CN103136734A - Restraining method on edge Halo effects during process of resetting projections onto convex sets (POCS) super-resolution image - Google Patents

Abstract

The invention discloses a restraining method on edge Halo effects during the process of resetting a projections onto convex sets (POCS) super-resolution image, and belongs to the technical field of image processing techniques. The restraining method on the edge Halo effects during the process of resetting the POCS super-resolution image comprises a step of reading the sequence of a low-resolution image, a step of acquiring a high-resolution initial estimated value of convex sets by applying of a wavelet bicubic interpolation, a step of carrying out motion estimation to achieve low-resolution image registration by applying of a pre-filtering sub pixel iteration method, a step of determining a point spread function with edge-preserving characteristics by comprehensively considering space position information and grey information restraint of high-resolution estimation image pixels, a step of determining a length-variable loose projection parameter value according to relevance between all low-resolution observation frames and reference frames, and a step of resetting super-resolution by applying of a POCS method and combining of the high-resolution initial estimated value, the point spread function with the edge-preserving characteristics and the length-variable loose projection parameter value. Therefore, the image space resolution is improved and the ideal high-resolution image is obtained. By means of the restraining method, the edge Halo effects in the reset image can be reduced to a very large extent.

Description

The inhibition method of edge Halo effect during the POCS super-resolution image reconstruction

Technical field

The invention belongs to image processing field, more specifically, the inhibition method of edge Halo effect when relating to the POCS super-resolution image reconstruction.

Background technology

Be accompanied by the high speed development of information age, high-definition picture has obtained using more and more widely.Yet due to limiting factors such as physical equipment cost and various imaging circumstances, the picture quality that causes people to obtain is lower, and resolution is not high.For this reason, some scholars have proposed the super-resolution rebuilding technology.Its purport is the deficiency that makes up hardware, improves the spatial resolution of image from the angle of software, strengthens the availability of image.Super-resolution technique has wide application.For example: in the military surveillance field, location, identification and pre-alerting ability and scouting precision in order to improve military target just need to improve image resolution ratio with high resolution technique; In astronomical remote sensing field, due to the restriction of satellite monitoring equipment cost, the image that photographs is not too clear, can carry out the high definition reconstruction to taking the image that obtains by the super-resolution rebuilding technology, thereby can reduce risk or cost; Monitor the field in bank, security, traffic safety, when having abnormal conditions to occur, can rebuild monitor video, thereby the recognition capability that improves the suspicion personnel helps case scout and crack; In the medical diagnosis field, due to the too little None-identified of pathology target, can obtain high-definition image by super-resolution technique, can accurately locate the pathology target, for making correct diagnostic result, the doctor provides directive function; At computer vision field, in order to save video compress transmission time and resource, generalized case can be transmitted low-resolution image, can utilize when needed super-resolution technique to carry out high definition reconstruct to it, obtains HD video.

Super-resolution image reconstruction utilizes a frame or multiframe low resolution degraded image (as fuzzy, distortion is by noise pollution) to obtain the technology of a vertical frame dimension image in different resolution exactly.Image super-resolution research starts from the research work of the Harris sixties in 20th century.The research of super resolution ratio reconstruction method mainly is divided into two classes: frequency domain method and spatial domain method.The spatial domain method comprises non-uniform spacing sample interpolation method, iteration back projection method, maximum a posteriori probability (Maximum a Posteriori, MAP) method, convex set projection (Projections Onto Convex Sets, POCS) method, mixing MAP/POCS method, regularization method for reconstructing, filtering method etc.POCS algorithm based on set theory is one of important method that solves the super-resolution image reconstruction problem.

The POCS super resolution ratio reconstruction method is based on sets theory and proposes.At first the method is defined as protruding constraint set with various prioris, and then any point begins to project to convex set from solution space, adopts the method for iteration correction, utilizes the constraint of other set to revise the position of this point, until reconstruct high-definition picture.Because the method is on the interpolation image grid, based on the process that PSF (point spread function) adopts fixing lax projective parameter to revise, the problem such as the image after therefore rebuilding exists jagged edge, edge Halo effect, and resolution is not high.

Summary of the invention

In view of this, the invention provides the method that suppresses edge Halo effect in the POCS super-resolution image reconstruction, the edge fog, edge Halo effect, the not high image quality issues of resolution that exist to solve the POCS method.

1.POCS during super-resolution image reconstruction, the inhibition method of edge Halo effect, is characterized in that, realizes according to the following steps successively in computing machine:

Step (1), computer initialization:

Set: image reading module, high resolving power initial estimate acquisition module, low-resolution image registration module and convex set projection POCS method are rebuild module;

Step (2), Halo effect in edge when suppressing according to the following steps the POCS super-resolution image reconstruction successively, edge Halo effect refers to the edge halo effect:

Step (2.1), described image reading module reads in sequence of low resolution pictures g under Same Scene from image library _i, 1≤i≤n, i are the sequence number of low-resolution image, n is the sum of low-resolution image, n ∈ [6,36], determine iterations L ∈ [6,15] according to n, the resolution of described low-resolution image is M * N, M ∈ [64,256], N ∈ [64,256], between adjacent low-resolution image, lap is arranged, have common coordinate system, with (x, y) denotation coordination point;

Step (2.2), described high resolving power initial estimate acquisition module, comprise the low-resolution reference frame and obtain submodule, wavelet decomposition submodule, bicubic interpolation submodule and wavelet inverse transformation submodule, the size of corresponding high-definition picture is qM * qN, q represents the multiple that resolution improves, q ∈ [2,4], wherein:

The low-resolution reference frame obtains submodule, optional described sequence of low resolution pictures wherein a frame low-resolution image as low-resolution reference frame g ₀, and preserve a low-resolution reference frame;

The wavelet decomposition submodule carries out the conversion of q layer scattering wavelet decomposition to described low-resolution reference frame,

Obtain low frequency subgraph

, horizontal direction the high frequency subgraph

, vertical direction the high frequency subgraph

, diagonal the high frequency subgraph

, the size of all subgraphs is M/q * Nq;

The bicubic interpolation submodule adopts bicubic interpolation method in the Matlab storehouse respectively to three described high frequency subgraphs With

Extrapolate, obtaining resolution is all the high frequency subgraph g of M * N ^H, g ^VAnd g ^D

The wavelet inverse transformation submodule is described low-resolution reference image g ₀Make low frequency subgraph, the described image g of three M * N ^H, g ^V, g ^DMake the high frequency subgraph and carry out wavelet inverse transformation, obtain the high-definition picture initial estimate f of qM * qN ₀, [f ₀]=IDWT[g ₀, g ^H, g ^V, g ^D, q];

Step (2.3), described low-resolution image registration module comprises: pre-service submodule, estimation submodule and motion vector compensation submodule, wherein:

The pre-service submodule utilizes low-pass filter F to carry out filtering, F=1/16 * [1,2,1 to all low-resolution images of the low-resolution reference frame that comprises preservation; 2,4,2; 1,2,1], the low-resolution image after the acquisition rough handling;

The estimation submodule is successively asked for described each low resolution observation frame g by following formula _i(x, y) and low-resolution reference frame g ₀Motion vector Z=[k between (x, y) ₁, k ₂, k ₃, k ₄]:

Z=G ^-1V, wherein:

$V=\left[\begin{array}{c}\mathrm{\&Sigma;}{R}_1(g_0-g_t)\\ \mathrm{\&Sigma;R}(g_0-g_t)\\ \mathrm{\&Sigma;}\frac{\&PartialD;g_t}{\&PartialD;x}(g_0-g_t)\\ \mathrm{\&Sigma;}\frac{\&PartialD;g_t}{\&PartialD;y}(g_0-g_t)\end{array}\right],$ $R=y\frac{{\&PartialD;g}_t}{\&PartialD;x}-x\frac{{\&PartialD;g}_t}{\&PartialD;y},$ ${\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1=x\frac{\&PartialD;g_t}{\&PartialD;x}+y\frac{{\&PartialD;g}_t}{\&PartialD;y},$

