece;

Step 2 is for arbitrary image block in the low resolution facial image of input

Carry out following substep and obtain corresponding reconstruct high-definition picture piece:

Step 2.1, get the image block of each low resolution people face sample image relevant position in the low resolution training set as sample point, set up low resolution people face sample block space, get the image block of each high-resolution human face sample image relevant position in the high resolving power training set as sample point, set up high-resolution human face sample block space, get the image block of estimating high-resolution human face image relevant position, the high-definition picture piece that obtains estimating; The initial value that makes iterations p is 1;

Among the embodiment, to certain the position image block in the low resolution facial image of input Set up low resolution people face sample block space X _j={ x _Ij| 1≤i≤N} is as low-resolution image piece dictionary, high-resolution human face sample block space Y _j={ y _Ij| 1≤i≤N} seeks the relevant position and estimates the high-definition picture piece as high-definition picture piece dictionary

Wherein ${\{y}_j^t\left(1\right)=\mathrm{Bicubic}\left(x_j^t\right)|1\≤j\≤M\};$

Step 2.2, the reconstruct high-definition picture piece that execution in step 2.2 obtains when utilizing last iteration calculates this iterative target of its correspondence, i.e. and the reconstruct high-definition picture piece of this iteration comprises following substep:

Step 2.2.1 is according to current reconstruct high-definition picture piece Calculate the high-definition picture piece dictionary Y of itself and relevant position _j={ y _Ij| the distance of each image block among 1≤i≤N}, and seek K image block of its middle distance minimum, i.e. K arest neighbors, method is as follows,

$\mathrm{dist}\left(p\right)={\left|\right|y_j^t\left(p\right)-y_{\mathrm{ij}}\left|\right|}_2,i=1,\·\·\·,N$

$N_K\left(y_j^t\left(p\right)\right)=\mathrm{support}\left({\mathrm{dist}|}_K\right)$

Wherein, dist (p) ∈ R ^N, dist (p) expression reconstruct high-definition picture piece With each image block y in the high-definition picture piece dictionary _IjDistance, R ^NExpression N dimension real number space; Dist| _KK value of minimum among the expression dist (p), || || ₂Represent two norms,

It is current reconstruct high-definition picture piece

With the index set of K image block of high-definition picture piece dictionary middle distance minimum, p is current iterations.K can be by those skilled in the art's default value, and in the present embodiment, K is made as 150.

For the first time during execution in step 2.2.1, current reconstruct high-definition picture piece

Employing is to the image block of input picture

What obtain estimates high-resolution human face image Y ^t(0) the corresponding high-definition picture piece of estimating in

Afterwards during execution in step 2.2.1, current reconstruct high-definition picture piece

The reconstruct high-definition picture piece that execution in step 2.2.3 obtains when being last iteration.

Step 2.2.2 seeks the set of step 2.2.1 gained

Middle K high-resolution image block be the image block of corresponding low resolution in low-resolution image piece dictionary respectively, and utilizes them to image block

Carry out linear reconstruction, obtain the optimum weights coefficient of reconstruct

Method is as follows:

Wherein, Return the value about function w (p) when obtaining minimum value of weights coefficient w (p), i.e. desired optimum weights coefficient

x _kBe one of K high-definition picture piece y _kThe image block of corresponding low resolution, w _k(p) corresponding x among the expression weights coefficient w (p) _kComponent, the inner product operation between two vectors of " ο " expression,

Expression is to two norms || || ₂The result ask square, τ is the regularization parameter that balance is rebuild the sum of errors local restriction, τ suggestion value 1e-5.

Step 2.2.3 is obtaining optimum weights coefficient

After, can rebuild the reconstruct high-definition picture piece of this iteration by following formula, i.e. new current reconstruct high-definition picture piece,

$y_j^t(p+1)=\underset{y_k{\∈N}_K\left(y_j^t\left(p\right)\right)}{\mathrm{\Σ}}{\hat{w}}_k\left(p\right)y_k$

Wherein,

It is the optimum weights coefficient that execution in step 2.2.2 asks in this iteration

Middle corresponding some high-resolution image block y _kComponent.

Step 2.3 is judged whether iterations reaches the value that sets in advance, otherwise is made p=p+1, carries out next iteration according to returning repeating step 2.2 at step 2.2.3 gained reconstruct high-definition picture piece in this iteration; Be with in this iteration in the reconstruct high-definition picture piece of step 2.2.3 gained reconstruct high-definition picture piece as final gained, finishing iteration;

Step 3, the high-definition picture piece that all weightings are reconstructed superposes according to the position, divided by the overlapping number of times of each location of pixels, reconstructs the high-resolution human face image then.

