Summary of the invention
The goal of the invention of the present invention is: for the problem of above-mentioned existence, it is provided that a kind of to noise robustness and can be accurately
Estimate the new super resolution ratio reconstruction method based on the convex combination improving kernel regression and Combined estimator of movable information.
A kind of based on the convex combination improving kernel regression and Combined estimator the super resolution ratio reconstruction method that the present invention proposes is
Utilize the significant two kinds of methods of Performance comparision in current MFSR method: method based on kernel regression and associated movement estimation and oversubscription
The method that resolution is rebuild.Method based on kernel regression is to be difficult to, for kinematic parameter between image sequence frame, the difficult problem accurately estimated
Proposing, it need not accurate estimation just can carry out super-resolution rebuilding to multiple image.The method is in super-resolution
Reconstruction field has outstanding performance, it is possible to while interpolation reconstruction, it is achieved the denoising of image.Application scenarios due to project
It is monitor video, and monitor video exists mostly, and resolution is relatively low and ambiguous problem, so using the kernel regression improved
Method carries out super-resolution rebuilding.Associated movement estimation has the biggest with the method for super-resolution rebuilding in terms of precise motion estimation
Advantage.This combined estimation method mainly has the process of sequencing to put estimation and super-resolution rebuilding the two
Complete together in a framework, optimized the result of estimation by the estimated value of super-resolution rebuilding, it is thus possible to obtain
Compared to the higher precision of existing motion estimation algorithm, promote the effect of super-resolution rebuilding the most in turn.Both sides
Method respectively has superiority, and is also respectively arranged with shortcoming, for method based on kernel regression, because it does not has motion-estimation step, so working as
When needing the objects in images rebuild to have Large Amplitude Motion, it is rebuild precision and can decline.And for combined estimation method
For, it is to have that interactive estimation and super-resolution rebuilding the two process can mutually bring the information of another process
Limit, so it is not unlimited to the lifting of image super-resolution rebuilding effect, namely for rebuilding the lifting of effect
It is not the most notable.
For problem above, the present invention proposes the framework of a kind of convex combination, can effectively merge both approaches, mutually
Make up respective inferior position, the most mutually promote respective advantage.Object is had significantly to transport in needing the image rebuild
Time dynamic, combined estimation method at this time just can provide the advantage in estimation accurately, and rebuilds the preferable core of effect
Homing method can also promote the reconstruction effect of associated movement method of estimation further.
The new method that the present invention proposes is embodied as comprising the following steps:
Step 1: the multiframe LR image sequence enforcement associated movement estimation of input is obtained with the method for super-resolution rebuilding
" associating high-resolution HR image Ijoint”;
Step 2: the multiframe LR image sequence enforcement to input obtains " kernel regression HR based on the method for reconstructing improving kernel regression
Image Ikernel”;
Step 3: use the convex combination framework that the present invention proposes, determines the region merged according to spatial domain fusion criterion, according to
Frequency domain criteria determines the frequency content needing to merge, and overall situation weight determines the weight distribution of the fusion to two kinds of methods,
The HR image obtained upper two steps by these three aspect is merged, and obtains the HR image of last reconstruction;Convex combination mode
Can be expressed as:
Icom=(1-A(r,s))·Ikernel+A(r,s)·Ijoint (1)
A(r,s)=α V W(r,s) (2)
Wherein V and W(r,s)Determine the Mixed Zone, spatial domain to two images and frequency domain hybrid frequency blending constituent respectively, entirely
Office's parameter alpha ∈ [0,1] determines associating HR image IjointWith kernel regression HR image IkernelMixing is how many.
Main novel point due to the present invention is the proposition of convex combination framework, so for two kinds of basic MFSR methods:
Based on improve kernel regression method for reconstructing and associated movement estimation and super resolution ratio reconstruction method be embodied as only do simple introduction
Do not do and go into seriously, the determination of convex combination framework is mainly discussed in detail.
The concrete process of multi-frame image super-resolution reconstruction method substantially can be divided into three basic steps: registration, interpolation,
Rebuild.According to different algorithms, three basic links can be carried out simultaneously, it is also possible to the most independent realization.Generally, multiframe oversubscription
Resolution is rebuild and is registrated by each low-resolution image firstly the need of by motion estimation process, recycling frequency domain or spatial domain weight
Build algorithm to be permeated by the low-resolution image registrated panel height image in different resolution.And associated movement estimation and Super-resolution reconstruction
The principal character of construction method is exactly that estimation link is carried out with super-resolution rebuilding link simultaneously, two process iterative estimates,
Interact, improve motion-estimation precision thus improve reconstruction effect.