$G=\left[\begin{array}{cccc}\mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1^2& \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="RR"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">RR</figure-callout>}}_1& \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;y}\\ \mathrm{\&Sigma;}{\mathrm{RR}}_1& \mathrm{\&Sigma;}{R}^2& \mathrm{\&Sigma;R}\frac{{\&PartialD;g}_t}{\&PartialD;x}& \mathrm{\&Sigma;R}\frac{\&PartialD;g_t}{\&PartialD;y}\\ \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;R}\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;}{\left(\frac{\&PartialD;g_t}{\&PartialD;x}\right)}^2& \mathrm{\&Sigma;}(\frac{{\&PartialD;g}_t}{\&PartialD;x}\&CenterDot;\frac{{\&PartialD;g}_t}{\&PartialD;y})\\ \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;y}& \mathrm{\&Sigma;R}\frac{{\&PartialD;g}_t}{\&PartialD;y}& \mathrm{\&Sigma;}(\frac{{\&PartialD;g}_t}{\&PartialD;x}\&CenterDot;\frac{\&PartialD;g_t}{\&PartialD;y})& \mathrm{\&Sigma;}{\left(\frac{\&PartialD;g_t}{\&PartialD;y}\right)}^2\end{array}\right]$

By taking progressively approaching to reality value of alternative manner,

S is iterations, when iteration error satisfies certain precision Th≤10 ^-3The time termination of iterations;

The motion vector compensation submodule, the coordinate of setting before the high-definition picture compensation be (x', y '), the coordinate after the high-definition picture compensation is (x, y), by following formula the described motion vector Z=[k that calculates ₁, k ₂, k ₃, k ₄] compensate to high resolving power initial estimation f ₀In (x, y), described f ₀F ₀The simple expression formula of (x, y):

$\left\{\begin{array}{c}x=x^{\&prime;}+k_1{x}^{\&prime;}+k_2{y}^{\&prime;}+k_3\\ y=y^{\&prime;}+k_1{y}^{\&prime;}-k_2{x}^{\&prime;}+k_4\end{array}\right.$

Step (2.4), described convex set projection POCS method is rebuild module, comprise: point spread function PSF obtains submodule, lax projective parameter self-adaptation is adjusted submodule and iterative projection correction submodule, reduces according to the following steps successively Halo effect in edge in reconstructed results:

Point spread function PSF obtains submodule, obtains according to the following steps successively the point spread function PSF with edge retention performance, it is characterized in that:

Be calculated as follows described high resolving power initial estimation codomain f ₀Euclidean distance in spatial domain in (x, y) between each pixel and supporting domain center pixel represents with h (x, y):

$h(x,y)=\exp \{-\frac{{(x-m)}^2+{(y-n)}^2}{2{\mathrm{\&delta;}}_s^2}\},$ Wherein,

δ _sBe the standard deviation of point spread function PSF, δ _s∈ [0.5,2], (m, n) are the coordinate of center pixel in 5 * 5 pixel supporting domains between each neighbor,

Be calculated as follows described high resolving power initial estimation codomain f ₀Gray-scale relation in (x, y) between each pixel

(| f ₀(x, y)-f ₀(m, n) |):

$w_{{\mathrm{\&delta;}}_r}\left(\right|f_0(x,y)-f_0(m,n)\left|\right)=\exp \{-\frac{{|f_0(x,y)-f_0(m,n)|}^2}{2{\mathrm{\&delta;}}_r^2}\},$ Wherein,

f ₀(x, y), f ₀(m, n) is respectively coordinate (x, y), the gray-scale value that (m, n) locates, δ _rBe standard deviation, δ _r∈ [0.8,15];

Obtain by following formula and considered the gray-scale relation between each pixel in described high resolving power initial estimation codomain

The point spread function with edge retention performance

$\tilde{h}(x,y)=U\&CenterDot;h(x,y)\&CenterDot;w_{{\mathrm{\&delta;}}_r}\left(\right|f_0(x,y)-f_0(m,n)\left|\right),$

U is normalization coefficient, U ∈ (0.001,1);

Lax projective parameter self-adaptation is adjusted submodule, and projective parameter lax according to the correlativity self-adaptation adjustment between low resolution observation frame and low-resolution reference frame according to the following steps successively obtains the lax projective parameter of variable length:

Be calculated as follows the correlativity Rl between each low resolution observation frame and low-resolution reference frame:

$\mathrm{Rl}=\frac{\underset{x,y}{\mathrm{\&Sigma;}}\left[g_i(x,y)g_0(x,y)\right]}{\sqrt{\underset{x,y}{\mathrm{\&Sigma;}}{\left[g_i(x,y)\right]}^2\underset{x,y}{\mathrm{\&Sigma;}}{\left[g_0(x,y)\right]}^2}},$ 0＜Rl＜1，

g _i(x, y) is the current observation frame of each low resolution, g ₀(x, y) is the low-resolution reference frame,

Set: lax projective parameter λ, λ=2Rl, described λ are the lax projective parameters of a variable length, have reflected that the current observation frame of low resolution participates in the degree of projection, this λ value is directly proportional to Rl;

Iterative projection correction submodule is successively according to the following steps to described high resolving power initial estimation codomain f ₀(x, y) carries out the correction of convex set iterative projection, in order to obtain high-definition picture by the iterations of setting:

Known: prior imformation and a priori bound value δ thereof _i, described prior imformation comprises positive definite, finite energy, the smooth any information of data consistency, according to data consistency information definition convex set C _{M, n, i}, (m, n) is low-resolution image g _iIn (x, y), coordinate points (x, y) is corresponding to the high resolving power estimated value

In coordinate, i is the sequence number of low-resolution image observation frame,

Set the estimation codomain in full resolution pricture iterative projection process

L ∈ [0, L), L is the total degree of iteration, L ∈ [6,15],

The estimation codomain in any one full resolution pricture iterative projection process

To convex set C _{M, n, i}On projection

For:

$P_{(x,y)}\left[{\hat{f}}_l(m,n)\right]={\hat{f}}_l(m,n)+\left\{\begin{array}{cc}(\mathrm{\&lambda;}\&CenterDot;r_i^{{\hat{f}}_l(m,n)}(x,y)-{\mathrm{\&delta;}}_0)\&CenterDot;{\tilde{h}}_i(x,y;m,n),& r_i^{{\hat{f}}_l(m,n)}(x,y)>{\mathrm{\&delta;}}_{\mathrm{<figure-callout\; id="0"\; label="i"\; filenames="HDA00002865662200081.png,HDA00002865662200091.png"\; state="\{\{state\}\}">i</figure-callout>}}\\ 0,& \left|r_i^{{\hat{f}}_l(m,n)}(x,y)\right|\&le;{\mathrm{\&delta;}}_i\\ (\mathrm{\&lambda;}\&CenterDot;r_i^{{\hat{f}}_l(m,n)}(x,y)+{\mathrm{\&delta;}}_0)\&CenterDot;{\tilde{h}}_i(x,y;m,n),& r_i^{{\hat{f}}_l(m,n)}(x,y)<-{\mathrm{\&delta;}}_i\end{array}\right.,$

Wherein, δ _i=c δ ₀, c ∈ [2,5], δ ₀∈ [0.1,3.8],

$C_{m,n,i}=\{{\hat{f}}_l(m,n):|r_i^{{\hat{f}}_l(m,n)}(x,y)|\&le;{\mathrm{\&delta;}}_i\},$

Be residual error function, $r_i^{{\hat{f}}_l(m,n)}(x,y)=g_i(x,y)-\underset{m,n}{\mathrm{\&Sigma;}}{\hat{f}}_l(m,n){\tilde{h}}_i(x,y;m,n),$

Be the high resolving power estimated value

Mid point (m, n) is located 5 * 5 supporting domains of gray-scale value and point (m, n) In restriction relation between each grey scale pixel value, expression formula is:

${\tilde{h}}_i(x,y;m,n)=U\&CenterDot;h(x,y;m,n)\&CenterDot;w_{{\mathrm{\&delta;}}_r}\left(\right|{\hat{f}}_l(x,y)-{\hat{f}}_l(m,n)\left|\right),$ U∈(0.001,1)

$h(x,y;m,n)=\exp \{-\frac{{(x-m)}^2+{(y-n)}^2}{2{\mathrm{\&delta;}}_s^2}\},$ δ _s∈[0.5]，

$w_{{\mathrm{\&delta;}}_r}\left(\right|{\hat{f}}_l(x,y)-{\hat{f}}_l(m,n)\left|\right)=\exp \{-\frac{{|{\hat{f}}_l(x,y)-{\hat{f}}_l(m,n)|}^2}{2{\mathrm{\&delta;}}_r^2}\},$ δ _r∈[0.8,15]，

Be in fact the point spread function with edge retention performance, obtain low-resolution image by the high-definition picture estimated value is carried out a DIFFUSION TREATMENT;

Order

By

Calculate

The like, Successively carry out iteration; If do not reach iterations,

The substitution residual error function is upgraded residual error

Otherwise iteration finishes, from

In block and exceed 8 bit image span [0,255] parts, the high-definition picture that obtains rebuilding.

The present invention compared with prior art has the following advantages:

1, the method takes full advantage of the characteristic that small echo has the maintenance details, adopts the Wavelet Bi-cubic Interpolation method to obtain the high resolving power initial estimate of convex set projection POCS, improves the soft edge problem that bilinear interpolation causes.

2, the method adopts pre-filtering sub-pix process of iteration to carry out estimation and realizes the low-resolution image registration, improves image registration accuracy.

3, the method organically combines the gray-scale relation between pixel in the Euclidean distance in spatial domain between pixel and codomain, utilize Gauss's codomain function to obtain to have the point spread function of edge retention performance, reduced to a great extent the edge Halo effect in the reconstructed results.

4, the method utilizes correlativity between low resolution observation frame and low-resolution reference frame to obtain the lax projective parameter of variable length, reduces motion estimation error to rebuilding the impact of picture strip.

Description of drawings

In order to be illustrated more clearly in embodiments of the invention or technical scheme of the prior art, the below does simple introduction with the accompanying drawing that needs in embodiments of the invention or description of the Prior Art to use.Obviously, the accompanying drawing in the following describes is only some embodiment in the present invention, for those skilled in the art, under the prerequisite of not paying creative work, can also obtain other accompanying drawings according to these accompanying drawings.

Fig. 1 is the structural representation of method disclosed by the invention.

Fig. 2 is the structural representation of high resolving power initial estimate acquisition module B disclosed by the invention.

Fig. 3 is the structural representation of low-resolution image registration module C disclosed by the invention.

Fig. 4 is the structural representation that POCS method disclosed by the invention is rebuild module D.

Fig. 5 is the disclosed method flow diagram of the embodiment of the present invention one.

Fig. 6 is the disclosed process flow diagram that obtains the high resolving power initial estimate of the embodiment of the present invention one.

Fig. 7 is the process flow diagram of the disclosed low-resolution image registration of the embodiment of the present invention one.

Fig. 8 is the process flow diagram that the disclosed POCS method of the embodiment of the present invention one is rebuild.

Fig. 9 is the basic principle schematic of the disclosed POCS method of the embodiment of the present invention one.

Figure 10 is the disclosed process flow diagram that obtains the PSF with edge retention performance of the embodiment of the present invention one.

Figure 11 is the process flow diagram that the disclosed lax projective parameter self-adaptation of the embodiment of the present invention one is adjusted.

Figure 12 is that the disclosed POCS super-resolution method of the embodiment of the present invention two is rebuild detail flowchart.

Figure 13 be the present invention under four kinds of different images sequences, adopt the POCS method of fixing lax projective parameter to rebuild high-definition picture PSNR(Y-PSNR) curve map: The PSNR curve map of Lena image is rebuild in representative;

The PSNR curve map of Cameraman image is rebuild in representative;

The PSNR curve map of Boats image is rebuild in representative; The PSNR curve map of House image is rebuild in representative.

Figure 14 be the present invention under four kinds of different images sequences, adopt the POCS method of the lax projective parameter of variable length to rebuild the PSNR(Y-PSNR of high-definition picture) curve map.

Figure 15 is the gray level center drawing in side sectional elevation that adopts respectively Lena image that existing POCS method and method disclosed by the invention rebuild.(a) for adopting existing POCS method to rebuild the gray level center drawing in side sectional elevation of Lena image, in figure, the part spike is marked by the dotted line circle.(b) for adopting the gray level center drawing in side sectional elevation of Lena image that method disclosed by the invention rebuilds.

Embodiment

The inhibition method of edge Halo effect during the POCS super-resolution image reconstruction comprises:

The image reading module is used for reading the pending low-resolution image of n frame that Same Scene has lap, n ∈ [6,36];

High resolving power initial estimate acquisition module, being used for that the low-resolution reference frame is carried out discrete wavelet decomposes, high frequency subgraph on the level that recycling bicubic interpolation method obtains decomposition, vertical, 3 directions of diagonal line carries out the high frequency extrapolation, adopts at last wavelet inverse transformation to obtain the high resolving power initial estimate.

The low-resolution image registration module be used for to adopt pre-filtering sub-pix process of iteration to carry out estimation and obtains motion vector, and the compensation motion vector that estimation is obtained is realized image registration in the high resolving power initial estimate.

The POCS method is rebuild module, is used for the continuous iterative projection of high resolving power estimated value to a plurality of prior imformation constrain sets, obtains the high-definition picture that meets the demands, realizes rebuilding a panel height image in different resolution from sequence of low resolution pictures.

Preferably, described high resolving power initial estimate acquisition module comprises:

The low-resolution reference frame obtains submodule, be used for optional sequence of low resolution pictures wherein a frame low-resolution image as the low-resolution reference frame, and preserve a;

The wavelet decomposition submodule, be used for the low-resolution reference picture breakdown be 1 low frequency subgraph with level, vertical, 3 directions of diagonal line on the high frequency subgraph, level under corresponding resolution, vertically, the high frequency subgraph of three directions of diagonal line reflects respectively the local edge of these three direction epigraphs;

The bicubic interpolation submodule is used for the high frequency subgraph on level, vertical, 3 directions of diagonal line is carried out the high frequency extrapolation, obtains the big or small high frequency subgraph that equals the low-resolution reference frame on 3 directions;

The wavelet inverse transformation submodule is used for the low-resolution reference frame is made low frequency subgraph, and the image on 3 directions of bicubic interpolation acquisition is carried out wavelet inverse transformation as the high frequency subgraph, obtains the initial estimate of high-definition picture.

Preferably, described low-resolution image registration module comprises:

The pre-service submodule utilizes low-pass filter F to carry out filtering to low-resolution image, F=1/16 * [1,2,1; 2,4,2; 1,2,1], the low-resolution image after the acquisition rough handling;

The estimation submodule is used for adopting the sub-pix process of iteration to carry out registration, obtains the motion vector between low-resolution image;

The compensation motion vector submodule, the compensation motion vector that is used for calculating is estimated to high resolving power, obtains the high resolving power initial estimate of registration.

Preferably, described POCS method reconstruction module comprises:

Point spread function obtains submodule, is used for adopting Gaussian function to obtain to have the point spread function of edge retention performance, it is characterized in that: at first calculate the Euclidean distance in the spatial domain between high resolving power estimated image pixel; Then investigate interior each pixel of 5 * 5 neighborhoods in the similarity on codomain and between center pixel according to Gaussian codomain function, obtain the relation between grey scale pixel value in codomain; At last, the gray-scale relation between pixel in the Euclidean distance in spatial domain between pixel and codomain is organically combined, obtain to have the point spread function of edge retention performance.

