CN105913388A - Priority constraint colorful image sparse expression restoration method - Google Patents

Abstract

The invention discloses a priority constraint colorful image sparse expression restoration method. The method comprises steps that step 1, a colorful image is mapped from an RGB color space to a YUV color space, and three-layer component restoration is respectively carried out; step 2, a Fast-ICA algorithm is utilized for training to acquire a complete dictionary; step 3, priority of all edge points of a broken area of a to-be-restored image is calculated, and the restoration priority sequence is determined; step 4, in combination with an SL0 algorithm, sparse reconstruction of the broken block is carried out; and step 5, edge update is carried out, and the step 3 and the step 4 are repeated till image restoration is accomplished. The method is advantaged in that image restoration quality is high, the edge structure after image restoration has relatively good continuity, and image integrity is further kept.

Description

The coloured image rarefaction representation restorative procedure of priority constraint

Technical field

The invention belongs to Digital Image Inpainting field, be specifically related to the priority in image repair technology about The coloured image rarefaction representation restorative procedure of bundle.

Background technology

The reparation of coloured image at present is based primarily upon and the restorative procedure of gray level image is directly extended to coloured image Tri-components of RGB in repair respectively.Traditional gray level image restorative procedure is divided into two classes: a class be based on The method of partial differential equation.The method substantially carries out the diffusion of information according to isophote direction, repairs effect Fruit presses close to the visual experience of people, but the method can only repair the image that damaged block is less.Another kind of be based on The method of textures synthesis.The method is to find best matching blocks to replace multiblock to be repaired, when looking in non-damaged area Less than producing erroneous matching during match block.Although at present the method has been carried out more improvement, but also It is to there will be the coupling of mistake and that weighted image block too much brings is fuzzy.Recently, a kind of new image sparse Method for expressing is paid attention in image processing field, and sparse theory is applied to the side of image repair by many scholars In method.

Summary of the invention

Finding under study for action, dependency and its structural complexity between RGB color model triple channel cause repairing Effect undesirable.During image repair, the less priority considering to repair block can make structural edge repair Multiple effect is undesirable.In order to preferably repair coloured image, coloured image is mapped by the present invention from rgb space Repair to yuv space, and utilize prioritization functions to determine the reparation order of damaged block.In conjunction with Fast-ICA Algorithm for Training dictionary and the advantage of rarefaction representation, the invention provides the coloured silk of a kind of priority constraint Color image rarefaction representation recovery technique.

The present invention takes techniques below scheme: the coloured image rarefaction representation restorative procedure of priority constraint, will Coloured image is mapped to the layering of YUV color space from RGB color and repairs；Utilize Fast-ICA algorithm Trained complete dictionary；By redefining the expression formula of gradient, optimize the calculating of priority so that knot Structure part is preferentially repaired, and uses smooth l₀Norm (SL0) algorithm carries out sparse reconstruct, ensures with this The seriality that structural edge is repaired；Carry out as follows:

The first step: coloured image is mapped to YUV color space by RGB color, to three layers of component respectively Repair；

Second step: utilize Fast-ICA Algorithm for Training to obtain complete dictionary；

3rd step: calculate image damaged area to be repaired edge priority a little, determine the excellent of reparation First order；

4th step: combine existing SL0 algorithm and damaged block is carried out sparse reconstruct, implement process as follows:

(1) initial value is setInitial surplus r₀=0, initial parameter σ=1 of function；Its Middle D^TFor the transposition of dictionary D, Y is for treating reconstruction signal.

(2) the optimal value direction of search is calculatedWherein Functionx_iIt it is the component of X.It is Function F_σ(X) local derviation on each component,

(3) damped Newton method X=X+ μ d, μ is step factor, in the present invention μ=1；

(4) Projected is utilized can to obtain X=X-D^T(DD^T)-¹(DX-Y), calculate surplus r=Y-DX simultaneously；

(5) meet when the surplus of adjacent 2 iteration | | r-r₀||₂During ＜ ε, perform next step, take ε=0.01；Otherwise, Make r₀=r, continues executing with step (2) and arrives (4).

