CN102254301B - Demosaicing method for CFA (color filter array) images based on edge-direction interpolation - Google Patents

Abstract

The invention discloses a demosaicing method for CFA (color filter array) images based on edge-direction interpolation. The demosaicing method for CFA images mainly solves the problems that the interpolation effect of the high-frequency part of an image is not good, and the false color effect is serious in the traditional demosaicing method. The demosaicing method for CFA images comprises the steps of: (1) inputting a CFA image to be demosaiced; (2) estimating the brightness; (3) interpolating a green channel; (4) respectively carrying out dual linear interpolation on a red channel and a blue channel; (5) respectively correcting the red channel, a green channel and the blue channel; and (6) outputting a colorful image. The demosaicing method for CFA images has the advantages of capability of better maintaining high-frequency information of the image, inhibiting the false color effect effectively and improving the visual effect of the demosaiced CFA image has short running time.

Description

CFA image demosaicing method based on the edge direction interpolation

Technical field

The invention belongs to technical field of image processing, further relate to the CFA image demosaicing method based on the edge direction interpolation in image-recovery technique field.The present invention carries out interpolation to the CFA image that mosaic effect is arranged, to obtain to have the coloured image of three passage full details of red, green, blue.The present invention can be used in the patrilineal line of descent with only one son in each generation sense camera, processes as the CFA image that post-processing algorithm is caught CCD or cmos sensor, recovers complete digital color image, thereby remedies owing to reducing losing of image color information that the camera hardware cost brings.

Background technology

Digital color image represents color value with the red, green, blue three primary colours usually.For cost consideration, what present most of camera adopted is single CCD or cmos sensor, by add a color filter array (color filer array, CFA) before sensor, only represents coloured image with a matrix.Only have a color value on each pixel, two other color value then comes interpolation according to its neighborhood information, and this interpolation technique is called as " demosaicing " (demosaicking).CFA image demosaicing is in fact an ill-condition problem, and namely the information by known former Fig. 1/3 recovers whole information.The CFA of Bayer (Bayer) pattern is owing to simply and efficiently, being be most widely used at present a kind of.In the CFA of Bayer pattern, there are two pixels to only have green value in every four adjacent pixels, other two are only had respectively red value, blue valve.

Existing CFA image demosaicing technology can be divided into linear interpolation techniques and non-linear interpolation technology simply.Linear interpolation is the simplest also the most representative bilinear interpolation and two cube interpolation.The advantage of bilinear interpolation and two cube interpolation techniques is that its realization is simple, and speed is fast; But its shortcoming also is apparent: in image, can produce serious artificial trace (such as color fringing, i.e. false colour effect), and especially more outstanding at the HFS of image.Compare linear interpolation techniques, the non-linear interpolation technology is more complicated, owing to having considered the interchannel correlativity of RGB, its interpolation successful is better than linear interpolation techniques.This class technology has a lot, for example, S.Pei and I.Tam are at article " Effective color interpolation in CCD color filter array using signal correlation " (Proc.Int.Conf.Image Process., Sep.2000, PP.488-491) the middle efficient interpolation technique that utilizes signal correction that proposes.The non-linear interpolation technology comprises that also some are based on the demosaicing technology of iterative algorithm, such as B.Gunturk, Deng people " Color plane interpolation using alternating projections " (IEEE Trans.Image Process. in article, vol.11, no.9, pp.997-1013, Sep.2002) the alternating projection method that proposes in, Wenmain Lu and Yap-peng Tan " Color filter array demosaicing:new method and performance measures " (IEEE Trans.Image Process. in article and for example, vol.12, no.10, pp.1194-1210, Oct.2003) the middle demosaicing method that proposes.A kind of coloured image demosaicing technology based on rarefaction representation has been proposed in recent years, such as people such as Julien Mairal at article " Sparse representation for color image restoration " (IEEE Trans.Image Process., vol.17, no.1, Jan.2008) the middle method of describing.Corresponding sparse dictionary be learnt and be trained to this method need to mass data, and then according to dictionary reconstruct coloured image, algorithm is more complicated.In a word, although the non-linear interpolation technology can obtain the high-quality color image, its complexity is high, and calculation cost is large.

The patented claim that Microsoft proposes " high-quality gradient that is used for the coloured image demosaicing is proofreaied and correct the linear interpolation " (applying date: on 03 15th, 2005, application number: 200510055929.5, publication number: disclose a kind of gradient calibration linear interpolation method and system for the coloured image demosaicing CN1722852).The value of required color is at first estimated by the method and system with existing linear interpolation techniques such as bilinear interpolation technology, then come the calculation correction item by the gradient of calculating required color on the current pixel, last this interpolation of linear combination and correction term are omitted color-values to produce on the pixel.The method directly affects with gradient and proofreaies and correct by the color-values that has the interpolation technique estimation now.The weak point of the method is: only utilized gradient to proofread and correct the result of bilinear interpolation, although than bilinear interpolation certain improvement has been arranged, still not ideal to image detail section processes effect, false colour effect is still serious.