Relate generally to three parameters in the embodiment of the invention, i.e. projection recently count K, regularization parameter τ and iterations.Experiment shows, when K gets 150, can obtain reconstruct effect preferably; When regularization parameter τ was between 1e-6～1e-3, our method can obtain stable performance, and is as far as possible little in order to guarantee reconstruction error simultaneously, parameter τ is decided to be 1e-5 obtains best effect; Along with the increase of iterations, experiment gained PSNR value also in continuous increase, when iterations increased all to a certain degree, PSNR value convergence was stable, the reduction computation complexity advises that iterations is decided to be 6 when obtaining best effects as far as possible.

In order to verify validity of the present invention, adopt extensive Chinese face database (the document 6:W.Gao of CAS-PEAL-R1, B.Cao, S.Shan, X.Chen, et al.The CAS-PEAL Large-Scale Chinese Face Database and Baseline Evaluations[J] .IEEE Trans.SMC (Part A), 2008,38 (1): 149-161) experimentize, select the neutrality expression of all 1040 individualities, the front face image under the normal illumination for use.Take human face region and it be cut into 112 * 100 pixels, more manual demarcate the people on the face five unique points (two centers, nose and two corners of the mouths) and carry out the affined transformation alignment, obtain original high-resolution human face image.The low resolution facial image by 4 times of Bicubic down-samplings of high-resolution human face image after again 4 times of Bicubic up-samplings obtain.Select 1000 at random as training sample, will remain 40 as test pattern.The reconstruct effect that we obtain the present invention and the overall face method (document 7 of Wang, X.Wang and X.Tang, " Hallucinating face by eigentransformation; " IEEE Trans.Systems, Man, and Cybernetics.Part C, vol.35, no.3, pp.425 – 434,2005.) and some methods based on the piece position compare, for example neighborhood embedding grammar (NE, document 2), least square method (LSR, document 3), rarefaction representation method (SR, document 5) and local constraint representation method (LcR, patent 1) etc.

Y-PSNR is adopted in experiment, and (Peak Signal to Noise Ratio PSNR) weighs the quality that contrasts algorithm, and SSIM then is the index of weighing two width of cloth figure similarities, and its value illustrates that more close to 1 the effect of image reconstruction is more good.Above method is handled mean P SNR and the SSIM value that obtains to whole 40 test patterns, overall face, and neighborhood embeds, least square, rarefaction representation, local restriction represent etc. that the mean P SNR value of method and the inventive method is followed successively by 26.53,27.90,28.17,28.27,28.84,29.33; Overall situation face, neighborhood embeds, least square, rarefaction representation, the average SSIM value of method such as local restriction and the inventive method is followed successively by 0.8247,0.8868,0.8975,0.8968,0.9083,0.9140.The inventive method improves 0.49 dB and 0.006 respectively than best algorithm (patent 1) in control methods on PSNR and SSIM value.This shows that the inventive method is compared than other existent method, effect has had significant raising.

Specific embodiment described herein only is that the present invention's spirit is illustrated.Those skilled in the art can make various modifications or replenish or adopt similar mode to substitute described specific embodiment, but can't depart from spirit of the present invention or surmount the defined scope of appended claims.

Claims (2)

1. the unreal structure method of people's face that embeds based on local restriction iteration neighborhood is characterized in that, comprises the steps:

Step 1, low resolution facial image to input carries out up-sampling and obtains estimating the high-resolution human face image, to the low resolution facial image of input, estimate all low resolution people face sample images in high-resolution human face image, the low resolution training set and all the high-resolution human face sample images in the high resolving power training set are divided overlapped image block respectively;

Step 2 for each image block in the low resolution facial image of input, is carried out following steps and is obtained reconstruct high-definition picture piece,

Step 2.1 is got the image block of each low resolution people face sample image relevant position in the low resolution training set as sample point, sets up low resolution people face sample block space; Get the image block of each high-resolution human face sample image relevant position in the high resolving power training set as sample point, set up high-resolution human face sample block space; Get the image block of estimating high-resolution human face image relevant position, obtain estimating the high-definition picture piece;

Step 2.2, calculate the current K of reconstruct high-definition picture piece on high-resolution human face sample block space of last iteration gained nearest image block, seek K the image block of this K image block in low resolution people face sample block space, adopt during execution in step 2.2 step 2.1 gained to estimate the high-definition picture piece as current reconstruct high-definition picture piece first; And utilize K image block in the low resolution people face sample block space that the low-resolution image piece of importing is carried out linear reconstruction, obtain optimum weights coefficient; Utilize K image block on optimum weights coefficient and the high-resolution human face sample block space, linear reconstruction obtains the reconstruct high-definition picture piece of this iteration; K is default value;

Step 2.3, the reconstruct high-definition picture piece current according to step 2.2 gained returns repeating step 2.2, reaches the value that sets in advance up to iterations;

Step 3 divides other corresponding reconstruct high-definition picture piece to superpose according to the position all images piece in the low resolution facial image of input, divided by the overlapping number of times of each location of pixels, reconstructs the high-resolution human face image then.

