CN102968775B - Low-resolution face image rebuilding method based on super-resolution rebuilding technology - Google Patents

Abstract

The invention relates to a low-resolution face image rebuilding method based on a super-resolution rebuilding technology and belongs to the field of image processing. The rebuilding method comprises enabling an original image to undergo normalization based on face three-point positioning, and obtaining a to-be-rebuilt normalization face image; then denoising a to-be-rebuilt low-resolution image by generating a training set, determining a learning sample by ambiguity estimation, and obtaining a finally rebuilt high-resolution image by locally linear embedding (LLE) learning. The adopted face image normalization method not only is accurate, but also is consistent with a training set normalization method, and improves accuracy of the low-resolution face image. The super-resolution face image rebuilding method based on low frequency components is adopted to manufacture large-size rebuilt image of the low-resolution face image, and rebuilding of the low-resolution face image is achieved well.

Description

Based on the method for reconstructing of the low-resolution face image of super-resolution rebuilding technology

Technical field

The invention belongs to image processing field, particularly a kind of method making low resolution human face rebuilding picture.

Background technology

Current, video monitoring obtains and develops rapidly, has played more and more important effect in security protection work.In the image of video monitoring, have recorded many figure informations relevant with case.But, because the case-involving facial image of video monitoring is often very little, therefore cannot differentiate case-involving people, cause the work of handling a case of many cases to be absorbed in predicament.The essential problem of this little facial image is the problem that face resolution is low.The reconstruction technique of low-resolution face image is the gordian technique that public security department's work of handling a case is badly in need of.Indication facial image of the present invention, is defined as front face image (lower same).Part (see People's Republic of China (PRC) industry standards of public safety GA/T893-2010 " security protection living things feature recognition application term ") under the crown comprising people, on chin bottom line, between left ear to auris dextra.

The patent No. is a kind of method that the patent of ZL 2,005 1 0067692.X discloses man face geometric size normalization: namely based on 3 method for normalizing of the face in the facial image recognition method of man face geometric size normalization.Determine the images of left and right eyes of face and the coordinate of gnathion in image, and generate the normalization low-resolution face image of the size of standard by 3 method for normalizing.

The method of super-resolution face image conventional is at present: utilize a data set comprising the low-resolution image of paired high-definition picture and correspondence thereof as training storehouse.Image containing low resolution face to be reconstructed is divided into some fritters, to each fritter in low-resolution image to be reconstructed, the linear combination training all low-resolution image correspondence position fritters in storehouse is expressed as by LLE, then use identical linear combination weight to combine the fritter of corresponding high-definition picture correspondence position in training storehouse, obtain the fritter of the super resolution image correspondence position after rebuilding.Spliced by the fritter of the super resolution image correspondence position after the reconstruction obtained, have a small amount of overlapping between block with block, overlapping part is averaged.Finally obtain the large figure of super-resolution after whole reconstruction.Although the method can be rebuild ultra-low resolution image, it directly carries out manifold learning in the time domain, and local geometry similarity is not high, and popular results of learning are poor.In addition be subject to illumination and noise effect comparatively large in its process of reconstruction, registration do not carried out to ultra-low resolution facial image, poor to the super-resolution image reconstruction effect in actual scene.

Summary of the invention

The object of the invention is the weak point for overcoming prior art, proposing a kind of method for reconstructing of the low-resolution face image based on super-resolution rebuilding technology, solve a reconstruction difficult problem for low-resolution face image preferably.

Based on a method for reconstructing for the low-resolution face image of super-resolution rebuilding technology, it is characterized in that, comprise the following steps:

1) normalization based on face three-point fix is carried out to original image, obtains normalization facial image to be reconstructed, specifically comprise:

1.1) cutting is carried out to the original image containing low-resolution face image, cut out the rectangular area comprising low-resolution face image, obtain low-resolution face image; This upper left, region, lower-left, upper right, 4, bottom right summit coordinate in original image are respectively: (m ₁, n ₁), (m ₁, n ₂), (m ₂, n ₁), (m ₂, n ₂), the horizontal ordinate of the point in m, n difference denotation coordination system and ordinate;

1.2) to 1.1) in the low-resolution face image that obtains carry out geometric ratio amplification, enlargement ratio is K × K, and after amplifying, the height of image is L, L=K × (n ₂-n ₁), k meets: K>=150/ (n ₂-n ₁);

1.3) low-resolution face image after amplifying is adopted be normalized based on the method for normalizing of face three-point fix, obtain a standard-sized normalization facial image to be reconstructed;

1.31) low-resolution face image a after amplification determines the coordinate (x of left epibulbar 1 A ₁₁, y ₁₁), the coordinate position (x of right epibulbar 1 B ₂₁, y ₂₁), be straight line L by A, B at 2 ₁, and determine lower jaw point C ₀coordinate (x ₀₁, y ₀₁), by enlargement ratio K, the coordinate of 3 is corresponded to former cutting image coordinate, be respectively (x ₁, y ₁), (x ₂, y ₂), (x ₀, y ₀);

1.32) carry out based on 3 method for normalizing of the face in the facial image recognition method of man face geometric size normalization to original image, obtain the normalization facial image to be reconstructed being of a size of H × W, wherein H, W meet following formula:

$H=\left\{\begin{array}{cc}64,& (n_2-n_1)\≥64\\ 32,& 64>(n_2-n_1)\≥32\\ 16,& 32>(n_2-n_1)\≥16\end{array}\right.,---\left(1\right)$

W=H×3/4，H ₀=5×H/12,H ₁=63×H/120，H ₂=7×H/120；

2) carry out obtaining 512 × 384 high-definition pictures based on the reconstruction of the super-resolution facial image of low frequency component to the normalization facial image to be reconstructed of standard:

2.1) training set is generated:

2.11) image of clear front face image as training set of 100 width will be not less than;

2.12) face three-point fix method for normalizing is adopted the every piece image in training set to be become the standard resolution facial image of 512 × 384;

2.13) 9-7 wave filter down-sampling is used to generate 32 × 24 to each standard-resolution image of 512 × 384,64 × 48,128 × 96,256 × 192,512 × 384 five kinds resolution samples clearly, then use gaussian pyramid wave filter down-sampling to generate the fuzzy resolution sample of 32 × 24 to each standard-resolution image of 512 × 384;