The people such as Zhang are at article " Color demosaicking via directional linear minimum mean square-error estimation " (IEEE Trans.Image Process., vol.14, no.12, pp.2167-2178, Dec.2005.) a kind of demosaicing method based on direction linear sowing square estimation of error of middle proposition.The method utilizes linear sowing square estimation of error (LMMSE) that red green and bluish-green difference signal is estimated from level and vertical both direction, then to each pixel with two green estimated value optimization fusions so that the variance of itself and neighborhood territory pixel is minimum, at last by difference signal linear sowing square estimation of error being reconstructed the full detail of three passages.The weak point of the method is: the method need to repeatedly be estimated and merge, although the treatment of details of image is made moderate progress, algorithm complex is high, and program runtime is long.

Summary of the invention

The present invention is directed to existing demosaicing algorithm to the defective of the HFS interpolation poor effect of image, a kind of CFA image demosaicing method based on the edge direction interpolation has been proposed, by take the brightness of accurate estimation as instructing the correction of green channel after edge direction interpolation and the interpolation, so that the interpolation result of the HFS of image is significantly improved, establishment false colour effect.

For achieving the above object, the present invention includes following key step:

(1) input one width of cloth is treated the CFA image of demosaicing;

(2) estimated brightness

2a) 9 * 9 wave filter Γ of design ₁With 5 * 5 wave filter Γ ₂

2b) use respectively wave filter Γ ₁And Γ ₂To input CFA image filtering, obtain and input image I after the filtering of the same size of CFA image _{Γ 1}And I _{Γ 2}, two width of cloth image co-registration are obtained a width of cloth luminance picture

(3) to the green channel interpolation

3a) calculate luminance picture

Gradient delta H on the horizontal and vertical directions, Δ V;

3b) relatively Δ H and Δ V are big or small, obtain the matrix of edge E of horizontal and vertical direction _hAnd E _v

3c) according to horizontal and vertical direction matrix of edge E _hAnd E _vJudge the interpolation direction and to the green channel interpolation;

(4) red and blue channel are carried out respectively bilinear interpolation;

(5) respectively the red, green, blue passage is revised;

(6) output coloured image.

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

The first, the present invention is by revising after edge direction interpolation and interpolation green channel as instructing take the brightness of accurate estimation so that the interpolation effect of the HFS of image is significantly improved, good visual effect, establishment false colour effect;

The second, interpolation of the present invention adopts linear method, and calculated amount is little, and travelling speed is fast, compares with some existing iterative algorithms based on the algorithm of linear sowing square estimation of error with prior art, and working time is short, has shortened the camera imaging time.

Description of drawings

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is prior art demosaicing design sketch;

Fig. 3 is demosaicing design sketch of the present invention.

Embodiment

Below in conjunction with accompanying drawing 1, the step that the present invention is realized is described in further detail:

Step 1 is inputted the CFA image I that a width of cloth is treated demosaicing _s

The CFA image for the treatment of demosaicing is the CFA image of Bayer (Bayer) pattern, and each pixel only has a color value in the three primary colours known in this image, all the other two color value the unknowns; In four pixels of every adjacent 2 * 2 image blocks of CFA image two known green values of pixel are arranged, two other pixel is known red value, blue valve respectively.

Step 2, estimated brightness

2a) according to 9 * 9 wave filter Γ of following formula design ₁With 5 * 5 wave filter Γ ₂

$\mathrm{\Γ}=A{(H^T{\mathrm{\σ}}^{-1}H+{\mathrm{\λ}}_1{M}_1^T{M}_1+{\mathrm{\λ}}_2{M}_2^T{M}_2)}^{-1}{H}^T{\mathrm{\σ}}^{-1}$

Wherein, Γ is the wave filter of required design, A is the transition matrix that the RGB coloured image is converted into brightness space, H is the decimation factor that the RGB color picture sampling becomes Bayer form A FA image, T is matrix transpose operator, and-1 is the matrix inversion operation symbol, and σ is noise variance, σ=0.00001, constant λ ₁=0.0008, λ ₂=0.02, M ₁And M ₂Be filtering matrix,

I ₃Be 3 * 3 unit matrixs, S ₁Be Hi-pass filter, Expression Kronecker (Kronecker) operator, M ₂By Hi-pass filter S ₂Obtain by the following formula computing:

$M_2=\left(\begin{array}{ccc}1.547& -0.577& -0.577\\ -0.577& 1.547& -0.577\\ -0.577& -0.577& 1.547\end{array}\right)\⊗S_2$

The matrix A size is got 81 * 243 unit matrixs in the following formula, and the matrix H size is taken as 81 * 243, calculates 81 * 81 matrixes according to formula, and this matrix the 41st row is taken out, and is arranged in 9 * 9 matrixes, i.e. wave filter Γ ₁

The matrix A size is taken as 75 * 25 unit matrixs in the following formula, and the matrix H size is taken as 25 * 75, calculates 25 * 25 matrixes according to formula.This matrix the 13rd row is taken out, be arranged in 5 * 5 matrixes, i.e. wave filter Γ ₂

2b) use respectively wave filter Γ ₁And Γ ₂To image I _sFiltering.

With the CFA image I _sRespectively with wave filter Γ ₁And Γ ₂Carry out two-dimensional convolution, obtain two width of cloth and image I _sImage I of the same size _{Γ 1}And I _{Γ 2}Two width of cloth image co-registration are obtained a width of cloth luminance picture

Fusion rule is as follows: when location of pixels index i and j are odd number or are even number,

Be the brightness value of pixel (i, j), I _{Γ 1}(i, j) is image I _{Γ 1}The value of middle pixel (i, j), otherwise,

I _{Γ 2}(i, j) is image I _{Γ 2}The value of middle pixel (i, j) has obtained a width of cloth luminance picture thus

Step 3 is to the green channel interpolation

3a) calculate luminance picture

Gradient delta H on the horizontal and vertical directions, Δ V, computing formula is as follows:

$\left\{\begin{array}{c}\mathrm{\ΔH}(i,j)=D_h(i,j)+\frac{1}{2}{D}_h(i,j-1)+\frac{1}{2}{D}_h(i,j+1)\\ \mathrm{\ΔV}(i,j)=D_v(i,j)+\frac{1}{2}{D}_v(i-1,j)+\frac{1}{2}{D}_v(i+1,j)\end{array}\right.$

$\left\{\begin{array}{c}{D}_h(i,j)=|\hat{L}(i,j)-\hat{L}(i,j-1)|+|\hat{L}(i,j)-\hat{L}(i,j+1)|+|2\hat{L}(i,j)-\hat{L}(i,j-2)-\hat{L}(i,j+2)|\\ D_v(i,.j)=|\hat{L}(i,j)-\hat{L}(i-1,j)|+|\hat{L}(i,j)-\hat{L}(i+1,j)|+|2\hat{L}(i,j)-\hat{L}(i-2,j)-\hat{L}(i+2,j)|\end{array}\right.$

Wherein, the location index of (i, j) expression pixel, namely this pixel is positioned at the capable j row of i, and Δ H (i, j) is the horizontal gradient of pixel (i, j), D _hBe the horizontal gradient operator, Δ V (i, j) is the VG (vertical gradient) of pixel (i, j), D _vBe the VG (vertical gradient) operator,

Be the brightness value of pixel (i, j), | * | the expression signed magnitude arithmetic(al).

3b) relatively Δ H and Δ V are big or small, obtain the matrix of edge E of horizontal and vertical direction according to following formula _hAnd E _v:

$e_h(i,j)=\left\{\begin{array}{c}1,\mathrm{\&Delta;H}(i,j)<\mathrm{\&Delta;V}(i,j)\\ 0.5,\mathrm{\&Delta;H}(i,j)=\mathrm{\&Delta;V}(i,j)\\ 0,\mathrm{\&Delta;H}(i,j)>\mathrm{\&Delta;V}(i,j)\end{array}\right.,$ $e_v(i,j)=\left\{\begin{array}{c}1,\mathrm{\&Delta;H}(i,j)>\mathrm{\&Delta;V}(i,j)\\ 0.5,\mathrm{\&Delta;H}(i,j)=\mathrm{\&Delta;V}(i,j)\\ 0,\mathrm{\&Delta;H}(i,j)<\mathrm{\&Delta;V}(i,j)\end{array}\right.,$

$\left\{\begin{array}{c}{E}_h(i,j)=e_h(i,j-2)+e_h(i,j)+e_h(i,j-2)\\ E_v(i,j)=e_v(i-2,j)+e_v(i,j)+e_v(i-2,j)\end{array}\right.$

Wherein, e _h(i, j) is the horizontal edge quantized value of pixel (i, j), and Δ H is horizontal gradient, and Δ V is VG (vertical gradient), e _v(i, j) is the vertical edge quantized value of pixel (i, j), E _h(i, j) is the terminal level rim value of pixel (i, j), E _v(i, j) is the final vertical edge value of pixel (i, j).