2. according to the described unreal structure method of people's face that embeds based on local restriction iteration neighborhood of claim 1, it is characterized in that:

If the low resolution facial image X to input ^tDividing gained image block collection is

Estimate high-resolution human face image Y ^t(0) dividing gained image block collection is

All high-resolution human face sample images in the high resolving power training set are divided respectively obtain high-definition picture piece collection Y={y _Ij| 1≤i≤N, 1≤j≤M} divides respectively all low resolution people face sample images in the low resolution training set and to obtain low-resolution image piece collection X={x _Ij| 1≤i≤N, 1≤j≤M}; Wherein, sign i represents the sequence number of low resolution people face sample image in the sequence number of high resolving power training set middle high-resolution people face sample image and the low resolution training set, piece position number on the sign j presentation video;

Be to estimate high-resolution human face image Y ^t(1) image block at position j place,

Be the low resolution facial image X of input ^tThe image block at j place, position; y _IjBe that i opens the image block at j place, picture position in the high resolving power training set, x _IjBe that i opens the image block at j place, picture position in the low resolution training set; The number of the number of low resolution people face sample image and high resolving power training set middle high-resolution people face sample image all is designated as N in the low resolution training set, and M is the piece number of every width of cloth image partitioned image piece;

In the step 2, to arbitrary image block in the low resolution facial image of input

Carry out following substep,

Step 2.1, the initial value that makes iterations p is 1; Image block to the arbitrary position in the low resolution facial image of input

Set up low resolution people face sample block space X _j={ x _Ij| 1≤i≤N} sets up high-resolution human face sample block space Y as low-resolution image piece dictionary _j={ y _Ij| 1≤i≤N} seeks the relevant position and estimates the high-definition picture piece as high-definition picture piece dictionary

Wherein ${\{y}_j^t\left(1\right)=\mathrm{Bicubic}\left(x_j^t\right)|1\≤j\≤M\};$

Step 2.2, the reconstruct high-definition picture piece that execution in step 2.2 obtains when utilizing last iteration calculates the reconstruct high-definition picture piece of this iteration, comprises following substep:

Step 2.2.1 is according to current reconstruct high-definition picture piece Calculate the high-definition picture piece dictionary Y with the relevant position _j={ y _Ij| the distance of each image block among 1≤i≤N}, and it is as follows to seek K minimum image block of distance,

$\mathrm{dist}\left(p\right)={\left|\right|y_j^t\left(p\right)-y_{\mathrm{ij}}\left|\right|}_2,i=1,\·\·\·,N$

$N_K\left(y_j^t\left(p\right)\right)=\mathrm{support}\left({\mathrm{dist}|}_K\right)$

Wherein, dist (p) ∈ R ^N, dist (p) expression reconstruct high-definition picture piece

With each image block y in the high-definition picture piece dictionary _IjDistance, R ^NExpression N dimension real number space; Dist| _KK value of minimum among the expression dist (p), It is current high-definition picture piece

Index set with K image block of high-definition picture piece dictionary middle distance minimum;

For the first time during execution in step 2.2.1, current reconstruct high-definition picture piece

The high-definition picture piece is estimated in employing Afterwards during execution in step 2.2.1, current reconstruct high-definition picture piece

The reconstruct high-definition picture piece that execution in step 2.2.3 obtains during for last iteration;

Step 2.2.2 seeks the set of step 2.2.1 gained

Middle K image block y _kCorrespondence image piece x in low-resolution image piece dictionary respectively _k, and to image block

Carry out linear reconstruction, obtain optimum weights coefficient

As shown in the formula,

Wherein, Return the value about function w (p) when obtaining minimum value of weights coefficient w (p), w _k(p) corresponding x among the expression weights coefficient w (p) _kComponent, τ is the regularization parameter that balance is rebuild the sum of errors local restriction;

Step 2.2.3 rebuilds the reconstruct high-definition picture piece of this iteration by following formula,

$y_j^t(p+1)=\underset{y_k\∈N_K\left(y_j^t\left(p\right)\right)}{\mathrm{\Σ}}{\hat{w}}_k\left(p\right)y_k$

Wherein,

It is the optimum weights coefficient of step 2.2.2 gained

Middle correspondence image piece y _kComponent;

Step 2.3 is judged whether iterations reaches the value that sets in advance, otherwise is made p=p+1, carries out next iteration according to returning repeating step 2.2 at step 2.2.3 gained reconstruct high-definition picture piece in this iteration; Be with in this iteration in the reconstruct high-definition picture piece of step 2.2.3 gained reconstruct high-definition picture piece as final gained, finishing iteration.

Images