Kernel regression is as a kind of non-parametric homing method needing not rely on data modeling itself, due to the spirit that it is stronger
Activity and ductility, be widely used in image processing field, such as: image interpolation, denoising, super-resolution rebuilding etc..Based on changing
Entering kernel regression method for reconstructing is the further improvement theoretical for classical kernel regression.Classical kernel regression is rebuild and is built upon multinomial
On the basis of interpolation reconstruction method, only considered the half-tone information of image.On the basis of classical kernel regression theory, utilize image
Partial structurtes feature also been proposed the multiframe super-resolution reconstruction method of a kind of controlled kernel regression.Controlled kernel regression method for reconstructing exists
While considering half-tone information, incorporate the structural information of image also by the controlled kernel function of structure, although to a certain extent
Improve the visual quality of image, but the calculating of local covariance matrix be to rebuild from classical kernel regression to calculate gradient,
Computation complexity is high, and the problem that the deviation that there is pixel overall gray value after interpolation reconstruction is bigger and empty.The present invention uses
A kind of based on improve kernel regression multiframe super resolution ratio reconstruction method.It uses local structure tensor to replace covariance matrix,
The geometry feature of image is described, it is possible to decrease the computation complexity of kernel regression algorithm.Further, in rebuilding for controlled kernel regression
The problem that image pixel overall gray value deviation is big and empty, the geometric distance function correction core letter of structure image neighborhood pixels
Number so that the image reconstructed has preferable visual quality.
Convex combination framework set up on above two method, the image utilizing kernel regression method to rebuild have preferable details and
The advantage that combined estimation method precise motion is estimated, the fusion of convex combination mode in addition.Make up mutually the bad of two kinds of methods existence
Gesture: kernel regression method lacks exercise estimation, and combined estimation method reconstructed image quality is relatively poor.Convex combination mode relies primarily on
In the reconstruction image I that kernel regression method obtainskernel, and in place of mixing for needs, the mainly spatial domain that compound mode needs is mixed
Close criterion V, frequency-domain frequency composition mixed criteria W(r,s)And the determination of overall situation mixed criteria α.The purpose of spatial domain mixed criteria V exists
In finding out the local needing two width image blend, say, that kernel regression method rebuilds HR image the place of bigger error, also
The multiframe LR image i.e. rebuild occurs in that the place of the object of Large Amplitude Motion.Briefly, V needs to lock kernel regression method
The appearance certain error region of HR image mixes with integrated processes HR image.Similarly, frequency-domain frequency composition mixed criteria W(r,s)Purpose be to find out the frequency components needing two width image blend, need exist for using image carried out steerable pyramid
Decompose.W(r,s)Task be lock kernel regression method HR image frequency domain occur the bigger loss of signal primary bands, the most just
It is to say that the information that the loss of these band signals needs integrated processes HR image to bring makes up.For overall situation mixed criteria α,
After mainly ensureing two width image blend, still there is this feature having of normal picture, say, that make mixed image inclined
Minimum from the degree with natural image.
In sum, owing to have employed technique scheme, the invention has the beneficial effects as follows:
(1) present invention proposes the convex combination mode of a kind of innovation, can obtain high-quality SR and rebuild image;
(2) due to the determination of three kinds of principles of convexity compound mode of the present invention, for improving the HR image that kernel regression method is rebuild
Bring the information of loss;
(3) due to the present invention, two kinds of MFSR methods of current main flow have been carried out effective mixing so that this SR algorithm for reconstructing
Have robustness to interframe Large Amplitude Motion and noise, reconstructed image quality is high.
Embodiment 1
Process 1, implements combined estimation method and generates high-definition picture Ijoint
(1.1) data such as the method for super-resolution rebuilding is firstly the need of obtaining motion estimation parameter, image blurring parameter are
Can rebuild.But these parameters owing to obtaining cannot ensure the most accurately, and the capital of error tiny in parameter
Process of reconstruction is exaggerated.Therefore, if the build-up effect of parameter error can be tried every possible means to reduce, then super-resolution rebuilding
Effect will be improved.So integrated processes to rebuild super-resolution rebuilding framework as follows:
WhereinIt is the HR image rebuild, HkIt is degenerate matrix,It is the prior information about unknown HR image, Q
(Hk) function adjusts degenerate matrix HkSo that degenerate matrix is more accurate.