2.14) to 64 × 48,128 × 96,256 × 192,512 × 384 four kinds clearly resolution sample 9-7 bank of filters carry out Laplacian pyramid and extract high fdrequency component and preserve, as high-definition picture wavelet coefficient sample;

2.15) to 32 × 24, 64 × 48, 128 × 96, 256 × 192 4 kinds clearly resolution sample carry out most neighbor interpolation and bicubic linear interpolation, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract their high fdrequency component and preserve, as low-resolution image wavelet coefficient sample clearly, most neighbor interpolation and bicubic linear interpolation are carried out to 32 × 24 fuzzy resolution samples simultaneously, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract its high fdrequency component and preserve, as 32 × 24 fuzzy low-resolution image wavelet coefficient samples,

2.2) denoising carried out to low-resolution image to be reconstructed and estimate to determine learning sample by blur level:

2.21) denoising is carried out to 32 × 24 low-resolution images to be reconstructed after normalization; Step is as follows:

Laplacian pyramid is carried out to image to be reconstructed, calculate the median σ of its all high frequency coefficient absolute value as noise criteria difference estimated value, according to classical denoise algorithm with λ σ for threshold value carries out denoising, λ is experience factor, and span is 0.05 < λ < 0.3;

2.22) carry out fog-level estimation, calculate fog-level coefficient:

$u=\frac{{\mathrm{\&Sigma;}}_h}{{\mathrm{\&Sigma;}}_i},---\left(2\right)$

∑ _hfor the high frequency coefficient absolute value sum of low-resolution image to be reconstructed, ∑ _ifor 32 × 24 low-resolution image pixel value sums to be reconstructed after denoising;

2.23) learning sample is chosen according to fog-level coefficient u:

U < 0.05, in time, learns to adopt 32 × 24 fuzzy samples, adopts 32 × 24 samples clearly during u >=0.05;

2.3) high-definition picture finally reconstructed is learnt by LLE:

2.31) amplify from 32 × 24 to 64 × 48:

2.31.1) using in 2.21) in the pixel value of 32 × 24 low-resolution images after the denoising that obtains directly as the low frequency coefficient of 64 × 48 high-definition pictures reconstructed;

2.31.2) carry out most neighbor interpolation and bicubic linear interpolation to 32 × 24 low-resolution images after denoising, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.31.3) 32 × 24 low-resolution images after denoising are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.31.4) utilize LLE to learn to obtain 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.31.2) in obtain high frequency coefficient composition vector, x _lR(i) be i-th 2.23) vector of the high frequency coefficient composition of relevant position fritter in the sample set determined, 2.23) get front k (5<k<8) and y in the fritter of sample set relevant position determined _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lR(i), altogether k row, 1 is all the vector of 1 for element; W (i) is k coefficient to be asked, and the solution of this extreme-value problem is w is the vector that w (i) forms; And then the high frequency coefficient vector calculating 64 × 48 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 64 × 48 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.31.5) repeat 2.31.4) obtain all 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters corresponding to the fritter that is partitioned into

2.31.6) by 2.31.5) 64 × 48 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 64 × 48 high-definition pictures;

2.31.7) the high and low frequency coefficient of 64 × 48 high-definition pictures obtained is carried out inverse transformation and just obtain 64 × 48 high-definition pictures;

2.32) amplify from 64 × 48 to 128 × 96:

2.32.1) will from 2.31) 64 × 48 low-resolution images that 64 × 48 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 128 × 96 high-definition pictures reconstructed;

2.32.2) carry out most neighbor interpolation and bicubic linear interpolation to 64 × 48 low-resolution images, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.32.3) 64 × 48 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.32.4) utilize LLE to learn to obtain 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.32.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 64 × 48 high frequency coefficient composition of resolution sample set relevant position fritter clearly, get front k (5<k<8) and y in the fritter of 64 × 48 resolution sample set relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 128 × 96 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 128 × 96 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.32.5) repeat 2.32.4) obtain all 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters corresponding to the fritter that is partitioned into

2.32.6) by 2.32.5) 128 × 96 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 128 × 96 high-definition pictures;

2.32.7) the high and low frequency coefficient of 128 × 96 high-definition pictures obtained is carried out inverse transformation and just draw 128 × 96 high-definition pictures;

2.33) amplify from 128 × 96 to 256 × 192:

2.33.1) will from 2.32) 128 × 96 low-resolution images that 128 × 96 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 256 × 192 high-definition pictures reconstructed;

2.33.2) most neighbor interpolation and bicubic linear interpolation are carried out to 128 × 96 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.33.3) 128 × 96 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.33.4) utilize LLE to learn to obtain 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.33.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 128 × 96 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k (5<k<8) and y in the fritter of 128 × 96 resolution sample relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 256 × 192 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 256 × 192 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.33.5) repeat 2.33.4) obtain all 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters corresponding to the fritter that is partitioned into

2.33.6) by 2.33.5) 256 × 192 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 256 × 192 high-definition pictures;

2.33.7) the high and low frequency coefficient of 256 × 192 high-definition pictures obtained is carried out inverse transformation and just draw 256 × 192 high-definition pictures;

2.34) amplify from 256 × 192 to 512 × 384:

2.34.1) will from 2.33) 256 × 192 low-resolution images that 256 × 192 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 512 × 384 high-definition pictures reconstructed;

2.34.2) most neighbor interpolation and bicubic linear interpolation are carried out to 256 × 192 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.34.3) 256 × 192 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.34.4) utilize LLE to learn to obtain 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.34.3) in obtain fritter corresponding in 2.34.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 256 × 192 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k (5<k<8) and y in 256 × 192 clearly resolution sample relevant position fritter _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 512 × 384 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 512 × 384 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.34.5) repeat 2.34.4) obtain all 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters corresponding to the fritter that is partitioned into

2.34.6) by 2.34.5) 512 × 384 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 512 × 384 high-definition pictures;

The high and low frequency coefficient of 512 × 384 high-definition pictures obtained is carried out 512 × 384 high-definition pictures that inverse transformation is just finally reconstructed.