(6) value of parameter σ is reduced further.As σ ＞ 10^-2Time, make σ=β σ (0 ＜ β ＜ 1, general β takes 0.7), r₀=0, return step (2)；Otherwise, stop iteration, obtain optimal value

5th step: update edge and repeat the 3rd step, the 4th step, until image completes to repair.

Preferably, the first step: first with the conversion formula between RGB color model and YUV color model, Image is mapped to yuv space from rgb space.In order to overcome the dependency between RGB color model triple channel And its structural complexity causes the undesirable of repairing effect, the present invention at YUV color space to breakage image Repair.YUV model is to represent colouring information by brightness and colourity, and wherein, Y represents brightness, U and V represents colourity, is mutually independent between three-component.Brightness Y in YUV model and the R in RGB model, The transformational relation of tri-color components of G, B can be expressed as:

Y=0.3R+0.59G+0.11B (1)

Chrominance information U and V are to be mixed according to different ratios by B-Y, R-Y, and Gamma calibrates it Conversion formula between its model rear is:

$\left[\begin{array}{c}Y\\ U\\ V\end{array}\right]=\left[\begin{array}{ccc}0.299& 0.587& 0.114\\ -0.147& -0.289& 0.436\\ 0.615& 0.515& -0.100\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]---\left(2\right)$

Preferably, second step: utilize Fast-ICA Algorithm for Training dictionary.In order to improve reconstruction accuracy, reduce The complexity that dictionary training calculates, it is ensured that the mistake completeness of training dictionary, it is simple to extend complete dictionary and make it have Body motility, can also improve because the interatomic mutual coherence of dictionary causes the undesirable of the sparse reparation in border simultaneously Phenomenon.The present invention uses Fast-ICA algorithm to realize dictionary training: assume one group of signal X={x₁,x₂,…x_m} For another set signal S={s₁,s₂,…s_nObservation.If the i-th value in signal X can be by the S of signal N isolated component linear hybrid form:

x_i=a_i1s₁+a_i2s₂+…+a_ins_nI=1,2 ..., j (3)

Wherein, a₁,a₂,…a_nElement for hybrid matrix A.Therefore, formula (3) rewritable one-tenth:

X=AS (4)

Wherein, hybrid matrix A and source signal S is unknown, and the purpose of Fast-ICA method estimates exactly Hybrid matrix S, by the inverse matrix of hybrid matrix A, obtains a split-matrix W so that Y=WX.By The signal Y that this obtains is exactly the estimated value of source signal S.

Algorithm implements step:

A image block one n × n observing matrix x of composition that () extracts, and carry out decentration process,

B () utilizes existing Fast-ICA method, calculating observation matrixCharacteristic vector E under dimension of m m degree With eigenvalue diagonal matrix D, wherein n ＞ m；

C () utilizes feature value vector E and eigenvalue diagonal matrix D, to observing matrixWhitening processing,

D () selects the initialization vector w of a unit norm_i；

E () updatesAnd it is standardized w_i←w_i/||w_i||；Wherein E{ } table Show and average, vectorFor vector w_iTransposition, | | w_i| | representing modulus value, g () is non-quadratic function.

If f () not yet restrains return step (e)；

G (), when i is less than needing to estimate number m of component, is repeated step (d)-(f), is obtained the separation of a m × m Matrix W.

In the present invention, non-quadratic function isY is function variable.

Preferably, the 3rd step: in order to make the reparation of structural edge have more continuity, use prioritization functions Determine the reparation order of area to be repaired, in order to ensure that edge can preferentially be repaired, improve its reliability. By improving gradient representation so that it is more marginal information can be obtained, increase the weight of data item D (p), Marginal information is made to be more prone to preferentially be repaired.The calculating of priority: image I to be repaired, Φ be image not Damaged area, Ω is image damaged area, ψ_pIt is to repair border multiblock to be repaired centered by a p.It is excellent First power is defined as follows:

P (p)=C (p) D (p) (5)

Wherein, C (p) is confidence item；D (p) is data item, respectively by formula (6) and formula (8) definition.