(1.2) combined estimation method thinks it is not completely self-contained between low-resolution image degenerate matrix, by setting up
Relation between degenerate matrix is as restrictive condition, thus ensures that degenerate matrix is the most accurate.The process rebuild is divided into two steps: the
One step, does not consider the factors such as fuzzy noise, associated movement estimation and data fusion, the information of all low-resolution images is all melted
It is combined in piece image;Second step, carries out post processing to the image after merging.Then the model that degrades of image becomes:
Yk=FkZ+Ek (4)
Wherein YkIt is low-resolution image, FkIt is motion estimation parameter matrix, EkIt it is noise.Z represents low-resolution image warp
Cross the image after data fusion:
Z=BkX+σk (5)
Wherein BkIt is fuzzy operator, σkIt it is noise.In formula (4), only consider low-resolution image and rebuild between image
Movement relation and noise, then set up the model of associated movement estimation and data fusion, solves the fusion image in formula (4)
Zk, solve last high-definition picture X finally according to formula (5).
Associated movement estimation includes two with the model of data fusion, data fit term and degenerate matrix limit entry.Data
Fit term employing robust iterative data fitting function:
The restrictive condition of degenerate matrix is as follows:
WhereinIt it is kth frame low-resolution image motion estimation parameter inverse of a matrix matrix.Wherein ε is constant, 0 to 1
Between.Weight in view of different low-resolution frames is distinguishing, this difference and the time gap of low-resolution image
Relevant.Adjacent frame correlation is relatively big, and the degenerate matrix therefore from consecutive frame is more credible, so needing to give bigger power
Being worth and frame apart from each other, possible picture material changes greatly, and degenerate matrix credibility is slightly smaller.Therefore there is employed herein index letter
Number as weight function, when two width low-resolution image time gaps farther out time, give less weights.(7) formula thinks each
Width low-resolution image can obtain the image needing to rebuild, different low-resolution images after " reversely " estimation
The approximation for reconstruction image after " reversely " estimation should be close.As shown in Figure 2.
(1.3) according to analysis above, the model such as following formula of associated movement estimation and data fusion is set up:
Then solve the problems referred to above with alternative iteration method, first carry out estimation, obtain F1,…FNInitial estimate, so
Rear hypothesis F1,…FNConstant, withIt is iterated for independent variable:
Wherein, ζ represents the step-length of iteration.According to formulaJudge the no end covered with clouds of whole iterative process.Wherein,
η is threshold value.After above-mentioned iteration stopping, the image Z after can being merged, sets up with drag:
Using gradient descent method to solve (10) formula, the cut-off condition of iteration is:The X finally obtained
It it is i.e. the HR image rebuild.
Process 2, implements the method for reconstructing based on improving kernel regression and generates high-definition picture Ikernel
It is the improvement theoretical to controlled kernel regression based on the method for reconstructing improving kernel regression, and controlled kernel regression theory can
Review in classical kernel regression theoretical.Classical kernel regression method, independent of the exponent number N returned, its essence is the local of pixel
Weighting procedure, this local linear filtering characteristic is the inherent limitation that traditional core returns.The mathematical model signal of kernel regression method
Figure is as shown in Figure 3.The kernel function of classical kernel regression theory contains only the spatial information of sampled point pixel, does not accounts for pixel
The impact on weights of the gray value distance.And in real image, usually contained different image texture details, the limit of image
The zoness of different such as edge, smooth, angle point reflect the variation characteristic that data are different.Therefore, on the basis of classical kernel regression theory,
Propose the controlled kernel regression method of estimation adding image Self-variation.The method not only relies on the space length of sampled point, also
Consider the Gray homogeneity between neighborhood sampled point, thus make controlled kernel regression method be provided with nonlinear filtering characteristic.
Obtained the partial signal construction of image by the half-tone information robustness analyzing local pixel value, this information incorporated kernel function,
The shape of the adaptive change kernel function of energy and size so that the shape of core and the partial structurtes feature being sized to image
Join.
But controlled kernel function theory is when calculating the structural information of each pixel, needs the picture interested to each
The local neighborhood of vegetarian refreshments is done covariance and is calculated;And the calculating of covariance has the phenomenon of double counting, calculate process more complicated,
And it is computationally intensive.Structure tensor, as the powerful of the local geometry analyzing image, has been successfully applied figure
The fields such as the estimation of the field of direction of picture, removal picture noise.Structure tensor is one and comprises each element neighborhood side in image
To the symmetrical matrix with strength information, it is possible to accurately portray direction and the structural strength of partial gradient, and amount of calculation is little,
Therefore the kernel regression method for reconstructing three dimensional structure tensor that improves replaces covariance matrix, not only can the structure letter of picture engraving
Breath, distinguishes different structural regions, and can reduce amount of calculation, effectively maintains the detailed information such as edge of image.