3c) according to horizontal and vertical direction matrix of edge E _hAnd E _vJudge the interpolation direction and to the green channel interpolation, circular is as follows:

When matrix E _hThe capable j column element of i E _hThe value of (i, j) is more than or equal to 2.5 o'clock, the green estimated value of pixel (i, j) Calculated by following formula:

$\hat{G}(i,j)=\hat{L}(i,j)-\frac{\underset{n\&PlusMinus;1}{\mathrm{\&Sigma;}}(\hat{L}(i,j+n)-c(i,j+n))}{2}-\frac{2(\hat{L}(i,j)-c(i,j))-\underset{l=\&PlusMinus;2}{\mathrm{\&Sigma;}}(\hat{L}(i,j+l)-c(i,j+l))}{4}$

Wherein, Be the green estimated value of pixel (i, j),

Be the brightness value of pixel (i, j), the known color value of c (i, j) expression pixel (i, j), n and l are constant, ∑ represents accumulating operation;

When matrix E _vThe capable j column element of i E _vThe value of (i, j) is more than or equal to 2.5 o'clock, the green estimated value of pixel (i, j)

Calculated by following formula:

$\hat{G}(i,j)=\hat{L}(i,j)-\frac{\underset{m=\&PlusMinus;1}{\mathrm{\&Sigma;}}(\hat{L}(i+m,j)-c(i+m,j))}{2}-\frac{2(\hat{L}(i,j)-c(i,j))-\underset{k=\&PlusMinus;2}{\mathrm{\&Sigma;}}(\hat{L}(i+k,j)-c(i+k,j))}{4}$

Wherein,

Be the green estimated value of pixel (i, j), Be the brightness value of pixel (i, j), the known color value of c (i, j) expression pixel (i, j), m and k are constant, ∑ represents accumulating operation;

Other situation, the green estimated value of pixel (i, j)

Calculated by following formula:

$\hat{G}(i,j)=\hat{L}(i,j)-\frac{\underset{m=\&PlusMinus;1,n=\&PlusMinus;1}{\mathrm{\&Sigma;}}(\hat{L}(i+m,j+n)-c(i+m,j+n))}{4}-\frac{2(\hat{L}(i,j)-c(i,j))-\underset{k=\&PlusMinus;2,l=\&PlusMinus;2}{\mathrm{\&Sigma;}}(\hat{L}(i+k,j+l)-c(i+k,j+l))}{8}$

Wherein,

Be the green estimated value of pixel (i, j),

Be the brightness value of pixel (i, j), the known color value of c (i, j) expression pixel (i, j), m, n, k and l are constant, ∑ represents accumulating operation.

Obtained thus a green channel image that all has green value in all pixel position.

Step 4 is carried out respectively bilinear interpolation to red and blue channel

4a) for the pixel (i, j) of known green value, the red estimated value of this pixel

Calculated by following formula:

${\hat{R}}_b(i,j)=\hat{G}(i,j)-\frac{1}{4}\underset{m,n\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i+m,j+n)-r(i+m,j+n)]$

Wherein,

Be the red estimated value of pixel (i, j),

Be the green value after pixel (i, the j) interpolation, m and n are integer, and r (i+m, j+n) is the known red value of pixel (i+m, j+n).

4b) for the pixel (i, j) that is positioned at known green value, the pixel (i, j-1) about this pixel and (i, j+1) are the pixel of known green value, their red estimated value

Calculated by following formula:

${\hat{R}}_{g1}(i,j)=g(i,j)-\frac{1}{2}\underset{n\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i,j+n)-r(i,j+n)]$

Wherein,

Be the red estimated value of pixel (i, j),

Be the green value after pixel (i, the j) interpolation, n is integer, and r (i, j+n) is the known red value of pixel (i, j+n).

4c) for the pixel (i, j) that is positioned at known red value, this pixel pixel (i-1, j) and (i+1, j) up and down is the pixel of known green value, their red estimated value

Calculated by following formula:

${\hat{R}}_{g2}(i,j)=g(i,j)-\frac{1}{2}\underset{m\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i+m,j)-r(i+m,j)]$

Wherein, Be the red estimated value of pixel (i, j), m is integer,

Be the green value after pixel (i+m, the j) interpolation, r (i+m, j) is the known red value of pixel (i+m, j).

Obtained thus a width of cloth all has red value in all pixel position red channel image.