$C\left(p\right)=\frac{\underset{q\∈{\mathrm{\ψ}}_p\∩\mathrm{\Φ}}{\Σ}C\left(q\right)}{\left|{\mathrm{\ψ}}_p\right|}---\left(6\right)$

Time initial, function C (p) is defined as:

$C\left(p\right)=\left\{\begin{array}{cc}0& \∀p\∈\mathrm{\Ω}\\ 1& \∀p\∈I-\mathrm{\Ω}\end{array}\right.---\left(7\right)$

Wherein, | ψ_p| being the area of multiblock to be repaired, q is the point of known pixels in multiblock to be repaired.

$D\left(p\right)=\frac{\▿I_p^{\⊥}\·n_p}{\mathrm{\β}}---\left(8\right)$

Wherein,It is the vertical direction of p point gradient direction, namely isophot curve direction, n_pIt it is the method for p point Vector, β is normalization factor.

If setting the gradient table at p point to be shown as:

$\▿I=\[I_x,I_y\]---\left(9\right)$

Wherein, I_x, I_yIt is respectively the local derviation in x, y direction.In conventional methods where, gradient be all based on European away from From being indicated:

$|\▿I|={({I_x}^2-{I_y}^2)}^{\frac{1}{2}}---\left(10\right)$

The representation of the improvement shown in formula (11) is used in the present invention:

$|\▿I|=\left|I_x\right|+\left|I_y\right|---\left(11\right)$

Preferably, the 4th step: due to smooth l₀Norm (SL0) algorithm has and need not estimate in advance signal Degree of rarefication, reconstruction accuracy advantages of higher.Therefore, use existing SL0 algorithm to complete sparse reconstruct.Specifically Realize process as follows:

(1) initial value is setInitial surplus r₀=0, initial parameter σ=1 of function；Its Middle D^TFor the transposition of dictionary D, Y is for treating reconstruction signal.

(2) the optimal value direction of search is calculatedWherein Functionx_iIt it is the component of X.It is Function F_σ(X) local derviation on each component,

(3) damped Newton method X=X+ μ d, μ is step factor, in the present invention μ=1；

(4) Projected is utilized can to obtain X=X=D^T(DD^T)^-1(DX-Y), calculate surplus r=Y-DX simultaneously；

(5) meet when the surplus of adjacent 2 iteration | | r-r₀||₂During ＜ ε, perform next step, take ε=0.01；Otherwise, Make r₀=r, continues executing with step (2) and arrives (4).

(6) value of parameter σ is reduced further.As σ ＞ 10^-2Time, make σ=β σ (0 ＜ β ＜ 1, general β takes 0.7), r₀=0, return step (2)；Otherwise, stop iteration, obtain optimal value

The present invention, from considering the overall successional angle of structure, discloses one and utilizes YUV color space model, In conjunction with priority constraint and the image repair method of block rarefaction representation.This restorative procedure is as follows: first Step: damaged coloured image is mapped to YUV color space by RGB color, to three layers of component respectively Repair；Second step: utilize Fast-ICA Algorithm for Training to obtain complete dictionary；3rd step: calculate to be repaired The edge of complex pattern damaged area priority a little, determine the priority of reparation；4th step: combine Smooth l₀Norm (SL0) algorithm carries out sparse reconstruct to damaged block；5th step: update edge and repeat three or four Step is until image completes to repair.The present invention has the advantage that compared with repairing algorithm with conventional color image The inventive method repairs picture quality height, and the marginal texture repairing image has preferable seriality, can protect again Hold the globality of image.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the present invention.

Fig. 2 is the yuv space component map of Lena coloured image.

Fig. 3 is Y, the training dictionary of U, V component.

Fig. 4 is the theory diagram repairing priority.

Fig. 5 is the gradient map of image.

Fig. 6 is the present invention comparison to the repairing effect of damaged stone image.

Fig. 7 is the present invention comparison to colored Lena breakage image repairing effect.

Fig. 8 is the present invention comparison to colored baboon breakage image repairing effect.

Fig. 9 is that word of the present invention removes effectiveness comparison.

Detailed description of the invention

Below in conjunction with the accompanying drawings the preferred embodiment of the present invention is elaborated.

In the present embodiment, Fig. 1 gives the flow chart of the present invention.