(2.1) due in kernel regression method the calculating of directional information need to calculate gradient, so improve kernel regression calculate
First method carries out gradient estimation.The operator of rim detection generally utilizes the direction template convolved image of image space to detect figure
The vertical edge of picture and horizontal edge, and edge direction is perpendicular to gradient direction, therefore can be carried out by edge detection operator
Gradient is estimated.Sobel operator comprises gaussian derivative and smooth characteristic simultaneously, therefore to making an uproar in the presence of low-resolution video
Sound has robustness, will not have undesirable effect subsequent result because of the interference of noise, and has and quickly calculate gradient
Characteristic.So using 3D-Sobel operator permissible with multiple image sequence frame convolutional calculation Grad in improving kernel regression method
The effective time reducing the big moment matrix of calculating.Fig. 4 is used 3D-Sobel operator template.
(2.2) technology on the basis of kernel regression reconstruction is built upon gradient and interpolation, it is thus achieved that after frame sequence gradient, needs
The direction of computation structure tensor obtains the interpolation kernel adapting to picture structure, and structure tensor matrix is from partial gradient derivation
Coming, the characteristic vector of structure tensor and eigenvalue can describe the energy principal direction of concrete neighborhood of pixel points, at sampling location P
The three dimensional structure tensor of point is as follows:
Wherein,The gradient of representative image, I representing input images, subscript x, y, t represent three gradient directions respectively, and W is
Partial analysis window, three dimensional structure tensor S (P) is symmetrical matrix, and after eigenvalue solves, three dimensional structure tensor is expressed as:
S (P)=λ1e1e1 T+λ2e2e2 T+λ3e3e3 T (12)
Wherein, λ1≥λ2≥λ3>=0 is the eigenvalue being arranged in decreasing order of matrix, e1, e2, e3It is that they characteristics of correspondence are vowed
Amount, the intensity of they representative image partial structurtes and direction.The kernel function then improved is:
XiRepresenting noise samples point, X is intended to the sampled point interested estimated, h referred to as overall situation smoothing parameter.
(2.3), in process of reconstruction, image intensity value deviation is excessive, utilizes the pixel interdependence principle of image, by image slices
The kernel function of the neighbouring geometric distance function above-mentioned improvement of insertion of element can suppress the generation of this phenomenon, and this function calculates
The value gone out can save in advance.Adjacency function is:
α is correction factor, the controlled kernel function expression formula that finally improve mutually multiplied with formula (13):
It is final that to solve form as follows:
Wherein,Represent the pixel value of point-of-interest to be estimated.Y be each sampled point gray value row to
Amount, XXRepresent sampled point each rank distance matrix, e1Be header element be 1, remaining is the column vector of 0.WXFor the controlled kernel function improved
KHThe diagonal matrix of ().
WX=diag [KH(X1-X),KH(X2-X),...,KH(XP-X)]×diag[K(X1),K(X2),...,K(XP)]
(17) (2.4) are analyzed by above, and the key step of the kernel regression multiframe super-resolution rebuilding improved is summarized as follows:
(1) input multiple image sequence
(2) 3D sobel operator is used to calculate the gradient of each interested pixel point;
(3) utilize formula (11) to calculate three-dimensional partial gradient structure tensor S (P), try to achieve three dimensional structure directional information;
(4) formula (14) is utilized to calculate the geometric distance functional value of the neighborhood territory pixel revising kernel function;
(5) solution step (3) and (4) obtained substitutes into formula (15), draws kernel function value K of improvementH(·);
(6) kernel function value K that step (5) is drawnH() substitutes into formula (17) and draws WX;
(7) W that step (6) is drawnXSubstitution formula (16), obtains final output image Ikernel。
Process 3, the Frank-Wolfe algorithm implementing the present invention generates high-definition picture Icom
Convex combination framework is set up on above two method, utilizes respective advantage, and the shortcoming making up the other side obtains more
High-quality reconstruction effect.With reference to Fig. 4, generate IcomTo implement process as follows:
(3.1) spatial domain mixed criteria V is determined:
Calculating the high-definition picture I that kernel regression method produceskernelTime, need to calculate sense with 3D sobel operator emerging
The gradient of interest pixel, can obtain the gradient of area-of-interest in the multiframe LR image sequence of input in the process, by
Grad in image reflects the place of the notable change of image intensity value to a certain extent.If so LR figure to be reconstructed
Area-of-interest exists the wing acutely agitated in the object of strenuous exercise, such as penguin picture, then every two field picture in Xiang
The Grad of corresponding region is compared other static regions and is had certain deviation.Therefore, kernel regression method is being used to calculate gradient
During value, whether the variance of multiple Grad of multiframe LR image same ROI pixel can exist as this region is significantly transported
One metric of animal body.Calculate the variance of the corresponding Grad of the LR image ROI region of all frames, and normalize to 0,
Between 1, the most just obtain a variance map (variance map) V.The region that variance map intermediate value is higher is corresponding
It is the biggest to there is the probability significantly moved in region, and the convex combination framework the most just proposed with greater need for the present invention adds more containing essence
The really information of estimation, also further illustrates and proposes the reasonability that variance this concept of map V mixes in spatial domain.Method exists
When spatial domain is implemented, choose V-value in variance map and exceed the region of certain threshold value and mix, and other regions keep constant.