4d) for the pixel (i, j) of known red value, the blue estimated value of this pixel Calculated by following formula:

${\hat{B}}_r(i,j)=\hat{G}(i,j)-\frac{1}{4}\underset{m,n\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i+m,j+n)-b(i+m,j+n)]$

Wherein,

Be the blue estimated value of pixel (i, j), m and n are integer,

Be the green value after pixel (i+m, the j+n) interpolation, b (i+m, j+n) is the known blue valve of pixel (i+m, j+n).

4e) for the pixel (i, j) that is positioned at known green value, the pixel (i, j-1) about this pixel and (i, j+1) are the pixel of known green value, their blue estimated value

Calculated by following formula:

${\hat{B}}_{g2}(i,j)=g(i,j)-\frac{1}{2}\underset{n\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i,j+n)-b(i,j+n)]$

Wherein, Be the blue estimated value of pixel (i, j), n is integer,

For pixel (i, j+n) is inserted

Green value after the value, b (i, j+n) are the known red value of pixel (i, j+n).

4f) for the pixel (i, j) that is positioned at known red value, this pixel pixel (i-1, j) and (i+1, j) up and down is the pixel of known green value, their red estimated value Calculated by following formula:

${\hat{B}}_{g1}(i,j)=g(i,j)-\frac{1}{2}\underset{m\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i+m,j)-b(i+m,j)]$

Wherein,

Be the blue estimated value of pixel (i, j), m is integer,

Be the green value after pixel (i+m, the j) interpolation, b (i+m, j) is the known blue valve of pixel (i+m, j).

Obtained thus a width of cloth all has red value in all pixel position blue channel image.

4b) for the pixel (i, j) of known R component, the estimated value of this pixel B component

Calculated by following formula.

${\hat{B}}_r(i,j)=\hat{G}(i,j)-\frac{1}{4}\underset{m,n\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i+m,j+n)-b(i+m,j+n)]$

For the pixel (i, j) that is positioned at known B component, the pixel (i, j-1) about this pixel and (i, j+1) are the pixel of known G component, then their B component estimated value

Calculated by following formula

${\hat{B}}_{g2}(i,j)=g(i,j)-\frac{1}{2}\underset{n\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i,j+n)-b(i,j+n)]$

For the pixel (i, j) that is positioned at known B component, this pixel pixel (i, j-1) and (i, j+1) up and down is the pixel of known G component, then the estimated value of their B component

Calculated by following formula.

${\hat{B}}_{g1}(i,j)=g(i,j)-\frac{1}{2}\underset{m\&Element;\{1,-1\}}{\mathrm{\&Sigma;}}[\hat{G}(i+m,j)-b(i+m,j)]$

Obtained thus a B component image that all has the B component information in all pixel position

Step 5 is revised the red, green, blue passage respectively

5a) calculate the horizontal and vertical gradient of green channel after the interpolation according to following formula:

$\left\{\begin{array}{c}{\mathrm{\&Delta;H}}_g(i,j)=|G(i,j-1)-G(i,j+1)|+\frac{1}{2}|2G(i,j)-G(i,j-1)-G(i,j+1)|\\ {\mathrm{\&Delta;V}}_g(i,j)=|G(i-1,j)-G(i+1,j)|+\frac{1}{2}|2G(i,j)-G(i-1,j)-G(i+1,j)|\end{array}\right.$

Wherein, Δ H _g(i, j) is green channel pixel (i, j) horizontal gradient value after the interpolation, and G (i, j) is green channel pixel (i, j) green value after the interpolation, Δ V _g(i, j) is green channel pixel (i, j) VG (vertical gradient) value after the interpolation;

To calculate according to following formula to the correction of known red value pixel in the green channel after the interpolation 5b):

$\tilde{G}(i,j)=r(i,j)+\left\{\begin{array}{c}\frac{w_2(g(i-1,j)-R(i-1,j))+w_7(g(i+1,j)-R(i+1,j))}{w_2+w_7},{\mathrm{\&Delta;H}}_g(i,j)>\mathrm{\&Delta;}{V}_g(i,j)\\ \frac{w_4(g(i,j-1)-R(i,j-1))+w_5(g(i,j+1)-R(i,j+1))}{w_4+w_5},{\mathrm{\&Delta;H}}_g(i,j)<{\mathrm{\&Delta;V}}_g(i,j)\\ \frac{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k(g(i+m_k,j+n_k)-R({i+m}_k,j+n_k))}{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k},{\mathrm{\&Delta;H}}_g(i,j)={\mathrm{\&Delta;V}}_g(i,j)\end{array}\right.$