Feature of the present invention and effect:

First the present invention adopts the method for normalizing of low-resolution face image, particularly have employed the method for normalizing of the low-resolution face image based on face three-point fix method for normalizing (see patent of invention: based on the facial image recognition method of man face geometric size normalization, patent of invention number: ZL 2,005 1 0067692.X,), the method for normalizing of this facial image is not only accurate, and consistent with the method for normalizing of training set, improve the accuracy of low-resolution face image.After obtaining normalized low-resolution face image, then adopt the method for reconstructing of the super-resolution facial image based on low frequency component, the large scale producing low-resolution face image rebuilds picture, solves a reconstruction difficult problem for low-resolution face image preferably.

Accompanying drawing explanation

Fig. 1 is low resolution face embodiment image of the present invention.

Fig. 2 is the low-resolution face image of the present embodiment cutting.

Fig. 3 is the enlarged image of the low resolution face of the cutting of the present embodiment.

Fig. 4 is the three-point fix image of the enlarged image of the low resolution face of the present embodiment.

Fig. 5 is the normalized image of the low-resolution face image based on face three-point fix of the present embodiment, and picture size is 32 × 24.

Fig. 6 is the reconstruction picture of the low-resolution face image of the present embodiment.

Fig. 7 is the true clear face image that the low resolution face of the present embodiment is corresponding.

Embodiment

The present invention propose the low-resolution face image based on super-resolution rebuilding technology method for reconstructing by reference to the accompanying drawings and embodiment be described in detail as follows:

The method for reconstructing of a kind of low-resolution face image based on super-resolution rebuilding technology that the present invention proposes, comprises the normalization of low-resolution face image and the super-resolution rebuilding based on low frequency component; Specifically comprise the following steps:

1) normalization based on face three-point fix is carried out to original image, obtain normalization facial image to be reconstructed

1.1) cutting is carried out to the original image containing low-resolution face image, cut out the rectangular image area comprising low-resolution face image, obtain low-resolution face image; This upper left, region, lower-left, upper right, 4, bottom right apex coordinate are respectively: (m ₁, n ₁), (m ₁, n ₂), (m ₂, n ₁), (m ₂, n ₂), the horizontal ordinate of m, n difference denotation coordination system mid point and ordinate;

1.2) to 1.1) low-resolution face image that obtains carries out geometric ratio amplification, and enlargement ratio is K × K, and after amplifying, the height of image is L, L=K × (n ₂-n ₁), k meets: K>=150/ (n ₂-n ₁);

1.3) low-resolution face image after amplifying is adopted be normalized based on the method for normalizing of face three-point fix, obtain a standard-sized facial image to be reconstructed;

1.31) low-resolution face image a after amplification determines the coordinate (x of left epibulbar 1 A ₁₁, y ₁₁), the coordinate position (x of right epibulbar 1 B ₂₁, y ₂₁), be straight line L by A, B at 2 ₁, and determine lower jaw point C ₀coordinate (x ₀₁, y ₀₁), by enlargement ratio k, the coordinate of 3 is corresponded to former cutting image coordinate, be respectively (x ₁, y ₁), (x ₂, y ₂), (x ₀, y ₀);

1.32) carry out, based on 3 method for normalizing of the face in the facial image recognition method of man face geometric size normalization, obtaining the normalization facial image to be reconstructed being of a size of H × W, specifically comprising (belonging to prior art) to original image:

1.32.1) calculated line L ₁with horizontal angle α; Straight line L1 and horizontal angle α is tried to achieve by following formula, wherein (x ₁, y ₁), (x ₂, y ₂) the corresponding right and left eyes spherical coordinates of difference:

the rotation process that the anglec of rotation is α is carried out to this facial image a, obtains facial image b; Rotation expression formula is as follows:

$\left[\begin{array}{c}{x}^{\&prime;}\\ y^{\&prime;}\end{array}\right]=\left[\begin{array}{cc}\cos \mathrm{\&alpha;}& \sin \mathrm{\&alpha;}\\ -\sin \mathrm{\&alpha;}& \cos \mathrm{\&alpha;}\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]$

In formula, x, y are the coordinate of input facial image, and x, y are the coordinate of facial image b;

1.32.2) on facial image b, calculate the coordinate position (x determining left epibulbar 1 C ₃, y ₃), the coordinate position (x of right epibulbar 1 D ₄, y ₄), be straight line L by C, D at 2 ₂, determine the coordinate position (x of the lower jaw point E of postrotational facial image ₅, y ₅);

1.32.3) specify the numerical value of the physical dimension of the facial image of geometric size normalization, wherein width is of a size of W, and height is of a size of H; Any point on regulation jaw rolls off the production line is H to the standard value of the vertical range of two lines ₀, the standard value to the vertical range of image lower frame is H ₁, two standard values being wired to the vertical range of image upper side frame are H ₂; H, W meet following formula:

$H=\left\{\begin{array}{cc}64,& (n_2-n_1)\&GreaterEqual;64\\ 32,& 64>(n_2-n_1)\&GreaterEqual;32\\ 16,& 32>(n_2-n_1)\&GreaterEqual;16\end{array}\right.,---\left(1\right)$

W=H×3/4，H ₀=5×H/12，H ₁=63×H/120，H ₂=7×H/120；

1.32.4) the vertical range h of E point to straight line L2 is obtained _y, and computed image scaling COEFFICIENT K=h _y/ H ₀; Wherein, E point is to the vertical range h of straight line L2 _yfor

$h_y=y_5-\frac{y_3+y_4}{2}$

1.32.5) according to scaling COEFFICIENT K, process is zoomed in or out to facial image b, be met gauged distance H ₀facial image c;

1.32.6) on facial image c, determine the coordinate position (x of left epibulbar 1 M ₆, y ₆), the coordinate position (x of right epibulbar 1 N ₇, y ₇), and the ordinate y of lower jaw point P ₈position; y ₈=MidPoint.y+H ₀;