Lax projective parameter self-adaptation is adjusted submodule, is used for obtaining the lax projective parameter of variable length.It is characterized in that: calculate the correlativity Rl between low resolution observation frame and low-resolution reference frame; According to 0＜Rl＜1, the projective parameter λ that relaxes is set to λ=2Rl; Projective parameter lax according to the adjustment of correlativity self-adaptation, the lax projective parameter of acquisition variable length.

Iterative projection correction submodule is used for the high resolving power initial estimate is constantly projected to a plurality of convex sets, by continuous iterative projection correction, obtains the full resolution pricture of rebuilding.

Below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is known intactly described.Obviously, described embodiment is only the present invention's part embodiment, rather than whole embodiment.Based on embodiments of the invention, other embodiment that those of ordinary skills obtain under the prerequisite of not making creative work belong to protection scope of the present invention.

There are the problems such as reconstructed image edge fog, edge Halo effect and ringing in the inhibition method of edge Halo effect when the invention discloses the POCS super-resolution image reconstruction with the algorithm that solves prior art.Its structure is as shown in Figure 1: image reading modules A, high resolving power initial estimate acquisition module B, low-resolution image registration module C, POCS method are rebuild module D.Wherein:

The image reading modules A is used for reading in the pending low-resolution image of n frame that has lap under Same Scene, n ∈ [6,36]; High resolving power initial estimate acquisition module B, being used for that the low-resolution reference frame is carried out discrete wavelet decomposes, high frequency subgraph on the level that recycling bicubic interpolation method obtains decomposition, vertical, 3 directions of diagonal line carries out the high frequency extrapolation, adopts at last wavelet inverse transformation to obtain the high resolving power initial estimate.Low-resolution image registration module C be used for to adopt pre-filtering sub-pix process of iteration to carry out estimation and obtains motion vector, and with the compensation motion vector estimating to obtain in the high resolving power initial estimate, realize image registration.The POCS method is rebuild module D, is used for the continuous iterative projection of high resolving power estimated value to a plurality of prior imformation constrain sets, obtains the high-definition picture that meets the demands, realizes from low-resolution sequence image reconstruction one panel height image in different resolution.Its detailed content can be referring to following corresponding embodiment.

Described high resolving power initial estimate acquisition module B structure as shown in Figure 2, comprising: the low-resolution reference frame obtains submodule B1, is used for a wherein frame low-resolution image of optional sequence of low resolution pictures as the low-resolution reference frame; Wavelet decomposition submodule B2 is used for the low-resolution reference frame is decomposed into high frequency subgraph on 1 low frequency subgraph and level, vertical, 3 directions of diagonal line; Bicubic interpolation module B3 is used for the image on 3 directions is carried out the high frequency extrapolation, obtains the high frequency subgraph that size on 3 directions equals the low-resolution reference frame; Wavelet inverse transformation module B4 is used for the low-resolution reference frame is made low frequency subgraph, the subgraph on 3 directions of bicubic interpolation acquisition is made the high frequency subgraph carry out wavelet inverse transformation, obtains the initial estimate of high-definition picture.

Described low-resolution image registration module C-structure comprises as shown in Figure 3: pre-service submodule C1, be used for sequence of low resolution pictures subject to registration is carried out the low-pass filtering pre-service, and obtain the image of preliminary noise reduction; Estimation submodule C2 is used for adopting the sub-pix process of iteration to carry out registration, obtains the motion vector between low-resolution image; Compensation motion vector submodule C3, the compensation motion vector that is used for calculating obtain the high resolving power initial estimate of registration to the high resolving power initial estimate.

Described POCS method is rebuild module D structure as shown in Figure 4, and comprising: point spread function obtains submodule D1, is used for adopting Gauss type function to obtain to have the point spread function of edge retention performance; Lax projective parameter self-adaptation is adjusted submodule D2, is used for obtaining the lax projective parameter of variable length; Iterative projection submodule D3 is used for the high resolving power estimated value is carried out the correction of convex set iterative projection, the full resolution pricture after obtaining to rebuild.Its detailed content can be referring to the following examples.

Its embodiment is as described below:

Embodiment one

Referring to Fig. 5, during the disclosed POCS super-resolution image reconstruction of the present embodiment, the flow process of the inhibition method of edge Halo effect comprises the following steps:

Step S1 reads the sequence of low resolution pictures g that has lap under Same Scene _i, i is the sequence number of low-resolution image observation frame, i ∈ [1, n], n are the low-resolution image sum, and n ∈ [6,36], resolution is M * N, M ∈ [64,256], N ∈ [64,256].

Step S2, an optional frame is as low-resolution reference frame g from n frame low-resolution image ₀, to g ₀Carry out Wavelet Bi-cubic Interpolation, obtain the high resolving power initial estimate f of qM * qN ₀(x, y), q represent the multiple that resolution improves, and q ∈ [2,4] is with (x) y denotation coordination point.

The POCS algorithm is exactly that the high resolving power estimated value that iteration each time produces is projected in next protruding constrain set, so the quality of high resolving power initial estimation image directly has influence on the quality of the high-definition picture of final reconstruction.The POCS algorithm adopts bilinear interpolation to obtain initial pictures usually, but bilinear interpolation can cause edge fog in the process that image amplifies.The characteristic that has the keep the edge information details due to Wavelet Bi-cubic Interpolation is therefore adopt the Wavelet Bi-cubic Interpolation method to obtain the high resolving power initial estimate in the present embodiment.

Step S3 carries out estimation and image registration to the n frame low-resolution image that reads in, and obtains motion vector Z=[k ₁, k ₂, k ₃, k ₄], and compensate in high resolving power initial estimation image, thereby realize image registration.

Image registration is that the same target to the current observation frame of low resolution and low-resolution reference frame aligns, and determines the matching relationship of corresponding point, to eliminate or to reduce the distortion that between image, target location difference and noise cause.Need the motion vector between each low resolution observation frame of accurate Calculation and reference frame before employing POCS method is rebuild, the accuracy of estimation of motion vectors has a strong impact on the quality of reconstructed image.At first the present invention utilizes low-pass filter F to carry out filtering to low-resolution image, F=1/16 * [1,2,1; 2,4,2; 1,2,1], then the low-resolution image after the acquisition rough handling uses the sub-pix process of iteration to carry out registration.

Step S4 utilizes the POCS method for reconstructing, and the continuous iterative projection of high resolving power estimated value in a plurality of prior imformation constrain sets, is realized rebuilding a panel height image in different resolution from sequence of low resolution pictures.

The present embodiment utilizes Gaussian function to obtain to have the point spread function of edge retention performance, reduces the edge Halo effect in reconstructed results; Utilize the correlativity between low-resolution reference frame and observation frame to obtain the lax projective parameter of variable length, thereby reduce the impact of motion estimation error; The high resolving power initial estimate that the Wavelet Bi-cubic Interpolation method is obtained carries out the correction of convex set iterative projection, thereby improves image resolution ratio, obtains the full resolution pricture of rebuilding.

In the present embodiment, referring to Fig. 6, obtaining specifically of the described high resolving power initial estimate of step S2 realized with step S21-S24, comprising:

Step S21, an optional wherein frame low-resolution image is as low-resolution reference frame g from sequence of low resolution pictures ₀, and preserve a low-resolution reference frame;

Step S22 carries out the conversion of q layer scattering wavelet decomposition to the low-resolution reference frame, obtains the high frequency subgraph of low frequency subgraph and level, vertical, 3 directions of diagonal line, and these subgraph resolution are M/q * N/q;

Step S23 utilizes in the Matlab storehouse bicubic interpolation respectively step S22 to be decomposed 3 high frequency subgraphs that obtain and carries out the high frequency extrapolation, and obtaining resolution is the high frequency subgraph of M * N;

Step S24 makes low frequency subgraph with the low-resolution reference frame, and the image that step S23 obtains is made the high frequency subgraph, carries out wavelet inverse transformation, obtains high-definition picture initial estimate f ₀(x, y).