The first step: first with the conversion formula between RGB color model and YUV color model, by image from Rgb space is mapped to yuv space.YUV model is to represent colouring information by brightness and colourity, wherein, and Y Representing brightness, U and V represents colourity, is mutually independent between three-component.Brightness Y in YUV model with R in RGB model, the transformational relation of tri-color components of G, B can be expressed as:

Y=0.3R+0.59G+0.11B (1)

Chrominance information U and V are to be mixed according to different ratios by B-Y, R-Y, and Gamma calibrates it Conversion formula between its model rear is:

$\left[\begin{array}{c}Y\\ U\\ V\end{array}\right]=\left[\begin{array}{ccc}0.299& 0.587& 0.114\\ -0.147& -0.289& 0.436\\ 0.615& 0.515& -0.100\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]---\left(2\right)$

Fig. 1 provides the YUV three-component figure of Lena coloured image.

Second step: utilize Fast-ICA Algorithm for Training dictionary.In order to improve reconstruction accuracy, reduce dictionary training The complexity calculated, it is ensured that the mistake completeness of training dictionary, it is simple to extend complete dictionary and make its concrete motility, Can also improve because the interatomic mutual coherence of dictionary causes the undesirable phenomenon of the sparse reparation in border simultaneously.This Bright employing Fast-ICA algorithm realizes dictionary training: assume one group of signal X={x₁,x₂,…x_mIt it is another set Signal S={s₁,s₂,…s_nObservation.If the i-th value in signal X can be independent by n in the S of signal Component linearly mixes:

x_i=a_i1s₁+a_i2s₂+…+a_ins_nI=1,2 ..., j (3)

Wherein, a₁,a₂,…a_nElement for hybrid matrix A.Therefore, formula (3) rewritable one-tenth:

X=AS (4)

Wherein, hybrid matrix A and source signal S is unknown, and the purpose of Fast-ICA method estimates exactly Hybrid matrix S, by the inverse matrix of hybrid matrix A, obtains a split-matrix W so that Y=WX.By The signal Y that this obtains is exactly the estimated value of source signal S.

Algorithm implements step:

A image block one n × n observing matrix x of composition that () extracts, and carry out decentration process,

B () utilizes existing Fast-ICA method, calculating observation matrixCharacteristic vector E under dimension of m m degree With eigenvalue diagonal matrix D, wherein n ＞ m；

C () utilizes feature value vector E and eigenvalue diagonal matrix D, to observing matrixWhitening processing,

D () selects the initialization vector w of a unit norm_i；

E () updatesAnd it is standardized w_i←w_i/||w_i||；Wherein E{ } table Show and average, vectorFor vector wi transposition, | | w_i| | representing modulus value, g () is non-quadratic function.

If f () not yet restrains return step (e)；

G (), when i is less than needing to estimate number m of component, is repeated step (d)-(f), is obtained the separation of a m × m Matrix W.

In the present invention, non-quadratic function isY is function variable.Fig. 3 is Y, and U, V divide The dictionary of amount.

3rd step: in order to make the reparation of structural edge have more continuity, use prioritization functions to determine to treat The reparation order of restoring area, in order to ensure that edge can preferentially be repaired, improves its reliability.By improving Gradient representation so that it is can obtain more marginal information, increases the weight of data item D (p) so that limit Edge information is more prone to preferentially be repaired.The calculating of priority: image I to be repaired, Φ are the non-damage zone of image Territory, Ω is image damaged area, ψ_pIt is to repair border multiblock to be repaired centered by a p.Its priority is fixed Justice is as follows:

P (p)=C (p) D (p) (5)

Wherein, C (p) is confidence item；D (p) is data item, respectively by formula (6) and formula (8) definition.

$C\left(p\right)=\frac{\underset{q\∈{\mathrm{\ψ}}_p\∩\mathrm{\Φ}}{\Σ}C\left(q\right)}{\left|{\mathrm{\ψ}}_p\right|}---\left(6\right)$

Time initial, function C (p) is defined as:

$C\left(p\right)=\left\{\begin{array}{cc}0& \∀p\∈\mathrm{\Ω}\\ 1& \∀p\∈I-\mathrm{\Ω}\end{array}\right.---\left(7\right)$

Wherein, | ψ_p| being the area of multiblock to be repaired, q is the point of known pixels in multiblock to be repaired.