1.32.7) facial image c is reduced to the normalization facial image d to be reconstructed of the standard of obtaining: cut out and go x coordinate in facial image c to be less than (x ₆+ x ₇)/2-W/2, be greater than (x ₆+ x ₇the part of)/2+W/2, and y coordinate is less than (y ₇-H ₂), be greater than (y ₈+ H ₁) part; If the width of image is less than W or is highly less than H after reducing, then adopt the method for interpolation, width mended W or highly mend H;

2) carry out obtaining 512 × 384 high-definition pictures based on the reconstruction of the super-resolution facial image of low frequency component to the normalization facial image to be reconstructed of standard:

2.1) generate training set:

2.11) image of clear front face image as training set of 100 width will be not less than;

2.12) face three-point fix method for normalizing is adopted the every piece image in training set to be become the standard resolution facial image of 512 × 384;

2.13) 9-7 wave filter down-sampling is used to generate 32 × 24 to each standard-resolution image of 512 × 384,64 × 48,128 × 96,256 × 192,512 × 384 five kinds resolution samples clearly, then use gaussian pyramid wave filter down-sampling to generate the fuzzy resolution sample of 32 × 24 to each standard-resolution image of 512 × 384;

2.14) to 64 × 48,128 × 96,256 × 192,512 × 384 four kinds clearly resolution sample 9-7 bank of filters carry out Laplacian pyramid and extract high fdrequency component and preserve, as high-definition picture wavelet coefficient sample;

2.15) to 32 × 24, 64 × 48, 128 × 96, 256 × 192 4 kinds clearly resolution sample carry out most neighbor interpolation and bicubic linear interpolation, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract their high fdrequency component and preserve, as low-resolution image wavelet coefficient sample clearly, most neighbor interpolation and bicubic linear interpolation are carried out to 32 × 24 fuzzy resolution samples simultaneously, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract its high fdrequency component and preserve, as 32 × 24 fuzzy low-resolution image wavelet coefficient samples

2.2) denoising carried out to low-resolution image to be reconstructed and estimate to determine learning sample by blur level:

2.21) denoising is carried out to 32 × 24 low-resolution images to be reconstructed after normalization; Step is as follows:

Laplacian pyramid is carried out to image to be reconstructed, calculate the median σ of its all high frequency coefficient absolute value as noise criteria difference estimated value, according to classical denoise algorithm with λ σ for threshold value carries out denoising, λ is experience factor, and span is 0.05 < λ < 0.3;

2.22) carry out fog-level estimation, calculate fog-level coefficient:

$u=\frac{{\mathrm{\&Sigma;}}_h}{{\mathrm{\&Sigma;}}_i}---\left(2\right)$

∑ _hfor the high frequency coefficient absolute value sum of low-resolution image to be reconstructed, ∑ _ifor 32 × 24 low-resolution image pixel value sums to be reconstructed after denoising;

2.23) learning sample is chosen according to fog-level coefficient u:

U < 0.05, in time, learns to adopt 32 × 24 fuzzy samples, adopts 32 × 24 samples clearly during u >=0.05;

2.3) high-definition picture of calcium fluoride is obtained by LLE study:

2.31) amplify from 32 × 24 to 64 × 48:

2.31.1) using in 2.21) in the pixel value of 32 × 24 low-resolution images after the denoising that obtains directly as the low frequency coefficient of 64 × 48 high-definition pictures reconstructed;

2.31.2) carry out most neighbor interpolation and bicubic linear interpolation to 32 × 24 low-resolution images after denoising, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.31.3) 32 × 24 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.31.4) utilize LLE to learn to obtain 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.31.2) in obtain high frequency coefficient composition vector, x _lR(i) be i-th 2.23) vector of the high frequency coefficient composition of relevant position fritter in the sample set determined, 2.23) get front k (5<k<8) and y in the fritter of sample set relevant position determined _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lR(i), altogether k row, 1 is all the vector of 1 for element; W (i) is k coefficient to be asked, and the solution of this extreme-value problem is w is the vector that w (i) forms; And then the high frequency coefficient vector calculating 64 × 48 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 64 × 48 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.31.5) repeat 2.31.4) obtain all 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters corresponding to the fritter that is partitioned into

2.31.6) by 2.31.5) 64 × 48 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 64 × 48 high-definition pictures;

2.31.7) the high and low frequency coefficient of 64 × 48 high-definition pictures obtained is carried out inverse transformation and just obtain 64 × 48 high-definition pictures;

2.32) amplify from 64 × 48 to 128 × 96:

2.32.1) will from 2.31) 64 × 48 low-resolution images that 64 × 48 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 128 × 96 high-definition pictures reconstructed;

2.32.2) carry out most neighbor interpolation and bicubic linear interpolation to 64 × 48 low-resolution images, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.32.3) 64 × 48 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.32.4) utilize LLE to learn to obtain 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.3.1.3) in obtain fritter corresponding in 2.32.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 64 × 48 high frequency coefficient composition of resolution sample set relevant position fritter clearly, get front k (5<k<8) and y in the fritter of 64 × 48 resolution sample set relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 128 × 96 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 128 × 96 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.32.5) repeat 2.32.4) obtain all 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters corresponding to the fritter that is partitioned into

2.32.6) by 2.32.5) 128 × 96 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 128 × 96 high-definition pictures;

2.32.7) the high and low frequency coefficient of 128 × 96 high-definition pictures obtained is carried out inverse transformation and just draw 128 × 96 high-definition pictures;

2.33) amplify from 128 × 96 to 256 × 192:

2.33.1) will from 2.32) 128 × 96 low-resolution images that 128 × 96 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 256 × 192 high-definition pictures reconstructed;

2.33.2) most neighbor interpolation and bicubic linear interpolation are carried out to 128 × 96 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.33.3) 128 × 96 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.33.4) utilize LLE to learn to obtain 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.33.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 128 × 96 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k (5<k<8) and y in the fritter of 128 × 96 resolution sample relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 256 × 192 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 256 × 192 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.33.5) repeat 2.33.4) obtain all 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters corresponding to the fritter that is partitioned into

2.33.6) by 2.33.5) 256 × 192 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 256 × 192 high-definition pictures;