The implementation process of step S21-S24 is elaborated as follows:

From the optional wherein frame low-resolution image of sequence of low resolution pictures as low-resolution reference frame g ₀, n ∈ [6,36], and preserve a;

It is carried out q layer scattering wavelet decomposition,

Obtain a low frequency subgraph

High frequency subgraph with horizontal direction

, vertical direction the high frequency subgraph

And the high frequency subgraph of diagonal , the resolution of all subgraphs is M/q * N/q;

Adopt bicubic interpolation pair

Extrapolate, obtain the high frequency subgraph g that resolution is M * N ^H, g ^V, g ^D

G ₀Make low frequency subgraph, g ^H, g ^V, g ^DMake the high frequency subgraph and carry out q layer wavelet inverse transformation, obtain high-definition picture initial estimate f ₀, [f ₀]=IDWT[g ₀, g ^H, g ^V, g ^D, q], f ₀F ₀The simple expression of (x, y).

In the present embodiment, referring to Fig. 7, the described low-resolution image registration of step S3 is specifically with step S -S33 is realized, comprising:

Step S31 utilizes low-pass filter F to carry out filtering, F=1/16[1,2,1 to all low-resolution images of the low-resolution reference frame that comprises preservation; 2,4,2; 1,2,1], obtain the low-resolution image of rough handling;

Step S32 adopts the sub-pix process of iteration to carry out registration, calculates the motion vector between each low resolution observation frame and low-resolution reference frame;

Step 33 is used for calculating compensation motion vector to the high resolving power initial estimate, realizes image registration, obtains the high resolving power initial estimate of registration.

The implementation process of step S31-S33 is elaborated as follows:

Utilize low-pass filter F to carry out filtering to low-resolution image, F=1/16 * [1,2,1; 2,4,2; 1,2,1], obtain the low-resolution image of rough handling;

Adopt the sub-pix process of iteration to carry out registration, successively ask for each low resolution observation frame g _i(x, y) and low-resolution reference frame g ₀Motion vector Z=[k between (x, y) ₁, k ₂, k ₃, k ₄]:

Z＝G ^-1V (1)

Wherein, $V=\left[\begin{array}{c}\mathrm{\&Sigma;}{R}_1(g_0-g_t)\\ \mathrm{\&Sigma;R}(g_0-g_t)\\ \mathrm{\&Sigma;}\frac{\&PartialD;g_t}{\&PartialD;x}(g_0-g_t)\\ \mathrm{\&Sigma;}\frac{\&PartialD;g_t}{\&PartialD;y}(g_0-g_t)\end{array}\right],$ $R=y\frac{{\&PartialD;g}_t}{\&PartialD;x}-x\frac{{\&PartialD;g}_t}{\&PartialD;y},$ ${\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1=x\frac{\&PartialD;g_t}{\&PartialD;x}+y\frac{{\&PartialD;g}_t}{\&PartialD;y},$

$G=\left[\begin{array}{cccc}\mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1^2& \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="RR"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">RR</figure-callout>}}_1& \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;y}\\ \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="RR"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">RR</figure-callout>}}_1& \mathrm{\&Sigma;}{R}^2& \mathrm{\&Sigma;R}\frac{{\&PartialD;g}_t}{\&PartialD;x}& \mathrm{\&Sigma;R}\frac{\&PartialD;g_t}{\&PartialD;y}\\ \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;R}\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;}{\left(\frac{\&PartialD;g_t}{\&PartialD;x}\right)}^2& \mathrm{\&Sigma;}(\frac{{\&PartialD;g}_t}{\&PartialD;x}\&CenterDot;\frac{{\&PartialD;g}_t}{\&PartialD;y})\\ \mathrm{\&Sigma;}{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HDA00002865662200081.png"\; state="\{\{state\}\}">R</figure-callout>}}_1\frac{\&PartialD;g_t}{\&PartialD;y}& \mathrm{\&Sigma;R}\frac{{\&PartialD;g}_t}{\&PartialD;y}& \mathrm{\&Sigma;}(\frac{{\&PartialD;g}_t}{\&PartialD;x}\&CenterDot;\frac{\&PartialD;g_t}{\&PartialD;y})& \mathrm{\&Sigma;}{\left(\frac{\&PartialD;g_t}{\&PartialD;y}\right)}^2\end{array}\right]$

$Z_{s+1}=G_s^{-1}{V}_s+Z---\left(2\right)$

According to formula (2), by alternative manner approaching to reality value progressively, s is iterations, when iteration error satisfies certain precision Th≤10 ^-3The time termination of iterations;

The coordinate of setting before the high-definition picture compensation be (x', y'), and the coordinate after high-definition picture compensates is (x, y), by formula (3) the described motion vector Z=[k that calculates ₁, k ₂, k ₃, k ₄] compensate to high resolving power initial estimation f ₀In (x, y),

In the present embodiment, referring to Fig. 8, the described POCS method of step S4 is rebuild and is specifically realized with step S41-S43, comprising:

Step S41 adopts Gaussian function that the gray-scale relation between pixel in the Euclidean distance in spatial domain between high resolving power estimated image pixel and codomain is organically combined, and obtains to have the point spread function PSF of edge retention performance;

Step S42 utilizes the correlativity between each low resolution observation frame and low-resolution reference frame, obtains the lax projective parameter of variable length, makes lax projective parameter along with the adjustment of correlativity size adaptation;

Step S43 carries out the correction of convex set iterative projection with the high resolving power initial estimate, obtains the full resolution pricture of rebuilding.

The implementation process of step S41-S43 is elaborated as follows:

As shown in Figure 9, the POCS algorithm is from given high resolving power initial estimate, constantly iterative projection in convex set until produce the high-definition picture that meets the demands.In the POCS method, each prior imformation is limited in the solution of POCS in a closed convex set.

As shown in figure 10, the following process of execution of obtaining of point spread function PSF:

The input size is Gauss's template of 5 * 5, and it is acted on pending image;

Calculate high resolving power initial estimation f by formula (4) ₀Euclidean distance between (x, y) pixel in the spatial domain:

$h(x,y)=\exp \{-\frac{{(x-m)}^2+{(y-n)}^2}{2{\mathrm{\&delta;}}_s^2}\}---\left(4\right)$

δ _s∈ [0.5,2], (m, n) are the coordinate of center pixel in 5 * 5 pixel supporting domains between each neighbor;

Calculate high resolving power initial estimation codomain f by formula (5) ₀Gray-scale relation in (x, y) between each pixel

(| f ₀(x, y)-f ₀(m, n) |):

$w_{{\mathrm{\&delta;}}_r}\left(\right|f_0(x,y)-f_0(m,n)\left|\right)=\exp \{-\frac{{|f_0(x,y)-f_0(m,n)|}^2}{2{\mathrm{\&delta;}}_r^2}\}---\left(5\right)$

Wherein, δ _rBe standard deviation, δ _r∈ [0.8,15];

According to formula (6) with the gray-scale relation between pixel in the Euclidean distance h (x, y) in spatial domain between pixel and codomain

Organically combine, obtain to have the point spread function of edge retention performance

$\tilde{h}(x,y)=U\&CenterDot;h(x,y)\&CenterDot;w_{{\mathrm{\&delta;}}_r}\left(\right|f_0(x,y)-f_0(m,n)\left|\right)---\left(6\right)$

Wherein, U is normalization coefficient, U ∈ (0.001,1).