(3.2) frequency-domain frequency composition mixed criteria W is determined(r,s):
Frequency-domain frequency composition mixed criteria W(r,s)Purpose determine that and need to carry out, at frequency domain, the primary bands that mixes.Due to
Will process image at frequency domain, the present invention is to improving the high-resolution rate image I that kernel regression method obtainskernelUse controlled gold word
Tower (steerable pyramid) algorithm decomposes frequency domain, obtains image IkernelDifferent scale and different directions frequency band point
Solve.Steerable pyramid algorithm can be provided by the decomposition of multiple dimensioned multi-direction frequency band image is combined localization (in space or
Frequency domain) expression.As shown in Figure 5.Represent direction number of frequency bands r ∈ (0, R) decomposed with R, S represents yardstick quantity s decomposed
∈ (0, S), on different scale S, can obtain the decomposition result of R different directions band, it is simple to we select more accurately
The direction frequency band of mixing is needed on different scale.By variance map V obtained in the previous step, choose the district having Large Amplitude Motion to change
Territory (assume V > 0.5), implement controllable pyramid decomposition and obtain the decomposition of R direction frequency band of S yardstick, then calculate these image districts
Territory average energy under the frequency band of different scale direction, is distributed one according to region energy under the frequency band of different scale direction
Important conclusion: signal be lost in that frequency band containing higher-energy occurs the most universal.And these contain the frequency band of higher-energy
Preferably corresponding to a certain partial-band in R direction band decomposition under different scale S.Introduce new variableRepresent
There is under yardstick S the part frequency band of higher-energy.λ span is between 0 to 1, controls the frequency chosen from total frequency band number R
The number of band.By analysis above, for obtaining at the location of pixels x in ROI region:
W in above formula(r,s)Choose and can effectively make two width image IkernelAnd IjointAlong image IkernelIn need most volume
Those frequency bands of external information mix.The mixing schematic diagram in block region is as shown in Figure 6.
(3.3) overall hybrid parameter α is determined:
The Main Function of overall situation hybrid parameter α is to ensure that mixed two width images are the width meeting natural image character
Real image.The statistical property of band reduction of fractions to a common denominator solution is analyzed by these needs by natural image: natural image is for bandpass filtering
The skirt response of device is height non-gaussian.And the departure degree for Gauss model can be with the peak value (Kurtosis) of response
Measure.And the most really have the peak value of investigation display natural image to be kept approximately constant over different frequency bands.Certain point
The peak value of cloth is defined as:
Wherein μ4Being the Fourth-order moment of distribution average, σ is the standard deviation of distribution.Defined by this, Gauss distribution
Peak value is 0.So choosing of optimum alpha make last mixing HR image IcomPeak value on different frequency bands has minimum
Change.I.e. α should meet:
WhereinIt is last mixing HR image IcomR direction is contained for comprising S yardstick and each yardstick
The peak value of the skirt response distribution of the band filter of frequency band.It it is the average of peak value on all frequency bands.Above formula uses
The fminsearch function of MATLAB is readily available solution.Finally implement mixing, (2) can be brought into according to the analysis of above two steps
Formula obtains A(r,s)(1) formula of bringing into again obtains last reconstruction HR image Icom。
The invention is not limited in aforesaid detailed description of the invention.The present invention expands to any disclose in this manual
New feature or any new combination, and the arbitrary new method that discloses or the step of process or any new combination.