$D\left(p\right)=\frac{\▿I_p^{\⊥}\·n_p}{\mathrm{\β}}---\left(8\right)$

Wherein,It is the vertical direction of p point gradient direction, namely isophot curve direction, n_pIt it is the method for p point Vector, β is normalization factor.

If setting the gradient table at p point to be shown as:

$\▿I=\[I_x,I_y\]---\left(9\right)$

Wherein, I_x, I_yIt is respectively the local derviation in x, y direction.In conventional methods where, gradient be all based on European away from From being indicated:

$\left|\▿I\right|={({I_x}^2+{I_y}^2)}^{\frac{1}{2}}---\left(10\right)$

The representation of the improvement shown in formula (11) is used in the present invention:

$\left|\▿I\right|=\left|I_x\right|+\left|I_y\right|---\left(11\right)$

Fig. 5 is the gradient map of image.It may be seen that the gradient method for expressing of the present invention is compared to traditional method More structural information can be obtained.Owing to illumination line strength determines the size of data item D (p), from Fig. 5 In can be seen that the illumination line strength of linearity structure division is substantially stronger than traditional method.Therefore, use This kind of expression formula improves the weight of data item D (p), makes structure division be easier preferentially and is repaired, from And improve image texture detail section repairing effect.

4th step: due to smooth l₀Norm (SL0) algorithm has the degree of rarefication that need not estimate signal in advance, Reconstruction accuracy advantages of higher.Therefore, use existing SL0 algorithm to complete sparse reconstruct.Implement process As follows:

(1) initial value is setInitial surplus r₀=0, initial parameter σ=1 of function；Its Middle D^TFor the transposition of dictionary D, Y is for treating reconstruction signal.

(2) the optimal value direction of search is calculatedWherein Functionx_iIt it is the component of X.It is Function F_σ(X) local derviation on each component,

(3) damped Newton method X=X+ μ d, μ is step factor, in the present invention μ=1；

(4) Projected is utilized can to obtain X=X-D^T(DD^T)-¹(DX-Y), calculate surplus r=Y-DX simultaneously；

(5) meet when the surplus of adjacent 2 iteration | | r-r₀||₂During ＜ ε, perform next step, take ε=0.01；Otherwise, Make r₀=r, continues executing with step (2) and arrives (4).

(6) value of parameter σ is reduced further.As σ ＞ 10^-2Time, make σ=β σ (0 ＜ β ＜ 1, general β takes 0.7), r₀=0, return step (2)；Otherwise, stop iteration, obtain optimal value

5th step: update edge and repeat third and fourth step until image completes to repair.

In order to check the repairing effect of inventive algorithm, image is simulated emulation, and relevant to other Repair algorithm and carry out contrast experiment.Emulation experiment is carried out under MATLAB environment.Image repair is being imitated When fruit is commented, in addition to using subjective assessment, it is also adopted by Y-PSNR (PSNR) simultaneously and carries out objective evaluation.

Fig. 6 is the present invention comparison to the repairing effect of damaged stone image.Four kinds of algorithms are to stone line The repairing effect of bar breakage image and the middle of each algorithm repair result.As can be seen from Figure, four kinds of calculations Method can complete the reparation to breakage.But structure tensor fills repairing method^[1]Mask is trained with Fast-ICA Method^[2]Mid portion at restoring area occurs in that the phenomenon of color tomography.Its reason is owing to structure tensor is filled Repairing method^[1]Mask method is trained with Fast-ICA^[2]Directly utilize mask mode at repair process and determine reparation block, Do not consider the structural information of image, cause lacking on texture seriality.In yuv space, its U, V component Mainly chrominance information, structural information therein is less.Therefore, in the reparation of U, V layer component, repair Effect is preferable so that overall repairing effect is enhanced, but the repairing effect in Y-component is still The most undesirable.The sparse reconstruction method of imperfect signal^[3]Introduce the reparation order of priority constraint image block, but it is excellent First level function cannot ensure that structure division is preferentially repaired, and causes edge the phenomenon of fracture occur.The present invention Algorithm improvement pri function, from fig. 6, it can be seen that the image more natural harmony that inventive algorithm is repaired.