As shown in figure 11, the implementation of lax projective parameter self-adaptation adjustment is as follows:

Calculate correlativity Rl between each low resolution observation frame and low-resolution reference frame according to formula (7):

$\mathrm{Rl}=\frac{\underset{x,y}{\mathrm{\&Sigma;}}\left[g_i(x,y)g_0(x,y)\right]}{\sqrt{\underset{x,y}{\mathrm{\&Sigma;}}{\left[g_i(x,y)\right]}^2\underset{x,y}{\mathrm{\&Sigma;}}{\left[g_0(x,y)\right]}^2}},$ 0＜Rl＜1(7)

Wherein, g _i(x, y) is the current observation frame of low resolution, and i is the sequence number of low-resolution image observation frame, g ₀(x, y) is the low-resolution reference frame;

The projective parameter λ that relaxes is set to:

λ＝2Rl (8)

Described λ is the lax projective parameter of a variable length, has reflected that the current observation frame of low resolution participates in the degree of projection, and this λ value is directly proportional to Rl.

Known prior imformation and a priori bound value δ thereof _i, described prior imformation comprises positive definite, finite energy, data consistency, slickness, according to data consistency definition convex set C _{M, n, i}:

$C_{m,n,i}=\{{\hat{f}}_l(m,n):|r_i^{{\hat{f}}_l(m,n)}(x,y)|\&le;{\mathrm{\&delta;}}_i\}---\left(9\right)$

$r_i^{{\hat{f}}_l(m,n)}(x,y)=g_i(x,y)-\underset{m,n}{\mathrm{\&Sigma;}}{\hat{f}}_l(m,n){\tilde{h}}_i(x,y;m,n)---\left(10\right)$

Wherein,

Be residual error function, (m, n) is low-resolution image g _iIn (x, y), coordinate points (x, y) is corresponding to the estimated value in high-definition picture iterative projection process

In coordinate, [0, L), L is the iteration total degree to l ∈, and L ∈ [6,15], i are the sequence number of low-resolution image observation frame.

Estimation codomain in any one full resolution pricture iterative projection process

To convex set C _{M, n, i}On projection For:

$P_{(x,y)}\left[{\hat{f}}_l(m,n)\right]={\hat{f}}_l(m,n)+\left\{\begin{array}{cc}(\mathrm{\&lambda;}\&CenterDot;r_i^{{\hat{f}}_l(m,n)}(x,y)-{\mathrm{\&delta;}}_0)\&CenterDot;{\tilde{h}}_i(x,y;m,n),& r_i^{{\hat{f}}_l(m,n)}(x,y)>{\mathrm{\&delta;}}_i\\ 0,& \left|r_i^{{\hat{f}}_l(m,n)}(x,y)\right|\&le;{\mathrm{\&delta;}}_i\\ (\mathrm{\&lambda;}\&CenterDot;r_i^{{\hat{f}}_l(m,n)}(x,y)+{\mathrm{\&delta;}}_0)\&CenterDot;{\tilde{h}}_i(x,y;m,n),& r_i^{{\hat{f}}_l(m,n)}(x,y)<-{\mathrm{\&delta;}}_i\end{array}\right.---\left(11\right)$

Wherein, δ _i=c δ ₀, c ∈ [2,5], δ ₀∈ [0.1,3.8],

Be the high resolving power estimated value

Mid point (m, n) is located 5 * 5 supporting domains of gray-scale value and (m, n)

In restriction relation between each grey scale pixel value, can calculate according to formula (6).

If j prior imformation arranged, j corresponding convex set C arranged _k, k ∈ [1, j].The solution space of POCS method is exactly the common factor C of these protruding constraint sets ₀If, C ₀Non-NULL, C ₀In any one element be all a feasible solution of problem.For given convex set C _kAnd corresponding convex set projection operator P _k, the sequence of iterations from sequence of low resolution pictures reconstruction one panel height image in different resolution is expressed as:

${\hat{f}}_{l+1}=P_j{P}_{j-1}\&CenterDot;\&CenterDot;\&CenterDot;P_1{\hat{f}}_l---\left(12\right)$

Wherein, j is the number of convex set, j ∈ [1,6], and more generally form adopts lax projection operator, can be expressed as:

${\hat{f}}_{l+1}=T_j{T}_{j-1}\&CenterDot;\&CenterDot;\&CenterDot;T_1{\hat{f}}_l---\left(13\right)$

Wherein, T _k=(1-λ _k) I+ λ _kP _k, I is vector of unit length, λ _kBe lax projection operator, be used for the speed of convergence of accelerating algorithm, 0＜λ _k＜2.Adopt lax projection operator, improved POCS convergence of algorithm stability.

Order By

Calculate

The like,

Successively carry out iteration; If do not reach iterations,

The substitution residual error function is upgraded residual error

Otherwise iteration finishes, from

In block and exceed 8 bit image span [0,255] parts, the high-definition picture that obtains rebuilding.

The present embodiment takes full advantage of small echo and has the characteristic that keeps grain details, adopts the Wavelet Bi-cubic Interpolation method to obtain the high resolving power initial estimate of convex set projection POCS, improves the soft edge problem that is caused by bilinear interpolation; Gray-scale relation between pixel in Euclidean distance in spatial domain between the high-definition picture pixel and codomain is organically combined, utilize Gauss's codomain function to obtain to have the point spread function of edge retention performance, reduce to a great extent the edge Halo effect in reconstructed results; Utilize correlativity between low resolution observation frame and low-resolution reference frame to obtain the lax projective parameter of variable length, reduce motion estimation error to rebuilding the impact of picture strip.

Embodiment two

Above-described embodiment is from theoretical side, method disclosed by the invention to be described in detail, and from theoretical side, its beneficial effect is described.Below, the present embodiment will compare the present invention from actual process from the result that prior art is rebuild the image of different scenes, support the present invention from practical application.

Referring to Figure 12, the detailed implementation that the POCS method is rebuild comprises: adopt the image of 256 * 256 pixels as original high-definition picture, original image is through generating the low-resolution image of 8 128 * 128 pixels after level, vertical translation and rotation and down-sampling; Might as well choose in low-resolution image the first frame and as the reference frame, by the small echo bilinear interpolation, the low-resolution reference frame be processed, obtain 256 * 256 high resolving power initial estimate; Realize image registration by estimation, the high-definition picture of one 256 * 256 of the row iteration backprojection reconstruction of going forward side by side.In the present embodiment, parameter is set to M=128, N=128, q=2, n=8, c=3, δ ₀=1.5, δ _s=1, δ _r=10, Th=10 ^-4, U=1, j=2, L=10.In order fully to support the present invention, the present embodiment is tested 6 kinds of different scene images in image library respectively, comprises Lena, Cameraman, Boats, House, letter ' a ', Peppers image.

For sufficient proof validity of the present invention, the present embodiment adopts Y-PSNR (Peak Signal to Noise Ratio, PSNR) to come evaluation map image quality quantitatively.The value of PSNR is larger, shows that picture quality is higher.Original high-definition picture f (x, y) and the high-definition picture of rebuilding

Between PSNR be defined as:

$\mathrm{PSNR}=10\&CenterDot;{\log }_{10}\frac{{255}^2\&CenterDot;\mathrm{qM}\&CenterDot;\mathrm{qN}}{\underset{m=1}{\overset{M}{\mathrm{\&Sigma;}}}\underset{n=1}{\overset{N}{\mathrm{\&Sigma;}}}{[f(x,y)-\hat{f}(x,y)]}^2}---\left(14\right)$

Wherein, M and N are respectively the length of low-resolution image and wide, and q is the multiple that resolution improves.

Form 1 adopts respectively prior art (the PSNR result of the high-definition picture of rebuilding based on the POCS method of bilinear interpolation-BL+POCS) and method disclosed by the invention for different images.These images comprise simulated series letter ' a ', Peppers, Lena, Boats, Cameraman, House image.By data in table as can be known, method disclosed by the invention has higher PSNR value, can effectively remove edge Halo effect, improves and rebuilds the high resolution graphics image quality.