$w_k=\frac{1}{{[1+{(r(i,j)-r(i+m_k,j+n_k))}^2]}^{1/2}}$

m _k∈{-2，-2，-2，0，0，2，2，2}，n _k∈{-2，0，2，-2，2，-2，0，2}，k＝1，2...8

Wherein,

Be the revised green value of pixel (i, j), r (i, j) is pixel (i, j) known red value, g (i-1, j) are the known green value of pixel (i-1, j), red value after R (i-1, j) pixel (i-1, the j) interpolation, w _kBe weights, m _kAnd n _kBe integer;

To calculate according to following formula to the correction of known blue valve pixel in the green channel after the interpolation 5c):

$\tilde{G}(i,j)=b(i,j)+\left\{\begin{array}{c}\frac{w_2(g(i-1,j)-B(i-1,j))+w_7(g(i+1,j)-B(i+1,j))}{w_2+w_7},{\mathrm{\&Delta;H}}_g(i,j)>\mathrm{\&Delta;}{V}_g(i,j)\\ \frac{w_4(g(i,j-1)-B(i,j-1))+w_5(g(i,j+1)-B(i,j+1))}{w_4+w_5},{\mathrm{\&Delta;H}}_g(i,j)<{\mathrm{\&Delta;V}}_g(i,j)\\ \frac{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k(g(i+m_k,j+n_k)-B({i+m}_k,j+n_k))}{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k},{\mathrm{\&Delta;H}}_g(i,j)={\mathrm{\&Delta;V}}_g(i,j)\end{array}\right.$

$w_k=\frac{1}{{[1+{(b(i,j)-b(i+m_k,j+n_k))}^2]}^{1/2}}$

m _k∈{-2，-2，-2，0，0，2，2，2}，n _k∈{-2，0，2，-2，2，-2，0，2}，k＝1，2...8

Wherein,

Be the revised green value of pixel (i, j), b (i, j) is pixel (i, j) known red value, g (i-1, j) are the known green value of pixel (i-1, j), red value after B (i-1, j) pixel (i-1, the j) interpolation, w _kBe weights, m _kAnd n _kBe integer.

To calculate according to following formula to the red channel correction after the interpolation 5d):

$\tilde{R}(i,j)=\tilde{G}(i,j)+\frac{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k(R(i+m_k,j+n_k)-\tilde{G}({i+m}_k,j+n_k))}{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k}$

$w_k=\frac{1}{{[1+{(\tilde{G}(i,j)-\tilde{G}({i+m}_k,j+n_k))}^2]}^{1/2}}$

m _k∈{-2，-2，-2，0，0，2，2，2}，n _k∈{-2，0，2，-2，2，-2，0，2}，k＝1，2...8

Wherein,

Be the revised red value of pixel (i, j),

Be the revised green value of pixel (i, j), w _kBe weights, m _kAnd n _kBe integer, R (i+m _k, j+n _k) be pixel (i+m after the interpolation _k, j+n _k) red value.

To calculate according to following formula to the blue channel correction after the interpolation 5e):

$\tilde{B}(i,j)=\tilde{G}(i,j)+\frac{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k(B(i+m_k,j+n_k)-\tilde{G}({i+m}_k,j+n_k))}{\underset{k=1}{\overset{8}{\mathrm{\&Sigma;}}}{w}_k}$

$w_k=\frac{1}{{[1+{(\tilde{G}(i,j)-\tilde{G}({i+m}_k,j+n_k))}^2]}^{1/2}}$

m _k∈{-2，-2，-2，0，0，2，2，2}，n _k∈{-2，0，2，-2，2，-2，0，2}，k＝1，2...8

Wherein, Be the revised blue valve of pixel (i, j),

Be the revised green value of pixel (i, j), w _kBe weights, m _kAnd n _kBe integer, B (i+m _k, j+n _k) be pixel (i+m after the interpolation _k, j+n _k) blue valve.

Step 6, the output coloured image.

Will

As three passages of RGB coloured image, obtain a width of cloth coloured image, Output rusults respectively.

Be described further below in conjunction with accompanying drawing 2,3 pairs of effects of the present invention of accompanying drawing.

Accompanying drawing 2 is demosaicing design sketchs of the prior art, and wherein, accompanying drawing 2 (a) is bilinear interpolation technology demosaicing design sketch; Accompanying drawing 2 (b) is alternating projection method demosaicing design sketch; Accompanying drawing 2 (c) is alternating projection method demosaicing design sketch; Accompanying drawing 2 (c) is iteration demosaicing method design sketch; Accompanying drawing 2 (d) is the method demosaicing design sketch that art is estimated based on the direction least mean-square error; Accompanying drawing 2 (e) is the method demosaicing design sketch of estimating continuously; Accompanying drawing 2 (f) is regularization method demosaicing design sketch.