Fig. 7 is the inventive algorithm comparison to colored Lena breakage image repairing effect.Four kinds of algorithms are to smooth The reparation in region, can meet the vision requirement of human eye.But in terms of grain details, structure tensor is filled Repairing method^[1]Mask method is trained with Fast-ICA^[2]Reparation all have much room for improvement, in the reparation of damaged area, the upper right corner, All occur in that the extension situation of color；In the reparation of the part of shoulder, structure tensor fills repairing method^[1]With Fast-ICA trains mask method^[2]Reparation occur in that the fracture at edge, also occur in that edge does not connects at cap portion Situation about connecing；In the reparation part of eyebrow, structure tensor fills repairing method^[1]Mask method is trained with Fast-ICA^[2] Reparation to make the eyebrow of extension create fuzzy.The sparse reconstruction method of imperfect signal^[3]Repair relatively at cap portion Good, but occur in that the discontinuous phenomenon of structure at top-right part and shoulder part.From Fig. 7 (f) it can be seen that Inventive algorithm has preferable repairing effect at reparation shoulder, eyebrow, the edge of medicated cap.

Fig. 8 is the present invention comparison to colored baboon breakage image repairing effect.Can from repairing result Going out, structure tensor fills repairing method^[1], Fast-ICA train mask method^[2]Reconstruction method sparse with imperfect signal^[3] Preferable at the repairing effect of nose smooth.But when repairing nose edge, structure tensor fills repairing method^[1] Occurring in that discontinuous, the seriality of structure is destroyed；Fast-ICA trains mask method^[2]Occur in that the extension of color Phenomenon, occurs in that the situation of edge breaks in some marginal portions, and repairing mark is the most obvious.Imperfect letter Number sparse reconstruction method^[3]Create the phenomenon of edge breaks when cheek lines are repaired, on hair is repaired, repair trace Mark is the most obvious.Owing to eyes have the texture of complexity and stronger edge, structure tensor fills repairing method^[1]、 Fast-ICA trains mask method^[2]Reconstruction method sparse with imperfect signal^[3]Repairing effect is poor.No matter the present invention exists There is preferable repairing effect smooth and marginal portion, and the reparation at eye portion also can substantially conform to people The visual effect of eye.

Fig. 9 is that word of the present invention removes effectiveness comparison.Four kinds of algorithms have removed preferably reparation for word Effect.But structure tensor fills repairing method^[1]Mask method is trained with Fast-ICA^[2]Contour completion has color The phenomenon extended, has obvious repairing mark, the sparse reconstruction method of imperfect signal^[3]Separate at the edge part repaired Show discontinuous phenomenon, and the present invention has had preferable repairing effect on border.

Annotation:

[1] structure tensor is filled repairing method and is referred to documentM,Kopriva I,Cichocki A. Inpainting color images in learned dictionary[C]//SignalProcessing Conference (EUSIPCO),2012Proceedings of the 20th European.IEEE,2012:66-70.

[2] Fast-ICA training mask method refers to documentM,Kopriva I.A comparison of dictionary based approaches to inpainting and denoising with an emphasis to independent component analysis learned dictionaries[J].Inverse Probl.Imaging,2011, 5(4):815-841.

[3] the sparse reconstruction method of imperfect signal refers to that document refers to document Xu Z, Sun J.Image inpainting by patch propagation using patch sparsity[J].Image Processing,IEEE Transactions on,2010,19(5):1153-1165.

Above the preferred embodiments of the present invention and principle are described in detail, the ordinary skill to this area For personnel, the thought provided according to the present invention, detailed description of the invention will change, and these Change and also should be regarded as protection scope of the present invention.

Claims (5)

1. the coloured image rarefaction representation restorative procedure of priority constraint, is characterized in that carrying out as follows:

Three layers of component are entered by the first step: by RGB color, coloured image is mapped to YUV color space respectively Row is repaired；

Second step: utilize Fast-ICA Algorithm for Training to obtain complete dictionary；

3rd step: calculate image damaged area to be repaired edge priority a little, determine the preferential of reparation Sequentially；

4th step: combine SL0 algorithm and damaged block is carried out sparse reconstruct；

5th step: update edge and repeat the 3rd step, the 4th step, until image completes to repair.