2.33.7) the high and low frequency coefficient of 256 × 192 high-definition pictures obtained is carried out inverse transformation and just draw 256 × 192 high-definition pictures;

2.34) amplify from 256 × 192 to 512 × 384:

2.34.1) will from 2.33) 256 × 192 low-resolution images that 256 × 192 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 512 × 384 high-definition pictures reconstructed;

2.34.2) most neighbor interpolation and bicubic linear interpolation are carried out to 256 × 192 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.34.3) 256 × 192 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.34.4) utilize LLE to learn to obtain 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.34.3) in obtain fritter corresponding in 2.34.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 256 × 192 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k (5<k<8) and y in 256 × 192 clearly resolution sample relevant position fritter _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 512 × 384 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 512 × 384 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.34.5) repeat 2.34.4) obtain all 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters corresponding to the fritter that is partitioned into

2.34.6) by 2.34.5) 512 × 384 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 512 × 384 high-definition pictures;

2.34.7) the high and low frequency coefficient of 512 × 384 high-definition pictures obtained is carried out 512 × 384 high-definition pictures that inverse transformation is just finally reconstructed;

Embodiment

1) normalization based on face three-point fix is carried out to original image, obtain normalization facial image to be reconstructed

1.1) cutting is carried out to the original image containing low-resolution face image, cut out the rectangular image area comprising low-resolution face image, obtain low-resolution face image, as shown in Figure 2, this upper left, region, lower-left, upper right, 4, bottom right apex coordinate are respectively: (898,526), (898,556), (924,526), (924,556);

1.2) to 1.1) low-resolution face image that obtains carries out geometric ratio amplification, and enlargement ratio is 6 × 6, and after amplifying, the height of image is L=180, as shown in Figure 3;

1.3) low-resolution face image after amplifying is adopted be normalized based on the normalized method of face three-point fix, obtain a standard-sized facial image to be reconstructed as shown in Figure 5;

1.31) low-resolution face image a after amplification determines the coordinate (63,104) of left epibulbar 1 A, the coordinate position (112,105) of right epibulbar 1 B, be straight line L by A, B at 2 ₁, and determine lower jaw point C ₀coordinate (87,174), as shown in Figure 4, corresponds to former cutting image coordinate by enlargement ratio 6 by the coordinate of 3, is respectively (908,543), (916,543), (912,555);

1.32) carry out, based on 3 method for normalizing of the face in the facial image recognition method of man face geometric size normalization, obtaining the normalization facial image to be reconstructed being of a size of 32 × 24, specifically comprising (belonging to prior art) to original image

1.32.1) calculated line L ₁with horizontal angle α;

Straight line L1 and horizontal angle α is tried to achieve by following formula, wherein (x ₁, y ₁), (x ₂, y ₂) the corresponding right and left eyes spherical coordinates of difference:

$\mathrm{\&alpha;}=\arctan \left(\frac{y_2-y_1}{x_2-x_1}\right)$

1.32.2) the rotation process that the anglec of rotation is α is carried out to this facial image a, obtain facial image b; Rotation expression formula is as follows:

$\left[\begin{array}{c}{x}^{\&prime;}\\ y^{\&prime;}\end{array}\right]=\left[\begin{array}{cc}\cos \mathrm{\&alpha;}& \sin \mathrm{\&alpha;}\\ -\sin \mathrm{\&alpha;}& \cos \mathrm{\&alpha;}\end{array}\right]\left[\begin{array}{c}x\\ y\end{array}\right]$

In formula, x, y are the coordinate of input facial image, and x, y are the coordinate of facial image b;

1.32.3) on facial image b, calculate the coordinate position (x determining left epibulbar 1 C ₃, y ₃), the coordinate position (x of right epibulbar 1 D ₄, y ₄), be straight line L by C, D at 2 ₂, determine the coordinate position (x of the lower jaw point E of postrotational facial image ₅, y ₅);

1.32.4) specify the numerical value of the physical dimension of the facial image of geometric size normalization, wherein width is of a size of 24, and height is of a size of 32; Any point on regulation jaw rolls off the production line is H to the standard value of the vertical range of two lines ₀, the standard value to the vertical range of image lower frame is H ₁, two standard values being wired to the vertical range of image upper side frame are H ₂;

1.32.5) the vertical range h of E point to straight line L2 is obtained _y, and computed image scaling COEFFICIENT K=h _y/ H ₀; Wherein, E point is to the vertical range h of straight line L2 _yfor

$h_y=y_5-\frac{y_3+y_4}{2}$

1.32.6) according to scaling COEFFICIENT K, process is zoomed in or out to facial image b, be met gauged distance H ₀facial image c;

1.32.7) on facial image c, determine the coordinate position (10,17) of left epibulbar 1 M, the coordinate e position (18,17) of right epibulbar 1 N, and the ordinate y of lower jaw point P ₈position; y ₈=29;

1.32.8) facial image c is reduced to the normalization facial image d to be reconstructed of the standard of obtaining: cut out and go x coordinate in facial image c to be less than (x ₆+ x ₇)/2-W/2, be greater than (x ₆+ x ₇the part of)/2+W/2, and y coordinate is less than (y ₇-H ₂), be greater than (y ₈+ H ₁) part; If the width of image is less than W or is highly less than H after reducing, then adopt the method for interpolation, width mended W or highly mend H;

2) carry out obtaining 512 × 384 high-definition pictures based on the reconstruction of the super-resolution facial image of low frequency component to the normalization facial image to be reconstructed of standard:

2.1) generate training set:

2.11) with the clear front face image of the 100 width image as training set;

2.12) face three-point fix method for normalizing is adopted the every piece image in training set to be become the standard resolution facial image of 512 × 384;

2.13) 9-7 wave filter down-sampling is used to generate 32 × 24 to each standard-resolution image of 512 × 384,64 × 48,128 × 96,256 × 192,512 × 384 five kinds resolution samples clearly, then use gaussian pyramid wave filter down-sampling to generate the fuzzy resolution sample of 32 × 24 to each standard-resolution image of 512 × 384;