The image of accompanying drawing 2 (b) derive from document " B.Gunturk; Y.Altunbasak; and R.Mersereau; " Color plane interpolation using alternating projections; " IEEE Trans.Image Process., vol.11, no.9, pp.997-1013, Sep.2002. ".

The image of accompanying drawing 2 (c) derive from document " Wenmain Lu and Yap-peng Tan; " Color filter array demosaicing:new method and performance measures; " IEEE Trans.Image Process., vol.12, no.10, pp.1194-1210, Oct.2003. ".

The image of accompanying drawing 2 (d) derive from document " L.Zhang and X.Wu; " Color demosaicking via directional linear minimum mean square-error estimation; " IEEE Trans.Image Process., vol.14, no.12, pp.2167-2178, Dec.2005. ".

The image of accompanying drawing 2 (e) derives from document " X.Li, " Demosaicing by successive approximation, " IEEE Trans.Image Process., vol.14, no.3, pp.370-379, Mar.2005. ".

Accompanying drawing 2 (f) is existing regularization method demosaicing design sketch, the method derive from document " D.Menon; G.Calvagno; " Regularization approaches to demosaicking; " IEEE Trans.Image Process., vol.18, no.10, Oct.2009 ".

Accompanying drawing 3 (d) is demosaicing design sketch of the present invention.

Contrast accompanying drawing 2 and accompanying drawing 3, the as a result figure that can find out the bilinear interpolation technology is the poorest, the false color that occurs is maximum, the false color that the as a result figure of alternating projection method and iteration demosaicing method occurs is also a lot, less based on the false color that the as a result figure of direction least mean-square error method of estimation and regularization method occurs, the false color that test findings figure of the present invention occurs is considerably less; Very approaching with the result of direction minimum mean square error method, but two width of cloth images carefully contrasted, can find out that residual false color of the present invention is less, visual effect of the present invention is better than other several method.

Claims (7)

1. the CFA image demosaicing method based on the edge direction interpolation comprises the steps:

(1) input one width of cloth is treated the CFA image of demosaicing;

(2) estimated brightness

2a) 9 * 9 wave filter Γ of design ₁With 5 * 5 wave filter Γ ₂

2b) use respectively wave filter Γ ₁And Γ ₂To input CFA image filtering, obtain and input image I after the filtering of the same size of CFA image _{Γ 1}And I _{Γ 2}, two width of cloth image co-registration are obtained a width of cloth luminance picture

(3) to the green channel interpolation

3a) calculate luminance picture

Gradient delta H on the horizontal and vertical directions, Δ V;

3b) relatively Δ H and Δ V are big or small, obtain the matrix of edge E of horizontal and vertical direction _hAnd E _v

3c) according to horizontal and vertical direction matrix of edge E _hAnd E _vJudge the interpolation direction and to the green channel interpolation;

(4) red and blue channel are carried out respectively bilinear interpolation;

(5) respectively the red, green, blue passage is revised;

(6) output coloured image.

2. the CFA image demosaicing method based on the edge direction interpolation according to claim 1, it is characterized in that: the wave filter step 2a) obtains according to following formula design:

$\mathrm{\&Gamma;}=A{(H^T{\mathrm{\&sigma;}}^{-1}H+{\mathrm{\&lambda;}}_1{M}_1^T{M}_1+{\mathrm{\&lambda;}}_2{M}_2^T{M}_2)}^{-1}{H}^T{\mathrm{\&sigma;}}^{-1}$

Wherein, Γ is the wave filter of required design, A is the transition matrix that the RGB coloured image is converted into brightness space, H is the decimation factor that the RGB color picture sampling becomes Bayer form A FA image, T is matrix transpose operator, and-1 is the matrix inversion operation symbol, and σ is noise variance, σ=0.00001, constant λ ₁=0.0008, λ ₂=0.02, M ₁And M ₂Be filtering matrix,

I ₃Be 3 * 3 unit matrixs, S ₁Be Hi-pass filter,

Expression Kronecker (Kronecker) operator, M ₂By Hi-pass filter S ₂Obtain by the following formula computing:

$M_2=\left(\begin{array}{ccc}1.547& -0.577& -0.577\\ -0.577& 1.547& -0.577\\ -0.577& -0.577& 1.547\end{array}\right)\&CircleTimes;S_2$

3. the CFA image demosaicing method based on the edge direction interpolation according to claim 1, it is characterized in that: the Image Fusion Rule step 2b) is as follows: when location of pixels index i and j are odd number or are even number,

Be the brightness value of pixel (i, j), I _{Γ 1}(i, j) is image I _{Γ 1}The value of middle pixel (i, j), otherwise,

I _{Γ 2}(i, j) is image I _{Γ 2}The value of middle pixel (i, j).