2. the coloured image rarefaction representation restorative procedure of priority constraint as claimed in claim 1, is characterized in that:

The first step: first with the conversion formula between RGB color model and YUV color model, by image from Rgb space is mapped to yuv space；YUV color model is to represent colouring information by brightness and colourity, wherein, Y represents that brightness, U and V represent colourity, is mutually independent between three-component；Brightness Y in YUV model It is expressed as with the transformational relation of tri-color components of R, G, B in RGB model:

Y=0.3R+0.59G+0.11B (1)

Chrominance information U and V are to be mixed according to different ratios by B-Y, R-Y, and Gamma calibrates it Conversion formula between rear model is:

$\left[\begin{array}{c}Y\\ U\\ V\end{array}\right]=\left[\begin{array}{ccc}0.299& 0.587& 0.114\\ -0.147& -0.289& 0.436\\ 0.615& 0.515& -0.100\end{array}\right]\left[\begin{array}{c}R\\ G\\ B\end{array}\right]---\left(2\right).$

3. the coloured image rarefaction representation restorative procedure of priority constraint as claimed in claim 2, is characterized in that: Second step: assume one group of signal X={x₁,x₂,…x_mIt is another set signal S={s₁,s₂,…s_nObservation, if I-th value in signal X can be formed by n isolated component linear hybrid in the S of signal:

x_i=a_i1s₁+a_i2s₂+…+a_ins_nI=1,2 ..., j (3)

Wherein, a₁,a₂,…a_nElement for hybrid matrix A；Therefore, formula (3) rewritable one-tenth:

X=AS (4)

Wherein, hybrid matrix A and source signal S is unknown, the purpose of Fast-ICA method be estimate mixed Close matrix S, by the inverse matrix of hybrid matrix A, obtain a split-matrix W so that Y=WX, thus The signal Y obtained is exactly the estimated value of source signal S.

4. the coloured image rarefaction representation restorative procedure of priority constraint as claimed in claim 3, its feature It is: obtain split-matrix W and implement step:

A image block one n × n observing matrix x of composition that () extracts, and carry out decentration process,

B () utilizes principal component analytical method calculating observation matrixCharacteristic vector E under dimension of m m degree and feature Value diagonal matrix D, wherein n ＞ m；

C () utilizes feature value vector E and eigenvalue diagonal matrix D, to observing matrixWhitening processing,

D () selects the initialization vector w of a unit norm_i；

E () updatesAnd it is standardized w_i←w_i/||w_i||；Wherein, G=yexp (-y²/2)；

If f () not yet restrains return step (e)；

G (), when i is less than needing to estimate number m of component, is repeated step (d)-(f), is obtained the separation of a m × m Matrix W.

5. the coloured image rarefaction representation restorative procedure of priority constraint as claimed in claim 4, is characterized in that: 3rd step: the calculating of priority: Ф is the non-damaged area of image, ψ_pIt is to repair border centered by a p Multiblock to be repaired, priority is defined as follows:

P (p)=C (p) D (p) (5)

Wherein, C (p) is confidence item；D (p) is data item, is respectively defined as:

$C\left(p\right)=\frac{\underset{q\∈{\mathrm{\ψ}}_p\∩\mathrm{\Φ}}{\Σ}C\left(q\right)}{\left|{\mathrm{\ψ}}_p\right|}---\left(6\right)$

$D\left(p\right)=\frac{\▿I_p^{\⊥}\·n_p}{\mathrm{\β}}---\left(7\right)$

Wherein, | ψ_p| it is the area of multiblock to be repaired,It is the vertical direction of p point gradient direction, namely etc. According to line direction, n_pBeing the normal vector of p point, β is normalization factor；

If setting the gradient table at p point to be shown as:

$\▿I=\[I_x,I_y\]---\left(8\right)$

Wherein, I_x, I_yIt is respectively the local derviation in x, y direction；

Improvement representation shown in employing formula (10):

$\left|\▿I\right|=\left|I_x\right|+\left|I_y\right|---\left(10\right).$