2.14) to 64 × 48,128 × 96,256 × 192,512 × 384 four kinds clearly resolution sample 9-7 bank of filters carry out Laplacian pyramid and extract high fdrequency component and preserve, as high-definition picture wavelet coefficient sample;

2.15) to 32 × 24, 64 × 48, 128 × 96, 256 × 192 4 kinds clearly resolution sample carry out most neighbor interpolation and bicubic linear interpolation, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract their high fdrequency component and preserve, as low-resolution image wavelet coefficient sample clearly, most neighbor interpolation and bicubic linear interpolation are carried out to 32 × 24 fuzzy resolution samples simultaneously, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract its high fdrequency component and preserve, as 32 × 24 fuzzy low-resolution image wavelet coefficient samples

2.2) denoising carried out to low-resolution image to be reconstructed and estimate to determine learning sample by blur level:

2.21) denoising is carried out to 32 × 24 low-resolution images to be reconstructed after normalization; Step is as follows:

Laplacian pyramid is carried out to image to be reconstructed, the median σ calculating its all high frequency coefficient absolute value as noise criteria difference estimated value, according to classical denoise algorithm with λ σ for threshold value carries out denoising, λ=0.15;

2.22) carry out fog-level estimation, calculate fog-level coefficient:

$u=\frac{{\mathrm{\&Sigma;}}_h}{{\mathrm{\&Sigma;}}_i}=0.03,---\left(1\right)$

∑ _hfor the high frequency coefficient absolute value sum of low-resolution image to be reconstructed, ∑ _ifor 32 × 24 low-resolution image pixel value sums to be reconstructed after denoising;

2.23) learning sample is chosen according to fog-level coefficient u:

Because u=0.03 < 0.05, study adopts 32 × 24 fuzzy samples;

2.3) the final high-definition picture rebuild is obtained by LLE study:

2.31) amplify from 32 × 24 to 64 × 48:

2.31.1) using in 2.2.1) in the pixel value of 32 × 24 low-resolution images after the denoising that obtains directly as the low frequency coefficient of 64 × 48 high-definition pictures reconstructed;

2.31.2) carry out most neighbor interpolation and bicubic linear interpolation to 32 × 24 low-resolution images after denoising, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.31.3) 32 × 24 low-resolution images are divided into the fritter that size is 5 × 10, between adjacent isles, have the overlap of 4 × 9;

2.31.4) utilize LLE to learn to obtain 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.31.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of the high frequency coefficient composition of i-th relevant position fritter in 32 × 24 fuzzy sample sets, get front k=6 and y in the fritter of 32 × 24 fuzzy sample set relevant positions _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lR(i), altogether k row, 1 is all the vector of 1 for element; W (i) is k coefficient to be asked, and the solution of this extreme-value problem is w is the vector that w (i) forms; And then the high frequency coefficient vector calculating 64 × 48 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 64 × 48 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.31.5) repeat 2.31.4) obtain all 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters corresponding to the fritter that is partitioned into

2.31.6) by 2.31.5) 64 × 48 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 64 × 48 high-definition pictures;

2.31.7) the high and low frequency coefficient of 64 × 48 high-definition pictures obtained is carried out inverse transformation and just obtain 64 × 48 high-definition pictures;

2.32) amplify from 64 × 48 to 128 × 96:

2.32.1) will from 2.31) 64 × 48 low-resolution images that 64 × 48 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 128 × 96 high-definition pictures reconstructed;

2.32.2) carry out most neighbor interpolation and bicubic linear interpolation to 64 × 48 low-resolution images, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.32.3) 64 × 48 low-resolution images are divided into the fritter that size is 5 × 10, between adjacent isles, have the overlap of 4 × 9;

2.32.4) utilize LLE to learn to obtain 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.32.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 64 × 48 high frequency coefficient composition of resolution sample set relevant position fritter clearly, get front k=6 and y in the fritter of 64 × 48 resolution sample set relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 128 × 96 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 128 × 96 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.32.5) repeat 2.32.4) obtain all 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters corresponding to the fritter that is partitioned into

2.32.6) by 2.32.5) 128 × 96 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 128 × 96 high-definition pictures;

2.32.7) the high and low frequency coefficient of 128 × 96 high-definition pictures obtained is carried out inverse transformation and just draw 128 × 96 high-definition pictures;

2.33) amplify from 128 × 96 to 256 × 192:

2.33.1) will from 2.32) 128 × 96 low-resolution images that 128 × 96 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 256 × 192 high-definition pictures reconstructed;

2.33.2) most neighbor interpolation and bicubic linear interpolation are carried out to 128 × 96 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.33.3) 128 × 96 low-resolution images are divided into the fritter that size is 5 × 10, between adjacent isles, have the overlap of 4 × 9;

2.33.4) utilize LLE to learn to obtain 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.33.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 128 × 96 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k=6 and y in the fritter of 128 × 96 resolution sample relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 256 × 192 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 256 × 192 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.33.5) repeat 2.33.4) obtain all 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters corresponding to the fritter that is partitioned into

2.33.6) by 2.33.5) 256 × 192 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 256 × 192 high-definition pictures;

2.33.7) the high and low frequency coefficient of 256 × 192 high-definition pictures obtained is carried out inverse transformation and just draw 256 × 192 high-definition pictures;

2.34) amplify from 256 × 192 to 512 × 384:

2.34.1) will from 2.33) 256 × 192 low-resolution images that 256 × 192 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 512 × 384 high-definition pictures reconstructed;

2.34.2) most neighbor interpolation and bicubic linear interpolation are carried out to 256 × 192 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.34.3) 256 × 192 low-resolution images are divided into the fritter that size is 5 × 10, between adjacent isles, have the overlap of 4 × 9;

2.34.4) utilize LLE to learn to obtain 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.34.3) in obtain fritter corresponding in 2.34.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 256 × 192 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k=6 and y in 256 × 192 clearly resolution sample relevant position fritter _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|\mathrm{\&epsiv;}{y}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,$ $\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 512 × 384 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 512 × 384 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.34.5) repeat 2.34.4) obtain all 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters corresponding to the fritter that is partitioned into

2.34.6) by 2.34.5) 512 × 384 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 512 × 384 high-definition pictures;

2.34.7) the high and low frequency coefficient of 512 × 384 high-definition pictures obtained is carried out 512 × 384 high-definition pictures that inverse transformation is just finally reconstructed, as shown in Figure 6;

Just said method, program according to the step of the amplification of the cutting of low-resolution face image, low-resolution face image, three-point fix normalization, reconstruction, storage, realize the reconstruction of the low-resolution face image of super-resolution rebuilding technology, obtain final reconstructed results as shown in Figure 6.True clear face image corresponding to low resolution face as shown in Figure 7.Can see from Fig. 6, Fig. 7, successful of the present invention.