Form 1: distinct methods is rebuild the PSNR contrast (db) of high-definition picture

Figure 13 is that the present invention adopts fixing lax projective parameter to rebuild the PSNR of high-definition picture, and horizontal ordinate is lax projective parameter value, and ordinate is the PSNR value, and by curve in figure as can be known, the PSNR of 4 width images is less than 28db.Figure 14 is the PSNR that the present invention adopts the lax projective parameter reconstructed image of variable length, horizontal ordinate is lax projective parameter value, and ordinate is the PSNR value, by curve in figure as can be known, the PSNR of 4 width images is greater than 28db, and wherein the PSNR value of Boats image has improved 1.4db.

Figure 15 is the gray level center drawing in side sectional elevation of the Lena high-definition picture that adopts respectively prior art and method disclosed by the invention and rebuild.(a) for adopting prior art to rebuild the gray level center drawing in side sectional elevation of Lena high-definition picture.Obvious spike (part with dashed lines circle marks) has appearred in edge, illustrates to have a large amount of Halo effects; (b) for adopting the gray level center drawing in side sectional elevation of Lena high-definition picture that method disclosed by the invention rebuilds.The spike of edge almost disappears, and has effectively suppressed edge Halo effect.This shows, the edge Halo effect in the high-definition picture that method disclosed by the invention is rebuild has almost disappeared, and illustrate that the present invention can effectively suppress POCS reconstructed image edge Halo effect, has improved the quality of reconstruction high-definition picture greatly.

Aspect processing speed, the present invention rebuilds 8 width low-resolution sequence images, and resolution improves 2 times as long as 294ms just can obtain a panel height image in different resolution, and visible method and system velocity ratio disclosed by the invention is very fast, has practical application.

In this instructions, each embodiment adopts the mode of going forward one by one to describe, the professional can also further recognize, each exemplary module and algorithm steps in conjunction with the disclosed embodiments description in the present invention, can realize with electronic hardware, computer software or combination both, for the interchangeability of hardware and software clearly is described, composition and the step of individual example described in general manner according to function in the above description.Actually or these functions are carried out with the hardware software mode, depend on application-specific and the design constraint of technical scheme.The professional and technical personnel can realize described function with diverse ways to each specific application, but this realization should not thought and exceeds scope of the present invention.

In conjunction with method or the algorithm steps that the disclosed embodiments in the present invention are described, can directly implement with hardware, the software module of processor execution or combination both.Software module can be placed in known any other forms of storage medium in random access memory (RAM), internal memory, ROM (read-only memory) (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technical field.

To the explanation of disclosed above-described embodiment, make this area professional and technical personnel can realize or use the present invention.Multiple modification to these embodiment will be apparent concerning those skilled in the art.In the present invention, defined General Principle without departing from the spirit and scope of the present invention, realizes in other embodiments.Therefore, the present invention can not be restricted to these embodiment shown in the present, but will meet the wide region consistent with principles of this disclosure and features of novelty.

Claims (2)

1.POCS during super-resolution image reconstruction, the inhibition method of edge Halo effect, is characterized in that, realizes according to the following steps successively in computing machine:

Step (1), computer initialization:

Set: image reading module, high resolving power initial estimate acquisition module, low-resolution image registration module and convex set projection POCS method are rebuild module;

Step (2), Halo effect in edge when suppressing according to the following steps the POCS super-resolution image reconstruction successively, edge Halo effect refers to the edge halo effect:

Step (2.1), described image reading module reads in sequence of low resolution pictures g under Same Scene from image library _i, 1≤i≤n, i are the sequence number of low-resolution image, n is the sum of low-resolution image, n ∈ [6,36], determine iterations L ∈ [6,15] according to n, the resolution of described low-resolution image is M * N, M ∈ [64,256], N ∈ [64,256], between adjacent low-resolution image, lap is arranged, have common coordinate system, with (x, y) denotation coordination point;

Step (2.2), described high resolving power initial estimate acquisition module, comprise the low-resolution reference frame and obtain submodule, wavelet decomposition submodule, bicubic interpolation submodule and wavelet inverse transformation submodule, the size of corresponding high-definition picture is qM * qN, q represents the multiple that resolution improves, q ∈ [2,4], wherein:

The low-resolution reference frame obtains submodule, optional described sequence of low resolution pictures wherein a frame low-resolution image as low-resolution reference frame g ₀, and preserve a low-resolution reference frame;

The wavelet decomposition submodule carries out the conversion of q layer scattering wavelet decomposition to described low-resolution reference frame,

Obtain low frequency subgraph , horizontal direction the high frequency subgraph

, vertical direction the high frequency subgraph

, diagonal the high frequency subgraph , the size of all subgraphs is M/q * N/q;

The bicubic interpolation submodule adopts bicubic interpolation method in the Matlab storehouse respectively to three described high frequency subgraphs

With

Extrapolate, obtaining resolution is all the high frequency subgraph g of M * N ^H, g ^VAnd g ^D

The wavelet inverse transformation submodule is described low-resolution reference image g ₀Make low frequency subgraph, the described image g of three M * N ^H, g ^V, g ^DMake the high frequency subgraph and carry out wavelet inverse transformation, obtain the high-definition picture initial estimate f of qM * qN ₀, [f ₀]=IDWT[g ₀, g ^H, g ^V, g ^D, q];

Step (2.3), described low-resolution image registration module comprises: pre-service submodule, estimation submodule and motion vector compensation submodule, wherein:

The pre-service submodule utilizes low-pass filter F to carry out filtering, F=1/16 * [1,2,1 to all low-resolution images of the low-resolution reference frame that comprises preservation; 2,4,2; 1,2,1], the low-resolution image after the acquisition rough handling;

The estimation submodule is successively asked for described each low resolution observation frame g by following formula _i(x, y) and low-resolution reference frame g ₀Motion vector Z=[k between (x, y) ₁, k ₂, k ₃, k4 _]:

Z=G ^-1V, wherein:

$V=\left[\begin{array}{c}\mathrm{\&Sigma;}{R}_1(g_0-g_t)\\ \mathrm{\&Sigma;R}(g_0-g_t)\\ \mathrm{\&Sigma;}\frac{\&PartialD;g_t}{\&PartialD;x}(g_0-g_t)\\ \mathrm{\&Sigma;}\frac{\&PartialD;g_t}{\&PartialD;y}(g_0-g_t)\end{array}\right],$ $R=y\frac{{\&PartialD;g}_t}{\&PartialD;x}-x\frac{{\&PartialD;g}_t}{\&PartialD;y},$ $R_1=x\frac{\&PartialD;g_t}{\&PartialD;x}+y\frac{{\&PartialD;g}_t}{\&PartialD;y},$

$G=\left[\begin{array}{cccc}\mathrm{\&Sigma;}{R}_1^2& \mathrm{\&Sigma;}{\mathrm{RR}}_1& \mathrm{\&Sigma;}{R}_1\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;}{R}_1\frac{\&PartialD;g_t}{\&PartialD;y}\\ \mathrm{\&Sigma;}{\mathrm{RR}}_1& \mathrm{\&Sigma;}{R}^2& \mathrm{\&Sigma;R}\frac{{\&PartialD;g}_t}{\&PartialD;x}& \mathrm{\&Sigma;R}\frac{\&PartialD;g_t}{\&PartialD;y}\\ \mathrm{\&Sigma;}{R}_1\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;R}\frac{\&PartialD;g_t}{\&PartialD;x}& \mathrm{\&Sigma;}{\left(\frac{\&PartialD;g_t}{\&PartialD;x}\right)}^2& \mathrm{\&Sigma;}(\frac{{\&PartialD;g}_t}{\&PartialD;x}\&CenterDot;\frac{{\&PartialD;g}_t}{\&PartialD;y})\\ \mathrm{\&Sigma;}{R}_1\frac{\&PartialD;g_t}{\&PartialD;y}& \mathrm{\&Sigma;R}\frac{{\&PartialD;g}_t}{\&PartialD;y}& \mathrm{\&Sigma;}(\frac{{\&PartialD;g}_t}{\&PartialD;x}\&CenterDot;\frac{\&PartialD;g_t}{\&PartialD;y})& \mathrm{\&Sigma;}{\left(\frac{\&PartialD;g_t}{\&PartialD;y}\right)}^2\end{array}\right]$