4. the CFA image demosaicing method based on the edge direction interpolation according to claim 1, it is characterized in that: the gradient calculation formula on the direction of horizontal and vertical step 3a) is as follows:

$\left\{\begin{array}{c}\mathrm{\&Delta;H}(i,j)=D_h(i,j)+\frac{1}{2}{D}_h(i,j-1)+\frac{1}{2}{D}_h(i,j+1)\\ \mathrm{\&Delta;V}(i,j)=D_v(i,j)+\frac{1}{2}{D}_v(i-1,j)+\frac{1}{2}{D}_v(i+1,j)\end{array}\right.$

$\left\{\begin{array}{c}{D}_h(i,j)=|\hat{L}(i,j)-\hat{L}(i,j-1)|+|\hat{L}(i,j)-\hat{L}(i,j+1)|+|2\hat{L}(i,j)-\hat{L}(i,j-2)-\hat{L}(i,j+2)|\\ D_v(i,.j)=|\hat{L}(i,j)-\hat{L}(i-1,j)|+|\hat{L}(i,j)-\hat{L}(i+1,j)|+|2\hat{L}(i,j)-\hat{L}(i-2,j)-\hat{L}(i+2,j)|\end{array}\right.$

Wherein, the location index of (i, j) expression pixel, namely this pixel is positioned at the capable j row of i, and Δ H (i, j) is the horizontal gradient of pixel (i, j), D _hBe the horizontal gradient operator, Δ V (i, j) is the VG (vertical gradient) of pixel (i, j), D _vBe the VG (vertical gradient) operator,

Be the brightness value of pixel (i, j), | * | the expression signed magnitude arithmetic(al).

5. the CFA image demosaicing method based on the edge direction interpolation according to claim 1 is characterized in that: the matrix of edge E of the horizontal and vertical direction step 3b) _hAnd E _vObtain according to following formula:

$e_h(i,j)=\left\{\begin{array}{c}1,\mathrm{\&Delta;H}(i,j)<\mathrm{\&Delta;V}(i,j)\\ 0.5,\mathrm{\&Delta;H}(i,j)=\mathrm{\&Delta;V}(i,j)\\ 0,\mathrm{\&Delta;H}(i,j)>\mathrm{\&Delta;V}(i,j)\end{array}\right.,$ $e_v(i,j)=\left\{\begin{array}{c}1,\mathrm{\&Delta;H}(i,j)>\mathrm{\&Delta;V}(i,j)\\ 0.5,\mathrm{\&Delta;H}(i,j)=\mathrm{\&Delta;V}(i,j)\\ 0,\mathrm{\&Delta;H}(i,j)<\mathrm{\&Delta;V}(i,j)\end{array}\right.$

$\left\{\begin{array}{c}{E}_h(i,j)=e_h(i,j-2)+e_h(i,j)+e_h(i,j-2)\\ E_v(i,j)=e_v(i-2,j)+e_v(i,j)+e_v(i-2,j)\end{array}\right.$

Wherein, e _h(i, j) is the horizontal edge quantized value of pixel (i, j), and Δ H is horizontal gradient, and Δ V is VG (vertical gradient), e _v(i, j) is the vertical edge quantized value of pixel (i, j), E _h(i, j) is the terminal level rim value of pixel (i, j), E _v(i, j) is the final vertical edge value of pixel (i, j).

6. the CFA image demosaicing method based on the edge direction interpolation according to claim 1 is characterized in that: the interpolation method step 3c): if matrix E _hThe capable j column element of i E _hThe value of (i, j) was carried out horizontal direction interpolation more than or equal to 2.5 o'clock to pixel (i, j); If matrix E _vThe capable j column element of i E _vThe value of (i, j) was carried out vertical direction interpolation more than or equal to 2.5 o'clock to pixel (i, j).

7. the CFA image demosaicing method based on the edge direction interpolation according to claim 1, it is characterized in that: the modification method described in the step (5): to pixel difference calculated level and the VG (vertical gradient) of interpolation in the green channel, if horizontal gradient is less than VG (vertical gradient), this pixel green value is carried out horizontal neighbors weighting correction, if horizontal gradient is greater than VG (vertical gradient), this pixel green value is carried out vertical neighborhood weighting correction, if horizontal gradient equals VG (vertical gradient), this pixel green value is carried out horizontal vertical neighborhood weighting correction; Red value and blue valve to interpolating pixel point all carry out horizontal vertical neighborhood weighting correction.

Images