Claims (1)

1. based on a method for reconstructing for the low-resolution face image of super-resolution rebuilding technology, it is characterized in that, comprise the following steps:

1) normalization based on face three-point fix is carried out to original image, obtains normalization facial image to be reconstructed, specifically comprise:

1.1) cutting is carried out to the original image containing low-resolution face image, cut out the rectangular area comprising low-resolution face image, obtain low-resolution face image; This upper left, region, lower-left, upper right, 4, bottom right summit coordinate in original image are respectively: (m ₁, n ₁), (m ₁, n ₂), (m ₂, n ₁), (m ₂, n ₂), the horizontal ordinate of the point in m, n difference denotation coordination system and ordinate;

1.2) to 1.1) in the low-resolution face image that obtains carry out geometric ratio amplification, enlargement ratio is K × K, and after amplifying, the height of image is L, L=K × (n ₂-n ₁), K meets: K>=150/ (n ₂-n ₁);

1.3) low-resolution face image after amplifying is adopted be normalized based on the method for normalizing of face three-point fix, obtain a standard-sized normalization facial image to be reconstructed;

1.31) low-resolution face image a after amplification determines the coordinate (x of left epibulbar 1 A ₁₁, y ₁₁), the coordinate position (x of right epibulbar 1 B ₂₁, y ₂₁), be straight line L by A, B at 2 ₁, and determine lower jaw point C ₀coordinate (x ₀₁, y ₀₁), by enlargement ratio K, the coordinate of 3 is corresponded to former cutting image coordinate, be respectively (x ₁, y ₁), (x ₂, y ₂), (x ₀, y ₀);

1.32) carry out based on 3 method for normalizing of the face in the facial image recognition method of man face geometric size normalization to original image, obtain the normalization facial image to be reconstructed being of a size of H × W, wherein H, W meet following formula:

$H=\left\{\begin{array}{cc}64,& (n_2-n_1)\&GreaterEqual;64\\ 32,& 64>(n_2-n_1)\&GreaterEqual;32\\ 16,& 32>(n_2-n_1)\&GreaterEqual;16\end{array}\right.,---\left(1\right)$

W＝H×3/4；

2) carry out obtaining 512 × 384 high-definition pictures based on the reconstruction of the super-resolution facial image of low frequency component to standard-sized normalization facial image to be reconstructed:

2.1) training set is generated:

2.11) image of clear front face image as training set of 100 width will be not less than;

2.12) face three-point fix method for normalizing is adopted the every piece image in training set to be become the standard resolution facial image of 512 × 384;

2.13) 9-7 wave filter down-sampling is used to generate 32 × 24 to each standard-resolution image of 512 × 384,64 × 48,128 × 96,256 × 192,512 × 384 five kinds resolution samples clearly, then use gaussian pyramid wave filter down-sampling to generate the fuzzy resolution sample of 32 × 24 to each standard-resolution image of 512 × 384;

2.14) to 64 × 48,128 × 96,256 × 192,512 × 384 four kinds clearly resolution sample 9-7 bank of filters carry out Laplacian pyramid and extract high fdrequency component and preserve, as high-definition picture wavelet coefficient sample;

2.15) to 32 × 24, 64 × 48, 128 × 96, 256 × 192 4 kinds clearly resolution sample carry out most neighbor interpolation and bicubic linear interpolation, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract their high fdrequency component and preserve, as low-resolution image wavelet coefficient sample clearly, most neighbor interpolation and bicubic linear interpolation are carried out to 32 × 24 fuzzy resolution samples simultaneously, get both average, carry out Laplacian pyramid by 9-7 bank of filters extract its high fdrequency component and preserve, as 32 × 24 fuzzy low-resolution image wavelet coefficient samples,

2.2) denoising carried out to low-resolution image to be reconstructed and estimate to determine learning sample by blur level:

2.21) denoising is carried out to 32 × 24 low-resolution images to be reconstructed after normalization; Step is as follows:

Laplacian pyramid is carried out to image to be reconstructed, the median σ calculating its all high frequency coefficient absolute value as noise criteria difference estimated value, according to classical denoise algorithm with λ _σfor threshold value carries out denoising, λ is experience factor, and span is 0.05 < λ < 0.3;

2.22) carry out fog-level estimation, calculate fog-level coefficient:

$u=\frac{{\mathrm{\&Sigma;}}_h}{{\mathrm{\&Sigma;}}_i},---\left(2\right)$

Σ _hfor the high frequency coefficient absolute value sum of low-resolution image to be reconstructed, Σ _ifor 32 × 24 low-resolution image pixel value sums to be reconstructed after denoising;

2.23) learning sample is chosen according to fog-level coefficient u:

During u < 0.05, study adopts 32 × 24 fuzzy samples, adopts 32 × 24 samples clearly during u >=0.05;

2.3) high-definition picture finally reconstructed is learnt by LLE:

2.31) amplify from 32 × 24 to 64 × 48:

2.31.1) using in 2.21) in the pixel value of 32 × 24 low-resolution images after the denoising that obtains directly as the low frequency coefficient of 64 × 48 high-definition pictures reconstructed;