By taking progressively approaching to reality value of alternative manner,

S is iterations, when iteration error satisfies certain precision Th≤10 ^-3The time termination of iterations;

The motion vector compensation submodule, the coordinate of setting before the high-definition picture compensation be (x', y'), the coordinate after the high-definition picture compensation is (x, y), by following formula the described motion vector Z=[k that calculates ₁, k ₂, k ₃, k ₄] compensate to high resolving power initial estimation f ₀In (x, y), described f ₀F ₀The simple expression formula of (x, y):

$\left\{\begin{array}{c}x=x^{\&prime;}+k_1{x}^{\&prime;}+k_2{y}^{\&prime;}+k_3\\ y=y^{\&prime;}+k_1{y}^{\&prime;}-k_2{x}^{\&prime;}+k_4\end{array}\right.$

Step (2.4), described convex set projection POCS method is rebuild module, comprise: point spread function PSF obtains submodule, lax projective parameter self-adaptation is adjusted submodule and iterative projection correction submodule, reduces according to the following steps successively Halo effect in edge in reconstructed results:

Point spread function PSF obtains submodule, obtains according to the following steps successively the point spread function PSF with edge retention performance, it is characterized in that:

Be calculated as follows described high resolving power initial estimation codomain f ₀Euclidean distance in spatial domain in (x, y) between each pixel and supporting domain center pixel represents with h (x, y):

$h(x,y)=\exp \{-\frac{{(x-m)}^2+{(y-n)}^2}{2{\mathrm{\&delta;}}_s^2}\},$ Wherein,

δ _sBe the standard deviation of point spread function PSF, δ _s∈ [0.5,2], (m, n) are the coordinate of center pixel in 5 * 5 pixel supporting domains between each neighbor,

Be calculated as follows described high resolving power initial estimation codomain f ₀Gray-scale relation in (x, y) between each pixel

$w_{{\mathrm{\&delta;}}_r}\left(\right|f_0(x,y)-f_0(m,n)\left|\right)=\exp \{-\frac{{|f_0(x,y)-f_0(m,n)|}^2}{2{\mathrm{\&delta;}}_r^2}\},$ Wherein,

f ₀(x, y), f ₀(m, n) is respectively coordinate (x, y), the gray-scale value that (m, n) locates, δ _rBe standard deviation, δ _r∈ [0.8,15];

Obtain by following formula and considered the gray-scale relation between each pixel in described high resolving power initial estimation codomain The point spread function with edge retention performance

$\tilde{h}(x,y)=U\&CenterDot;h(x,y)\&CenterDot;w_{{\mathrm{\&delta;}}_r}\left(\right|f_0(x,y)-f_0(m,n)\left|\right),$

U is normalization coefficient, U ∈ (0.001,1);

Lax projective parameter self-adaptation is adjusted submodule, and projective parameter lax according to the correlativity self-adaptation adjustment between low resolution observation frame and low-resolution reference frame according to the following steps successively obtains the lax projective parameter of variable length:

Be calculated as follows the correlativity Rl between each low resolution observation frame and low-resolution reference frame:

$\mathrm{Rl}=\frac{\underset{x,y}{\mathrm{\&Sigma;}}\left[g_i(x,y)g_0(x,y)\right]}{\sqrt{\underset{x,y}{\mathrm{\&Sigma;}}{\left[g_i(x,y)\right]}^2\underset{x,y}{\mathrm{\&Sigma;}}{\left[g_0(x,y)\right]}^2}},$ 0＜Rl＜1，

g _i(x, y) is the current observation frame of each low resolution, g ₀(x, y) is the low-resolution reference frame,

Set: lax projective parameter λ, λ=2Rl, described λ are the lax projective parameters of a variable length, have reflected that the current observation frame of low resolution participates in the degree of projection, this λ value is directly proportional to Rl;

Iterative projection correction submodule is successively according to the following steps to described high resolving power initial estimation codomain f ₀(x, y) carries out the correction of convex set iterative projection, in order to obtain high-definition picture by the iterations of setting:

Known: prior imformation and a priori bound value δ thereof _i, described prior imformation comprises positive definite, finite energy, the smooth any information of data consistency, according to data consistency information definition convex set C _{M, n, i}, (m, n) is low-resolution image g _iIn (x, y), coordinate points (x, y) is corresponding to the high resolving power estimated value In coordinate, i is the sequence number of low-resolution image observation frame,

Set the estimation codomain in full resolution pricture iterative projection process

L ∈ [0, L), L is the total degree of iteration, L ∈ [6,15],

The estimation codomain in any one full resolution pricture iterative projection process

To convex set C _{M, n, i}On projection

For:

$P_{(x,y)}\left[{\hat{f}}_l(m,n)\right]={\hat{f}}_l(m,n)+\left\{\begin{array}{cc}(\mathrm{\&lambda;}\&CenterDot;r_i^{{\hat{f}}_l(m,n)}(x,y)-{\mathrm{\&delta;}}_0)\&CenterDot;{\tilde{h}}_i(x,y;m,n),& r_i^{{\hat{f}}_l(m,n)}(x,y)>{\mathrm{\&delta;}}_i\\ 0,& \left|r_i^{{\hat{f}}_l(m,n)}(x,y)\right|\&le;{\mathrm{\&delta;}}_i\\ (\mathrm{\&lambda;}\&CenterDot;r_i^{{\hat{f}}_l(m,n)}(x,y)+{\mathrm{\&delta;}}_0)\&CenterDot;{\tilde{h}}_i(x,y;m,n),& r_i^{{\hat{f}}_l(m,n)}(x,y)<-{\mathrm{\&delta;}}_i\end{array}\right.,$

Wherein, δ _i=c δ ₀, c ∈ [2,5], δ ₀∈ [0.1,3.8],

$C_{m,n,i}=\{{\hat{f}}_l(m,n):|r_i^{{\hat{f}}_l(m,n)}(x,y)|\&le;{\mathrm{\&delta;}}_i\},$

Be residual error function, $r_i^{{\hat{f}}_l(m,n)}(x,y)=g_i(x,y)-\underset{m,n}{\mathrm{\&Sigma;}}{\hat{f}}_l(m,n){\tilde{h}}_i(x,y;m,n),$

Be the high resolving power estimated value

Mid point (m, n) is located 5 * 5 supporting domains of gray-scale value and point (m, n)

In restriction relation between each grey scale pixel value, expression formula is:

${\tilde{h}}_i(x,y;m,n)=U\&CenterDot;h(x,y;m,n)\&CenterDot;w_{{\mathrm{\&delta;}}_r}\left(\right|{\hat{f}}_l(x,y)-{\hat{f}}_l(m,n)\left|\right),$ U∈(0.001,1)

$h(x,y;m,n)=\exp \{-\frac{{(x-m)}^2+{(y-n)}^2}{2{\mathrm{\&delta;}}_s^2}\},$ δ _s∈[0.5，]，

$w_{{\mathrm{\&delta;}}_r}\left(\right|{\hat{f}}_l(x,y)-{\hat{f}}_l(m,n)\left|\right)=\exp \{-\frac{{|{\hat{f}}_l(x,y)-{\hat{f}}_l(m,n)|}^2}{2{\mathrm{\&delta;}}_r^2}\},$ δ _r∈[0.8,15]，

Be in fact the point spread function with edge retention performance, obtain low-resolution image by the high-definition picture estimated value is carried out a DIFFUSION TREATMENT;

Order

By

Calculate

The like,

Successively carry out iteration; If do not reach iterations,

The substitution residual error function is upgraded residual error

Otherwise iteration finishes, from In block and exceed 8 bit image span [0,255] parts, the high-definition picture that obtains rebuilding.

2. the inhibition method of edge Halo effect during POCS super-resolution image reconstruction according to claim 1 is characterized in that: choose the first frame in sequence of low resolution pictures as the reference frame.

Images