2.31.2) carry out most neighbor interpolation and bicubic linear interpolation to 32 × 24 low-resolution images after denoising, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.31.3) 32 × 24 low-resolution images after denoising are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.31.4) utilize LLE to learn to obtain 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.31.2) in obtain high frequency coefficient composition vector, x _lR(i) be i-th 2.23) vector of the high frequency coefficient composition of relevant position fritter in the sample set determined, 2.23) get front k (5<k<8) and y in the fritter of sample set relevant position determined _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|{\mathrm{\&epsiv;y}}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lR(i), altogether k row, 1 is all the vector of 1 for element; W (i) is k coefficient to be asked, and the solution of this extreme-value problem is w is the vector that w (i) forms; And then the high frequency coefficient vector calculating 64 × 48 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 64 × 48 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.31.5) repeat 2.31.4) obtain all 2.31.3) in the high frequency coefficient of 64 × 48 high resolving power fritters corresponding to the fritter that is partitioned into;

2.31.6) by 2.31.5) 64 × 48 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 64 × 48 high-definition pictures;

2.31.7) the high and low frequency coefficient of 64 × 48 high-definition pictures obtained is carried out inverse transformation and just obtain 64 × 48 high-definition pictures;

2.32) amplify from 64 × 48 to 128 × 96:

2.32.1) will from 2.31) 64 × 48 low-resolution images that 64 × 48 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 128 × 96 high-definition pictures reconstructed;

2.32.2) carry out most neighbor interpolation and bicubic linear interpolation to 64 × 48 low-resolution images, get both average, carrying out Laplacian pyramid by 9-7 bank of filters obtains its high frequency coefficient;

2.32.3) 64 × 48 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.32.4) utilize LLE to learn to obtain 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.32.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 64 × 48 high frequency coefficient composition of resolution sample set relevant position fritter clearly, get front k (5<k<8) and y in the fritter of 64 × 48 resolution sample set relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|{\mathrm{\&epsiv;y}}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 128 × 96 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 128 × 96 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.32.5) repeat 2.32.4) obtain all 2.32.3) in the high frequency coefficient of 128 × 96 high resolving power fritters corresponding to the fritter that is partitioned into;

2.32.6) by 2.32.5) 128 × 96 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as a result, obtains the high frequency coefficient of 128 × 96 high-definition pictures;

2.32.7) the high and low frequency coefficient of 128 × 96 high-definition pictures obtained is carried out inverse transformation and just draw 128 × 96 high-definition pictures;

2.33) amplify from 128 × 96 to 256 × 192:

2.33.1) will from 2.32) 128 × 96 low-resolution images that 128 × 96 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 256 × 192 high-definition pictures reconstructed;

2.33.2) most neighbor interpolation and bicubic linear interpolation are carried out to 128 × 96 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.33.3) 128 × 96 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.33.4) utilize LLE to learn to obtain 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.31.3) in obtain fritter corresponding in 2.33.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 128 × 96 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k (5<k<8) and y in the fritter of 128 × 96 resolution sample relevant positions clearly _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|{\mathrm{\&epsiv;y}}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 256 × 192 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 256 × 192 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.33.5) repeat 2.33.4) obtain all 2.33.3) in the high frequency coefficient of 256 × 192 high resolving power fritters corresponding to the fritter that is partitioned into;

2.33.6) by 2.33.5) 256 × 192 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 256 × 192 high-definition pictures;

2.33.7) the high and low frequency coefficient of 256 × 192 high-definition pictures obtained is carried out inverse transformation and just draw 256 × 192 high-definition pictures;

2.34) amplify from 256 × 192 to 512 × 384:

2.34.1) will from 2.33) 256 × 192 low-resolution images that 256 × 192 high-definition pictures that obtain of step walk as this, its pixel value is directly as the low frequency coefficient of 512 × 384 high-definition pictures reconstructed;

2.34.2) most neighbor interpolation and bicubic linear interpolation are carried out to 256 × 192 low-resolution images, get both average, carry out Laplacian pyramid by 9-7 bank of filters and extract and obtain its high frequency coefficient;

2.34.3) 256 × 192 low-resolution images are divided into the fritter that size is s × t, s span is 4 ~ 8, t span is 8 ~ 12, has (s-1) × the overlap of (t-1) between adjacent isles;

2.34.4) utilize LLE to learn to obtain 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters of fritter correspondence position that is partitioned into:

Make y _lRfor single in 2.34.3) in obtain fritter corresponding in 2.34.2) in obtain high frequency coefficient composition vector, x _lRi () is the vector of i-th 256 × 192 high frequency coefficient composition of resolution sample relevant position fritter clearly, get front k (5<k<8) and y in 256 × 192 clearly resolution sample relevant position fritter _lRthe sample fritter that square error is minimum, is calculated as follows extreme-value problem:

$W=\underset{w\left(i\right)}{\arg \min }{\left|\right|{\mathrm{\&epsiv;y}}_{\mathrm{LR}}-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{LR}}\left(i\right)\left|\right|}^2,\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)=1---\left(3\right)$

Make G=(ε y _lR1 ^t-X) ^t(ε y _lR1 ^t-X), wherein X be often classified as x _lRi (), altogether k row, 1 is all the vector of 1 for element, and w (i) is the individual coefficient to be asked of k, and the solution of this extreme-value problem is w is the vector that w (i) forms, and then the high frequency coefficient vector calculating 512 × 384 high-definition picture fritters that study obtains is:

$y_{\mathrm{HR}}=\frac{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}w\left(i\right)x_{\mathrm{HR}}\left(i\right)}{\mathrm{\&epsiv;}}---\left(4\right)$

Wherein x _hRi () is x _lRthe high frequency coefficient vector of i 512 × 384 clear resolution sample fritters that () is corresponding;

If carry out photo-irradiation treatment, in (3), (4) two formulas if do not carry out photo-irradiation treatment, ε=1 in (3), (4) two formulas;

2.34.5) repeat 2.34.4) obtain all 2.34.3) in the high frequency coefficient of 512 × 384 high resolving power fritters corresponding to the fritter that is partitioned into

2.34.6) by 2.34.5) 512 × 384 high resolving power fritters that obtain of learning splice, and between adjacent isles, lap gets all averages participating in overlapping fritter as net result, obtains the high frequency coefficient of 512 × 384 high-definition pictures;

2.34.7) the high and low frequency coefficient of 512 × 384 high-definition pictures obtained is carried out 512 × 384 high-definition pictures that inverse transformation is just finally reconstructed.